text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Indian Regional Movie Dataset for Recommender Systems
Prerna Agarwal∗
IIIT Delhi
[email protected]
Richa Verma∗
IIIT Delhi
[email protected]
arXiv:1801.02203v1 [cs.IR] 7 Jan 2018
ABSTRACT
Indian regional movie dataset is the first database of regional Indian
movies, users and their ratings. It consists of movies belonging to
18 different Indian regional languages and metadata of users with
varying demographics. Through this dataset, the diversity of Indian
regional cinema and its huge viewership is captured. We analyze
the dataset that contains roughly 10K ratings of 919 users and
2,851 movies using some supervised and unsupervised collaborative
filtering techniques like Probabilistic Matrix Factorization, Matrix
Completion, Blind Compressed Sensing etc. The dataset consists of
metadata information of users like age, occupation, home state and
known languages. It also consists of metadata of movies like genre,
language, release year and cast. India has a wide base of viewers
which is evident by the large number of movies released every
year and the huge box-office revenue. This dataset can be used for
designing recommendation systems for Indian users and regional
movies, which do not, yet, exist. The dataset can be downloaded
from https://goo.gl/EmTPv6.
KEYWORDS
Movie dataset, Recommender System, Metadata, Indian Regional
Cinema, Crowd Sourcing, Collaborative Filtering
1
INTRODUCTION
Recommendation systems [17, 18, 19, 20, 33] utilize user ratings to
provide personalized suggestions of items like movies and products.
Some popular brands that provide such services are Amazon [34],
Netflix, IMDb, BarnesAndNoble [35], etc. Collaborative Filtering
(CF) and Content-based (CB) recommendation are two commonly
used techniques for building recommendation systems. CF systems
[21, 22, 23]operate by gathering user ratings for different items in
a given domain and compare various users, their similarities and
differences, to determine the items to be recommended. Contentbased methods recommend items by comparing representations of
content that a user is interested in to representations of content
that an item consists of.
There are several datasets like MovieLens [1] and Netflix [2] that are
available for testing and bench-marking recommendation systems.
We present an Indian regional movie dataset on similar lines. India
has been the largest producer of movies in the world for the last
∗ Prerna Agarwal (now working with IBM Research) and Richa Verma (now working
with TCS Research ) have equal contribution
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
Conference’17, Washington, DC, USA
© 2016 Copyright held by the owner/author(s). 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
DOI: 10.1145/nnnnnnn.nnnnnnn
Angshul Majumdar
IIIT Delhi
[email protected]
few years with a lot of diversity in languages and viewers. As per
the UNESCO cinema statistics [9], India produces around 1,724
movies every year with as many as 1,500 movies in Indian regional
languages. India’s importance in the global film industry is largely
because India is home to Bollywood in Mumbai. There’s a huge base
of audience in India with a population of 1.3 billion which is evident
by the fact that there are more than two thousand multiplexes in
India where over 2.2 billion movie tickets were sold in 2016. The box
office revenue in India keeps on rising every year. Therefore, there
is a huge need for a dataset like Movielens in Indian context that can
be used for testing and bench-marking recommendation systems
for Indian Viewers. As of now, no such recommendation system
exists for Indian regional cinema that can tap into the rich diversity
of such movies and help provide regional movie recommendations
for interested audiences.
1.1
Motivation
As of now, Netflix and Movielens datasets do not have a comprehensive listing of regional productions as the clipping shows in Figure
1( borrowed from [39]). Therefore, a substantial source of such a
data comprising movies of various regions, varying languages and
genres encompassing a wider folklore is strongly needed that could
provide such data in a suitable format required for building and
benchmarking recommendation systems.
To capture the diversity of Indian regional cinema, popular websites
like Netflix are trying to shift focus towards it [36, 37]. The goal is
to bring some of the greatest stories from Indian regional cinema
on a global platform. Through this, viewers are exposed to a wide
variety of new and diverse stories from India. As a result of this
initiative, Indian regional cinema will be available across countries.
Building a recommendation system using a dataset of such movies
and their audience can prove to be useful in such situations. Here,
we present such a dataset which is the first of its kind.
1.2
Contributions
• Web portal for data collection: A web portal where a user
can sign up by filling details like email, date of birth, gender,
home town, languages known and occupation. The user
can then provide rating to movies as like/dislike.
• Indian Regional Cinema Dataset: It is the first dataset of
Indian Regional Cinema which contains ratings by users
for different regional movies along with user and movie
metadata. User metadata is collected while signing up
on the portal. Movie metadata consists of genre, release
year, description, language, writer, director, cast and IMDb
rating.
• Detailed analysis of the dataset using some supervised and
unsupervised Collaborative Filtering techniques.
such as Blind Compressed Sensing, Probabilistic Matrix Factorization, Matrix completion, Supervised Matrix Factorization are used
on our dataset to provide benchmarking results. These techniques
are chosen over others because these techniques have proven to
provide better accuracy in recent works [6].
3
INDIAN REGIONAL MOVIE DATASET
This is the first dataset of Indian regional cinema which covers
movies of 18 different regional languages and a variety of user
ratings for such movies. It consists of 919 users with varying demographics and 2,851 movies with different genres. It has 10K ratings
from 919 users.
3.1
Metadata Information
The data for movies has been scraped from IMDb [3]. IMDb has a
collection of Indian movies spanning across multiple Indian regional
languages and genres. Each movie is associated with the following
metadata.
Figure 1: Amazon and Netflix: Focus on Regional Films
2
RELATED WORK
MovieLens [1] is a web-based portal that recommends movies. It
uses the film preferences of its users, collected in the form of movie
ratings and preferred genres and then utilizes some collaborative
filtering techniques to make movie recommendations. The Department of Computer Science and Engineering at the University of
Minnesota houses a research lab known as Grouplens Research that
created Movielens in 1997 [38]. Indian Regional Cinema dataset is
inspired from Movielens. The primary goal was to collect data for
performing research on providing personalized recommendations.
MovieLens released three datasets for testing recommendation systems: 100K, 1M and 10M datasets. They have released 20M dataset
as well in 2016. In the dataset, users and movies are represented
with integer IDs, while ratings range from 1 to 5 at a gap of 0.5.
Netflix released a training data set for their contest, Netflix Prize [8],
which consists of about 100,000,000 ratings for 17,770 movies given
by 480,189 users. Each rating in the training dataset consists of
four entries: user, movie, date of grade, grade. Users and movies
are represented with integer IDs, while ratings range from 1 to 5.
These datasets are largely for hollywood movies and TV series, and
their viewers. They are not designed for those user communities
which are inclined towards watching Indian regional cinema.
From the view point of recommender systems, there have been a lot
of work using user ratings for items and metadata to predict their
liking and disliking towards other items [4, 5, 6, 11]. Many unsupervised and supervised collaborative filtering techniques have been
proposed and benchmarked on movielens dataset. Here, in this
paper, we have chosen few popular techniques such as user-user
similarity to establish baseline and then other deeper techniques
• Movie id: Each movie has a unique id for its representation.
• Description: Description of the movie for users.
• Language: Language(s) used in the movie. A movie may
have been released in multiple regional languages. The
distribution is shown in Table 1.
• Release date: Date of release of the movie.
• Rating count: As per IMDb, to judge the popularity of the
movie.
• Crew: Director, writer and cast of the movie.
• Genre: Movie genre. It can be one or multiple out of 20
genres available on IMDb. The number of movies for each
genre are shown in Figure 2.
Table 1: Language Distribution
Languages
Hindi
Bengali
Assamese
Tamil
Nepali
Punjabi
Rajasthani
Malayalam
Bhojpuri
Kannada
Haryanvi
Manipuri
Urdu
Marathi
Telugu
Oriya
Gujarati
Konkani
Movie Count
615
582
22
313
51
150
18
346
26
303
3
8
129
204
338
98
49
6
User Count
902
28
9
30
9
78
14
16
21
11
18
4
23
14
18
6
7
4
Figure 2: Genre Distribution
For better recommendations, it is important to include the factors
which influence user ratings the most. Following is the metadata
information collected for a user:
• User id: A user will have a unique id for its representation.
• Languages: The languages known by the users. Its count is
shown in Table 1.
• State: The state of India that the user belongs to. The region
wise distribution is shown in Figure 3.
Figure 4: Age Distribution
Figure 3: Region-wise Distribution
• Age: Date of birth of the user is taken as input to calculate
the age of the user. The distribution is shown in Figure 4.
• Gender: Denotes the gender of the user. Gender distribution of the data is shown in Figure 5.
• Occupation: It denotes the occupation of the user. It can be
any one out of student, self-employed, service, retired and
others. Its distribution is shown in Figure 6.
Figure 5: Gender Distribution
Table 2: Comparison of Datasets
Dataset
Our Dataset
Movielens 100K
Movielens 1M
Movie Count
2851
1700
6000
User Count
919
1000
4000
Rating Count
10,000
100,000
1,000,000
Sparsity(%)
99.96
99.94
99.96
Release Year
2017
4/1998
2/2003
Figure 6: Occupation Distribution
3.2
MovieLens vs Indian Regional Cinema
Dataset
The key difference between the presented dataset and movielens is
that the latter does not contain movies from Indian regional cinema.
Movielens only has few Hindi and Urdu movies. Also, our data has
been collected mainly from the viewers of regional movies in India.
The user metadata, thus, collected can be used to recommend more
relevant movies for such audiences.
Also, the MovieLens datasets are biased towards a certain category
of users. They contain data only from users who have rated at
least twenty movies. The datasets do not include the data of those
users who could not find enough movies to rate or did not find the
system easy enough to use. There is a possibility that there is a
fundamental difference between such users and the other users in
the datasets. Our dataset makes no such distinction among users
based on the number of movies that they have rated.
To make the process of rating multiple movies easier for a user, we
have used the concept of binary rating for movies where, a user
can either ”like” or ”dislike” a movie denoted by ”1” and ”-1” in
the dataset. On the other hand, MovieLens uses a 10-point scale
for rating (from 0 to 5). A basic comparison of these datasets are
shown in Table 2. The table indicates the number of users, movies,
ratings, release year and the sparsity of datasets.
3.3
Figure 7: Sign Up form on Portal
information. The user can then login to the portal to rate movies as
either like or dislike and the responses are recorded as shown in
Figure 8.
Dataset Collection
For the collection of user information and movie ratings, a web
portal named Fickscore [15] is created where users can sign up
filling in all details as shown in Figure 7. The user has to provide
the preferred languages so that the portal can ask users to rate the
movies of their preferred languages.
While signing up, the user is prompted to fill up the metadata
Figure 8: Rating Movies on Portal
4
UNSUPERVISED COLLABORATIVE
FILTERING TECHNIQUES
To analyze the dataset, some unsupervised techniques are used
such as user-user similarity, item-item similarity, Matrix Factorization, Probabilistic Matrix Factorization, Blind Compressed Sensing
etc. The main advantage of using such techniques is the ease of
implementation and their incremental nature. On the other hand,
it is human data dependent its performance decreases on increase
of sparsity of data. These techniques cannot address the cold start
problem i.e., when a new user or item adds in the dataset whose
ratings are not available because these use ratings of users to make
predictions. Bias correction is performed on the dataset by calculating global mean, user bias and item bias and then the above
techniques are used to predict the rating of a new user for an item.
4.1
User and Item-based similarity
In user-user model, a similarity matrix A is calculated, each entry
Ai j indicates the score computed by cosine similarity between a
user i and another user j. It denotes how much similar are two
users i and j, higher the score higher is the similarity. Similarly, in
item-item model, each entry Ai j of the similarity matrix ’A’ denotes
the cosine similarity score between an item i and another item j.
Higher the score, the two items are more similar.
Cosine similarity can be calculated for two users u and u’ using the
following equations:
Í
∀j ru, j ru 0, j
Su,u 0 = qÍ
(1)
2 Í
2
∀j ru, j ∀j ru 0, j
Where, r i, j denotes the rating by i t h user for j t h item.
Prediction for u t h user for j t h item is done as:
Õ
wu,u 0 ru,u 0
ru,ˆ j =
(2)
u 0 ∈N (u), j ∈R(u 0 )
Where, wu,u 0 is the normalized similarity weight, ru,u 0 is the
rating by u’ user for j t h item.
Similar to user-user similarity, item-item similarity is calculated
by computing cosine similarity between two items and ratings are
predicted in the similar way.
4.2
Matrix factorization
There are some hidden traits (latent factors) of liking/disliking of
users which may depend on the pattern of their ratings. Users and
movies are mapped by this model to a joint latent factor space. Each
item i and user j is associated with vector q and p respectively which
measures the possessiveness of an item or user for those factors.
The dot product qTi pu denotes the liking of a user for a specific item
which approximates the rating rui [10]. Computing the mapping of
each user and item to factor vectors is a major challenge. Imputation
can prove to be expensive as it noticeably increases the amount
of data during calculation. To model the observed ratings directly
with regularization, the following equation is used:
Õ
min
(rui − qTi pu )2 + λ(||qi || 2 + ||pu || 2 )
(3)
∗ ∗
q .p
(u,i ∈k )
Here, k is the set of those user-item pairs in the training set for
which rui is known.
The system uses the already observed ratings to fit a model on them
and uses that model to predict the new ratings.
The intuition behind using matrix factorization to analyze this
dataset is that there should be some latent features that determine
how a user rates an item. For example, two users may give high
ratings to a certain movie if they both like the actors/actresses of
the movie, or if the movie is an action movie, which is a genre
preferred by both users. Hence, if we can discover these latent
features, we should be able to predict the rating given by a certain
user to a certain item, because the features associated with the user
should match with the features associated with the item.
4.3
Probabilistic Matrix Factorization
It can handle large datasets because it scales linearly with the number of observations in the dataset. Let Ri j represent the rating of
a user i for a movie j. Let U and V be latent feature matrices for
user and movie, respectively. The column vectors are denoted as
Ui , representing user-specific latent feature vectors, and Vj , representing movie-specific latent feature vectors. [4]. The log posterior
is maximized over movie and user features with hyper parameters
using the following equation:
N M
N
M
λv Õ
λu Õ
1 ÕÕ
Ii j (Ri j − UiT Vj )2 +
||Ui ||f2 r ob +
||Vj ||f2 r ob
2 i=1 j=1
2 i=1
2 j=1
(4)
Where, λu and λv are the regularization parameters for user and
item respectively. A local minimum of the equation can be computed by gradient descent in U and V . The model performance is
measured by computing mean average error (MAE) and root mean
squared error (RMSE) on the test set.
This model can be viewed as a probabilistic extension of the SVD
model, since if all ratings have been observed, the objective given
by the equation reduces to the SVD objective in the limit of prior
variances going to infinity. This technique better addresses the
sparsity and scalability problems and thus improves prediction
performance. It gives an intuitive rationale for recommendation.
4.4
Blind Compressed Sensing
A dense user item matrix is not a reasonable assumption as each
user will like/dislike a trait to certain extent [24]. However, any
item will possess only a few of the attributes and never all. Hence,
the item matrix will ideally have a sparse structure rather than a
dense one as formulated in earlier works.
The objective of this approach is to find the user and item latent
factor matrices. As per the approach, user latent factor matrix can
be dense but the same does not logically follow for the item latent
factor matrix. The sparsity of the item latent factor matrix increases
the recommendation accuracy significantly. [5]. The following
equation is minimized:
min ||Y − A(UV )|| 2 + λu ||U ||f r ob + λv ||V ||f r ob
(U ,V )
(5)
Where λu and λv are regularization parameters for user and item
respectively. A is the binary mask matrix and Y is the rating matrix.
U and V is the user latent matrix and item latent matrix respectively
which were assumed to be dense in earlier models.
Table 3: Unsupervised Techniques
Techniques
Errors
User-User similarity
Item-Item similarity
Matrix Factorization
Probabilistic Matrix Factorization
Blind Compressed Sensing
Matrix Completion
Movielens 100K
MAE RMSE
0.6980 1.026
0.744
1.061
0.828
1.128
0.7564 0.9639
0.7356 0.9409
0.8324 1.102
Our Dataset
MAE
RMSE
0.5307 1.03
0.648
1.049
0.471
0.971
0.481
0.9372
0.463
0.9612
0.4827 0.9264
Movielens 1M
MAE RMSE
0.607
0.8810
0.671
0.9196
0.6863 0.8790
0.7241 0.9127
0.6917 0.8789
0.7196 0.9102
Table 4: Supervised Technique
Technique
Errors
Supervised Matrix Factorization
4.5
Movielens 100K
MAE RMSE
0.7199 0.9196
Matrix Completion
Matrix completion involves filling up the missing entries of a partially observed matrix. It aims to compute the matrix with the
lowest rank or, if the rank of the completed matrix is known, a
matrix of rank r that matches the known entries. A popular approach for solving the problem is nuclear-norm-regularized (NN)
matrix [7] as shown in the following equation.
min ||B − R||f2 r ob + λ||R||N N
B = Rk −1 + M T (Y − M.Rk −1 )
(7)
and, M is the binary mask. R is the rating matrix imputed and Y is
the original rating matrix.
5
SUPERVISED COLLABORATIVE FILTERING
TECHNIQUES
To analyze the dataset, some supervised techniques are used such
as supervised Matrix Factorization. The main advantage of using
supervised methods is that whenever a new user or new item comes
in, it can make predictions for them as well which unsupervised
techniques fail to do [28, 29, 30, 31]. This is also called as cold start
problem. These are scalable and are dependent on the metadata
information of user and item because of which it gives more accurate predictions as it establishes relation well. Bias correction is
performed on the dataset by calculating user bias and item bias and
then the above technique is used to calculate the rating of a new
user for an item.
5.1
Supervised Matrix Factorization
The task of predicting ratings becomes difficult largely because
of the sparsity of the ratings available in the database of a recommender system. Therefore, using the knowledge related to users
demographics and item categories can enhance prediction accuracy
[25, 26, 27, 32]. Classes are formed as per users age group, gender
and occupation. A user can belong to multiple classes at a time.
Class label information is important to learn the latent factor vectors of users and movies in a supervised environment, in a way
that they are consistent with the class label information available.
Movielens 1M
MAE RMSE
0.6709 0.8567
Class label information puts in additional constraints which results
in reducing the search space as a result of which determinacy of
the problem is reduced.
Mathematically, within the matrix factorization framework, additional information of user metadata (U) and item metadata (V) can
be used and the following equation can be minimized [6].
min
(U ,V ,A,C)
(6)
Where,
Our dataset
MAE
RMSE
0.4367 0.9283
||Y − M(UV )||f r ob + λu ||U ||f r ob + λv ||V ||f r ob +
µu ||W − UC ||f r ob + µv ||Q − AV ||f r ob
(8)
Where, Wi j = 1 if user i belongs to class j else 0. C is the linear
map from latent factor space to classification domain. Q is the
class information matrix created similar to W . Other variables
have their usual meanings. Introducing supervised learning into
the latent factor model helps in improving the prediction accuracy
by reducing the problem of rating matrix sparsity. The value of
regularization parameters are determined using l-curve technique
[16]
6
EXPERIMENTS AND RESULTS
Three different datasets are used to compare the results of supervised and unsupervised collaborative filtering techniques used to
predict user ratings. The datasets used for experiments are Movielens 100K, MovieLens 1M and our dataset of Indian regional movies.
For error calculation, Mean absolute error (MAE) and Root mean
squared error (RMSE) is calculated between the actual ratings and
the predicted ratings. The datasets are divided into 5 folds for evaluation. The ratings are binarized into like/dislike (1/-1) labels for
experiments. Results of different techniques on these datasets are
shown in Table 3 and Table 4.
As the values in the Table 2 indicate, the basic cosine similarity measures between users and movies perform fairly well on all datasets.
The minimum MAE values result from the experiments using our
dataset. Since the sparsity of regional cinema dataset and Movielens
1M dataset is is very high (as indicated in Table 2), techniques like
Probabilistic Matrix Factorization and Blind Compressed Sensing
perform better than other basic similarity measures and among
them the least MAE is again shown for our dataset.
To use metadata information, the information is encoded in the
form of one hot vector of 1’s and 0’s where in case of languages, multiple 1’s can be present in the vector. Since supervised techniques
uses both user and item metadata they outperform unsupervised
collaborative filtering techniques. Among all three datasets, minimum MAE is shown on our dataset. This shows that the Indian
Regional Cinema dataset can prove to be useful for building and
benchmarking recommendation systems in Indian context, which
has the most diverse languages and demographics.
7
CONCLUSION AND FUTURE WORK
India is one of the country where not only varying languages are
present, it’s population’s demographics are also very diverse in
nature. Therefore, Indian regional cinema has a lot of diversity
when it comes to the number of languages and the demographics of
the viewers. There are thousands of such movies that are produced
annually and there is a huge community of people who watch them.
Therefore, a recommender system for Indian regional movies is
needed to address the preferences of the growing number of their
viewers. This dataset has around 10K ratings by Indian users, along
with their demographic information. We believe that this dataset
could be used to design, improve and benchmark recommendation
systems for Indian regional cinema. We plan to release the dataset
after its publication. We further want to release another version
of this dataset with more number of ratings and users, which will
help to improve the current state of recommender systems for the
Indian audience.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
MovieLens: https://movielens.org/
Netflix: https://www.netflix.com/in/
IMDb: http://www.imdb.com/
A. Mnih and R. Salakhutdinov, Probabilistic matrix factorization, in Proc. Adv.
Neural Inf. Process. Syst., 2007, pp. 1257-1264.
A. Gogna and A. Majumdar. (2015). Blind compressive sensing framework for
collaborative filtering. Available: http://arxiv.org/abs/1505.01621
A. Gogna and A. Majumdar, A Comprehensive Recommender System Model: Improving Accuracy for Both Warm and Cold Start Users, in IEEE Access, vol. 3, no. ,
pp. 2803-2813, 2015.
T. Hastie, R. Mazumder, J. Lee, R. Zadeh, Matrix Completion and Low-Rank SVD
via Fast Alternating Least Squares, https://arxiv.org/abs/1410.2596.
Netflix Prize: http://www.netflixprize.com/
http://uis.unesco.org/en/news/record-number-films-produced
Yehuda Koren, Matrix Factorization Techniques for Recommender Systems, Published by the IEEE Computer Society, August 2009.
Xiaoyuan Su and Taghi M. Khoshgoftaar, A survey of collaborative filtering
techniques, in Advances in artificial intelligence, vol. 2009, pp. 4, 2009.
Thomas Hofmann, Latent semantic models for collaborative filtering, in ACM
Transactions on Information Systems (TOIS), vol. 22, no. 1, pp. 89-115, 2004.
Gediminas Adomavicius and Alexander Tuzhilin, Toward the next generation
of recommender systems: A survey of the state-of-the-art and possible extensions,
in Knowledge and Data Engineering, IEEE Transactions on, vol. 17, no. 6, pp.
734-749, 2005.
Sivan Gleichman and Yonina C Eldar, Blind compressed sensing, in Information
Theory, IEEE Transactions on, vol. 57, no. 10, pp. 6958-6975, 2011.
Flickscore: http://flickscore.iiitd.edu.in
C. L. Lawson and R. J. Hanson, Solving Least Squares Problems, vol. 161. Englewood Cliffs, NJ, USA: Prentice-Hall, 1974.
M. F. Hornick and P. Tamayo, Extending recommender systems for disjoint
user/item sets: The conference recommendation problem, IEEE Trans. Knowl. Data
Eng., vol. 24, no. 8, pp. 1478-1490, Aug. 2012.
Q.Liu, E.Chen, H.Xiong, Y.Ge, Z.Li and X.Wu, A cocktail approach for travel
package recommendation, IEEE Trans. Knowl. Data Eng.,vol. 26, no. 2, pp. 278293, Feb. 2014.
Y. Koren and R. Bell, Advances in collaborative filtering, in Recommender Systems
Handbook. New York, NY, USA: Springer, 2011, pp. 145-186.
X. Su and T. M. Khoshgoftaar, A survey of collaborative filtering techniques, Adv.
Artif. Intell., vol. 2009, Jan. 2009, Art. ID 4.
[21] R. M. Bell and Y. Koren, Improved neighborhood-based collaborative filtering, in
Proc. KDD-Cup Workshop, 2007, pp. 7-14.
[22] C.Desrosiers, and G.Karypis, A comprehensive survey of neighborhood based
recommendation methods, in Recommender Systems Handbook. New York, NY,
USA: Springer, 2011, pp. 107-144.
[23] J. Wang, A. P. de Vries, and M. J. T. Reinders, Unifying user-based and item-based
collaborative filtering approaches by similarity fusion, in Proc. 29th Annu. Int.
ACM SIGIR Conf. Res. Develop. Inf. Retr., 2006, pp. 501-508.
[24] S.Gleichman and Y.C.Eldar, Blind compressed sensing,IEEETrans. Inf. Theory, vol.
57, no. 10, pp. 6958-6975, Oct. 2011.
[25] H.Ma, D.Zhou, C.Liu, M.R.Lyu, and I.King, Recommender systems with social
regularization, in Proc. 4th ACM Int. Conf. Web Search Data Mining, 2011, pp.
287-296.
[26] P. Massa and P. Avesani, Trust-aware recommender systems, in Proc.ACM Conf.
Recommender Syst., 2007, pp. 17-24.
[27] X.Tang, Y.Xu, and S.Geva, Learning higher order interactions for user and item profiling based on tensor factorization, in Proc. 20th Int. Conf. Intell. User Interfaces,
2015, pp. 213-224.
[28] Q.Gu, J.Zhou, and C.Ding,Collaborative filtering:Weighted non negative matrix
factorization incorporating user and item graphs, in Proc. SDM, 2010, pp. 199-210.
[29] S.-T. Park, D. Pennock, O. Madani, N. Good, and D. DeCoste, Naive filterbots for
robust cold-start recommendations, in Proc. 12th ACM SIGKDD Int. Conf. Knowl.
Discovery Data Mining, 2006, pp. 699-705.
[30] A. L. V. Pereira and E. R. Hruschka, Simultaneous co-clustering and learning to
address the cold start problem in recommender systems, Knowl.-Based Syst., vol.
82, pp. 11-19, Jul. 2015.
[31] Z.-K. Zhang, C. Liu, Y.-C. Zhang, and T. Zhou, Solving the cold-start problem
in recommender systems with social tags, EPL (Europhys. Lett.), vol. 92, no. 2, p.
28002, Nov. 2010.
[32] A.-T. Nguyen, N. Denos, and C. Berrut, Improving new user recommendations
with rule-based induction on cold user data, in Proc. ACM Conf. Recommender
Syst., 2007, pp. 121-128.
[33] G. Adomavicius and A. Tuzhilin, Toward the next generation of recommender
systems: A survey of the state-of-the-art and possible extensions, IEEE Trans. Knowl.
Data Eng., vol. 17, no. 6, pp. 734-749, Jun. 2005.
[34] Amazon: http://www.amazon.in
[35] BarnesAndNoble: https://www.barnesandnoble.com
[36] https://www.clickondetroit.com/entertainment/
netflix-steps-up-its-battle-with-amazon-in-india
[37] http://www.financialexpress.com/industry/technology/
the-desi-content-battleground/798145/
[38] https://grouplens.org/datasets/movielens/
[39] http://timesofindia.indiatimes.com/companies/
amazon-to-join-bollywood-film-industry-hires-consultants-to-create-a-blueprint\
-for-a-hindi-film-studio/articleshow/58176571.cms
| 2cs.AI
|
Weighted Message Passing and Minimum Energy Flow for
Heterogeneous Stochastic Block Models with Side Information
T. Tony Cai1 , Tengyuan Liang2 and Alexander Rakhlin1
arXiv:1709.03907v1 [math.ST] 12 Sep 2017
1
2
Department of Statistics, The Wharton School at the University of Pennsylvania
Econometrics and Statistics, The University of Chicago Booth School of Business
Abstract
We study the misclassification error for community detection in general heterogeneous stochastic
block models (SBM) with noisy or partial label information. We establish a connection between the
misclassification rate and the notion of minimum energy on the local neighborhood of the SBM. We
develop an optimally weighted message passing algorithm to reconstruct labels for SBM based on the
minimum energy flow and the eigenvectors of a certain Markov transition matrix. The general SBM
considered in this paper allows for unequal-size communities, degree heterogeneity, and different
connection probabilities among blocks. We focus on how to optimally weigh the message passing to
improve misclassification.
1 Introduction
The stochastic block model (SBM), or planted partition model, is a celebrated model that captures the
clustering or community structure in large networks. Fundamental phase transition phenomena and
limitations for efficient algorithms have been established for the “vanilla” SBM, with equal-size communities [9, 10, 27, 30, 31, 24, 1, 16, 2, 11]. However, when applying the algorithms to real network datasets,
one needs to carefully examine the validity of the vanilla SBM model. First, real networks are heterogeneous and imbalanced; they are often characterized by unequal community size, degree heterogeneity,
and distinct connectivity strengths across communities. Second, in real networks, additional side information is often available. This additional information may come, for instance, in the form of a small
portion of revealed community memberships, or in the form of node features, or both. In this paper, we
aim to address the above concerns by answering the following questions:
Algorithm For a general stochastic block model that allows for heterogeneity and contains noisy or
partial side information, how to utilize this information to achieve better classification performance?
Theory What is the transition boundary on the signal-to-noise ratio for a general heterogeneous stochastic block model? Is there a physical explanation for the optimal misclassification error one can achieve?
1.1 Problem Formulation
We define the general SBM with parameter bundle (n, k, N ∈ Rk ,Q ∈ Rk×k ) as follows. Let n denote the
number of nodes and k the number of communities. The vector N = [n 1 , n 2 , . . . , n k ]T denotes the number
of nodes in each community. The symmetric matrix Q = [Q i j ] represents the connection probability: Q i j
1
is the probability of a connection between a node in community i to a node in community j . Specifically,
one observes a graph G(V, E ) with |V | = n, generated from SBM as follows. There is a latent disjoint parS
tition that divides V = kl=1 Vl into k communities. Define `(·) : V → [k] to be the label (or, community)
of a node v. For any two nodes v, u ∈ V , there is an edge between (u ↔ v) ∈ E with probability Q `(u),`(v) .
The goal is to recover the latent label `(v) for each node v. Here we consider the following kinds of heterogeneity: unequal size communities (represented by [n i ]), different connection probabilities across
communities (as given by [Q i j ]), and degree heterogeneity (due to both [n i ] and [Q i j ]).
We study the problem when either noisy or partial label information is available in addition to the
graph structure and show how to “optimally” improve the classification result (in terms of misclassification error). We argue that this is common for many practical problems. First, in real network datasets, a
small portion of labels (or, community memberships) is often available. Second, a practitioner often has
certain initial guess of the membership, either through training regression models using node features
and partially revealed labels as side information, or running certain clustering algorithms (for example,
spectral clustering using non-backtracking matrix, semi-definite programs or modularity method) on a
subset or the whole network. We will show that as long as these initial guesses are better than random
assignments, one can “optimally weigh” the initial guess according to the network structure to achieve
small misclassification error.
Formally, the noisy (or partial) information is defined as a labeling `˜prior on the nodes of the graph
with the following stochastic description. The parameter δ quantifies either (a) the portion of randomly
revealed true labels (with the rest of entries in `˜prior missing), or (b) the accuracy of noisy labeling `˜prior ,
meaning
1−δ
P(`˜prior (v) = `(v)) =
+ δ,
k
and when `˜prior (v) 6= `(v), each label occurs with equal probability.
1.2 Prior Work
In the literature on vanilla SBM (equal size communities, symmetric case), there are two major criteria —
weak and strong consistency. Weak consistency asks for recovery better than random guessing in a sparse
random graph regime (p, q ³ 1/n), and strong consistency requires exact recovery for each node above
the connectedness theshold (p, q ³ log n/n). Interesting phase transition phenomena in weak consistency for SBM have been discovered in [10] via the insightful cavity method from statistical physics.
Sharp phase transitions for weak consistency have been thoroughly investigated in [9, 30, 31, 32, 27].
In particular for k = 2, spectral algorithms on the non-backtracking matrix have been studied in [27] and
the non-backtracking walk in [32]. In these two fundamental papers, the authors resolved the conjecture
on the transition boundary for weak consistency posed in [10]. Spectral algorithms as initialization and
belief propagation as further refinement to achieve better recovery was established in [31]. Recent work
of [3] establishes the positive detectability result down to the Kesten-Stigum bound for all k via a detailed
analysis of a modified version of belief propagation. For strong consistency, [1, 16, 17] established the
phase transition using information-theoretic tools and semi-definite programming (SDP) techniques. In
the statistics literature, [40, 14] investigated the misclassification rate of the standard SBM.
For the general SBM with connectivity matrix Q, [15, 5, 7] provided sharp non-asymptotic upper
bound analysis on the performance of a certain semi-definite program. They investigated the conditions
on Q for a targeted recovery accuracy, quantified as the loss (as a matrix norm) between the SDP solution
and the ground truth. The results are more practical for heterogeneous real networks. However, for the
analysis of SDP to work, these results all assume certain density gap conditions, i.e., max1≤i < j ≤k Q i j <
2
min1≤i ≤r Q i i , which could be restrictive in real settings. Our technical approach is different, and does
not require the density gap conditions. Moreover, we can quantify more detailed recovery guarantees,
for example, when one can distinguish communities i , j from l , but not able to tell i , j apart. In addition,
our approach can be implemented in a decentralized fashion, while SDP approaches typically do not
scale well for large networks.
For SBM with side information, [20, 6, 38] considered SBM in the semi-supervised setting, where
the side information comes as partial labels. [20] considered the setting when the labels for a vanishing
fraction of the nodes are revealed, and showed that pushing below the Kesten-Stigum bound [22, 21] is
possible in this setting, drawing a connection to a similar phenomenon in k-label broadcasting processes
[29]. In addition, [6, 38] studied linearized belief propagation and misclassification error on the partially
labeled SBM.
The focus of this paper is on local algorithms, which are naturally suited for distributed computing
[25] and provide efficient solutions to certain computationally hard combinatorial optimization problems on graphs. For some of these problems, they are good approximations to global algorithms [23,
13, 36, 35]. The fundamental limits of local algorithms have been investigated, in particular, in [28] in
the context of a sparse planted clique model. We also want to point out that [34] studied the local belief propagation and characterized the expected fraction of correctly labeled vertices using fixed point
analysis of the density evolution, in the case of vanilla SBM with side information.
Finally, we briefly review broadcasting processes on trees. Consider a Markov chain on an infinite
tree rooted at ρ with branching number b. Given the label of the root `(ρ), each vertex chooses its label
by applying the Markov rule M to its parent’s label, recursively and independently. The process is called
broadcasting process on trees. One is interested in reconstructing the root label `(ρ) given all the nth level leaf labels. Sharp reconstruction thresholds for the broadcasting process on general trees for
the symmetric Ising model setting (each node’s label is {+, −}) have been studied in [12]. [33] studied
a general Markov channel on trees that subsumes k-state Potts model and symmetric Ising model as
special cases, and established non-census-solvability below the Kesten-Stigum bound. [18] extended
the sharp threshold to robust reconstruction, where the vertex’ labels are contaminated with noise. The
transition thresholds proved in the above literature correspond to the Kesten-Stigum bound b|λ2 (M )|2 =
1 [22, 21].
1.3 Our Contributions
The main results of the present paper are summarized as follows.
Weighted Message Passing We propose a new local algorithm – Weighted Message Passing (WMP) –
that can be viewed as linearized belief propagation with a novel weighted initialization. The optimal
weights are jointly determined by the minimum energy flow that captures the imbalance of local treelike neighborhood of SBM, and by the second eigenvectors of the Markov transition matrix for the label
broadcasting process. As we will show, these initializations are crucial for the analysis of general SBM
that is heterogeneous and asymmetric.
For the technical contribution, we provide non-asymptotic analysis on the evolution of WMP messages. For general number of communities, it is challenging to track the densities of WMP messages
during evolution. We overcome the difficulty through introducing carefully chosen weights and then
prove concentration-of-measure phenomenon on messages.
Misclassification Error We establish a close connection between the misclassification error and a notion
called minimum energy through the optimally weighted message passing algorithm. In fact, we show
3
that asymptotically almost surely, the misclassification error of WMP Err(`ˆwmp ) satisfies
µ
ˆ
Err(`wmp ) ≤ exp −
¶
1
,
2E∗ (θ −2 )
where E∗ (θ −2 ) is defined as the minimum energy based on the local tree-like neighborhood, with θ 2 chosen as the conductance level on the edges of the tree. Intuitively, the smaller the energy is, the better the
misclassification error one can achieve. This result provides a physical interpretation for the misclassification error. In return, the above upper bound provides a principled way of choosing the optimal
weights as to minimize the energy determined by the Thomson’s principal [26]. This approach is key to
dealing with asymmetric and imbalanced local neighborhoods.
Transition Boundary We show that the Kesten-Stigum bound is the sharp boundary for local algorithms
on the signal-to-noise ratio for the general heterogeneous SBM. Define the following quantities
£
¤−1
K := diag(QN ) Qdiag(N ), M := Qdiag(N )
(1)
θ := λ2 (K ), λ := λ1 (M ),
and SNR := λθ 2 ,
(2)
where N ,Q are defined in Section 1.1, and λi (·) denotes the i -th eigenvalue. Then the Kesten-Stigum
bound SNR = 1 is the threshold for local algorithms. Above it, the minimum energy E∗ (θ −2 ) is finite,
which asserts a valid upper bound on the misclassification error. Below it, the minimum energy diverges
and WMP fails. In fact, we show that below the threshold, no local algorithm can perform significantly
better than random guessing.
Set Identification When the number of communities k ≥ 3, we define a notion of set identification to
describe, for two disjoint sets (of communities) S, T ⊂ [k], whether one can distinguish S from T . This
notion subsumes as a special case the classic identification when S, T are singletons. However, it describes more general cases when one cannot distinguish the communities inside S and T , but is able to
distinguish S and T . We provide a mathematical description of this fact using the structure of eigenvectors for the Markov transition matrix K defined in (1). Further, we show that one can weigh the labels in
the “most informative direction” by initializing WMP according to the second eigenvectors.
1.4 Organization of the Paper
The paper is organized as follows. Section 2 reviews the background, definitions, and theoretical tools
that will be employed to solve the general SBM. To illustrate the main idea behind the theoretical analysis
better, we split the main result into two sections. Section 3 resolves the k = 2 case, where we emphasize
the derivation of WMP as a linearized belief propagation, and, more importantly, detail the initialization
of WMP according to minimum energy flow. Then we establish the connection between misclassification
and energy. In Section 4, we focus on the general k ≥ 3 case, where we incorporate an additional layer of
weights on the labels introduced by the eigenvectors of the Markov transition matrix. We then describe
the mathematical treatment of set identification. Discussions on the gap between local and global algorithms for growing k, and on how WMP utilizes the asymmetry follow in the end. Section 5 considers the
numerical performance of the proposed algorithm. The proofs of the main results are given in Section 6.
4
2 Preliminaries
2.1 Tree, Branching Number, Flow and Energy
Let T t (o) denote the tree up to depth t with root o. For a node v, the set of children is denoted by C (v),
children at depth d denoted by C d (v), and the parent of v is denoted by P (v). We use |v| to denote the
depth of v relative to o. If we view a tree as an electrical network, one can define the current flow and
energy on the tree [26]. Later in the paper we will show the close connection between these notions and
the misclassification error.
Definition 1 (Electric Flow). A unit flow i(·) : V → R on a tree T is called a valid unit flow if i(
for any v
i(
v) =
X
i(
o) = 1 and
u).
u∈C (v)
Definition 2 (Energy and Resistance). The energy E(r, i) of a unit flow i at resistance level r > 0 is defined
as
E(r, i) :=
X
i(
v)2 r|v| .
v∈T
The minimum energy E∗ (r) is
E∗ (r) := inf E(r, i),
i
where the infimum is over all valid unit flows. Denote the minimum energy flow as i∗ .
When assigning resistance rd to edges that are d -depth away from the root, the energy enjoys the
natural physical interpretation. We also remark that for a given resistance level, one can calculate the
minimum energy flow i∗ on the tree using Thomson’s principal. We identify the reciprocal of resistance
level with the conductance level.
Now we are ready to define the branching number of a tree T through minimum energy.
Definition 3 (Branching Number). The branching number br(T ) can be defined as
br(T ) := sup{r : E(r) < ∞} = sup{r : inf
X
i v∈T
i(
v)2 r|v| < ∞}.
It is well known that the branching number not only captures the growth rate of the tree, but also the
more detailed structure, such as imbalance [26].
2.2 Broadcasting Trees and SBM
When viewed locally, stochastic block models in the sparse regime share similarities with a label broadcasting process on a Galton-Watson tree. In fact, the local neighborhood of SBM can be coupled with
a broadcasting tree with high probability as n → ∞. This phenomenon has been investigated in studying the detectability and reconstruction threshold for vanilla SBM (equal-size communities, symmetric
case), as in [30].
Let us formally define the label broadcasting process conditioned on a tree T (o).
5
Definition 4 (Label Broadcasting). Given a tree T (o), the k-broadcasting process on T with the Markov
transition matrix K ∈ Rk×k describes the following process of label evolution. Conditioning on a node v
and its label `(v) ∈ [k], the labels of children u ∈ C (v) are sampled independently from
P(`(u)|`T|v| (o) ) = P(`(u)|`(v)) = K `(v),`(u) ,
where the first equality is the Markov property.
Let us review the definition of the multi-type Galton-Watson tree. We shall only consider the Poisson
branching process.
Definition 5 (Multi-type Galton-Watson Tree). Consider a k-types Galton-Watson process with the mean
matrix M ∈ Rk×k . For a node v, given its type `(v) = i , the number of type j children of v enjoys a
Poisson(mi j ) distribution, independently of other types. Start the process recursively for t generations
from root o. The tree T t (o) is called a multi-type Galton-Watson tree.
2.3 Notation
The moment generating function (MGF) for a random variable X is denoted by Ψ X (λ) = Ee λX . For
asymptotic order of magnitude, we use a(n) = O (b(n)) to denote that ∀n, a(n) ≤ C b(n) for some universal constant C , and use O ∗ (·) to omit the poly-logarithmic dependence. As for notation -, %: a(n) - b(n)
a(n)
if and only if lim b(n)
≤ C , with some constant C > 0, and vice versa. The square bracket [·] is used to
n→∞
represent the index set [k] := [1, 2, . . . , k]; in particular when k = 2, [2] := {+, −} for convenience.
x
−e −x
Recall that the hyperbolic tangent is tanh x = ee x +e
−x . The message-passing algorithm in the following
sections involves a non-linear update rule defined through a function
f θ1 ,θ2 (x) := log
for 0 < θ1 , θ2 < 1. Note that the derivative f θ0
1 ,θ2
(0) =
1 + θ1 tanh x2
1 − θ2 tanh x2
,
(3)
θ1 +θ2
2 .
3 Two Communities
In this Section we will illustrate the main results for the case of two, possibly imbalanced, communities.
We motivate the weighted message passing algorithm, and its relation to minimum energy flow. We
investigate the connection between misclassification and minimum energy, as well as the corresponding
transition threshold for general SBM.
6
3.1 Main Algorithmic and Theoretical Results
This section serves as an informal summary of the results for k = 2. As a start, we introduce the following
weighted message passing (WMP) Algorithm 1.
Algorithm 1: Weighted Message Passing
Data: Graph G(V, E ) with noisy label information `˜prior . Parameters: neighborhood radius t̄ and
conductance level θ̄ 2 .
Result: The labeling for each node o ∈ V .
for each node o ∈ V , do
Open the tree neighborhood T t̄ (o) induced by the graph G(V, E ) ;
Layer t̄ : for every node u ∈ C t̄ (o) with distance t̄ to the root on T t̄ (o), initialize its message
M (u, 0) = θ̄ −2|u| · i∗ (
u) · sign[`˜prior (u)],
where i∗ ( u) is the minimum energy flow to u calculated via Thomson’s principal on T t̄ (o)
with conductance level θ̄ 2 ;
for t = 1, . . . t̄ , do
Layer t̄ − t : for every node u ∈ C t̄ −t (o), calculate the message M (u, t ) through the
linearized update rule
X
M (u, t ) =
θ̄M (v, t − 1).
v∈C (u)
end
Output `ˆwmp (u) = sign[M (o, t̄ )].
end
We remark that WMP can run in parallel for all nodes due to its decentralized nature. For fixed depth
t̄ and sparse SBM (when n maxi , j Q i j - log n), the algorithm runs in O ∗ (n) time.
The following theorem is a simplified version of Theorems 2 and 3 below:
Theorem 1 (General SBM: k = 2). Consider the general stochastic block model G(V, E ) with parameter
bundle (n, k = 2, N ,Q), with either partial or noisy label information `˜prior with parameter 0 < δ < 1.
Assume that n maxi , j Q i j - n o(1) . For any node o ∈ V and its depth t leaf labels `˜prior (C t (o)), define the
worst-case misclassification error of a local estimator σt (o) : `˜prior (C t (o)) → {+, −} as
Err(σt ) := max P (σt (o) 6= `(o)|`(o) = l ) .
l ∈{+,−}
(4)
Define
µ
¶
1 n 1Q 11 − n 2Q 12 n 2Q 22 − n 1Q 21
+
4 n 1Q 11 + n 2Q 12 n 1Q 21 + n 2Q 22
µ·
¸¶
n 1Q 11 n 2Q 12
λ := λ1
.
n 1Q 21 n 2Q 22
θ̄ :=
Let E∗ (θ̄ −2 ) be the minimum energy on T t (o) with conductance level θ̄ 2 as t → ∞.
The transition boundary for this general SBM depends on the value
SNR = λθ̄ 2 .
7
(5)
(6)
On the one hand, if λθ̄ 2 > 1, the WMP Algorithm 1, denoted as `ˆwmp , enjoys the following upper bound on
misclassification
¶
µ
1
1
∧ ,
lim sup lim sup Err(`ˆwmp ) ≤ exp −
(7)
2
n→∞
t →∞
2E∗ (1/θ̄ 2 )
for any fixed δ > 0. On the other hand, if λθ̄ 2 < 1, for any local estimator σt that uses only label information
on depth t leaves, the minimax misclassification error is lower bounded by
1
lim inf lim inf inf Err(σt ) = .
t →∞ n→∞ σt
2
(8)
Remark 1. We remark that Algorithm 1 is stated for the case when noisy label information is known for
all nodes in layer t̄ . For the case of partial label information, there are two options to modify the initialization of the algorithm: (1) view the partial label information with parameter δ as the noisy label
information on layer t̄ only, with P(`˜prior (u) = `(u)) = δ + (1 − δ) 12 — with probability δ, the label is revealed exactly, and with probability 1 − δ, the label is decided using coin-flip — then proceed with the
algorithm; (2) view the partial information as on each layer there is a δ portion of nodes whose label
is shown exactly. Call the set of these nodes V l (T t̄ (o)). Then we need to initialize the message M (u)
P
for all u ∈ V l (T t̄ (o)) first before using the recursion M (u) = v∈C (u) θ̄M (v). It can be shown that these
two treatments enjoy similar asymptotic performance in terms of misclassification error, above the SNR
threshold. However, the latter performs better numerically for fixed depth tree as it utilizes more information.
We decompose the proof of Theorem 1 into several building steps: (1) conditioned on the local tree
structure, prove concentration-of-measure on WMP messages when label propagates according to a
Markov transition matrix K ; (2) for a typical tree instance generated from multi-type Galton-Watson process, establish connection among the misclassification rate, transition boundary and minimum energy
through the concentration result; (3) show that in the sparse graph regime of interest, the local neighborhood of general SBM can be coupled with a multi-type Galton-Watson with Markov transition matrix
" nQ
#
n Q
1
K :=
11
2
n 1 Q 11 +n 2 Q 12
n 1 Q 21
n 1 Q 21 +n 2 Q 22
12
n 1 Q 11 +n 2 Q 12
n 2 Q 22
n 1 Q 21 +n 2 Q 22
for label broadcasting (the explicit expression based on Eq. (1)). We remark that (3) follows similar proof
strategy as in [30], where the coupling for vanilla SBM has been established. The lower bound follows
from Le Cam’s testing argument, and the difficulty lies in analyzing the distance between measures recursively on the local tree.
Remark 2. When the local tree is regular and symmetric and λθ̄ 2 > 1, the minimum energy can be evaluated exactly as
1
E∗ (θ̄ −2 ) =
,
2
λθ̄ − 1
¡
¢
−1
which implies that misclassification error takes the exponentially decaying form exp − SNR
. Hence,
2
the result provides a detailed understanding of the strength of the SNR and its effect on misclassification,
i.e., the inference guarantee. More concretely, for the vanilla SBM in the regime p = a/n, q = b/n, the
boundary is SNR =
n(p−q)2
2(p+q)
> 1, which is equivalent to the boundary
(a − b)2
>1
2(a + b)
8
ˆ < 1/n →
for weak consistency in [32, 27]. In addition, one observes that SNR > 1 + 2 log n implies Err(`)
0, which asserts strong consistency. This condition on SNR is satisfied, for instance, by taking p =
a log n/n, q = b log n/n in vanilla SBM and computing the relationship between a, b to ensure SNR =
n(p−q)2
2(p+q)
> 1 + 2 log n. This relationship is precisely
p
s
p
p
a− b
2(a + b)
1
> 1+
· p
p
p > 1.
2 log n
2
a+ b
The above agrees with the threshold for strong recovery in [1, 16].
3.2 Weighted Message Passing & Minimum Energy Flow
In this section, we will motivate our proposed weighted message passing (WMP) from the well-known
belief propagation (BP) on trees. There are two interesting components in the WMP Algorithm 1: the
linearization part, and the initialization part. We will discuss each one in details in this section.
Recall the Definition 4 of the label broadcasting process on tree T (o) with k = 2. For convenience, let
us denote the Markov transition matrix K to be
#
"
K=
1+θ1
2
1−θ2
2
1−θ1
2
1+θ2
2
.
(9)
The BP algorithm is the Bayes optimal algorithm on trees given the labels of leaves. Define for a node
u ∈ V the BP message as
P(`(u) = +)|`obs (T t (u)))
B (u, t ) := log
,
P(`(u) = −|`obs (T t (u)))
which is the posterior logit of u’s label given the observed labels `obs (T t (u))). Using Bayes rule and conditional independence, one can write out the explicit evolution for BP message through f θ1 ,θ2 in (3)
Ã
B (u, t ) =
X
v∈C (u)
=
X
log
1 + θ1 tanh B (v,t2 −1)
!
1 − θ2 tanh B (v,t2 −1)
f θ1 ,θ2 (B (v, t − 1)) ,
(10)
v∈C (u)
with θ1 , θ2 as in Markov transition matrix K . While the method is Bayes optimal, the density of the messages B (u, t ) is difficult to analyze, due to the blended effect of the dependence on revealed labels and
the non-linearity of f θ1 ,θ2 . However, the WMP Algorithm 1 — a linearized BP — shares the same transition threshold with BP, and is easier to analyze. Above a certain threshold, the WMP succeeds, which
implies that the optimal BP will also work. Below the same threshold, even the optimal BP will fail, and
so does the WMP. The updating rule for WMP messages M (u, t ) is simply a replacement of Eq. (10) by its
linearized version,
X θ1 + θ2
M (v, t − 1).
M (u, t ) =
2
v∈C (u)
The initialization of the WMP messages on the leaves M (u, 0) whose labels have been observed is
crucial to the control of the misclassification error of the root node, especially for general SBM with
heterogeneous degrees. For general SBM, one should expect to initialize the messages according to the
detailed local tree structure, where the degree for each node could be very different. It turns out that
9
the optimal misclassification for WMP is related to a notion called the minimum energy E∗ . Moreover,
the optimal initialization for leaf message u is proportional to the minimum energy flow i∗ ( u) on the
local tree, with conductance level θ̄ 2 . In plain language, i∗ ( u) provides a quantitative statement of the
importance of the vote u has for the root. Note that for imbalanced trees, i∗ could vary significantly from
node to node, and can be computed efficiently given the tree structure T t (o) for a specified conductance
level.
3.3 Concentration, Misclassification & Energy
We now prove the concentration-of-measure phenomenon on WMP messages. Through the concentration, we will show the close connection between misclassification and energy. We will first state the result
conditioned on the tree structure T t (o).
Lemma 1 (Concentration on Messages). Recall the label broadcasting process with Markov transition
kernel K ∈ R2×2 on tree T t̄ (o). Assume the MGF of messages on leaves M (u, 0) satisfies the following
h
i
λ2 σ2 (u)
0
E e λM (u,0) |`(u) = + ≤ e λµ0 (u,+) e 2
h
i
λ2 σ2 (u)
0
E e λM (u,0) |`(u) = − ≤ e λµ0 (u,−) e 2
for any λ, with parameter
¸
·
µ0 (u, +)
∈ R2 ,
µ0 (u) =
µ0 (u, −)
σ20 (u) ∈ R.
Define the following updating rules for a node v
µt (v) =
θ̄K µt −1 (u)
X
(11)
u∈C (v)
σ2t (v) =
½
¸ ¾
·
µt −1 (u, +) − µt −1 (u, −) 2
2
θ̄ σt −1 (u) +
.
2
u∈C (v)
X
2
(12)
Then the following concentration-of-measure holds for the root message M (o, t̄ ):
M (o, t̄ )|`(o)=+ ≥ µt̄ (o, +) − x · σt̄ (o)
M (o, t̄ )|`(o)=− ≤ µt̄ (o, −) + x · σt̄ (o)
2
both with probability 1 − exp(− x2 ).
µ (o,+)+µt̄ (o,−)
2
In addition, if we choose t̄
fication error is upper bounded by
as the cut-off to provide classification `ˆwmp , then the misclassiÃ
exp −
[µt̄ (o, +) − µt̄ (o, −)]2
8σ2t̄ (o)
!
.
(13)
The above Lemma provides an expression on the classification error. The next Theorem will show
that with the “optimal” initialization for WMP, the misclassification error is connected to the minimum
energy.
10
Theorem 2 (Connection between Misclassification & Energy). Define the current flow
i(
v) =
θ̄ 2|v| [µt −|v| (v, +) − µt −|v| (v, −)]
[µt (o, +) − µt (o, −)]
.
Then it is a valid unit flow on T t (o), and the following equation holds
σ2t (o)
h
[µt (o,+)−µt (o,−)]
2
i2 = (1 + o t (1))
X
i(
¡
¢|v|
v)2 θ̄ −2
= (1 + o t (1))Et (i, θ̄ −2 )
v∈T t (o)
when limt →∞ Et (i, θ̄ −2 ) < ∞. Moreover, if we choose µ0 (v) so that i is the minimum energy flow, then under
the condition
br[T (o)]θ̄ 2 > 1,
we have E∗ (θ̄ −2 ) < ∞ and
lim inf h
t →∞
i
σ2t (o)
[µt (o,+)−µt (o,−)]
2
i2 ≤
X
i∗ (
¡
¢|v|
v)2 θ̄ −2
= E∗ (θ̄ −2 ).
(14)
v∈T (o)
Remark 3. The above Theorem 2 and Lemma 1 together state the fact that if br[T (o)]θ̄ 2 > 1, E∗ (θ̄ −2 ) is
finite, and the optimal initialization of WMP enjoys the asymptotic misclassification error bound of
µ
¶
1
exp −
.
2E∗ (θ̄ −2 )
Qualitatively, the smaller the minimum energy is, the smaller the misclassification error is, and it decays
exponentially. On the contrary, if the minimum energy is infinite (br[T (o)]θ̄ 2 < 1), the misclassification
error bound for WMP becomes vacuous. Another remark is that when the tree is regular, the minimum
energy takes the simple form E∗ (θ̄ −2 ) =
asymptotic misclassification error.
θ̄ 2 −1
1
, which implies the upper bound exp(− br[T (o)]
) on
2
br[T (o)]θ̄ 2 −1
3.4 Below the Threshold: Limitation of Local Algorithms
In this section, we will show that the SNR threshold (for WMP algorithm) is indeed sharp for the local
algorithm class. The argument is based on Le Cam’s method. Let us prove a generic lower bound for any
fixed tree T t (o), and for the k = 2 label broadcasting process with transition matrix K (as in Eq. (9)).
Theorem 3 (Limitation of Local Algorithms). Recall the label broadcasting process with Markov transition
kernel K on tree T t (o). Consider the case when noisy label information (with parameter δ) is known on
the depth-t layer leaf nodes. Denote the following two measures π+
, π−
as distributions on leaf labels
`
`
T t (o)
given `(o) = +, − respectively. Under the condition
T t (o)
br[T (o)]θ̄ 2 < 1,
2
4δ
2
if log(1 + 1−δ
2 ) ≤ 1 − br[T (o)]θ̄ , the following equality on total variation holds
³
2
lim d TV
π+
`T
t →∞
t
, π−
`T
(o)
11
´
t (o)
= 0.
Furthermore, the above equation implies
lim inf sup P (σt (o) 6= `(o)|`(o) = l ) =
t →∞ σt l ∈{+,−}
1
2
where σt (o) : `˜prior (C t (o)) → {+, −} is any estimator mapping the prior labels in the local tree to a decision.
The above theorem is stated under the case when the noisy label information is known and only
known for all nodes in layer t . One can interpret the result as, below the threshold br[T (o)]θ̄ 2 < 1, one
cannot do better than random guess for the root’s label based on noisy leaf labels at depth t as t → ∞.
The proof relies on a technical lemma on branching number and cutset as in [37]. We would like to
4δ2
2
remark that the condition log(1 + 1−δ
2 ) ≤ 1 − br[T (o)]θ̄ can be satisfied when δ is small.
4 General Number of Communities
In this section, we will extend the algorithmic and theoretical results to the general SBM for any fixed k
or growing k with a slow rate (with respect to n). There are several differences between the general k case
and the k = 2 case. First, algorithmically, the procedure for general k requires another layer of weighted
aggregation besides the weights introduced by minimum energy flow (according to the detailed tree irregularity). The proposed procedure introduces the weights on the types of labels (k types) revealed,
and then aggregates the information in the most “informative direction” to distinguish the root’s label.
Second, the theoretical tools we employ enable us to formally describe the intuition that in some cases
for general SBM, one can distinguish the communities i , j from k, but not being able to tell i and j apart.
We will call this the set identification.
4.1 Summary of Results
We summarize in this section the main results for general SBM with k unequal size communities, and
introduce the corresponding weighted message passing algorithm (WMP).
We need one additional notation before stating the main result. For a vector w ∈ Rk , assume there
are m unique values for w l , l ∈ [k]. Denote by S i , 1 ≤ i ≤ m, the sets of equivalent values associated with
w — for any l , l 0 ∈ [k], w l = w l 0 if and only if l , l 0 ∈ S i for some i ∈ [m]. Denote w S i to be the equivalent
value w l , l ∈ S i .
Theorem 4 (General SBM: k communities). Consider the general stochastic block model G(V, E ) with
parameter bundle (n, k, N ,Q), with either partial or noisy label information `˜prior with parameter 0 < δ <
1. Assume that n maxi , j Q i j - n o(1) . For any node o ∈ V and its depth t leaf labels `˜prior (C t (o)), define the
set misclassification error of a local estimator σt (o) : `˜prior (C t (o)) → [k] as,
ErrS,T (σt ) := max {P (σt (o) ∈ S|`(o) ∈ T ) , P (σt (o) ∈ T |`(o) ∈ S)} ,
where S, T ⊂ [k] are two disjoint subsets. Define
£
¤−1
K := diag(QN ) Qdiag(N ), M = Qdiag(N )
θ := λ2 (K ), λ := λ1 (M ).
(15)
(16)
(17)
Let E∗ (1/θ 2 ) be the minimum energy on T t (o) with conductance level θ 2 as t → ∞. Assume that K is
symmetric and denote V ∈ Rk to be the space spanned by the second eigenvectors of K . Choose any w ∈
V, w ⊥ 1 as the initialization vector in WMP Algorithm 2.
12
On the one hand, when λθ 2 > 1, the WMP Algorithm 2 initialized with w outputs `ˆwmp that can distinguish the indices set S i , 1 ≤ i ≤ m
µ
lim sup lim sup max ErrS i ,S j (`ˆwmp ) ≤ exp −
t →∞
n→∞ i , j ∈[m]
for any fixed δ > 0, where R 2 =
maxi , j |w S i −w S j | .
¶
R2
,
2E∗ (1/θ 2 )
(18)
mini , j |w S i −w S j |
On the other hand, if λθ 2 < 1, for any t -local estimator σt that only based on layer t ’s noisy labels, the
minimax misclassification error is lower bounded by
lim inf lim inf inf
t →∞
n→∞
sup
σt i , j ∈[k],i 6= j
Erri , j (σt ) ≥
1
.
2k
(19)
The proof for general k case requires several new ideas compared to the k = 2 case. Let us first explain
the intuition behind some quantities here. Again we focus on the case when the network is sparse, i.e.
n maxi , j Q i j - n o(1) . According to the coupling Proposition 1, one can focus on the coupled multi-type
Galton-Watson tree, for a shallow local neighborhood of a node o. K ∈ Rk×k then denotes the transition
kernel for the label broadcasting process on the tree, and λ denotes the branching number of the multitype Galton-Watson tree. The transition threshold λθ 2 = 1, also called Kesten-Stigum bound, has been
well-studied for reconstruction on trees [21, 22, 29, 18]. Our contribution lies in establishing the connection between the set misclassification error, minimum energy flow, as well as the second eigenvectors of
K . This is done through analyzing Algorithm 2 (to be introduced next) with a novel initialization of the
messages, using both minimum energy flow and the eigenvectors of K .
Remark 4. One distinct difference between the general k case and the k = 2 case is the notion of set
misclassification error, or set identification. This formalizes the intuition that for general SBM that is
asymmetric and imbalanced, it may be possible to distinguish communities i , j from community l , yet
not possible to tell i and j apart. The above Theorem provides a mathematical description of the phenomenon, for any initialization using vectors in the eigen-space corresponding to the second eigenvalue.
The key new ingredient compared to the Algorithm 1 is the introduction of additional weights w ∈ Rk
13
on the labels. The choice of w will become clear in a moment.
Algorithm 2: Weighted Message Passing for Multiple Communities
Data: Same as in Algorithm 1 and an additional weight vector w ∈ Rk .
Result: The labeling for each node o ∈ V .
for each node o ∈ V , do
Open the tree neighborhood T t̄ (o) ;
Layer t̄ : for every node u ∈ C t̄ (o), initialize its message
M (u, 0) = θ −2|u| · i∗ (
u) · w `˜prior (u) ,
where w `˜prior (u) denotes the `˜prior (u)-th coordinate of the weight vector w, i∗ (
u) is the
minimum energy flow ;
Initialize parameters µ0 (u) ∈ Rk , σ20 (u) ∈ R as
µ0 (u, l ) = δ · θ −2|u| i∗ ( u) · w l , for l ∈ [k]
¢2
¡
σ20 (u) = θ −2|u| i∗ ( u) · max |w i − w j |2
i , j ∈[k]
for t = 1, . . . t̄ , do
Layer t̄ − t : for every node u ∈ C t̄ −t (o), update message M (u, t ) through the linearized rule
X
M (u, t ) =
θM (v, t − 1).
v∈C (u)
Update the parameters µt (u) ∈ Rk , σ2t (u) ∈ R
µt (u) =
X
θK µt −1 (v)
v∈C (u)
σ2t (u) =
θ 2 σ2t −1 (v) +
v∈C (u)
X
max |µt −1 (v, i ) − µt −1 (v, j )|
i , j ∈[k]
2
2
.
end
Output `ˆwmp (o) = arg minl ∈[k] |M (o, t̄ ) − µt̄ (o, l )|.
end
4.2 Vector Evolution & Concentration
As in the k = 2 case, we establish the recursion formula for the parameter updates. However, unlike
the k = 2 case, for a general initialization µ0 , it is much harder to characterize µt (u), σ2t (u) analytically,
and thus relate the misclassification error to the minimum energy. We will show that this goal can be
achieved by a judicious choice of µ0 . We will start with the following Lemma that describes the vector
evolution and concentration-of-measure.
Lemma 2 (Concentration, general k). Recall the label broadcasting process with Markov transition kernel
14
K ∈ Rk×k on tree T t̄ (o). Assume the MGF of messages on the leaves M (u, 0) satisfies, for any ` ∈ [k]
h
i
λ2 σ2 (u)
0
E e λM (u,0) |`(u) = l ≤ e λµ0 (u,l ) e 2
for any λ, with parameter
µ0 (u) = [µ0 (u, 1), . . . , µ0 (u, k)] ∈ Rk ,
σ20 (u) ∈ R.
Define the following updating rules for a node v
X
µt (v) =
θK µt −1 (u)
u∈C (v)
max |µt −1 (u, i ) − µt −1 (u, j )| 2
X 2
i , j ∈[k]
.
σ2t (v) =
θ σ2t −1 (u) +
2
u∈C (v)
The following concentration-of-measure holds for the root message M (o, t̄ ):
M (o, t̄ )|`(o)=l ∈ µt̄ (o, l ) ± x · σt̄ (o)
2
with probability 1 − 2 exp(− x2 ). In addition, if we we classify the root’s label as
`ˆwmp (o) = arg min |M (o, t̄ ) − µt̄ (o, l )|,
l ∈[k]
then the worst-case misclassification error is upper bounded by
exp(−
mini , j ∈[k] |µt̄ (o, i ) − µt̄ (o, j )|2
8σ2t̄ (o)
).
(20)
Remark 5. Unlike the k = 2 case, in general it is hard to quantitatively analyze this evolution system
for µt (u), σ2t (u). The main difficulty stems from the fact that the coordinates that attain the maximum of
maxi , j ∈[k] |µt −1 (u, i )−µt −1 (u, j )| vary with u, t . Hence, it is challenging to provide sharp bounds on σ2t (u).
In some sense, the difficulty is introduced by the instability of the relative ordering of the coordinates of
the vector µt (u) for an arbitrary initialization.
As will be shown in the next section, one can resolve this problem by initializing µ0 (u, l ), l ∈ [k] in a
“most informative” way. This initialization represents the additional weights on label’s types beyond the
weights given by the minimum energy flow.
4.3 Additional Weighting via Eigenvectors
We show in this section that the vector evolution system with noisy initialization is indeed tractable if we
weigh the label’s type according to the second right eigenvector of K ∈ Rk×k .
Theorem 5 (Weighting by Eigenvector). Assume that the second eigenvalue θ = λ2 (K ) of the Markov transition kernel K is a real, and denote the associated second right eigenvector by w ∈ Rk , kwk = 1, w T 1 = 0.
Denote the minimum energy flow on tree T (o) with conductance level θ 2 by i∗ . In the case of noisy label
information with parameter δ, if we initialize
µ0 (u, l ) = δ · θ −2|u| i∗ (
15
u) · w l , for l ∈ [k],
¡
and σ20 (u) = θ −2|u| i∗ (
bounded by
u)
¢2
· maxi , j ∈[k] |w i − w j |2 , then the worst case misclassification error is upper
lim sup
t →∞
with R =
max
i , j ∈[k],i 6= j
P(`ˆwmp (o) = i |`(o) = j ) ≤ exp(−
R2
2E∗ (θ −2 )
)
mini , j |w i −w j |
maxi , j |w i −w j | .
Remark 6. Observe that the upper bound becomes trivial when mini , j |w i − w j | = 0. In this case, one can
easily modify in the proof of Theorem 5 so that the following non-trivial guarantee for set misclassification error holds. Assume w has m distinct values, and denote the set S i , 1 ≤ i ≤ m to be the distinct value
sets associated with w. Then one has the following upper bound on the set misclassification error
lim sup
t →∞
with R S =
max
i , j ∈[m],i 6= j
P(`ˆwmp (o) ∈ S i |`(o) ∈ S j ) ≤ exp(−
R S2
2E∗ (θ −2 )
)
(21)
mini , j |w S i −w S j |
maxi , j |w S i −w S j | .
4.4 Lower Bound: Sharp Threshold
In this section we provide a new lower bound analysis through bounding the χ2 distance to the “average
measure”. The lower bound shows that the transition boundary λθ 2 = 1 achieved by WMP is sharp for
any k. To the best of our knowledge, the first lower bound for general k case is achieved in [18] using a
notion of weighted χ2 distance. For completeness of the presentation, we provide here a different proof
using the usual χ2 distance. In addition, our approach admits a clear connection to the upper bound
analysis through matrix power iterations.
Theorem 6 (Limitation for Local Algorithms, k-communities). Recall the label broadcasting process with
Markov transition kernel K on tree T t (o). Assume K ∈ Rk×k is symmetric. Consider the case when noisy
label information (with parameter δ) is known on the depth-t layer leaf nodes. Under the condition
br[T (o)]θ 2 < 1
and kδ2 (
1
δ+ 1−δ
k
+
1
1−δ
k
) < 1 − br[T (o)]θ 2 , we have
1
1
lim inf inf max P(σt (o) 6= `(o)|`(o) = l ) ≥ (1 − ).
t →∞ σt l ∈[k]
2
k
where σt (o) : `˜prior (C t (o)) → [k] is any estimator mapping the prior labels on leaves in the local tree to a
decision. The above inequality also implies
lim inf inf
t →∞
max
σt i , j ∈[k],i 6= j
P(σt (o) = i |`(o) = j ) ≥
1
.
2k
1
The above result shows that even belief propagation suffers the error at least 2k
in distinguishing i , j ,
which is within a factor of 2 from random guess. We remark in addition that the condition kδ2 ( 11−δ +
δ+
1
1−δ
k
2
) < 1 − br[T (o)]θ can be satisfied when δ is small.
16
k
4.5 Further Discussion
Local versus Global Algorithms In the balanced case with k equal size communities, and p, q denoting
the within- and between-community connection probabilities, the Kesten-Stigum threshold for local
algorithm class takes the following expression
SNR :=
n(p − q)2
k 2 (q +
p−q
k )
= 1.
However, it is known that the limitation for global algorithm class for growing number of communities is
log n
log k
SNR ³ O ( k ) ([3], weak consistency) and SNR ³ O ( k ) ([8], strong consistency). Therefore, as k grows,
there is an interesting gap between local and global algorithms in terms of SNR. An interesting direction
is to determine whether one can solve the problem down to the information-theoretic threshold O ∗ ( k1 )
with computationally efficient algorithms.
5 Numerical Studies
We apply the message passing Algorithm 1 to the political blog dataset [4] (with a total of 1222 nodes)
in the partial label information setting with δ portion randomly revealed labels. In the literature, the
state-of-the-art result for a global algorithm appears in [19], where the misclassification rate is 58/1222 =
4.75%. Here we run a weaker version of our WMP algorithm as it is much easier to implement and does
not require parameter tuning. Specifically, we initialize the message with a uniform flow on leaves (minimum energy flow that corresponds to a regular tree). We will call this algorithm approximate message
passing (AMP) within this section.
We run AMP with three different settings δ = 0.1, 0.05, 0.025, repeating each experiment 50 times. As
a benchmark, we compare the results to the spectral algorithm on the (1−δ)n sub-network. We focus on
the local tree with depth 1 to 5, and output the error for message passing with each depth. The results are
summarized as box-plots in Figure 1. The left figure illustrates the comparison of AMP with depth 1 to 5
and the spectral algorithm, with red, green, blue boxes corresponding to δ = 0.025, 0.05, 0.1, respectively.
The right figure zooms in on the left plot with only AMP depth 2 to 4 and spectral, to better emphasize the
difference. Remark that if we only look at depth 1, some of the nodes may have no revealed neighbors.
In this setting, we classify this node as wrong (this explains why depth-1 error can be larger than 1/2).
We present in this paragraph some of the statistics of the experiments, extracted from the above
Figure 1. In the case δ = 0.1, from depth 2-4, the AMP algorithm produces the mis-classification error
rate (we took the median over the experiments for robustness) of 6.31%, 5.22%, 5.01%, while the spectral
algorithm produces the error rate 6.68%. When δ = 0.05, i.e. about 60 node labels revealed, the error
rates are 7.71%, 5.44%, 5.08% with depth 2 to 4, contrasted to the spectral algorithm error 6.66%. In a
more extreme case δ = 0.025 when there are only ∼ 30 node labels revealed, AMP depth 2-4 has error
10.20%, 5.71%, 5.66%, while spectral is 6.63%. In general, the AMP algorithm with depth 3-4 uniformly
beats the vanilla spectral algorithm. Note that our AMP algorithm is a distributed decentralized algorithm that can be run in parallel. We acknowledge that the error ∼ 5% (when δ is very small) is still
slightly worse than the state-of-the-art degree-corrected SCORE algorithm in [19], which is 4.75%.
17
0.16
0.75
0.12
delta
0.025
0.50
0.05
0.1
error_rate
error_rate
delta
0.025
0.05
0.1
0.08
0.25
0.04
0.00
depth1
depth2
depth3
depth4
depth5
spectral
depth2
depth_AMP
depth3
depth4
spectral
depth_AMP
Figure 1: AMP algorithm on Political Blog Dataset.
6 Technical Proofs
We will start with two useful results. The first one is a coupling proposition. The proof follows exactly the
same idea as in Proposition 4.2 in [30]. The intuition is that when the depth of the tree is shallow, the SBM
in the sparse regime can be coupled to a Galton-Watson tree with Poisson branching (as there are many
nodes outside the radius R for the Poisson-Multinomial coupling, when R small). We want to prove a
more general version for SBM with unequal size communities. The proof is delayed to Appendix 7.
Proposition 1. Let R = R(n) = b 4 log[2np10 +2 log n] log nc, where p 0 = maxi , j Q i j . Denote (T, σT ) to be the
multi-type Galton-Watson tree (with Poisson branching) with mean matrix Qdiag(N ) and label tran£
¤−1
sition kernel K = diag(QN ) Qdiag(N ). Denote G R as the neighborhood of depth up to R induced
by the graph G, for a particular node. There exists a coupling between (G R , `G R ) and (T, σT ) such that
(G R , `G R ) = (TR , σTR ) with high probability as n → ∞. Here the tree equivalence is up to a label preserving
homomorphism.
Lemma 3 (Hoeffding’s Inequality). Let X be any real-valued random variable with expected value E X = 0
and such that a ≤ X ≤ b almost surely. Then, for all λ > 0,
µ 2
¶
h
i
λ (b − a)2
E e λX ≤ exp
.
8
Proof of Lemma 1. Recall the linearized message passing rule that “approximates” the Bayes optimal algorithm:
X
θ1 + θ2
M (u, t ) =
θ̄ · M (v, t − 1), where θ̄ =
.
2
v∈C (u)
Let us analyze the behavior of the linearized messages M (u, t ) for a particular node u. The proof
follows by induction on t . The case t = 0 follows from the assumption about µ0 (u), σ20 (u) and Chernoff
18
bound. Now, assume that the induction premise is true for t − 1. Note that
h
i
E e λM (u,t ) |`(u) = +
h
i
Y
=
E e λθ̄M (v,t −1) |`(u) = +
v∈C (u)
½ h
h
i 1−θ ¾
i 1+θ
1
1
+ E e λθ̄M (v,t −1) |`(v) = −
E e λθ̄M (v,t −1) |`(v) = +
2
2
v∈C (u)
½
¾
Y (λθ̄)2 σ2t −1 (v) λθ̄µ (v,+) 1 + θ1
1 − θ1
t −1
2
≤
e
e
+ e λθ̄µt −1 (v,−)
2
2
v∈C (u)
=
Y
≤
Y
e (λθ̄)
2
2 σt −1 (v)
2
e λθ̄[µt −1 (v,+)
1+θ1
2
+µt −1 (v,−)
1−θ1
2
] (λθ̄)2
e
[µt −1 (v,+)−µt −1 (v,−)]2
8
,
v∈C (u)
where the last step uses the Hoeffding’s Lemma. Rearranging the terms,
λ2 θ̄ 2
i
P
λM (u,t )
λ v∈C (u) θ̄〈K 1· ,µt −1 (v)〉
|`(u) = + ≤ e
e
E e
h
= e λµt (u,+) e
λ2 σ2
t (u)
2
P
(
·
¸ )
µt −1 (v,+)−µt −1 (v,−) 2
2
v∈C (u) σt −1 (v)+
2
2
,
where K 1· denotes the first row of transition matrix K . Clearly, same derivation holds with `(u) = −.
Applying the Chernoff bound and optimizing over λ, one arrives at the exponential concentration bound.
Induction completes.
To upper bound the
error, simply plug in the standardized absolute values of the
¯ misclassification
¯
¯ µt̄ (o,+)−µt̄ (o,−) ¯
difference, namely x = ¯
¯.
2σ (o)
t̄
Remark 7. Now let us propose the choice of µ0 (u) and σ20 (u) for the case of noisy label information with
parameter δ. In WMP algorithm, choose M (u, 0) = c(u) sign(`˜prior ) with factor c(u) that depends on the
node u. Using simple Hoeffding’s concentration for Bernoulli r.v., one has
µ0 (u, +) = c(u)δ, µ0 (u, −) = −c(u)δ,
and σ20 (u) = c(u)2 .
Proof of Theorem 2. Using the result of Lemma 1, the proof analyzes evolution of
σ2t (o)
h
i
[µt (o,+)−µt (o,−)] 2
2
.
First, let us derive the expression for µt (o, +) − µt (o, −). Denoting w = [1, −1]T , it is easy to verify that
w T K = θ̄w T . We have,
µt (o, +) − µt (o, −) =
X
θ̄w T K µt −1 (v) =
v∈C (o)
X
2
= θ̄
X
θ̄ 2 w T µt −1 (v)
v∈C (o)
[µt −1 (v, +) − µt −1 (v, −)].
v∈C (o)
19
Using the above equation recursively, one can easily see that for any d , 1 ≤ d ≤ t ,
X
µt (o, +) − µt (o, −) = θ̄ 2d
[µt −d (v, +) − µt −d (v, −)].
(22)
v∈C d (o)
Now for σ2t (o) for σ2t (ρ), one has
σ2t (o) = θ̄ 2
¸ ¾
½
·
µt −1 (v, +) − µt −1 (v, −) 2
2
σt −1 (v) +
2
v∈C (o)
X
which can be written, in turn, as
¸ ¾
½
·
µt −2 (u, +) − µt −2 (u, −) 2
2
θ̄
θ̄
σt −2 (u) +
2
v∈C (o)
u∈C (v)
·
¸2
X
µt −1 (v, +) − µt −1 (v, −)
+ θ̄ 2
2
v∈C (ρ)
·
¸
·
¸
X
X
X
µt −2 (u, +) − µt −2 (u, −) 2
µt −1 (v, +) − µt −1 (v, −) 2
4
2
= . . . . . . + θ̄
+ θ̄
2
2
v∈C (ρ) u∈C (v)
v∈C (o)
¸2
·
X
X 2|v| µt −|v| (v, +) − µt −|v| (v, −)
+
θ̄ 2t σ20 (u).
=
θ̄
2
t
v∈T t (o)
u∈C (o)
2
X
2
X
Using the above equation one can bound
P
σ2t (o)
h
[µt (o,+)−µt (o,−)]
2
i2 =
2|v|
v∈T t (o) θ̄
h
h
i
µt −|v| (v,+)−µt −|v| (v,−) 2
2
[µt (o,+)−µt (o,−)]
2
i2
u∈C t (o) θ̄
P
+h
2t
σ20 (u)
i2
[µt (o,+)−µt (o,−)]
2
£
¤2
θ̄ 2|v| µt −|v| (v, +) − µt −|v| (v, −)
+R
=
£
¤2
[µt (o, +) − µt (o, −)]
v∈T t (o)
¢2
¡
X θ̄ 2|v| [µt −|v| (v, +) − µt −|v| (v, −)]
θ̄ −2|v| + R
=
¡
¢2
[µt (o, +) − µt (o, −)]
v∈T t (o)
X
where the remainder
u∈C t (o) θ̄
P
R= h
Recall the definition of
i(
v) =
2t
(23)
σ20 (u)
i2 .
[µt (o,+)−µt (o,−)]
2
θ̄ 2|v| [µt −|v| (v, +) − µt −|v| (v, −)]
[µt (o, +) − µt (o, −)]
.
It is clear from Eq.(22) that i is a valid unit flow, in the sense of Definition 1. Continuing with Eq. (23), one
has
inf h
i
σ2t (o)
[µt (o,+)−µt (o,−)]
2
i2 ≤
X
i∗ (
v)2 θ̄ −2|v| + R
v∈T t (o)
= Et (i∗ , θ̄ −2 ) + R.
20
(24)
Let us now estimate R:
u∈C t (o) θ̄
P
R= h
i∗ (
u)2 θ̄ −2t · max h
t
i∗ (
u)2 θ̄ −2t
u∈C (o)
u∈C t (o)
X
=
σ20 (u)
i2
[µt (o,+)−µt (o,−)]
2
X
≤
2t
u∈C t (o)
σ20 (u)
i
[µ0 (u,+)−µ0 (u,−)] 2
2
1
.
δ2
The last step is because for noisy label information with parameter δ,
σ20 (u)
h
[µ0 (u,+)−µ0 (u,−)]
2
i2 =
1
.
δ2
P
In the case when limt →∞ Et (i∗ , θ̄ −2 ) < ∞, we know u∈C t (o) i∗ ( u)2 θ̄ −2t = Et (i∗ , θ̄ −2 ) − Et −1 (i∗ , θ̄ −2 ) → 0.
Therefore, R = δ12 o t (1).
Going back to Eq. (24), to minimize the LHS (ratio between noise and signal), one needs to make sure
that i = i∗ , the minimum energy flow. Therefore, the optimal strategy is to initialize µ0 (u) according to
i∗ ( u). Thus, if we choose
µ0 (u, +) = δθ̄ −2|u| i∗ (
u), µ0 (u, −) = δθ̄ −2|u| i∗ (
u),
we obtain
lim inf h
t →∞
i
σ2t (o)
i
[µt (o,+)−µt (o,−)] 2
2
= E∗ (θ̄ −2 ).
From Definition 3,
E∗ (θ̄ −2 ) < ∞ iff θ̄ −2 < br[T (o)].
Proof of Theorem 5. Note that by Perron-Frobenius Theorem, we have |θ| = |λ2 (K )| < 1. Thanks to the
choice of w,
E[M 0 (u)|`(u) = l ] = δθ −2|u| i∗ (
u)w l +
1 − δ −2|u| ∗
θ
i (
k
u)w T 1 = δθ −2|u| i∗ (
u)w l .
Let us first derive the formula for µt (o) ∈ Rk under the chosen initialization µ0 (u). We claim that
µt −|v| (v) = δ · θ −2|v| i∗ (
v) · w.
Proof is via induction. The base case |u| = t is exactly the choice of the initialization. Let us assume for
|u| > |v| the claim is true, and prove for v:
X
µt −|v| (v) =
θK µt −1 (u)
u∈C (v)
=
X
θK w · δθ −2|v|−2 i∗ (
u)
u∈C (v)
=
X
θ 2 w · δθ −2|v|−2 i∗ (
u∈C (v)
21
v) = δ · θ −2|v| i∗ (
v) · w,
completing the induction.
Now let us bound σ2t (o). Observe that in our derived formula for µt −|v| (v), all the coordinates are
proportional to w. In other words, µt −|v| (v) stays in the direction of w for all v. This greatly simplifies
the expression for σ2t (o). We have
σ2t (o) =
θ 2|v|
X
·
maxi , j ∈[k] |µt −|v| (v, i ) − µt −|v| (v, j )|
=δ
·
maxi , j ∈[k] |w(i ) − w( j )|
·
+
u∈C t (o)
¸2
2
maxi , j ∈[k] |w(i ) − w( j )|
i∗ (
X
v∈T t (o)
¸2
X
2
Plugging in the definition R =
X
+
2
v∈T t (o)
2
¸2
i∗ (
θ 2t σ20 (u)
v)2 θ −2|v|
v)2 θ −2|v| .
v∈C t (o)
mini , j |w i −w j |
maxi , j |w i −w j | , under the condition
br[T (o)]θ 2 > 1,
we have E(i∗ , θ −2 ) < ∞, and
σ2t̄ (o)
h
mini , j ∈[k] |µt̄ (o,i )−µt̄ (o, j )|
2
i2 =
1
1
E(i∗ , θ −2 ) + 2 2 o t (1).
2
R
δ R
Proof of Theorem 6. Recall that π(`∂Tt (o)∩Tt −|u| (u) |`(u) = i ) denotes the probability measure on the leaf labels on depth t , given `(u) = i . For a node u, when there is no confusion, we abbreviate the measure
π(`∂Tt (o)∩Tt −|u| (u) |`(u) = i ) as πu (i ). According to Perron-Frobenius Theorem, there is a unique left eigenvector for K with eigenvalue 1, denote this by w ∈ Rk . Under the assumption K being symmetric, we
P
know that w = k1 1. Denote π̄u = kj=1 w( j )πu ( j ).
Let us bound the d χ2 (πu (i )||π̄u ) by deriving a recursive bound:
"
Ã
!#
k
k X
k
X
X
X
£
¤
log 1 + d χ2 (πu (i )||π̄u ) =
log 1 + d χ2
K i l πv (l )||
w( j )K j l πv (l )
v∈C (u)
"
=
X
Ã
log 1 + d χ2
v∈C (u)
l =1
j =1 l =1
k
X
!#
K i l πv (l )||π̄v
l =1
since w T K = w T . By definition, the above expression is
¤2
Z £Pk
X
X
K
π
(l
)
−
π̄
v
v
i
l
l =1
log 1 +
π̄
v
v∈C (u) v∈C (u)
¤2
Z £Pk
X
X
K
(π
(l
)
−
π̄
)
v
v
i
l
l =1
=
log 1 +
π̄
v
v∈C (u) v∈C (u)
£Pk
¤2
X Z
K (πv (l ) − π̄v )
l =1 i l
≤
.
π̄v
v∈C (u)
22
Now we know that
£Pk
¤2
k
X Z X
K (πv (l ) − π̄v )
£
¤
l =1 i l
.
log 1 + d χ2 (πu (i )||π̄u ) ≤
π̄v
i =1
v∈C (u) i =1
k
X
Recall the following fact that for any z 1 , z 2 , . . . z k ≥ 0,
log(1 +
k
X
zi ) ≤
k
X
log(1 + z i ).
i =1
i =1
Using this fact the lower bound the LHS, we reach
"
log 1 +
k
X
i =1
#
d χ2 (πu (i )||π̄u ) ≤
¤2
£Pk
k
X Z X
K (πv (l ) − π̄v )
l =1 i l
π̄v
i =1
Z
X
kK (πv (·) − π̄v 1)k2
≤
π̄v
v∈C (u)
Z
X
k(πv (·) − π̄v 1)k2
≤ θ2
π̄v
v∈C (u)
v∈C (u)
k
X X
= θ2
v∈C (u) i =1
d χ2 (πv (i )||π̄v )
where the last two lines use the fact that πv (·) − π̄v 1 ⊥ 1, therefore kK (πv (·) − π̄v 1)k2 ≤ θ 2 kπv (·) − π̄v 1k2 .
We will need the the following Lemma that describes the branching number through the cutset.
Lemma 4 ([37], Lemma 3.3). Assume br[T ] < λ. Then for all ² > 0, there exists a cutset C such that
µ ¶
X 1 |x|
≤²
x∈C λ
(25)
µ ¶|x|−|v|
1
≤ 1.
x∈C ∩T (v) λ
(26)
and for all v such that |v| ≤ maxx∈C |x|,
X
Here the notation |v| denotes the depth of v.
P
Let us use the cutset argument to prove ki=1 d χ2 (πu (i )||π̄u ) → 0 when |u| → ∞. Fix any λ such
that θ −2 > λ > br[T (o)]. For any ² small, the above Lemma claims the existence of cutset C ² such that
Eq. (25) and (26) hold. Let us prove through induction on maxx∈C ² |x| − |v| that for any v such that
|v| ≤ maxx∈C ² |x|, we have
k
X
i =1
with the choice η = kδ2 (
1
δ+ 1−δ
k
+
d χ2 (πv (i )||π̄v ) ≤ η
1
1−δ
k
µ ¶|x|−|v|
1
≤ η.
x∈C ² ∩T (v) λ
X
(27)
). First for the base case, the claim is true because of the choice of η.
23
Preceding with the induction, assume for v such that maxx∈C ² |x|−|v| = t −1 equation (29) is satisfied,
and let us prove for v : maxx∈C ² |x| − |u| = t . We recall the linearized recursion
"
#
k
k
X
X X
log 1 +
d χ2 (πu (i )||π̄u ) ≤ θ 2
d χ2 (πv (i )||π̄v )
i =1
v∈C (u) i =1
µ ¶|x|−|v|
1
v∈C (u) x∈C ² ∩T (v) λ
µ ¶|x|−|u|
X
1
2
= θ λ·η
λ
x∈C ² ∩T (u)
≤ θ2
Using the assumption θ 2 λ <
η
X
η
X
1
1+η , the above can be upper bounded by
P
¡ 1 ¢|x|−|u|
x∈C ² ∩T (u) λ
1+η
≤
¡ 1 ¢|x|−|u|
x∈C ² ∩T (u) λ
¡ ¢|x|−|u|
P
1 + η x∈C ² ∩T (u) λ1
η
P
¡ ¢|x|−|u|
P
where the last inequality uses the fact that x∈C ² ∩T (u) λ1
< 1. Now we know that
"
#
Pk
k
2
X
i =1 d χ (πu (i )||π̄u )
≤ log 1 +
d χ2 (πu (i )||π̄u )
P
1 + ki=1 d χ2 (πu (i )||π̄u )
i =1
¡ ¢|x|−|u|
P
η x∈C ² ∩T (u) λ1
≤
¡ ¢|x|−|u| .
P
1 + η x∈C ² ∩T (u) λ1
By monotonicity of x/(1 + x) we have proved the induction claim holds as
µ ¶|x|−|u|
k
X
X
1
d χ2 (πu (i )||π̄u ) ≤ η
.
x∈C ² ∩T (u) λ
i =1
Take ² → 0, λ → br[T (o)]. Define t ² := min{|x|, x ∈ C ² }, it is also easy to see from equation (25) that
µ ¶t ²
µ ¶
X 1 |x|
1
log(1/²)
≤
≤ ² ⇒ t² >
→ ∞.
λ
log λ
x∈C ² λ
Putting things together, under the condition
η ≤ 1 − br[T (o)]θ 2 ,
we have
µ ¶|x|
k
X
1
1X
η
lim
d χ2 (πu (i )||π̄u ) ≤ · lim
= 0.
t →∞ k
k ²→0 x∈C ² ∩T (o) λ
i =1
Finally, we invoke the multiple testing argument Theorem 2.6 in [39]).
Lemma 5 ([39], Proposition 2.4, Theorem 2.6). Let P 0 , P 1 , . . . , P k be probability measures on (X , A ) satisfying
k
1X
d χ2 (P j , P 0 ) ≤ kα∗
k i =1
then we have for any selector ψ : X → [k]
1
1
max P i (ψ 6= i ) ≥ (1 − α∗ − ).
i ∈[k]
2
k
24
Plugging in the result with P 0 = π̄o and P i = πo (i ), we conclude that
1
1
lim inf inf max P(σ(o) 6= `(o)|`(o) = l ) ≥ (1 − ).
σ l ∈[k]
t →∞
2
k
Proof of Theorem 1. Given Proposition 1, Theorem 2 and Theorem 3, the proof of Theorem 1 is simple.
By Proposition 1, one can couple the local neighborhood of SBM with multi-type Galton Watson process
asymptotically almost surely as n → ∞, where the label transition matrix is
"
K :=
n 1 Q 11
n 1 Q 11 +n 2 Q 12
n 1 Q 21
n 1 Q 21 +n 2 Q 22
n 2 Q 12
n 1 Q 11 +n 2 Q 12
n 2 Q 22
n 1 Q 21 +n 2 Q 22
#
.
³
´
For the upper bound, Theorem 2 shows that the misclassification error is upper bounded by exp − ∗ 1 −2
E (θ̄ )
as the depth of the tree goes to infinity. Note if we first send n → ∞, due to Proposition 1, the coupling
is valid even when R → ∞ with a slow rate log n/ log log n. Therefore, the upper bound on misclassification error holds. One can establish the lower bound using the same argument together with Theorem 3. Finally, for the expression on transition boundary, we know that condition on non-extinction, the
branching number for this coupled multi-type Galton Watson tree is λ1 (Qdiag(N )) almost surely. Proof
is completed.
7 Additional Proofs
Proof of Lemma 2. The proof logic here is similar to the k = 2 case. Again, we analyze the message M (u, t )
for a particular node u. Use induction on t for the claim
h
i
λ2 σ2
t (u)
E e λM (u,t ) |`(u) = l ≤ e λµt (u,l ) e 2 .
The case for t = 0 follows from the assumption about µ0 (u), σ20 (u) and Chernoff bound.
Assume that the induction is true for t − 1, and prove the case for t . Note that
i
h
E e λM (u,t ) |`(u) = l
h
i
Y
=
E e λθM (v,t −1) |`(u) = l
v∈C (u)
(
=
≤
Y
h
i
k
X
E e λθM (v,t −1) |`(v) = i K l i
v∈C (u)
i =1
Y
(λθ)2
e
σ2 (v)
t −1
2
Y
k
X
e
λθµt −1 (v,i )
)
Kl i
i =1
v∈C (u)
≤
(
)
e (λθ̄)
2
2 σt −1 (v)
2
e λθ[
Pk
i =1 µt −1 (v,i )K l i ]
v∈C (u)
25
2
e (λθ)
maxi , j ∈[k] |µt −1 (v,i )−µt −1 (v, j )|2
8
,
where the last step uses the Hoeffding’s Lemma. Rearrange the terms, one can see that the above equation implies
λ2 θ 2
h
i
P
λM (u,t )
λ v∈C (u) θ〈K l · ,µt −1 (u)〉
e
E e
|`(u) = l ≤ e
= e λµt (u,l ) e
λ2 σ2
t (u)
2
P
(
¯
¯2 )
¯µ
¯
2
¯ t −1 (v,+)−µt −1 (v,−) ¯
¯
v∈C (u) σt −1 (v)+maxi , j ∈[k] ¯
2
2
,
where K l · denotes the l −row of transition matrix K . Apply the Chernoff bound to optimize over λ, one
can arrive the exponential concentration bound. Induction completes.
To upper bound the misclassification error, simply plug in
|x| =
mini , j ∈[k] |µt̄ (o, i ) − µt̄ (o, j )|
2σt̄ (o)
.
Proof of Theorem 3. We will gave the proof of Theorem 3 (for the δ noisy label information case) here.
Define the measure π+
on the revealed labels, for a depth t tree rooted from o with label `(o) = +
`
T t (o)
(and similarly define π−
`
T t (o)
). We have the following recursion formula
π+
`T t (o)
=
Y
v∈C (o)
·
¸
1 − θ1 −
1 + θ1 +
π`T (v) +
π`T (v) .
t −1
t −1
2
2
R µ2
Recall that the χ2 distance between two absolute continuous measures µ(x), ν(x) is d χ2 (µ, ν) = ν d x −
1, and we have the total
variation distance between these two measures is upper bounded by the χ2
¡
¢ q
¡
¢
distance d TV µ, ν ≤ d χ2 µ, ν .
Let us upper bound the symmetric version of χ2 distance defined as
n
³
´
³
−
−
+
2 π
D Tt (o) := max d χ2 π+
,
π
,
d
χ
`T (o) , π`T
`T (o) `T (o)
t
t
t
´o
t (o)
(abbreviate as D t (o) when there is no confusion), we have the following recursion
´i
h
³
−
log 1 + d χ2 π+
,
π
`T t (o) `T t (o)
¶¸
·
µ
X
1 − θ2 +
1 − θ1 −
1 + θ2 −
1 + θ1 +
2
π`T (v) +
π`T (v) ,
π`T (v) +
π`T (v)
=
log 1 + d χ
t −1
t −1
t −1
t −1
2
2
2
2
v∈C (o)
1 + θ1 +
1 − θ1 −
1 − θ2 +
1 + θ2 −
d χ2
π`T (v) +
π`T (v) ,
π`T (v) +
π`T (v)
t −1
t −1
t −1
t −1
2
2
2
2
³
´2
Z
π+
− π−
`T t −1 (v)
`T t −1 (v)
2
= θ̄
dx
1−θ2 +
1+θ2 −
π
π
+
2
2
`T (v)
`T t −1 (v)
" t −1
#
Z ³
´2 1 − θ
1
1
1 + θ2
2
2
+
−
≤ θ̄
π`T (v) − π`T (v)
+
dx
t −1
t −1
2 π+
2 π−
`
`
µ
T t −1 (v)
2
≤ θ̄ D Tt −1 (v) ,
26
T t −1 (v)
¶
where the second to last step follows from Jensen’s inequality for function 1/x. Now we have the following
recursion relationship
X
log(1 + D Tt (o) ) ≤
log(1 + θ̄ 2 · D Tt −1 (v) ).
v∈C (o)
Invoke the following fact,
log(1 + θ 2 x)
≤ (1 + η) log(1 + x) for all 0 ≤ x ≤ η, ∀θ,
θ2
whose proof is in one line
x
log(1 + θ 2 x)
≤ x ≤ (1 + η)
≤ (1 + η) log(1 + x).
θ2
1+x
Thus if D Tt −1 (v) ≤ η, ∀v ∈ C (o), then the following holds
X
log(1 + D Tt (o) ) ≤ (1 + η)θ̄ 2
log(1 + D Tt −1 (v) ).
(28)
v∈C u (ρ)
Denoting
d Tt (o) := log(1 + D Tt (o) ),
Equation (28) becomes
d Tt (o) ≤ (1 + η)θ̄ 2
X
v∈C u (ρ)
d Tt −1 (v) .
We will again need the Lemma 4 that describes the branching number through the cutset. Fix any λ
such that θ̄ −2 > λ > br[T (o)]. For any ² small, Lemma 4 claims the existence of cutset C ² such that Eq. (25)
and (26) holds. Let’s prove through induction on maxx∈C ² |x|−|v| that for any v such that |v| ≤ maxx∈C ² |x|,
we have
µ ¶|x|−|v|
X
η
1
η
d TC² (v) ≤
≤
.
(29)
1 + η x∈C ² ∩T (v) λ
1+η
Note for the start of induction v ∈ C ² ,
d TC² (v) = log(1 +
4δ2
η
)<
.
2
1−δ
1+η
Now precede with the induction, assume for u such that maxx∈C ² |x|−|u| = t −1 equation (29) is satisfied,
η
let’s prove for v : maxx∈C ² |x| − |v| = t . Due to the fact for all u ∈ C (v), d TC² (u) ≤ 1+η ⇒ D TC² (u) ≤ η, we can
recall the linearized recursion
X
d TC² (v) ≤ (1 + η)θ̄ 2
d T≤C² (u)
u∈C (v)
"
µ ¶|x|−|u| #
X
η
1
≤ (1 + η)θ̄
1
+
η
λ
x∈C ² ∩T (u)
u∈C (v)
µ ¶|x|−|u|+1
X
X
η
1
2
· (1 + η)θ̄ λ
≤
1+η
u∈C (v) x∈C ² ∩T (u) λ
µ ¶|x|−|v|
X
X
1
≤ ηθ̄ 2 λ
λ
u∈C (v) x∈C ² ∩T (u)
µ ¶|x|−|v|
µ ¶|x|−|v|
X
X
1
η
1
2
≤ ηθ̄ λ
≤
,
1 + η x∈C ² ∩T (v) λ
x∈C ² ∩T (v) λ
2
X
27
if θ̄ 2 λ ≤
1
1+η .
So far we have proved for any v, such that |v| ≤ maxx∈C ² |x|
µ ¶|x|−|v|
X
1
η
η
d T≤C² (v) ≤
≤
1 + η x∈C ² ∩T (v) λ
1+η
which implies D T≤C² (v) ≤ η
so that the linearized recursion (28) always holds. Take ² → 0, λ → br[T (o)]. Define t ² := min{|x|, x ∈ C ² },
it is also easy to see from equation (25) that
µ ¶t ²
µ ¶
X 1 |x|
1
log(1/²)
≤
≤ ² ⇒ t² >
→ ∞.
λ
log λ
x∈C ² λ
Putting things together, under the condition
¶
µ
4δ2
≤ 1 − br[T (o)]θ̄ 2 ,
log 1 +
1 − δ2
we have
lim D Tt (o) = lim D TC² (o) ≤
²→0
t →∞
µ ¶|x|
X
1
η
· lim
= 0.
1 + η ²→0 x∈C ² ∩T (o) λ
Proof of Proposition 1. The proof is a standard exercise following the idea from Proposition 4.2 in [30].
First, let’s recall Bernstein inequality. Consider X ∼ Binom(n, p 0 ), then the following concentration inequality holds
t2
P(X ≥ np 0 + t ) ≤ exp(−
).
2(np 0 + t /3)
p
Hence if we plug in t = 32 log n + 2np 0 log n, we know
st o.
|∂G 1 | ≤ X ≤ np 0 +
q
2
log n + 2np 0 log n ≤ 2np 0 + 2 log n
3
with probability at least 1 − n −1 .
Now, through union bound, we can prove that
¡
¢
P ∀r ≤ R, |∂G r | ≤ (2np 0 + 2 log n)r ≥ 1 −C · (2np 0 + 2 log n)R n −1 ≥ 1 − O(n −3/4 ).
And we know that on the same event,
|∂G r | ≤ n 1/4 , ∀r ≤ R.
It is clear that bad events that G R is not a tree (with cycles) for each layer is bounded above by
p 02 |∂G r | + p 0 |∂G r |2 . Take a further union bound over all layers, we know this probability is bounded
by O(n −1/8 ) provided p 0 = o(n −5/8 ).
Now we need to recursively use the Poisson-Binomial coupling (to achieve Poisson-Multinomial coupling). The following Lemma is taken from [30] (Lemma 4.6).
Lemma 6. If m, n are positive integers then
kBinom(m,
c 2m
m
c
) − Poisson(c)kT V ≤ O( 2 + c| − 1|)
n
n
n
28
Now we condition on all the good events up to layer G r −1 , which happens with probability at least 1−
n
−n −3/4 . We can couple the next layer for nodes in ∂G r . Take a node v ∈ ∂G r as an example. Assume
i
it is of color i , then the number of color j nodes in his children follows Binom(|V>r
|, p i j ). Comparing to
the Poisson version Poisson(n i p i j ), we know with probability at least
−1/8
i
1 − O(n i p i2j + p i j |V>r
− n i |),
i
− n i | ≤ |∂G r |.
one can couple the Poisson and Binomial in the same probability space. Note that |V>r
Repeat this recursively, and use the union bound, we can couple (G R , `G R ) = (TR , `TR ) with probability at
least 1 − O(k maxi (n i )p 02 + kp 0 n 1/4 )n 1/4 log n = 1 − o(1).
Therefore if np 0 = n o(1) and k - log n, we have the bad event (when we cannot couple) happens with
probability going to 0 as n → ∞. And if p 0 = n o(1) , we can allow R to grow to infinity at a slow rate as
log n
R - log[n o(1) +log n] .
Acknowledgements
The authors want to thank Elchanan Mossel for many valuable discussions.
References
[1] E. Abbe, A. S. Bandeira, and G. Hall. Exact recovery in the stochastic block model. arXiv preprint
arXiv:1405.3267, 2014.
[2] E. Abbe and C. Sandon. Community detection in general stochastic block models: fundamental
limits and efficient recovery algorithms. arXiv preprint arXiv:1503.00609, 2015.
[3] E. Abbe and C. Sandon. Detection in the stochastic block model with multiple clusters: proof of
the achievability conjectures, acyclic bp, and the information-computation gap. arXiv preprint
arXiv:1512.09080, 2015.
[4] L. A. Adamic and N. Glance. The political blogosphere and the 2004 us election: divided they blog.
In Proceedings of the 3rd international workshop on Link discovery, pages 36–43. ACM, 2005.
[5] T. T. Cai and X. Li. Robust and computationally feasible community detection in the presence of
arbitrary outlier nodes. The Annals of Statistics, 43(3):1027–1059, 2015.
[6] T. T. Cai, T. Liang, and A. Rakhlin. Inference via message passing on partially labeled stochastic
block models. arXiv preprint arXiv:1603.06923, 2016.
[7] Y. Chen, X. Li, and J. Xu. Convexified modularity maximization for degree-corrected stochastic block
models. arXiv preprint arXiv:1512.08425, 2015.
[8] Y. Chen and J. Xu. Statistical-computational tradeoffs in planted problems and submatrix localization with a growing number of clusters and submatrices. arXiv preprint arXiv:1402.1267, 2014.
[9] A. Coja-Oghlan. Graph partitioning via adaptive spectral techniques. Combinatorics, Probability
and Computing, 19(02):227–284, 2010.
29
[10] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborová. Asymptotic analysis of the stochastic block
model for modular networks and its algorithmic applications. Physical Review E, 84(6):066106, 2011.
[11] Y. Deshpande, E. Abbe, and A. Montanari. Asymptotic mutual information for the two-groups
stochastic block model. arXiv preprint arXiv:1507.08685, 2015.
[12] W. Evans, C. Kenyon, Y. Peres, and L. J. Schulman. Broadcasting on trees and the ising model. Annals
of Applied Probability, pages 410–433, 2000.
[13] D. Gamarnik and M. Sudan. Limits of local algorithms over sparse random graphs. In Proceedings
of the 5th conference on Innovations in theoretical computer science, pages 369–376. ACM, 2014.
[14] C. Gao, Z. Ma, A. Y. Zhang, and H. H. Zhou. Achieving optimal misclassification proportion in
stochastic block model. arXiv preprint arXiv:1505.03772, 2015.
[15] O. Guédon and R. Vershynin. Community detection in sparse networks via grothendieck’s inequality. Probability Theory and Related Fields, 165(3-4):1025–1049, 2016.
[16] B. Hajek, Y. Wu, and J. Xu. Achieving exact cluster recovery threshold via semidefinite programming.
arXiv preprint arXiv:1412.6156, 2014.
[17] B. Hajek, Y. Wu, and J. Xu. Achieving exact cluster recovery threshold via semidefinite programming:
Extensions. arXiv preprint arXiv:1502.07738, 2015.
[18] S. Janson and E. Mossel. Robust reconstruction on trees is determined by the second eigenvalue.
Annals of probability, pages 2630–2649, 2004.
[19] J. Jin. Fast community detection by score. The Annals of Statistics, 43(1):57–89, 2015.
[20] V. Kanade, E. Mossel, and T. Schramm. Global and local information in clustering labeled block
models. arXiv preprint arXiv:1404.6325, 2014.
[21] H. Kesten and B. P. Stigum. Additional limit theorems for indecomposable multidimensional galtonwatson processes. The Annals of Mathematical Statistics, pages 1463–1481, 1966.
[22] H. Kesten and B. P. Stigum. A limit theorem for multidimensional galton-watson processes. The
Annals of Mathematical Statistics, 37(5):1211–1223, 1966.
[23] J. Kleinberg. The small-world phenomenon: An algorithmic perspective. In Proceedings of the
thirty-second annual ACM symposium on Theory of computing, pages 163–170. ACM, 2000.
[24] F. Krzakala, C. Moore, E. Mossel, J. Neeman, A. Sly, L. Zdeborová, and P. Zhang. Spectral redemption in clustering sparse networks. Proceedings of the National Academy of Sciences, 110(52):20935–
20940, 2013.
[25] N. Linial. Locality in distributed graph algorithms. SIAM Journal on Computing, 21(1):193–201,
1992.
[26] R. Lyons and Y. Peres. Probability on trees and networks, 2005.
[27] L. Massoulié. Community detection thresholds and the weak ramanujan property. In Proceedings
of the 46th Annual ACM Symposium on Theory of Computing, pages 694–703. ACM, 2014.
30
[28] A. Montanari. Finding one community in a sparse graph. Journal of Statistical Physics, 161(2):273–
299, 2015.
[29] E. Mossel. Reconstruction on trees: beating the second eigenvalue. Annals of Applied Probability,
pages 285–300, 2001.
[30] E. Mossel, J. Neeman, and A. Sly. Stochastic block models and reconstruction. arXiv preprint
arXiv:1202.1499, 2012.
[31] E. Mossel, J. Neeman, and A. Sly. Belief propagation, robust reconstruction, and optimal recovery
of block models. arXiv preprint arXiv:1309.1380, 2013.
[32] E. Mossel, J. Neeman, and A. Sly. A proof of the block model threshold conjecture. arXiv preprint
arXiv:1311.4115, 2013.
[33] E. Mossel and Y. Peres. Information flow on trees. The Annals of Applied Probability, 13(3):817–844,
2003.
[34] E. Mossel and J. Xu. Local algorithms for block models with side information. In Proceedings of the
2016 ACM Conference on Innovations in Theoretical Computer Science, pages 71–80. ACM, 2016.
[35] H. N. Nguyen and K. Onak. Constant-time approximation algorithms via local improvements. In
Foundations of Computer Science, 2008. FOCS’08. IEEE 49th Annual IEEE Symposium on, pages 327–
336. IEEE, 2008.
[36] M. Parnas and D. Ron. Approximating the minimum vertex cover in sublinear time and a connection
to distributed algorithms. Theoretical Computer Science, 381(1):183–196, 2007.
[37] R. Pemantle and J. E. Steif. Robust phase transitions for heisenberg and other models on general
trees. Annals of Probability, pages 876–912, 1999.
[38] A. Saade, F. Krzakala, M. Lelarge, and L. Zdeborová. Fast randomized semi-supervised clustering.
arXiv preprint arXiv:1605.06422, 2016.
[39] A. B. Tsybakov. Introduction to nonparametric estimation, volume 11. Springer Series in Statistics,
2009.
[40] A. Y. Zhang and H. H. Zhou. Minimax rates of community detection in stochastic block models.
arXiv preprint arXiv:1507.05313, 2015.
31
| 10math.ST
|
Delft University of Technology
Software Engineering Research Group
Technical Report Series
Formalizing typical crosscutting concerns
Marius Marin
Report TUD-SERG-2006-010
SERG
TUD-SERG-2006-010
Published, produced and distributed by:
Software Engineering Research Group
Department of Software Technology
Faculty of Electrical Engineering, Mathematics and Computer Science
Delft University of Technology
Mekelweg 4
2628 CD Delft
The Netherlands
ISSN 1872-5392
Software Engineering Research Group Technical Reports:
http://www.se.ewi.tudelft.nl/techreports/
For more information about the Software Engineering Research Group:
http://www.se.ewi.tudelft.nl/
c copyright 2006, Software Engineering Research Group, Department of Software Technology, Faculty
of Electrical Engineering, Mathematics and Computer Science, Delft University of Technology. All rights
reserved. No part of this series may be reproduced in any form or by any means without prior written
permission of the publisher.
SERG
Marin – Formalizing typical crosscutting concerns
Formalizing typical crosscutting concerns
Marius Marin
Software Evolution Research Lab
Delft University of Technology
The Netherlands
[email protected]
ABSTRACT
We present a consistent system for referring crosscutting
functionality, relating crosscutting concerns to specific implementation idioms, and formalizing their underlying relations through queries. The system is based on generic crosscutting concerns that we organize and describe in a catalog.
We have designed and implemented a tool support for
querying source code for instances of the proposed generic
concerns and organizing them in composite concern models.
The composite concern model adds a new dimension to the
dominant decomposition of the system for describing and
making explicit source code relations specific to crosscutting
concerns implementations.
We use the proposed approach to describe crosscutting
concerns in design patterns and apply the tool to an opensource system (JHotDraw).
1.
INTRODUCTION
Aspect-oriented programming languages provide mechanisms to enable modularization of crosscutting concerns,
such as pointcut and advice or introductions [14, 7, 13].
These mechanisms address general symptoms of crosscuttingness, like scattering (i.e., a concern lacks localization being spread over several modules) and tangling (i.e., a module
implements a core concern, but also crosscutting aspects).
These symptoms and language mechanisms alone, however, do not ensure consistency for defining and describing
crosscutting concerns. To understand concerns specific to
logging mechanisms, design patterns, or implementation of
business rules, one has to rely on unconstrained descriptions
or sample code examples [10, 16, 20, 1]. Moreover, similar
concerns are sometimes described by different names (e.g.,
contract enforcement [1] and policy enforcement [16]).
The absence of a consistent and coherent system for defining and referring crosscutting concerns hinders their comprehension and identification in existing code (a.k.a. aspect
mining). The wide variety of examples describing crosscutting functionality and proposed aspect solutions does not
distinguish common properties to group similar concerns
Technical Report TUD-SERG-2006-010
c copyright 2006, Software Engineering Research Group, Department of
Software Technology, Faculty of Electrical Engineering, Mathematics and
Computer Science, Delft University of Technology. All rights reserved. No
part of this series may be reproduced in any form or by any means without
prior written permission of the publisher.
TUD-SERG-2006-010
and to separate between different groups. Hence, it is difficult to define requirements for aspect mining techniques, to
describe and compare their results [17, 2] without an answer
to the research question:
What are the crosscutting concerns that we try to identify and how to consistently describe and reason about these
concerns?
To address this question we have previously outlined a
classification system for crosscutting concerns based on sorts,
and proposed an informal set of sorts [18, 19]. Crosscutting concern sorts are generic and atomic concerns described
by their specific symptoms (i.e., implementation idiom in a
non-aspect-oriented language), and a (desired) aspect mechanism to modularize a sort’s instances with an aspect-based
solution.
This work provides a formalized description of the crosscutting concern sorts and turns the list of sorts into a catalog. The catalog describes each sort as a relation between
code elements and provides query templates to formalize
these relations. The catalog is built on an in-depth analysis
of crosscutting concerns reported in literature and practical experience with source code exploration and analysis for
identification of crosscutting concerns in software systems of
considerable size like PetStore, JHotDraw and Tomcat [17].
The contributions of the paper lie in several directions:
A catalog of generic crosscutting concerns to ensure consistency and coherence in addressing crosscutting functionality. The paper describes unified criteria for grouping crosscutting concern as relations between code elements and provides detailed examples for each entry
in the catalog.
A query component and templates for each sort to formalize the relation underlined by the concerns of the
sort. The templates support the development of query
libraries for crosscutting concerns. Moreover, the formalization of the queries identifies relations relevant
for describing crosscutting concerns and allows for extensions of existing query languages for capturing such
concerns.
Crosscutting concerns documented using sort queries
for the catalog of design patterns described by Gamma
et al. [8] and for a popular open-source system of relevant size. This documentation is available for further
references as a common benchmark for aspect mining [2]. The paper also discusses in significant extent
a number of relevant concerns in this system and how
sorts help in describing and documenting them.
1
SERG
Marin – Formalizing typical crosscutting concerns
Advanced tool support for querying sort instances in code
under analysis and building concern models to describe
crosscutting relations in software systems. A concern
model allows for grouping sort instances described by
queries into composite concerns. This model adds a
new dimension to the traditional representation based
on a dominant decomposition of an object-oriented
system [25]. This dimension relates relevant program
elements in the context of a concern implementation
and makes this relation explicit.
The organization of relevant relations describing generic
crosscutting concerns and the proposed formalization using
queries is aimed at supporting crosscutting concern comprehension. Nevertheless, the descriptions of concerns using
sorts contribute a first step towards refactoring to aspectoriented programming by documenting typical crosscutting
relations. Aspect-oriented languages could use the contributions of this paper to further support the refactoring of
concern sorts to aspect solutions.
The rest of the paper is organized as follows: The next
section presents related work on describing crosscutting concerns and how the existing approaches compare to the one
proposed in this paper. In Section 3 we introduce the classification of crosscutting concerns on sorts, followed in Section 4 by the query component and the template to describe
the sorts. Then, we present a catalog of crosscutting concerns sorts together with examples of relevant concerns and
template queries for each sort.
In Section 5 we look at how the proposed sorts cover the
crosscutting elements occurring in the implementation of design patterns. Then, in Sections 6-7, we present the tool
support and the results of applying the tool to an opensource system for describing instances of crosscutting concerns and organizing them in a composite model. We conclude with lessons learnt, opportunities for improvements
and directions for future work.
2.
RELATED WORK
2.1 Support for defining and describing crosscutting concerns
The Feature Exploration and Analysis Tool (FEAT) and
the Concern Manipulation Environment (CME) provide tool
support for grouping program elements related by a crosscutting concern implementation and organizing them into
more complex, composite structures.
FEAT organizes program elements that implement a concern in concern graphs [22]. The user can add elements to a
concern graph by investigating the incoming and outgoing
relations to and from an element that is part of the concern implementation. The elements in a concern graph are
classes, methods or fields connected by a call, read, write,
check, create, declare, or superclass relation.
Although the tool allows to add relations to the graph describing a concern, the focus is on the elements participating in the implementation of the concern. The navigation
for understanding a concern and incrementally building its
graph representation is from a root (class) element to other
elements in the relation chain. That is, a concern is described by its elements, and an element is described by its
relations. Unlike FEAT, the sorts-based approach that we
propose uses relations as the main representation of a con-
2
cern and builds concern models based on these relations.
CME allows for more complex queries than FEAT to define relationships between code elements, especially by providing options for restricting the query domain [11]. The
query and its definition can be saved as an element in the
CME concern model to describe a concern intensionally. The
output of the query can also be saved to describe a concern
extensionally.
However, neither CME nor FEAT provide a coherent,
standard way to document crosscutting functionality: no
consistency is imposed for documenting instances of the
same or similar concerns and no distinction is made between different concerns documented by the same or similar
queries. This prevents uniformity in addressing and describing crosscutting functionality and reflects upon the efficiency
of documenting concerns with, for instance, CME queries:
what is the right query for describing a particular crosscutting concern, like, for example, a simple pre-condition
check? Is this query the same as the other queries documenting pre-condition checks in the same system? Would
it be possible (and desirable) to distinguish between a standard query documenting calls for pre-condition checks and a
query to document logging calls? What queries are needed
to consistently document complex relations like a design pattern implementation?
2.2 Refactoring to aspects
The work on refactoring to aspect-oriented programming
shows a significant number of examples of crosscutting concerns.
The present approaches to aspect refactoring can generally be distinguished by their granularity. The group of aspect refactorings proposed by Laddad covers a significant
number of situations where crosscutting functionality could
occur in an application [16, 15]. Some of these refactorings are low level ones, closely associated with symptoms of
a large variety of concern implementations, such as extract
method calls into aspects or extract interface implementation. Yet most of the refactorings are relatively complex
and include design patterns, transactions management, or
business rules. These complex refactorings generally involve
more than one concern to be refactored. Transaction management, for instance, implies to take care of committing or
rolling back a transaction for a given operation, to ensure
that the same connection is used for all the updates, or that
there are no undesired calls on connection objects.
In the literature describing this group of refactorings, there
is no categorization of the various refactorings proposed nor
a specific classification of the concerns involved 1 .
Hannemann et al. [9, 10] propose role-based refactorings
for object-oriented design patterns [8]. The refactoring relies
on a library of abstract descriptions of the patterns and their
role elements, and instructions to refactor to an aspect solution after mapping code elements to the abstract roles. This
approach is a step further towards defining generic, abstract
solutions to typical problems that involve crosscutting functionality, such as applying design patterns. However, design
patterns implementations only define the context into which
the crosscutting occurs, they are not concerns. Moreover,
1
Work in [16, 1] mentions an aspect classification based on
the phase of the software lifecycle at which the aspects occur:
development and production aspects. This classification is
not related to the discussion proposed here.
TUD-SERG-2006-010
SERG
Marin – Formalizing typical crosscutting concerns
the design patterns typically exhibit multiple crosscutting
concerns, some of them sharing common properties, as we
shall see in Section 3.1.
Finer-grained refactorings have been proposed in form of
code transformations catalogs [20] and AspectJ laws [3].
These transformations can occur as steps in the aspect refactoring of a crosscutting concern, but they are oblivious to the
refactored concern. They describe mechanics of migrating
Java specific units to AspectJ ones (e.g., Extract Fragment
into Advice, Move Method/Field from Class to Inter-type).
2.3 Query languages and tools
From the discussed tools and approaches to describing
and organizing crosscutting concerns, only CME provides
its own query language: Panther is a pattern-matching language that allows to search for program units and query
relationships between them [24]. This language, however,
lacks a precisely defined syntax, as well as a complete implementation.
Alike CME, JQuery is a code browser developed as an
Eclipse plug-in. JQuery uses a logic query language (TyRuBa)
similar to Prolog [4]. The TyRuBa predicates supported by
JQuery cover all the relationships defined by FEAT and include many others. For example, JQuery supports a number
of predicates for checking the type of an argument; the recognized types are compilation units, class members, errors,
tasks, warnings, etc. Among the supported binary relations
some are not present neither in FEAT nor in Panther, such
as relations for thrown exceptions.
Despite being more flexible than CME, JQuery provides
a less friendly query syntax. For example, a simple query
for returning the classes implementing a specified interface
can be written in Panther like:
sourceof(relationship
implements(type *, interface FigureChangeListener)); .
The same query in JQuery is more difficult to write and
understand:
interface(?I), name(?I,FigureChangeListener),
implements(?C, ?I); .
Sextant is a tool similar to JQuery, which also enables
to query different kinds of artifacts [5]. The tool stores an
XML representation of a software’s artifacts and uses the
XQuery 2 language to define queries across these artifacts.
However, this is less relevant to our approach, which is aimed
at describing crosscutting relations in source code.
Eclipse IDE provides an advanced programming interface
for querying source code relations and various views for visualization of results. This support is more mature than the
one in CME and provides a significantly better performance
(speed and resources) than JQuery. Moreover, both CME
and JQuery build their own internal representations of the
code to be analyzed, parallel with the one available through
Eclipse. This implies recompilation of the code to be analyzed by the two tools, and additional use of resources.
From the available options, we have chosen to use a pseudo
query language that resembles CME’s Panther language for
formalizing the sort queries. This provides us with a more
elegant syntax to ease understanding. However, we have implemented our tool support as a plug-in that uses Eclipse’s
2
www.w3.org/TR/xquery
TUD-SERG-2006-010
Figure 1: FigureChanged Observer
infrastructure for querying source code and reporting results. This solution provides us with improved performance.
3. CROSSCUTTING CONCERN SORTS
The (dominant) decomposition of a software system provided by object-oriented languages permits to make certain
relations explicit in source code. Such relations include inheritance, call relationst, encapsulation of properties and
containment of actions, etc. Not all relations between program elements, however, can be made explicit by objectoriented languages: the relation between a set of methods
consistently invoking the same action as part of a common
requirement is hindered in the code and possibly made explicit by comments and documentation. Similarly, classes
that implement multiple roles (defined by different interfaces) relate secondary roles to a primary one in a given
context. We refer to such relations “hidden” in the code
as crosscutting concerns and further look at how these can
be categorized for a consistent reference. The aim of this
work is to provide a coherent system for referring crosscutting functionality, to relate crosscutting concerns to specific
implementation idioms and to formalize their underlying relations.
3.1 Motivating example
Design patterns, and most notably the Observer pattern,
are typical examples of problems exhibiting crosscutting functionality [10]. A standard implementation of the Observer
pattern introduces specific relations through the participantroles and the collaborations described by Gamma et al. [8].
The classes that participate in an implementation of the
pattern play either the role of the Observer or that of the
Subject. The two roles define specific members, typically abstracted in an interface definition: The Subject role knows
its observers and defines methods to allow attaching and detaching Observer objects, as well as a notification method
to inform the observers of changes in the Subject’s state.
Actions that change the state of the Subject consistently
invoke the notification method. The Observer defines a specific method to receive notifications from the Subject and
update its own state consistently.
3
SERG
Marin – Formalizing typical crosscutting concerns
class of concerns that share a number of properties:
• an intent (behavioral, design or policy requirements),
• a specific implementation idiom in an (object-oriented)
language and
• a (desired) aspect language mechanism that supports
the modularization of the sort’s concrete instances.
Figure 2: Concern model for FigureChanged Observer instance
Figure 1 shows an instance of the Observer pattern in a
drawing application. The Figure elements play the Subject
role and declare role-specific members like the changed()
method to notify observers, the methods for attaching and
detaching observer objects, and fields in concrete classes to
store the references to the observers. The FigureChangeListener interface defines the Observer role.
The atomic crosscutting relations that occur due to the
implementation of the Observer pattern, like the consistent
notification of observers or the implementation of multiple
roles by the classes participant in the pattern, are not, however, specific to this pattern only. Implementation of multiple roles occurs in other patterns as well, like the Visitor,
or in mechanisms for implementing persistence, like Java’s
Serializable mechanism. Such relations exhibit a common
implementation idiom, namely implementation of multiple
interfaces or direct implementation of members that can be
abstracted in an interface definition.
Similarly, the idiom for the consistent notification consists
of scattered method calls to a common functionality. This
idiom is shared by other known crosscutting concerns, like,
for example, consistent tracing operations.
The common relations and idioms specific to crosscutting
concerns like consistent notification or tracing suggest that
we can define generic crosscutting concerns, which describe
common properties. We call these generic crosscutting concerns, sorts. The two sorts in an implementation of the
Observer pattern are Role superimposition and Consistent
behavior. The Role superimposition, for example, has two
instances in the Observer pattern, specific to the two roles
for the participants in the pattern.
3.2 Starting point
This section summarizes our past contribution to defining
sorts of crosscutting concerns.
A crosscutting concern sort is a generic description of a
4
Examples of aspect mechanisms include pointcut and advice or introduction, as in AspectJ [13, 1].
Concrete implementations of a crosscutting concern sort
in source code represent sort instances.
Together with defining the sorts, we have also provided an
informal list of canonical sorts with a focus on their refactoring to aspect-oriented programming [18].
In the present paper we address the sorts as un-modularized
generic relations between program elements. To capture
and make explicit these relations, we provide a query component for describing sorts.
We revise the previously proposed list of sorts and provide a formalized description of the sorts. Furthermore, we
organize the sorts in a catalog that describes each sort in
detail.
Two of the sorts in the previously proposed list have been
merged as they differed only by the refactoring solutions.
Soares et al. [23] proposed a mechanism for introduction of
throws clauses for addressing the Exception propagation sort
described in Section 4.4. This is very much an alternative
to the AspectJ approach that uses exception softening. The
distinction between the various refactoring approaches is less
important to our current focus on describing concerns by
their defining relations.
3.3 A concern model based on crosscutting
concern sorts
Crosscutting concern sorts are, by definition, atomic elements. The atomicity of a sort provides a consistent granularity level for classifying crosscutting concerns by common
properties, but also gives the level of complexity for the
relations to be expressed by a sort instance. To relate sort
instances in more complex relations, like participation in design patterns implementation, we use a composite concern
model. A concern model organizes concerns in a hierarchical
structure. Relations describing sort instances are always leaf
elements and children of a composite concern model. Each
model can be a child of a super-model. A concern model for
a complex relation R can be formalized as:
ConcernM odel(R) =
n
[
i=1
SIi ∪
m
[
SRj
j=1
This model describes a relation R as a composition of sort
instances SIi and composite sub-relations SRj . Figure 2
shows the concern model for the Observer implementation
previously discussed. The composite FigureChanged model
groups instances of sorts like Role superimposition and Consistent behavior that participate in the implementation of
the Observer pattern for figure changes. A sort instance is
described by a a given name and an associated query together with the results of this query. A concern model is
also described by a given name.
The FigureChanged relation is part of parent, customdefined relations, like the one grouping all the instances of
TUD-SERG-2006-010
SERG
Marin – Formalizing typical crosscutting concerns
or Member and a Type, shows the type of a field member
or parameter, or the return type of a method. Note that
declares relations can also be specified by using wildcards
and qualified names, e.g., p.C.* for all members declared by
class C of package p.
In addition to these predicates, a query definition also allows for the transitive closure operator (+) and two “wildcards”:
• * matches names of identifiers (element names or return types) or modifiers (visibility);
• .. matches any parameters or arguments in a method
or constructor.
Figure 3: Relationships relevant to sort queries
the Observer pattern in the JHotDraw project. In this case,
the project corresponds to the top-level concern model.
The next section introduces the query component for describing relations specific to crosscutting concern sorts. It
is up to the software engineer to organize these instances in
more complex concern models that reflect her or his design
decisions, and to provide appropriate identifier names.
4.
DESCRIBING CROSSCUTTING CONCERN
SORTS
4.1 The query model
The proposed query model is aimed at providing a standard, formalized description of the underlying relations of
the crosscutting concern sorts. The model consists of a
generic query definition to describe the query model, and
a set of query templates (sort queries) to describe the relations specific to each of the sorts.
A sort query is a binary relation between two sets of elements, the source and the target context. A context is a
(restricted) set of program elements, which can also be expressed as the result of a query. The two contexts in the
definition of a sort query represent the end-points of the
sort’s relation.
The generic query describing the model can be expressed
as:
relation_id (<source-ctx>, <target-ctx>);
The end-points contexts can be selected by the sourceof
and targetof operators, respectively. Elements in the set of
results of a query sort are (e1,e2) tuples, where:
(e1 in source-ctx) && (e2 in target-ctx) &&
relation_id(e1,e2)
The two elements, e1 and e2, are program elements, such
as a class or a method. The relation between them, relation id, is a (binary) relation between program elements,
like call or inheritance relations, which can be extracted by
a static source code analysis.
The relevant relations for sort queries are shown in Figure 3. The same relations describe the list of predicates used
by the sort queries. The type relation, between a Parameter
TUD-SERG-2006-010
Variables are allowed and used to save (partial) results of
a query, such as the set of elements that gives the end-point
of a query. The variable can further serve as an end-point
for another query. The notation used to declare a variable
is <variable name>.
The intersection of variables and end-point sets is shown
through the && operator.
4.2 Template for the sort catalog entries
We describe the crosscutting concern sorts and their query
component by an extended template that comprises the following elements:
Intent of the sort, to give a generic description of the sort’s
concerns;
Relation underlined by the sort, to describe the type of
relation between the elements implementing instances
of the sort;
Idiom specific to the sort implementation in an objectoriented (particularly Java) language;
Query associated with the sort to make explicit the relation
between the elements implementing a sort’s instance
and to formalize the sort’s definition;
Example to show a concrete instance of the sort;
Other instances to provide additional examples of (possible) occurrences of the sort.
4.3 Examples of sort instances
For exemplification of sorts we refer to a number of concerns identified in well known applications or libraries. Where
examples of sorts instances are available, we refer to the
JHotDraw application, which we have used as a case-study
for documenting crosscutting concern instances using sort
queries.
JHotDraw is a drawing editor for bi-dimensional graphics,
developed as an open source project 3 with involvement from
the authors of the “Design Pattern - Elements of Reusable
Object-Oriented Software” book [8]. One of the application’s goal is to provide a show case for good use of design
patterns in the development of graphical applications.
4.4 The catalog
This section presents a catalog of crosscutting concern
sorts described by the template introduced in Section 4.2.
4.4.1 Method Consistent Behavior
3
jhotdraw.org
5
SERG
Marin – Formalizing typical crosscutting concerns
Intent
Execute consistently a specific action for a number of method
elements as part of their required functionality.
Relation
The method elements share a (secondary) concern, which
they implement through the consistent invocation of the action executing the desired functionality. The common action
invocation shows the relation between the callers as participants in the implementation of a crosscutting concern.
Idiom
Scattered method calls to the method implementing the
common action to be executed consistently.
Query
The sort query takes two arguments: a seed-element to define the source context, and the method implementing the
action to be executed consistently by the elements in the
source context. This method is the unique element of the
target context.
<context> = pckg.I+ || (project Proj) ||
type (pckg.Cls) || packge pckg;
We first define the (source) context based on the seedelement passed to the query, and save the elements of the
context into a variable. The context definition shows several possibilites to use structural relations for describing the
elements in the context; for example, the context could consist of the elements in the hierarchy of the type I of package
pckg (pckg.I+), or the elements of the project Proj, or members of the type Cls of package pckg, or elements declared
in package pckg, etc. The context seed is specific to each
of these cases, and that is, the type I, the project Proj, the
type Cls, and the package pckg, respectively.
The formalization of the context is similar to defining
pointcuts in languages like AspectJ [13]. However, the context consists of program elements rather than execution
points 4 .
<selcallers> = <context> && sourceof(
invokes(method *, * p.C.m(..)));
CB(contextElem, m) = invokes(<selcallers>, * p.C.m(..));
Next, we save in the <setcallers>variable all the method
members in the source context that invoke elements in the
target context (in this case, the method m).
The query returns the invocation relationships between
the elements in the source context and the method-action
in the target context.
Example
The notification mechanism in the previous example of the
Observer pattern instance is an example of consistent behavior: the actions that change the state of the Subject object
have to consistently call the notification method for allowing
the observers to update their state accordingly.
A different example of this sort’s instance is common to
transaction management, a complex concern aimed at ensuring data integrity; this implies that an operation is committed only if it is fully completed and roll-backed otherwise. In a banking transfer operation, for example, both the
debit and credit operation have to be successful at the same
4
The context definition shown for this query is common to
several sorts to be discussed next. We will refer to the <context>variable, without redefining it.
6
time for keeping a consistent state of the data. Java provides various alternatives to transaction management, like
JDBC transactions and Java Transaction API (JTA) 5 . A
JTA transaction, for example, implies that methods implementing the transaction logic consistently invoke dedicated methods of the javax.transaction.UserTransaction interface: the begin method at the beginning and the commit
(or rollback) at the end to demarcate a JTA transaction.
These invocations represent instances of the consistent behavior sort covering methods whose operations have to be
under transaction management.
Other instances
Logging of exception throwing events in a system; Wrapping
of service level exceptions of business services into application level exceptions[17]; Credentials checking as part of the
authorization mechanism[16]; etc.
4.4.2 Contract Enforcement
Intent
Comply with design by contract rules, such as executing
consistent condition checks to ensure that the assumptions
about a method are confirmed.
Relation
A set of elements are related by common condition checks.
Idiom
Similar to Consistent behavior, instances of this sort are implemented as scattered calls to methods checking conditions
as part of contract enforcement.
Query
The sort is similar to the Consistent behavior sort with the
difference that the action (condition check) is not part of the
caller-elements’ functionality; that is, a method can fulfill
all its requirements in the absence of the calls checking the
conditions. Hence, the two sorts differ only in intent. The
query documents this intent by its own description while the
same query template as for Consistent behavior applies to
this sort too.
Example
The use of assertions and assertions-like calls is a typical example of instances of this sort. An example of Contract enforcement instance is available through the JHotDraw drawing editor application. The application defines a hierarchy
for Command operations to be executed as response to user
actions, like menu items selection. The Command elements
are participants in the implementation of the Command pattern, and define a no-argument execute method to carry
out the specific action. Commands can be executed only if
a valid view is present in the drawing editor. The Command
elements check this condition through a consistent (check)
call before proceeding with their execution.
Other instances
Pre- and post-conditions complying with design by contract
rules.
5
http://java.sun.com/j2ee/1.4/docs/tutorial/doc/
TUD-SERG-2006-010
SERG
4.4.3 Entangled roles (Interfacing layer)
Intent
Extend an element (method or type) with a secondary role
or responsibility, which is entangled with its primary concern.
Relation
The crosscuttingness of the concerns of this sort resides in
the tight coupling between elements implementing different
roles, like, for instance, relations between graphical user interface (GUI) elements and their associated model or actioncontroller elements. The GUI element knows its associated
element and sends messages to it in order to be able to provide a response to an action or report about its state. The
interfacing task of the GUI element results in its methods
being crosscut by the logic of the model element. The methods’ implementations address a different concern than the
interface declaring them. The graphical elements store references to their model and mirror its state.
The relation specific to this sort is established between
elements in the context of a (interfacing) class that invoke
actions in the interfaced type through an object reference.
This reference is typically stored in a field of the interfacing
class.
Idiom
Redirection of calls to a specific reference.
Query
//the type of the interfaced object
<interfacedType> = sourceof (declares(*, field p.C.field));
ER(C,field)=references(class p.C, <interfacedType>));
Example
A common design practice for graphical user interface (GUI)
elements, like Java Swing components, is to use a ModelView-Controller(MVC) design where the View and the Controller are implemented by the same class. The class for the
graphical component keeps a reference to the model object,
which determines the component’s state. Hence, the graphical component is a wrapper around the model’s state that
“mirrors” the model’s state and delegates actions to this
model.
Swing components like buttons and menu items implement the described behavior (e.g., AbstractButton). The
methods for setting or reporting about their state, such as
selected armed, or enabled, redirect their calls to the model
whose state they mirror.
The controller functionality in the same components delegates calls to its associated command-action. The action
can also play a model role: in the JHotDraw drawing editor,
a menu item is enabled or disabled if the Command object
associated with the item is or not executable.
Other instances
Swing GUI elements (like classes extending AbstractButton);
Invoker-Command relation in the Command pattern.
Marin – Formalizing typical crosscutting concerns
Intent
Define an interfacing layer to an object (add functionality
or change the context) and forward the calls to dedicated
methods of the object.
Relation
The redirection layer acts as a front-end interface that assumes calls and redirects them to dedicated methods of a
specific reference, (with or without) executing additional
functionality. The consistent (yet, method specific) redirection logic crosscuts this layer’s methods.
Unlike the previous sort, instances of this sort do not
query a model for its state, but typically add functionality dynamically to the calls. This sort is a specialization of
the Interfacing layer sort.
The relation specific to this sort is between the redirecting layer and the target object, and resides in consistent
redirection of calls between pair methods.
Idiom
Identical logic in a number of methods that consistently redirect their calls to dedicated methods of a specific type.
Query
//C-component; D-decorator
<compType> = sourceof(declares(*, field p.D.field));
//all methods of C
<compMehtods> = targetof(declares(<compType>, method *));
//all decorator (D) methods
<decorMehtods> = targetof(declares(type D, method *));
RL(D,field) = (invokes(<decorMehods>,<compMethods>)) &&
(compares(<decorMehods>,<compMethods>));
The relation is fairly complex and connects the context
of the redirection layer and that of the type of the targetobject.
The query takes as arguments the type element of the
redirecting layer, and the field in this type that stores the
reference to the object to redirect to. The two arguments
define the source context, as the redirecting class, and the
target context, as the type of the reference receiving the
redirected calls.
The query returns invocation relations between methods
with the same name in the two contexts. Most relevant for
this relation are the methods in the redirecting layer that
consistently forward their calls.
Example
Implementations of the Decorator pattern are typical examples of instances of this sort. Decorator instances in JHotDraw allow to attach elements like borders to Figure objects.
The decorators for figures extend DecoratorFigure that defines the set of methods to consistently forward their calls
to the stored reference of the decorated object. Subclasses
of DecoratorFigure can override its methods to dynamically
extend its functionality.
Other instances
Implementations of the Decorator, Adapter, and Facade patterns[17].
4.4.5 Add variability
Intent
4.4.4 Redirection layer
TUD-SERG-2006-010
Use method-objects to pass a method as an argument.
7
SERG
Marin – Formalizing typical crosscutting concerns
Relation
Instances of the sort implement a consistent mechanism of
building and passing method-objects as method arguments.
Method-objects are (typically) objects of a type declaring
one method. The methods expecting arguments of this type
only need and invoke the specific method for the passed
object.
This mechanism is a substitute for passing pointers to
methods as arguments. Languages like Java use this mechanism, which is also referred as closures or functors or function objects, to achieve a behavior similar to the use of callback funtions.
The sort and its query describe a contract between clientcallers and server-callees that make use of method objects
as a substitute to passing references to methods.
Idiom
A number of actions in a system require arguments of a
method-object type for invoking the type’s unique method.
The clients conforming to the expectations consistently build
and pass wrapper objects for the expected method.
Query
//methods expecting args of MethodObject Type
<target> = sourceof(params(method *, type MethObjT));
<mCreateMethObj> = sourceof(invokes(method *, MethObjT.new(..)));
<source> = <mCreateMethObj> &&
sourceof(invokes(method *, <target>));
//method-variables of MethodObject Type
<varsInMethods> = targetof(declares(<source>, var MethObjT *));
//methods being called with the previous variables as args
<target> = sourceof(args(method *, <varsInMethods>));
AV(MethObjT) = invokes(<source>,<target>);
The query documents the set of methods that create methodobjects of a specified type, and pass them as arguments to
methods declaring parameters of the same type. The target end-point of the relation consists of methods declaring
parameters of a method-object type. The source context
is given by the intersection between (1)the set of methods
creating objects of a method-object type and (2) the set of
callers of the methods in the target context.
Example
Simple commands in JHoDraw, which do not save state,
implement a number of operations like opening, printing
and closing a drawing (application). The implementation
follows the Command pattern and consists of associating
menu items with instances of anonymous command classes
that declare and implement an execute method to carry
out the operation. These instances are method-objects and
their execute method simply invokes a dedicated method
that carries out the open (or close or etc.) operation. The
Command object is hence used as a wrapper for the execute
method to be invoked in response to menu item selection
actions.
Other example instances are related to practices specific
to functional programming [12], like the use of closures (functors) that can be implemented in Java using method objects
like, for instance, Runnable objects. These objects only implement a run method and are used by common Java mechanisms, like thread safety. Consider, for example, Component elements (like Swing objects) that need to execute in
a specific thread, i.e. the event dispatching thread, to avoid
deadlocks during painting the graphical components. Two
8
Java dedicated methods, invokeLater and invokeAndWait,
ensure that these components execute in the special thread.
The two methods expect an argument of type Runnable
whose run method contains the code accessing functionality
of the graphical (Swing) component to be executed.
Other instances
Similar behavior as for the Swing components is present in
the implementation and use of graphical elements of the
Eclipse Standard Widget Toolkit(SWT). The IRunnableWithProgress interface, for instance, has to be implemented
by classes with long-running operations for displaying a progress
bar. The interface defines a run method, and instances of
the implementing classes are passed as arguments to a dedicated method of a specialized class for running long-running
operations.
SWT also defines an user-interface thread from which
the SWT API methods to be called. Runnable objects are
passed to this thread for (a)synchronous execution.
The Visitor pattern declares a specific Visitable role with
an accept method to be invoked by visitor methods receiving arguments of type Visitable. Implementations of the
pattern can rely on a (Visitable) method object to pass the
accept method to interested clients.
Laddad proposes the worker object pattern and an implementation in AspectJ of the pattern [16]. The worker object
is a method-object: an instance of a class implementing an
worker method. Examples include Asynchronous method
execution, Authorization using Java Authentication and Authorization Service (JAAS) API, etc.
4.4.6 Expose context
Intent
Expose the caller’s context to a callee by passing information
to each method in the call stack of that callee.
Relation
Instances of this sort are implemented by methods that are
part of a call chain where an additional parameter is declared
and used to pass a context along the chain. The declaration
of additional parameters to pass context is crosscutting.
The elements related by this sort are the caller-method
receiving an argument with a specific name (and type) and
the callee-methods to which the caller passes the argument.
To get a full description of the context passing, a transitive
closure operator has to be applied to the relation associated
with this sort.
Idiom
Methods participating in a call chain declare additional parameters to pass a specific context required for fulfilling their
(secondary) requirements.
Query
<callees> = targetof(invokes(method p.C.caller(..), method *));
<selCallees> = <callees> &&
sourceof(args(method *, name argName))) &&
sourceof(args(method *, type ArgType)));
EC(caller, argName) = invokes(method p.C.caller(..),<selCallees>);
The query returns the relations between the input caller
method and the callees that are passed an argument with
the name specified in the query.
TUD-SERG-2006-010
SERG
Marin – Formalizing typical crosscutting concerns
Example
Intent
To monitor progress evolution for long-running operations in
Eclipse applications, one can use a IProgressMonitor object.
The long-running operations are passed a reference to the
monitor class through an additional parameter. The operation invokes methods of this reference to indicate progress,
like the worked(int) method to indicate that a given number of work units of the executing task have been completed.
Long-running sub-operations receive the same reference to
the monitor and use it to report progress.
Make the relationship between classes explicit (through nested
classes) to superimpose a role (to a hierarchy).
Other instances
Transaction management, Authorization, the Wormhole pattern [16];
4.4.7 Role superimposition
Intent
Implement a specific secondary role or responsibility.
Relation
A number of elements share a common role, other than their
defining hierarchy. The sort is specific to classes that participate in multiple collaborations and hence implement multiple roles [21]. The concrete instances occur as multiple
interfaces (or methods that can be abstracted into an interface) implementations. The crosscutting element resides
in the tangling of multiple roles in the class implementing
them.
Relation
A number of elements share a common role by enclosing
support classes of a specific type.
Complex roles can be implemented through nested, support classes. The nesting mechanism enforces and makes
explicit the relationship between the role of the enclosing
class and the one implemented by the support class.
Instances of this sort occur as an alternative to multiple
roles implementations: two hierarchies can interact by having common classes (classes that implement elements from
both hierarchies) or by having elements from one hierarchy
as support classes for the elements in the second hierarchy.
Idiom
Hierarchies interaction through containment of nested classes.
Query
<implementors> = sourceof(
implements(class p.EnclosingC+.*, type Role));
SC(EnclosingC, Role) = implements(<implementors>, type Role);
The query associated with this sort reports about nested
classes within a given context that implement a specific role.
A common context is a class hierarchy to which the role is
superimposed through the nested classes.
Idiom
Example
Implementation (and definition) of members separable in
distinct interface definitions; common instances occur as
classes that implement multiple interfaces.
In the JHotDraw application there are two hierarchies that
interact through support classes: the members of one hierarchy enclose members of the second, supporting hierarchy.
The main hierarchy, Command, defines command elements
for executing various application-specific activities like, copy
and paste, or operations for setting the attributes of a figure. The second hierarchy, Undoable, defines operations for
undo-ing and redo-ing the results of executing a command.
Typically, each Command class encloses its associated Undo
class.
Query
//role = class or interface
<implementors> = sourceof(implements (*, type Role));
<selectedImpls> = <context> && <implementors>
RSI(Role,contextElem) = implements(<selectedImpls>, type Role);
Role superimposition documents elements that implement
or extend the interface or class defining a specific role.
Example
Examples of role superimposition are available through implementations of design patterns defining specific roles, like
the Observer pattern previously discussed, or the Visitor
pattern.
Implementing persistence is also possible through role superimposition: the Figure elements in a drawing application,
like JHotDraw, implement a Storable interface that defines
the methods for a (figure) object to write and read itself to
and from a file. Each figure implements these methods in a
specific way to provide persistence and recovery of drawings
over work session.
Other instances
Implementations of design patterns that define specific roles;
Implementations of multiple interfaces with dedicated, specific roles, e.g., Serializable, Cloneable, etc.
4.4.8 Support classes for role superimposition
TUD-SERG-2006-010
Other instances
Specialized iterators for various Collection types.
4.4.9 Policy enforcement
Intent
Impose a (restriction) policy between groups of elements in
the system.
Relation
The sort describes a restrictive relation that limits certain
type of interactions between a source and a target set of
elements.
Concerns of this sort implement relationship policies between sets of elements that can be described in a source and
target context, respectively. These policies cannot be enforced by language mechanisms, like, for instance, visibility.
They are crosscutting because they have to be consistently
documented and followed.
9
SERG
Marin – Formalizing typical crosscutting concerns
Idiom
Example
Requirements specifications available through documentation, source code comments, etc.
The file reading (writing) operations in JHotDraw implied
by the drawing persistence concern, like read(StorableInput),
throw exception of the java.io.IOException type if the reading or parsing operation fails. The callers of these methods
re-throw the same exception to their callers, up to the driver
application that catches the exception and prints an error
message.
Query
PE(srcContextElem, targetContextElem) =
references(<src_context>, <tgt_context>);
The query documents (reference) relations between elements in a source and target context, respectively. If the enforced policy prohibits relations between the two contexts,
the query will check the policy and report those elements
that break it. This control query can be used as a complement to testing components.
Example
Sun provides a list of restrictions imposed to enterprise beans6 ;
one of these restrictions states that: “Enterprise beans should
not use the java.awt package to create a user interface ... and
stop the Java virtual machine and ...”.
Other instances
Typical examples of instances of this sort include interaction restrictions for (formalized) sets of program elements,
such as packages, and classes implementing specific functionality. Eichberg et al. also discuss several examples of
enforced policies [6]. Some of their examples, however, could
be addressed by conditions checks, such as methods returning non-null values.
4.4.10 Exception propagation
Intent
Propagate an exception for which no appropriate response
is available.
Relation
Concerns of this sort have specific a consistent propagation
(re-throwing) mechanism of the checked exceptions thrown
by methods that do not have appropriate answers to these
exceptions.
Similar to context passing, the described relation is part
of a call chain. The caller-elements implement the consistent (enforced) logic of re-throwing exceptions if not able to
handle them. The transitive closure operator applied to this
relation provides a full description of the elements in the call
chain that re-throw exceptions.
Idiom
Declare throws clauses in the definition of a method for passing the responsibility of catching the exception to the callers.
Query
<callers> = sourceof(invokes(*, p.C.m(..)));
<throw> = sourceof(throws(method *, type p.ExceptionType));
<source> = <throw> && <callers> && <context>;
EP(m,ExceptionType,contextElem)=throws(<source>, p.ExceptionType);
The query reports relations between methods in a given
context and the thrown exception of the specified type.
6
java.sun.com/blueprints/qanda/ejb_tier/
restrictions.html
10
Other instances
Checked exceptions in Java
4.4.11 Design enforcement
Intent
Enforce design, such as classes in an hierarchy must declare
no-arguments constructors.
Relation
A number of elements share a design logic (e.g., no-args
constructors).
The sort describes concerns ensuring design rules compliance for program elements in a defined (formalized) context.
The implementation of these concerns lacks language support, and relies on documentation, specifications, comments,
etc.
Idiom
Requirements specifications available through documentation, source code comments, etc.
Query
//method
DE(contextElem,m)=declares(<context>, member * m(..));
The design enforcement documents a relation between a
set of elements that (have to) declare specific members with
various signatures.
If member m is a constructor, the new keyword will be
used instead of the method name. The member can be a
field as well.
Example
The documentation for the Storable interface in JHotDraw
to be implemented by elements that can be stored and restored to and from a file states that “objects that implement
the interface ... have to provide a default constructor with no
arguments”. Similar to JavaBeans design, the no-argument
constructor will be invoked to create objects after being read
from the file.
Other instances
Serialization rules in Java require that (special) serializable
classes implement special methods (writeObject(..) and
readObject(..)) with specific signatures. These methods
are not declared by the Serializable interfaces but the design
requirements are provided through the documentation of the
interface.
JavaBeans design (providing classes with no-argument constructors) is similar to the discussed persistence example.
4.4.12 Dynamic behavior enforcement
TUD-SERG-2006-010
SERG
Intent
Enforce rules for object use, like before-use initialization and
post-use clean-up.
Relation
Shared logic of manipulating objects of a certain type.
Instances of the sort implement rules for object use, like
the use of lifecycle methods.
Idiom
Requirements specifications available through documentation, source code comments, etc.
Query
<context> = targetof(contains(type C, method *));
DBE(C,field)=set(<context>, C.field) || get(<context>, C.field);
We assume a (crosscutting) solution aimed at checking
that certain methods of an object are executed in an expected order. The solution consists of declaring a field which
to be set to a different (integer) value by each executing
method of the field’s object. The value of the field cannot
be lower with more than one unit than the value to which
a method is going to set the field. That is, each method
checks the value of the counter field before executing in order to ensure proper order of the execution.
Example
(De-)Activation of tools in the JHotDraw graphical editor
requires that the deactivate method is always used to cleanup resources when a tool has finished its execution and the
user switches to a different tool. This method should be the
last called for a given tool.
Other instances
Lifecycle concern [17];
5.
COVERAGE
This section investigates how the sorts are able to cover
the crosscutting concerns present in complex structures, commonly referred in literature for their crosscutting properties.
To this end we consider the list of design patterns for which
Hannemann and Kiczales reported improved modularity due
to proposed aspect-oriented (AspectJ [1]) solutions [10].
Table 1 shows the list of these design patterns. Each
pattern corresponds to a composite concern model whose
crosscutting elements are described by sort instances relations. The documented crosscutting elements in the design
patterns have been identified by examining the description
of the patterns and the sample code in Gamma et al. [8], as
well as the aforementioned AspectJ solutions 7
5.1 Design patterns
Implementations of the Adapter pattern could use either
multiple roles or object composition to adapt a class to an
interface expected by clients. In the first case, the Adaptee
role is super-imposed to the class implementing the Adapter
7
The solutions are available as a set of simple examples
of design patterns implementations in Java and AspectJ
at www.cs.ubc.ca/~jan/AODPs/. The crosscutting concerns
discussed in this paper are, however, not limited to these
examples only.
TUD-SERG-2006-010
Marin – Formalizing typical crosscutting concerns
functionality. The Adapter class implements both a Target
interface and (extends) the Adaptee, which is an instance of
the Role superimposition sort.
The solution relying on object composition would typically use delegation from the Adapter to a stored reference
of the Adaptee object. This is an instance of the Redirection
layer sort.
The State pattern comprises a number of crosscutting elements: The Context role is super-imposed and has specific
members for maintaining a reference to the object defining
the current state; second, the notification of changes of the
current state to be stored in the Context object is an instance of the Consistent behavior sort. The third element is
an instance of the redirection layer: the Context object forwards the received calls to the methods of the object storing
the current state.
The crosscuttigness occurring in the implementation of
the Decorator pattern can be described by the Redirection
layer sort. The methods in the decorator class consistently
redirect their calls to dedicated methods in the decorated
class, through the stored reference to the decorated object.
Decorators are typical examples of instances of the Redirection layer sort.
The crosscutting element of the Proxy pattern implementation resides in the consistent forwarding of the calls to
the reference of the real subject class, stored by the Proxy
object.
Another crrosscutting concern occurs in protection proxies as an instance of Consistent behavior; this consists of
a method call that checks the access permissions before executing the forwarding operation. Some implementations
also consistently check if the proxy’s subject has been initialized.
This check is part of the method for accessing
the reference to the subject. The method is invoked by the
actions in the proxy that forward their calls.
The Visitor and Composite patterns are often used in
combination [8, 10]. Both patterns define roles that in various implementations are super-imposed like, for instance,
the Visitable role. The roles we chose to document by sort
instances, Composite and Visitable, are those defining rolespecific methods.
Certain implementations could make use of method objects for passing a reference of the Visitable object and its
accept method to the visitors’ methods. Such an implementation could be documented by an instance of the Add
varaibility sort. We show in Appendix A an AspectJ solution for the Visitor pattern, different from the one of Hannemann and Kiczales [10], that addresses the Add variability
concern.
The Command pattern presents several participants, like
the Command ojects, the Invokers that require the execution of the command, and the Receivers to carry out the
requests. These participants play specific roles in the implementation of the pattern; however, not all of them declare
specific members and end up as super-imposed.
The Invoker participant keeps a reference to an associated
command to which it forwards the requests. Invokers like
buttons and menu items in graphical user interfaces are often interfacing the command object by mirroring its state
through their own display (e.g., enabled buttons correspond
to commands that can be executed with the active configuration, etc.).
In some implementations, the Invoker role is super-imposed
11
Marin – Formalizing typical crosscutting concerns
SERG
Design pattern Composition of sort instances
Adapter
Adapter = RSI(Adaptee, contextElem) + RL(Adapter, adapteeReference);
State
State = RSI(Context, contextElem1) +
CB(contextElemStateChanger, Context.changeState(State)) +
RL(Context, stateReference));
Decorator
Decorator = RL(Decorator, componentReference);
Proxy
Proxy = RL(Proxy, fieldRefRealSubject);
Protection proxies:
document the consistent behavior of checking the credentials:
CB(contextElem, checkAccessPermission());
Visitor
Visitor = RSI(VisitableElement, contextElem);
Specific implementations:
Visitor = AV(VisitableElement);
Command
Command = RSI(Receiver, contextElem1) +
ER(Invoker, commandReference) +
RSI(Invoker, contextElem2) +
CB(invokersContext, Command.execute()));
Certain implementations using Command for method objects:
AV(Command);
Composite
Composite = RSI(Composite, contextElem);
RSI(Leaf, contextElem2)) - not crosscutting
Iterator
Iterator = RSI(Aggregate, contextElem);
Flyweight
Flyweight = RSI(Flyweight, contextElem1) +
CB(contextElem2, FlyweightFactory.getFlyweight));
Memento
Memento = RSI(Originator, contextElem1) +
CB(careTakerContextElem1, Originator.createMemento));
Strategy
Strategy = RSI(Context, contextElem);
sometimes, we could also have:
RSI(Strategy, contextElem1);
Mediator
Mediator = RSI(Colleague, contextElem) +
CB(contextElem, notifyMediator));
Chain of Respon- ChainOfResponsibility = RSI(Handler, contextElem1) +
sibility
RL(Handler, successorReference));
Prototype
Prototype = RSI(Prototype, contextElem);
In some languages, like C++, copy constructors are required:
DE(contextCloneableObjs, CloneableType.new(const CloneableType&));
A similar instance can be used for requiring implementation
of the Object.clone method in Java
Singleton
Singleton = RSI(Singleton, contextElem1) +
DE(contextElemSingleton, private Singleton.new(..)) +
CB(contextElem2, Singleton.instance());
Observer
Observer = RSI(Observer, contextElem1) + RSI(Subject, contextElem2) +
CB(notify, contextElem3)+
CB(attachObserver, contextElem1)+ CB(dettachObserver, contextElem1);
Table 1: Design patterns as compoostion of sort-instances.
through interfaces that declare methods dedicated to handling events that trigger execution of commands. The implementations of these methods in the command invokers
consistently call the operation in the Command class. The
behavior indicates an instance of the Consistent behavior
sort.
Particular implemenations of the Command pattern can
serve as method-objects. In this case, the class defines just
an execute method and does not save state. The Command
object is passed around for clients to access the functionality of its unique method. The use as method object is an
instance of the Add variability sort.
12
The only crosscutting element occurring in the implementation of the Iterator pattern, is the super-imposed Aggregate role. The role defines the CreateIterator() method
to create an iterator object for traversing the elements of
the aggregate (structure).
The concerns documented for the Flyweight pattern comprise a Role superimposition instance for the Flyweight role,
and a Consistent behavior for obtaining references to a (new)
flyweight object. This behavior consists of calling the accessor method in the factory class for the flyweight instances,
instead of attempting to build new flyweight objects. A similar behavior is discussed below for the Singleton pattern.
TUD-SERG-2006-010
SERG
The refactoring of Memento pattern to AspectJ uses the
introduction mechanism for superimposing the Originator
role. In addition to this, we document a Consistent behavior instance, namely acquiring a memento object before
performing the operation that changes the state.
The Strategy pattern defines two roles, the Strategy and
the (Strategy)Context. Most commonly, the Context is a
super-imposed role, maintaining a reference to the Strategy
object (and defining methods to access the reference) and,
possibly, delegating requests from its clients to the Strategy.
The Mediator pattern implies a super-imposed role (Colleague) to store and access the reference to the Mediator
class. Moreover, each change in the coleague class results
in a consistent notification of the mediator for coordinating
the other colleague-classes. In some implementations, the
Mediator role could also be super-imposed.
The participants in the responsibility chain are superimposed the Handler role, which defines the method for handling specific requests, and the reference to the next Handler
in the chain. The handler-methods check the request and
consistenly pass it to the following handler in the chain. Because the one-to-one association between the passing and
receiving methods in the chain, the call passing behavior is
documented by a Redirection layer instance.
One crosscutting element in a typical implementation of
the Prototype pattern is due to the super-imposition of the
Prototype role that declares the clone method to allow objects to generate copies of themselves. In some languages,
like C++, the Prototype must declare a copy constructor for
cloning. (A copy constructor receives as unique parameter
a constant reference to an object of the class’ type.) This
would be an instance of the design enforcement sort.
In Java, the cloning is realized through the clone method
in the Object class, which is extended by all Java classes.
The class implementing an overriding method of clone has
to implement the Cloneable interface to indicate to the clone
method that it is legal to make copies of the fields of the
Cloneable class. This is a design enforcement sort instance.
The Singleton pattern has specific the access method to
the unique instance of the singleton class, which has to be
used by clients instead of calling a constructor. In some
implementations, the method is declared by an interface that
defines the super-imposed role of the Singleton.
Singletons have special designs, most notably they have to
declare the constructor as private for not allowing constructor calls from outside the class. This Design enforcement
instance cannot be specified other than by comments.
Similar to the Flyweight pattern, we use a Consistent behavior query to document the references to the method returning the unique instance. This behavior shows the rule
for accessing the singleton’s functionality.
The Observer pattern is documented as a composition of
Consistent behavior and Role superimposition intances, as
discussed in the previous sections. In addition to the consistent behavior of notifying changes in the Subject’s state, we
also document the mechanisms for registration and deregistration of observers.
Marin – Formalizing typical crosscutting concerns
composite concern models 8 . The concern model supported
by SoQueT is similar to that in CME, with the differences
that complex concerns are expressed as compositions of sort
instances [11].
SoQueT provides three main user-interface components:
The interface to define a query for a specific sort based on a
list of pre-defined templates is shown at the bottom of Figure 4. The template guides the user in querying for elements
that pertain to a concrete sort-instance.
The results of the query are displayed in the Search Sorts
view, also shown in Figure 4. The view provides a number of options for navigating and investigating the results,
like display and organization layouts, sorting and filtering
options, source code inspection, etc.
The Concern model view allows to organize sort instances
in composite concerns described by their user-defined names.
The concern model is a connected graph, defining a view
over the system that is complementary to the Eclipse’s standard package explorer. The graph is displayed as a tree hierarchy, with sort instances as leaf elements. A sort instance
element can be expanded to display the results of its associated query. The node tree representing a sort instance is
labeled with a user defined name and the description of the
associated query. Note that queries can be associated only
to sort instances and not to a composite concern.
The tool introduces the concept of virtual interface to support documenting sorts like Role superimposition. Because
some super-imposed roles do not have dedicated interfaces
but are tangled within other interface or class declarations,
the tool allows to define virtual interfaces for these roles.
The user can select in a graphical interface those members
of the multi-role type that define the analyzed role.
For some of the sorts that rely on a reference to a type
for specifying relations, like Redirection or Interfacing layer,
the referred type can be specified as either the type of a field
or the return type of a method in the referring type.
7. SORTS IN PRACTICE
We have documented crosscutting concerns in the JHotDraw application to test the suitability of the sorts and their
associated queries for describing and documenting such concerns in real systems.
JHotDraw is a relevant case for this task as it has been
developed as a show-case for applying design patterns, and
as we have seen, many of these patterns involve crosscutting
functionality. Moreover, the application has been proposed
and used as a common benchmark for aspect mining [17, 2].
We focus the discussion in this section on various instances
of design patterns in JHotDraw and how they can be modeled in SoQueT. The model for the documented concerns
in JHotDraw, including all those described in this paper,
is available for download 9 . This model can be loaded into
SoQueT and used to support concern understanding and exploration for the selected application, as well as a reference
for aspect mining.
7.1 Design patterns in JHotDraw
6.
TOOL SUPPORT
SoQueT (SOrts QUEry Tool) is the tool support we built
to query for sort instances and organize related instances in
TUD-SERG-2006-010
Figure 5 shows a number of core elements in the JHotDraw application and their collaborations. The figure is
8
9
http://swerl.tudelft.nl/view/AMR/SoQueT
http://swerl.tudelft.nl/view/AMR/SoQueT
13
SERG
Marin – Formalizing typical crosscutting concerns
Figure 4: SoQueT views
aimed at supporting the understanding of the pattern implementations discussed next.
Strategy
The Strategy pattern in JHotDraw is implemented by a number of elements, like Connector s and Locator s. Connectors
(e.g., ChopEllipseConnector, PolyLineConnector, etc.) define strategies for locating connection point on a figure, while
locators (ElbowTextLocator, PolyLineLocator, etc.) locate a
handle position on a figure.
The Strategy role is a primary role for the interfaces defining it and the classes implementing these interfaces. The
Context strategy role, however, is super-imposed, through
implementations of the ConnectionFigure and LocatorHandle interfaces respectively.
We document the pattern-specific roles by instances of the
Role superimposition sort: a sort query asks for all implementors of ConnectionFigure to describe the Context, and a
second query documents the Strategy in a similar way. We
group these two instances in a composite concern bearing
the name of the pattern instance, ConnectionStrategy.
Another instance of the pattern implements the update
strategy for drawing views. The Painter interface defines
the strategy in the context of the DrawingView hierarchy.
The DrawingView interface, however, defines multiple roles
and participates in multiple implementations of the Strategy
pattern for which it defines the context role. To document
the role of interest, we use a virtual interface definition. The
virtual interface definition comprises the members that are
role-specific, namely the members to refer (maintain and
allow access to) the Strategy object.
14
Visitor and Composite
The two patterns can be discussed together for the JHotDraw case, too. The Visitor pattern in JHotDraw allows to,
for instance, insert or delete figures into or from composite
figures, such as drawings. To realize this functionality, the
top level interface for figure elements (Figure) defines the
Visitable role. The role is tangled with the other concerns
common to figures, such as drawing the figure.
The one member of Figure specific to the Visitable role is
visit(FigureVisitor)10 . This will be part of a virtual interface definition for the super-imposed role of Visitable for
figure elements. Additionally, we can document the Visitor
role defined by the FigureVisitor interface.
One of the two implementations of the method for accepting visitors corresponds to composite figures and passes the
visitor to all the child-figures. Composite figures in JHotDraw extend the CompositeFigure abstract class, which defines and implements the method for manipulating childfigures. These methods are specific to the Composite role,
in the pattern with the same name, but tangled with the
definition of figure-specific methods. A new virtual interface definition for Role superimposition, which to include
the methods for children manipulation, helps us to document this crosscutting concern.
The Add variability instance discussed for the Visitor pattern in Section 5 is not applicable to this implementation of
the pattern.
10
The name of the method could be misleading, as it corresponds to the accept method in the GoF desciption of the
pattern
TUD-SERG-2006-010
SERG
Decorator
Marin – Formalizing typical crosscutting concerns
JHotDraw supports decoration of figures with borders and
animations through the use of the Decorator pattern. An
abstract class, DecoratorFigure, defines the decorator interface and the set of methods forwarding the requests to the
decorated figure. Concrete decorators extend the interface
and override some of the methods to add specific functionality.
The pattern’s implementation is a standard example of
Redirection layer sort instance. We document the instance
accordingly, by specifying in the query template the DecoratorFigure type and its field that stores the reference to
the decorated Figure. The query returns all the methods
in the Decorator that consistently redirect their calls to the
reference of the decorated figure.
The specification and the results of the query are also
shown in Figure 4.
DrawingEditor mediator by using the Observer pattern: the
two types of elements register themselves as listeners of the
mediator and receive notifications of changes of other colleagues. The DrawingEditor interface defines the Subject
role in the context of the Observer pattern for allowing the
mediator to communicate with the colleague objects. We
document this role by using (a virtual interface definition
for) Role Superimposition. This also documents the Mediator role. The calls to the notification mechanism (the
figureSelectionChanged method in DrawingEditor ) are an
instance of the Consistent behavior sort. The context for the
Consistent Behavior instance is given by the union of the
three hierarchies defining the Colleague roles: Tool, Command, and DrawingView.
This is one of the several instances of the Observer pattern that we have documented, one of them being shown in
Figure 1, for listeners to changes in figure elements.
State
Prototype
The Tool elements in JHotDraw realize an implementation
of the State pattern in the the context of DrawingView s.
The Tool and DrawingView interfaces define the State and
Context roles, respectively, which we describe by simple instances of the Role Superimposition sort.
The collaborations in the implementation of this pattern
instance are, however, more complex: A DrawingView knows
its DrawingEditor and through this it gets access to the active Tool object. The tools consistently notify the editor at
the end of their interaction by invoking the editor’s toolDone
method. The notification results in re-setting the active tool
to the default one. The consistent (notification) behavior is
a concern that crosscuts the tool elements. To describe the
concern using sorts, we query for the calls to the editor’s
toolDone method and restrict the query’s context to the
Tool hierarchy.
Most of the documented calls originate from mouse events
in the Tool classes. The mouse events are initiated by the
objects defining the State context, like a DrawingView forwards all its input events to the editor’s active tool. The
forwarding operation is an instance of the Redirection layer
sort. To document this concern, we query for the references
from mouse events listeners in the context class to the return type of the method providing the access point to the
tool (i.e., Tool DrawingView.tool()). This query, however,
will return no result because of the signature mismatch between the redirecting (delegator) method, and the method
to redirect to: for example, the mouseReleased(..) method
of the listener calls the tool’s mouseUp(..) method, and
mousePressed calls mouseDown(..). This is a limitation of
the tool in documenting redirection behavior.
One implementation of the Prototype pattern covers Figure elements and specific Tool elements, namely CreationTool s to create new figures from a specified prototype. The
Figures explicitly (re-)declare the inherited Object’s clone
method; hence, documenting the Prototype role using sort
instances only requires to define a virtual interface declaring
this method. To allow cloning using the Java mechanism,
the Figure interface has to extend Java’s Cloneable interface. Although the interface does not declare any member,
the extension declaration is part of a crosscutting concern
that we document by a Role superimposition instance realized through the Cloneable interface in the context of the
Figure hierarchy.
Other client types using (particular) Figure prototypes
include ConnectionTool and ConnectionHandle.
Mediator and Observer
The DrawingEditor element discussed above is participant
in an implementation of the Mediator pattern by defining
the interface to coordinate the different objects that participate in an editor. The participating elements (i.e., the
Colleague objects) are Tool, Command, or DrawingView objects. Each of the three types defines accessor methods for
the mediator object: these methods are part of the Colleague role and we document them as instances of the Role
Superimposition sort by declaring virtual interfaces for the
methods of interest.
The Tool and Command elements communicate with the
TUD-SERG-2006-010
Command
The Command pattern is implemented by 40 elements in the
Command hierarchy; half of these are anonymous classes.
The Command classes implement the Command role in the
pattern. The role can be described as an instance of the
Role superimposition sort.
The Receiver functionality varies for the different commands: some of them have associated a specialized Receiver, while others directly implement the command’s logic
(e.g., AlignCommand ). A common receiver for commands
are the Figure or Drawing elements as, for instance, for
BringToFront and SendToBack-Command s. However, many
commands do not carry out only a single forwarding, and
the actions they delegate to are not always dedicated to the
delegating command; that is, describing the Receiver role as
super-imposed is often a per-case decision.
We document the Receiver role, for instance, in the DrawApplication class. The class implements a method to create the standard menus of the JHotDraw drawings editor.
Each menu is associated an anonymous command whose
execute method delegates execution to dedicated actions
in the DrawApplication class. These actions define the Receiver role, which we describe in a sort-query using a virtual
interface definition (in the context of the same class).
A group of Command-Invoker elements consists of MenuItems and Buttons. These elements interface the associated commands by listening for action events and triggering the execution of commands. We document this behav-
15
SERG
Marin – Formalizing typical crosscutting concerns
ior with instances of the Interface layer sort. Furthermore,
the Command invokers implement the ActionListener interface whose unique method, actionPerformed, consistently
invokes the execute method of the associated command.
This is an instance of Consistent behavior.
The Command hierarchy also exhibits several other crosscutting concerns, less relevant to the pattern itself: The
named commands conduct a pre-condition check before execution, and a consistent notification at the end of their
execution. These concerns are instances of the Contract enforcement and Consistent behavior sorts, respectively. Commands make use of support classes to implement undo functionality, as discussed in Section 4. The undo support also
requires a proper, consistent initialization, which is a consistent behavior crosscutting the execute methods of the
Command classes.
Adapter
The Adapter pattern is implemented by the Handle elements, which adapt Figures to a common interface accessed
by Tool objects. Handles allow tools, like selection tools,
to manipulate figures. The implementation of the Adapter
pattern is based on object composition: the handle element
stores a reference to its owning figure and defines an accessor
method to this reference. The tool client (HandleTracker )
delegates mouse events to specific methods in the Handle
interface, which in turn translates the events into actions
directed to the handle’s figure.
The identified crosscutting elements to be described by
sort queries include the Adapter role in Handle classes to
access the reference to the Adaptee (i.e., Figure). We use for
this a virtual interface definition that declares the accessors
method for the Adaptee reference.
Although the implementation of handles is aimed at translating events into actions for figure elements, it does not
follow consistent rules for redirecting functionality to the
adaptee. Such a rule is apparent in the implementation of
the client tool: mouse events are redirected to dedicated
methods declared by the Handle interface; however, each
mouse event (up, drag, down) delegates to a method with a
different signature (invokeStart, invokeStep and invokeEnd).
Due to limitations in the tool implementation (and the query
template for the redirection layer sort), the concern remains
undocumented.
Nevertheless, Handles act as an interfacing layer for figures: the actions carried out by handles are in fact operations in the Figure objects. For instance, the PolygonScaleHandle for scaling and rotating PolygonFigures relies on the
scaleRotate method in the Figure class to answer mousedrag events.
Iterator and Singleton
The FigureEnumertor class participates in the implementation of multiple patterns. The class acts as an Adapter for
the standard Java Iterator, implements the Singleton pattern, and the Aggregate role in the Iterator pattern.
The Singleton role is documented through a virtual interface declaring the singleton member and the method to access it. We further document the references to this method
as a consistent behavior for accessing the functionality of
a FigureEnumerator object. A design enforcement sort instance indicates that the class should declare (only) a private
constructor. However, this implementation of the pattern
16
Figure 5: Collaborations in JHotDraw
uses a public constructor for resetting the Singleton. This
is rather atypical.
The access method to the unique instance of FigureEnumerator is also part of the Agregate role in the context of
the Iterator pattern, which is documented by a Role superimposition instance.
8. DISCUSSION
The experiments conducted for describing crosscutting
concerns as sort instances in complex cases, like design patterns, proved that sorts were effective in most of the cases.
The analysis carried out on different sample implementations of design patterns [8, 10], as well as on real systems,
like JHotDraw, shows that sorts cover well the crosscuting
concerns described in literature by various authors. Moreover, we covered a larger variety of concerns with the examples shown in the proposed catalog.
The sort queries and their tool support can be improved
in several cases. The query for the Redirection layer sort,
for instance, was not flexible enough for the cases where the
methods in the redirecting layer had different names than
the target methods. This limitation is due to the name
matching criteria, also common in many aspect-oriented languages, such as AspectJ. However, this matching is a common practice and we expect it to be present in most cases.
The formalization of the context to restrict the domain
of a relation can benefit from more flexible definitions as
well. Although the tool support for sort queries allows to
define contexts as enumerations of program elements, like
classes, packages, etc., elegant formalization of contexts give
a clearer description of the concern. Since context definitions are often similar to pointcuts definitions in aspectoriented languages, improved support for defining pointcuts
is equally useful for expressing contexts.
We mainly used structural relations for formalizing the
definition of a query’s context; however, many aspect (query)
languages also allow for definitions that use name-based criteria, such as all methods whose name starts with get. Another extension could consist of using support for source
code annotations to relate elements in a context by a common intent, like modifiers of the Subject state in implementations of the Observer pattern.
TUD-SERG-2006-010
SERG
The catalog of crosscutting concern sorts can contribute to
aspect refactoring efforts and extension of query languages
for aspect-oriented programming. For instance, sorts like
Design enforcement cannot be expressed by CME’s query
language or aspect languages like AspectJ.
The sort instances allow to group elements participating
in relevant crosscutting relations, which are not explicit in
source code. In this respect, the sorts are modular units
comparable with aspects. However, sorts are mainly aimed
at supporting crosscutting concern comprehension by describing atomic elements in a standard, consistent way. Such
elements can be associated template refactoring solutions
based on their description provided by sort queries. This is
a first step towards refactoring concerns to more complex
aspect solutions.
The sorts describe crosscutting concerns both intentionally and extentionally. The extent of the concern consits of
the elements captured by the sort query. The intent is given
by the query itself. New elements to be added to a system
should be aware of the intent of the concerns documented
by sort queries and be consistent with existent rules and
policies. To this end, the user should be able to query the
concern model for concerns that “touch” program elements
of interest to a development or maintainance task.
9.
Marin – Formalizing typical crosscutting concerns
[6]
[7]
[8]
[9]
[10]
[11]
CONCLUSIONS
This paper proposed a system for addressing crosscutting
functionality in source code based on crosscutting concern
sorts. Such a system can provide consistency and coherence
for referring and describing crosscutting concerns. As a result, sorts are useful in program comprehension and areas
like aspect mining and refactoring.
We have described crosscutting concern sorts as relations
between sets of program elements and formalized these relations through a query component. We have organized the
sorts in a catalog and discussed each sort in significant detail,
describing specific implementation idioms and examples of
concrete instances. Sorts have been assessed for crosscutting
relations present in design patterns and in a real application
system.
[12]
[13]
[14]
[15]
[16]
10. ACKNOWLEDGMENTS
I thank Arie van Deursen, Juri Memmert and anonymous
reviewers for comments and feedback on this and earlier
versions of this work.
11. REFERENCES
[1] The AspectJ Team. The AspectJ Programming Guide.
Palo Alto Research Center, 2003. Version 1.2.
[2] M. Ceccato, M. Marin, K. Mens, L. Moonen,
P. Tonella, and T. Tourwé. A Qualitative Comparison
of Three Aspect Mining Techniques. In Proceedings
13th International Workshop on Program
Comprehension (IWPC 2005), Los Alamitos, CA,
2005. IEEE Computer Society.
[3] L. Cole and P. Borba. Deriving Refactorings for
AspectJ. In Proc. Int. Conf. on Aspect-Oriented
Software Development (AOSD), Mar. 2005.
[4] P. Deransart, A. Ed-Dbali, and L. Cervoni. Prolog,
The Standard : Reference Manual. Springer Verlag,
1996.
[5] M. Eichberg, M. Haupt, M. Mezini, and T. Schäfer.
Comprehensive software understanding with sextant.
TUD-SERG-2006-010
[17]
[18]
[19]
[20]
[21]
In ICSM ’05: Proceedings of the 21st IEEE
International Conference on Software Maintenance
(ICSM’05), pages 315–324, Washington, DC, USA,
September 2005. IEEE Computer Society.
M. Eichberg, M. Mezini, T. Schäfer, C. Beringer, and
K.-M. Hamel. Enforcing system-wide properties. In
P. Strooper, editor, 2004 Australian Software
Engineering Conference, pages 158–167, Melbourne,
Australia, April 2004. IEEE Computer Society.
R. E. Filman, T. Elrad, S. Clarke, and M. Akşit,
editors. Aspect-Oriented Software Development.
Addison-Wesley, Boston, 2005.
E. Gamma, R. Helm, R. Johnson, and J. Vlissides.
Design Patterns: Elements of Reusable
Object-Oriented Software. Addison-Wesley, Reading,
MA, 1994.
J. Hannemann. Role-Based Refactoring of Crosscutting
Concerns. PhD thesis, Faculty of Graduate Studies,
University of British Columbia, Nov. 2005.
J. Hannemann and G. Kiczales. Design Pattern
Implementation in Java and AspectJ. In Proceedings
of the 17th Annual ACM conference on
Object-Oriented Programming, Systems, Languages,
and Applications (OOPSLA), pages 161–173, Boston,
MA, 2002. ACM Press.
W. Harrison, H. Ossher, S. M. S. Jr., and P. Tarr.
Concern Modeling in the Concern Manipulation
Environment. In IBM Research Report RC23344,
Yorktown Heights, NY, 2004. IBM Thomas J. Watson
Research Center.
J. Hughes. Why functional programming matters.
Comput. J., 32(2):98–107, 1989.
G. Kiczales, E. Hilsdale, J. Hugunin, M. Kersten,
J. Palm, and W. G. Griswold. An overview of aspectj.
In ECOOP ’01: Proceedings of the 15th European
Conference on Object-Oriented Programming, pages
327–353, London, UK, 2001. Springer-Verlag.
G. Kiczales, J. Lamping, A. Mendhekar, C. Maeda,
C. Lopes, J.-M. Loingtier, and J. Irwin.
Aspect-oriented programming. In 11th Europeen Conf.
Object-Oriented Programming, volume 1241 of LNCS,
pages 220–242. Springer Verlag, 1997.
R. Laddad. Aspect-Oriented Refactoring.
www.theserverside.com, Dec. 2003.
R. Laddad. AspectJ in Action - Practical Aspect
Oriented Programming. Manning Publications Co.,
Greenwich, CT, 2003.
M. Marin, A. van Deursen, and L. Moonen. Identifying
Aspects using Fan-In Analysis. In Proceedings of the
11th Working Conference on Reverse Engineering
(WCRE2004)., pages 132–141, Los Alamitos, CA,
2004. IEEE Computer Society Press.
M. Marin, L.Moonen, and A. van Deursen. A
classification of crosscutting concerns. In Proceedings
International Conference on Software Maintenance
(ICSM 2005). IEEE Computer Society, 2005.
M. Marin, L. Moonen, and A. van Deursen. An
approach to aspect refactoring based on crosscutting
concern types. In Int. Workshop on the Modeling and
Analysis of Concerns in Software, ICSE. Software
Engineering Notes (volume 30, issue 4), 2005.
M. Monteiro and J. Fernandes. Towards a Catalog of
Aspect-Oriented Refactorings. In International
Conference on Aspect-Oriented Software Development,
2005.
D. Riehle and T. Gross. Role model based framework
design and integration. In Proceedings of the 13th
17
Marin – Formalizing typical crosscutting concerns
[22]
[23]
[24]
[25]
18
SERG
ACM SIGPLAN OOPSLA:, pages 117–133, New
York, NY, USA, 1998. ACM Press.
M. Robillard and G. Murphy. Concern Graphs:
Finding and Describing Concerns. In Proc. Int. Conf.
on Software Engineering (ICSE). IEEE, 2002.
S. Soares, E. Laureano, and P. Borba. Implementing
Distribution and Persistence Aspects with AspectJ. In
Proc. 17th Conf. on Object-oriented programming,
systems, languages, and applications. ACM Press,
2002.
P. Tarr, W. Harrison, and H. Ossher. Pervasive Query
Support in the Concern Manipulation Environment.
In IBM Research Report RC23343, Yorktown Heights,
NY, 2004. IBM Thomas J. Watson Research Center.
P. Tarr, H. Ossher, W. Harrison, and J. Stanley
M. Sutton. N degrees of separation: multi-dimensional
separation of concerns. In ICSE ’99: Proceedings of
the 21st international conference on Software
engineering, pages 107–119, Los Alamitos, CA, USA,
1999. IEEE Computer Society Press.
TUD-SERG-2006-010
SERG
Marin – Formalizing typical crosscutting concerns
APPENDIX
A.
AN ASPECTJ SOLUTION TO THE VISITOR PATTERN
This example shows a simplified AspectJ solution for a pricing visitor: any call to the getPrice method of a (Composite)Equipment object are advised for building a Visitable object and pass it to visitors that just invoke the accept method of
the passed argument.
pointcut price(Equipment equipment):
call(int Equipment+.getPrice())
&& target(equipment);
int around(final Equipment equipment) : price(equipment) {
Visitable visitable = new Visitable() {
public void accept(EquipmentPriceVisitor v) {
if(equipment instanceof CompositeEquipment) {
List subcomponents = equipment.getComponents();
for(int i=0; i<subcomponents.size(); i++)
(((Equipment)subcomponents.get(i))).getPrice();
}
v.addToPrice(proceed(equipment));
}
};
visitor.visitEquipment(visitable);
return visitor.getTotalPrice();
}
TUD-SERG-2006-010
19
Marin – Formalizing typical crosscutting concerns
20
SERG
TUD-SERG-2006-010
TUD-SERG-2006-010
ISSN 1872-5392
SERG
| 2cs.AI
|
1
On the state space and dynamics selection in linear
stochastic models: a spectral factorization approach
arXiv:1802.00253v1 [cs.SY] 1 Feb 2018
Augusto Ferrante and Giorgio Picci
Abstract—Matrix spectral factorization is traditionally described as finding spectral factors having a fixed analytic pole
configuration. The classification of spectral factors then involves
studying the solutions of a certain algebraic Riccati equation
which parametrizes their zero structure. The pole structure of
the spectral factors can be also parametrized in terms of solutions
of another Riccati equation. We study the relation between the
solution sets of these two Riccati equations and describe the
construction of general spectral factors which involve both zeroand pole-flipping on an arbitrary reference spectral factor.
I. I NTRODUCTION
An important and widely used class of models in control
engineering and signal processing describes an m-dimensional
observed random signal {y(t)} as output of a linear system
driven by white noise:
x(t + 1) = Ax(t) + Bw(t)
(1)
y(t) = Cx(t) + Dw(t)
where A ∈ Rn×n , B ∈ Rn×m , C ∈ Rm×n , D ∈ Rm×m ,
w is a normalized white noise. The n-dimensional signal x
is the state vector. The basic steps for the constructions of
models of the form (1) from observations of {y(t)} lead to
the following three problems which in various forms permeate
all linear systems and control theory:
1) Estimate the spectral density Φy (z) of y, see [5], [8], [9],
[15]–[17] and references therein.
2) Compute a stochastically minimal1 spectral factor of
Φy (z), i.e. a matrix transfer function W (z) such that
Φy (z) = W (z)W ⊤ (z −1 ),
(2)
see [2], [3] and references therein.
3) Fix a minimal realization W (z) = C(zI − A)−1 B + D
to provide a parametrization of the model (1).
The literature on these topics being enormous we have chosen
to quote only a few recent papers in which one can find a
more extensive bibliography. The study of models (1) of the
signal y without a priori constraints of causality or analiticity is
exposed in the recent book [11]. The objective of this paper is
to continue the analysis and study in more depth the relations
among different models (1) which are in a sense equivalent as
they serve to represent the same process but may have different
system-theoretic structure and properties.
A. Ferrante and G. Picci are with the Department of Information Engineering, University of Padova, via Gradenigo 6/B, 35131 Padova, Italy; e-mail:
[email protected], [email protected]
1 Stochastic minimality means that we are only interested in models of
minimal complexity so that we only consider spectral factors W (z) of
minimal McMillan degree.
Indeed, representations (1) have several degrees of freedom.
The most obvious (and least interesting) one is the choice
of basis in the input and in the state space. In particular,
the matrices A, B, C, D in step 3. are determined up to a
transformation of the form T −1 AT, T −1 BU, CT, DU where
T is an arbitrary invertible matrix and U is an arbitrary
orthogonal matrix. Once these degrees of freedom are factored
out, we are left with two more interesting objects:
A. The state space as a coordinate free representative of a
model (1)
B. The (dynamical) causality structure (related in particular the
choice of direction of the time arrow) of equivalent models.
One of the key result of stochastic realization theory (see [11])
is that these two choices correspond, respectively, to the choice
of zeros and poles of the spectral factor W (z) in (2). Each pole
configuration of the spectral factor corresponds to a certain
causality structure so that, once this configuration is fixed,
one is left with the choice of the zero structure of the spectral
factor, which just means choosing a (minimal) state space of
the realization.
Matrix spectral factorization is traditionally described as
finding spectral factors having a fixed analytic pole configuration so that all corresponding models are causal, and
classifying different models corresponds to parametrizing all
possible zero structures of W . However, a zero structure fixes,
independent of causality, a possible minimal state space2 for
y. Hence, once a minimal state space (i.e. the zero structure
of W ) is fixed, there is a whole family of possible causality
structures which can be parametrized by the allowed pole
locations of a spectral factor W .
If some minimal reference spectral factor is fixed, minimal
spectral factorization can be seen as a zero- or pole- flipping
transformation performed on the reference factor. In this
paper we analyse the interplay between this two operations
in relation to the solutions sets of two families of algebraic
Riccati equations. We derive closed-form formulas that allow
to compute the model corresponding to a given causality structure and state space. This may be viewed as the completion of
an endeavour first undertaken in [12] in continuous time but
not pushed to the final consequences. Here we shall address
the discrete-time situation and give a complete solution.
Although our main motivation is stochastic modelling, our
contribution can also be viewed as related to algebraic Riccati
equations and to spectral factorization. Both have important
applications in several areas of control, signal processing and
2 We stress that the choice of the state space must not be confused with
the choice of basis in Rn .
2
system theory.
Some technical assumptions of this paper could probably
be weakened however most probably at the expense of clarity.
For pedagogical reasons we have decided to work in a setting
which reduces complications to a minimum.
II. BACKGROUND ON SPECTRAL FACTORIZATION AND
A LGEBRAIC R ICCATI EQUATIONS
Let Φ(z) be a m × m rational spectral density matrix of
a regular stationary process, where regularity is meant in the
sense explained in [11, Sec. 6.8] and let
W (z) := C(zI − A)−1 B + D
(3)
be a minimal realization of a minimal square spectral factor
of Φ(z) so that Φ(z) = W (z)W (z)∗ , where W (z)∗ :=
W (z −1 )⊤ is the conjugate transpose. By regularity the matrix
D is non singular, [6]; it will be assumed to be symmetric
and positive definite: this rules out the uninteresting degree of
freedom corresponding to multiplying a spectral factor on the
right side by a constant orthogonal matrix.
By regularity the numerator matrix Γ := A − BD−1 C is
non-singular (see Theorem 6.8.2 in [11]). In this paper we
shall moreover assume that both A and Γ are unmixed. Note
in particular that we do not assume analiticity of W (z) outside
of the unit disk. For the relevant definitions and facts about
spectral factorization in this context we shall refer to Chap 16
of the book [11].
Definition 2.1: Let Wi (z) ; i = 1, 2 be minimal spectral
factors of the same rational spectral density. We shall say that
W1 (z) and W2 (z) have the same pole structure if they admit
a state space realization with the same state transition matrix.
Likewise, we say that W1 (z) and W2 (z) have the same zero
structure if they admit a state space realization with the same
numerator matrix.
In classical spectral factorization one assumes that a the state
matrix A has all eigenvalues inside the unit circle and one aims
at classifying all different minimal spectral factors having a
fixed (analytic) pole structure, in terms of their zero structure,
equivalently, in terms of invariant subspaces for the transpose
of the numerator matrix Γ. It is well-known that this involves
the study of an algebraic Riccati equation. In the present
context we have the following result, which has appeared in
several places in the literature.
Proposition 2.1: Let W0 (z) := C(zI − A)−1 B + D be a
minimal realization of a square reference spectral factor.
1) There is a one-to-one correspondence between symmetric
solutions of the homogeneous algebraic Riccati equation
P = ΓP Γ⊤ − ΓP C ⊤ (DD⊤ + CP C ⊤ )−1 CP Γ⊤ (4)
and minimal spectral factors of Φ(z) having the same
pole structure of W0 (z). This correspondence is defined
by the map assigning to each solution P the spectral
factor
WP (z) := C(zI − A)−1 BP + DP
where
BP := (BD⊤ + AP C ⊤ )(DD⊤ + CP C ⊤ )−1/2 ;
DP := (DD⊤ + CP C ⊤ )1/2 .
(5)
(6)
2) There is a one-to-one correspondence between symmetric
solutions of (4) and Γ⊤ -invariant subspaces which is
defined by the map assigning to each solution P the Γ⊤ invariant subspace ker(P ).
For a proof we shall just refer the reader to Corollary 16.5.7
and Lemma 16.5.8 in [11] where the equation differs by
an inessential change of sign. A similar Riccati equation
although in a different context is studied in [14].3
In particular, let P+ be the unique non singular solution
of (4), then the corresponding Γ⊤ −invariant subspace ker P+
is trivial and the zeros of W0 (z) are all flipped to reciprocal
positions. This follows from standard Riccati theory. We shall
denote the corresponding spectral factor by W+ (z).
Zero-flipping can also be visualized as right-multiplication
of W0 (z) by a suitable square all-pass function so as to
preserve minimality. The entailed factorization of WP (z) is
in turn uniquely identified by the existence of a Γ⊤ −invariant
subspace [4].
On the other hand, we have the following fact which
describes the pole-flipping relation among spectral factors
keeping a fixed zero structure. The result can be traced back
to Theorem 16.4.2 of [11].
Proposition 2.2: Let W0 (z) := C(zI − A)−1 B + D be a
minimal realization of a square reference spectral factor.
1) There is a one-to-one correspondence between symmetric
solutions of the algebraic Riccati equation
Q = A⊤ QA − A⊤ QB (I + B ⊤ QB)−1 B ⊤ QA ,
(7)
and minimal normalized spectral factors having the same
zero structure of W0 (z). This correspondence is defined
by the map assigning to each solution Q the spectral
factor
WQ (z) := CQ (zI − AQ )−1 BQ + DQ ,
where
(8)
∆Q := I + B ⊤ QB ,
⊤
CQ := C − D∆−1
Q B QA ,
⊤
AQ := A − B∆−1
Q B QA ,
BQ :=
DQ :=
(9)
−1/2
B∆Q U ,
−1/2
D∆Q U ,
and U is the orthogonal matrix
−1/2 ⊤
U := (D∆Q
−1/2
) ((D∆Q
−1/2 ⊤ −1/2
)(D∆Q
) )
which is selected in such a way that DQ is symmetric
and positive definite.
2) There is a one to one correspondence between symmetric
solutions of (7) and A-invariant subspaces which is
defined by the map assigning to each solution Q the Ainvariant subspace ker(Q).
3 Any solution P can actually be seen as the difference say X − X of
0
two arbitrary solutions of an equivalent Riccati equation parametrizing the
minimal spectral factors which is defined directly in terms of a realization of
Φ and does not involve a reference spectral factor, see [11, Sect. 16.5]. Here
X0 is kept fixed as a reference solution and Γ describes the zero structure of
the reference spectral factor W0 .
3
Proof. That the zero structures of WQ (z) and of W0 (z)
coincide is the content of Theorem 16.4.5 in [11]. The rest
is readily checked.
III. C OMBINING
POLE AND ZERO FLIPPING
We want to understand the combination of zero- and poleflipping leading to an arbitrary minimal square spectral factor
W . To this end let’s consider the spectral factor WQ (z) defined
in (8) as a reference spectral factor and describe the zeroflipping process on WQ (z). By direct computation we easily
find that the numerator matrix of WQ (z) is the same of the
numerator matrix of W0 (z), i.e. the matrix Γ. Hence the
Riccati equation (4) corresponding to WQ (z) takes the form
⊤
⊤ −1
⊤
) CQ PQ Γ⊤
PQ = ΓPQ Γ⊤ − ΓPQ CQ
(DQ DQ
+ CQ PQ CQ
(10)
−1/2
where CQ is as defined in (9) and DQ := D∆Q U . Notice
that, since equations (4) and (10) involve the same matrix Γ
and each symmetric solution of either equation is uniquely
attached to a Γ⊤ -invariat subspace [13], the map assigning
to each solution P of (4) the solution PQ of (10) such that
ker(P ) = ker(PQ ) is a one to one correspondence between
the set P and the set PQ of symmetric solutions of (4) and
(10).
Our main contribution is to analyze the relations between
P and PQ and to provide an explicit formula to compute
the solution PQ from a give pair P, Q. In this way once
parametrized the solutions of (4) and (7), we do not need to
solve (10) and we have a closed-form formula for the spectral
factor with assigned pole and zero structure, or equivalently
for the model with assigned state-space and causality structure.
We may of course consider a dual path to transform W0
into W by taking instead the zero-flipped WP as a reference
and flipping poles by considering a solution QP of a Riccati
equation similar to (7) so as to make the following diagram
commutative:
P
W0 −−−−→ WP
Q
Qy
y P
(11)
WQ −−−−→ W
PQ
The resulting spectral factor should have been denoted WP Q
but the simplified notation W here should not be cause of
confusion.
It is well-known that both (4) and (10) have a unique
non-singular solution which we denote by P+ and PQ+ ,
respectively. The relation between these two solutions is the
content of the following lemma.
Lemma 3.1: The nonsingular solutions P+ and PQ+ are
related by the formula
−1
PQ+
= Q + P+−1 .
(12)
Proof. It is immediate to check that P+−1 is the (unique)
solution of the discrete-time Lyapunov equation
P+−1 − Γ⊤ P+−1 Γ + C ⊤ D−⊤ D−1 C = 0.
(13)
−1
Similarly, PQ+
is the (unique) solution of the discrete-time
Lyapunov equation
−1
−1
⊤ −⊤ −1
PQ+
− Γ⊤ PQ+
Γ + CQ
DQ DQ CQ = 0.
(14)
−1
Therefore, the difference ∆ := PQ+
− P+−1 is the (unique)
solution of the discrete-time Lyapunov equation
⊤ −⊤ −1
∆ − Γ⊤ ∆Γ + CQ
DQ DQ CQ − C ⊤ D−⊤ D−1 C = 0. (15)
We now compute
RQ
⊤ −⊤ −1
DQ DQ CQ − C ⊤ D−⊤ D−1 C
:= CQ
=
⊤
C ⊤ D−⊤ ∆Q D−1 C + A⊤ QB∆−1
Q B QA
−C ⊤ D−⊤ B ⊤ QA − A⊤ QBD−1 C
=
−C ⊤ D−⊤ D−1 C
C ⊤ D−⊤ B ⊤ QBD−1 C − C ⊤ D−⊤ B ⊤ QA
⊤
−A⊤ QBD−1 C + A⊤ QB∆−1
Q B QA
=
C ⊤ D−⊤ B ⊤ QBD−1 C − C ⊤ D−⊤ B ⊤ QA
−A⊤ QBD−1 C + A⊤ QA − Q
=
Γ⊤ QΓ − Q
This equation, together with (15) gives
∆ − Γ⊤ ∆Γ = Q − Γ⊤ QΓ
and, by uniqueness,
follows.
−1
∆ := PQ+
− P+−1 = Q,
(16)
so that (12)
Since ker P+ = {0}, all zeros of the corresponding spectral
factor, denoted by the symbol W+ (z), are those of W0 flipped
to their reciprocals. The same happens for WQ whatever solution Q of (7) is chosen. In particular, denoting the nonsingular
solution of (7) by Q+ , all poles of the corresponding spectral
factor, say W̄0 (z), will be the reciprocals of those of W0 (z).
The commutative diagram (11) takes on the form
P+
W0 −−−−→ W+
[QP ]+
Q+ y
y +
W̄0 −−−−−→ W̄+
[PQ+ ]+
where [QP+ ]+ and [PQ+ ]+ are the invertible solutions of the
Riccati equations which are respectively flipping the poles of
W+ (z) and the zeros of W̄0 (z). Hence, both poles and zeros
of W̄+ are the reciprocals of those of W0 (z). This corresponds
to “total” flipping of singularities.
We would now like to derive an explicit formula generalizing (12) to a generic solution PQ expressed as a function of P
and Q. To this end we shall use the following lemma which
is a particular case of [1, Theorem 2.2]. An analogous result
is Statement 1. (iii) of Theorem 3.1 in [7] although referring
to the specific case of all-pass functions.
Lemma 3.2: Any solution P of the Riccati equation (4)
corresponding to a Γ⊤ -invariant subspace S can be expresses
by the formula
†
(17)
P = (I − ΠS )P+−1 (I − ΠS )
4
where † denotes Moore-Penrose pseudoinverse and ΠS is the
orthogonal projector onto the subspace S = ker P .
We are now ready to present our main result.
Theorem 3.1: Let P be an arbitrary solution of (4). Then
the unique solution PQ of (10) such that ker(P ) = ker(PQ )
can be expressed by the formula
PQ = [P P † QP P † + P † ]†
(18)
wich generalizes (12).
Proof. Since (I − ΠS ) projects onto the range space of P ,
a basic property of the Moore-Penrose pseudoinverse [10, P.
421] implies that (I−ΠS ) = P P † so that (17) can be rewritten
†
P = P P † P+−1 P P † and hence
P † = P P † P+−1 P P † .
(19)
Now, since P and PQ have the same kernel they also have the
same image so that the orthogonal projectors on this image
may be written in two ways as:
I − ΠS = P P † = PQ PQ† .
(20)
Thus, the analog of formula (17) for PQ yields
i†
h
i† h
−1
−1
PQ = (I − ΠS )PQ+
(I − ΠS ) = P P † PQ+
(21)
PP†
where PQ+ is the only non-singular solution of (10) (such a
solution corresponds to the Γ⊤ -invariant subspace {0}).
Hence, after inserting (12), we get PQ = [P P † (Q +
P+−1 )P P † ]† , and, finally, by using (19) we obtain the following explicit expression for PQ depending only on P and
Q:
PQ = [P P † QP P † + P † ]† .
(22)
Finally, let us consider two arbitrary Γ⊤ and A- invariant
subspaces X and Y which is to say two arbitrary zero and
pole flipping transformations of the singularities of W0 (z) or,
equivalently, an arbitrary state space and causality configuration for the model (1). Suppose we want to compute the
corresponding minimal spectral factor W (z), or equivalently
the corresponding model (1). Let P and Q be the solutions of
the Riccati equations (4) and (7) corresponding to the invariant
subspaces X and Y and consider the left lower path in the
commutative diagram (11) so that the zero flipping is done
after a pole flipping defined by Q. The relevant Riccati solution
PQ is given in formula (18) so that the desired realization of
W (z) can be explicitly written in closed form as
WQ (z) := CQ (zI − AQ )−1 BPQ + DPQ
(23)
where
⊤
⊤
⊤
⊤ −1/2
+ AQ PQ CQ
)(DQ DQ
+ CQ PQ CQ
)
BPQ := (BQ DQ
⊤
⊤ 1/2
+ CQ PQ CQ
) ,
DPQ := (DQ DQ
(24)
PQ is given by (18) and AQ , BQ , CQ , DQ are given by (9).
Naturally, an analogous procedure would work by following
the upper right path; i.e. computing first P and then performing
the appropriate pole flipping defined by QP .
C ONCLUSION
We have discussed the classification of general (not necessarily analytic) square spectral factors in terms of the solutions
of two algebraic Riccati equations. We have also described
the construction of general spectral factors which involve
both zero- and pole-flipping on an arbitrary reference spectral
factor.
R EFERENCES
[1] D. Alpago and A. Ferrante, Families of Solutions of Algebraic
Riccati Equations. Submitted to Automatica. Preprint available at
arXiv:1801.09557, 2018.
[2] G. Baggio, and A. Ferrante. On Minimal Spectral Factors with Zeroes
and Poles lying on Prescribed Regions. IEEE Trans. Automatic Control.
Vol. AC-61(8):2251–2255, DOI: 10.1109/TAC.2015.2484330, 2016.
[3] G. Baggio, and A. Ferrante. On the Factorization of Rational DiscreteTime Spectral Densities. IEEE Trans. Automatic Control. Vol. AC61(4):969–981, DOI: 10.1109/TAC.2015.2446851, 2016.
[4] H. Bart, I. Gohberg and R. Kaashoek, Minimal Factorization of Matrix
and Operator Functions, Operator Theory 1, Birkhäuser Verlag, Basel
1984.
[5] A. Ferrante, C. Masiero, and M. Pavon. Time and Spectral Domain
Relative Entropy: A New Approach to Multivariate Spectral Estimation.
IEEE Trans. Automatic Control. Vol. AC-57(10):2561–2575, 2012.
[6] A. Ferrante, G. Picci, and S. Pinzoni. Silverman Algorithm and the
Structure of Discrete-Time Stochastic Systems. Linear Algebra and its
Applications. Vol. 351–352:219–242, 2002.
[7] A. Ferrante and G. Picci, Representation and Factorization of DiscreteTime Rational All-Pass Functions, IEEE Transactions on Automatic
Control, vol. 62, No. 7, July 2017.
[8] T. Georgiou. Spectral analysis based on the state covariance: the
maximum entropy spectrum and linear fractional parameterization. IEEE
Trans. Aut. Control, 47:1811–1823, 2002.
[9] T. Georgiou. Relative entropy and the multivariable multidimensional
moment problem. IEEE Trans. Inform. Theory, 52:1052–1066, 2006.
[10] R. A. Horn and C.R. Johnson Matrix Analysis Cambridge U.P. 1985.
[11] A. Lindquist and G. Picci, Linear Stochastic Systems: A Geometric
Approach to Modeling, Estimation and Identification, Springer series
in Contemporary Mathematics, Heidelberg, Germany: Springer Verlag,
2015.
[12] G. Picci and S. Pinzoni: Acausal Models and Balanced realizations of
stationary processes Linear Algebra and its Applications, vol. 205–206,
pp. 957–1003, 1994.
[13] J. C. Willems, Least squares stationary optimal control and the algebraic
Riccati equation, IEEE Transactions on Automatic Control, vol. 16: 621634, 1971.
[14] Harald K. Wimmer, A Parametrization of Solutions of the Discretetime Algebraic Riccati Equation Based on Pairs of Opposite Unmixed
Solutions SIAM Journal on Control and Optimization Vol. 44 n. 6, 2006.
[15] M. Zorzi. A new family of high-resolution multivariate spectral
estimators. IEEE Trans. Autom. Control, 59(4):892–904, Apr. 2014.
[16] M. Zorzi. Multivariate Spectral Estimation based on the concept of
Optimal Prediction. IEEE Trans. Autom. Control, 60:1647–1652, Jun.
2015.
[17] M. Zorzi. Rational approximations of spectral densities based on
the alpha divergence. Mathematics of Control, Signals, and Systems,
26(2):259–278, 2014.
| 3cs.SY
|
arXiv:1404.0352v1 [math.AG] 1 Apr 2014
CHERN CHARACTERS FOR TWISTED MATRIX
FACTORIZATIONS AND THE VANISHING OF THE HIGHER
HERBRAND DIFFERENCE
MARK E. WALKER
Abstract. We develop a theory of “ad hoc” Chern characters for twisted
matrix factorizations associated to a scheme X, a line bundle L, and a regular
global section W ∈ Γ(X, L).
As an application, we establish the vanishing, in certain cases, of hR
c (M, N ),
the higher Herbrand difference, and, ηcR (M, N ), the higher codimensional analogue of Hochster’s theta pairing, where R is a complete intersection of codimension c with isolated singularities and M and N are finitely generated Rmodules. Specifically, we prove such vanishing if R = Q/(f1 , . . . , fc ) has only
isolated singularities, Q is a smooth k-algebra, k is a field of characteristic 0,
the fi ’s form a regular sequence, and c ≥ 2.
Such vanishing was previously established in the general characteristic, but
graded, setting in [16].
Contents
1. Introduction
2. Chern classes for generalized matrix factorizations
2.1. The Jacobian complex
2.2. Matrix factorizations
2.3. Connections for matrix factorizations and the Atiyah class
2.4. Ad hoc Hochschild homology
2.5. Supports
2.6. Chern classes of matrix factorizations
2.7. Functorality of the Chern character and the top Chern class
3. Relationship with affine complete intersections
3.1. Euler characteristics and the Herbrand difference for hypersurfaces
3.2. Matrix factorizations for complete intersections
4. Some needed results on the geometry of complete intersections
5. The vanishing of Chern characters
5.1. The vanishing of the top Chern class
5.2. The Polishchuk-Vaintrob Riemann-Roch Theorem
5.3. Vanishing of the higher h and η Invariants
Appendix A. Relative Connections
A.1. The classical Atiyah class
A.2. The vanishing of the classical Atiyah class and connections
References
Date: April 2, 2014.
1
2
4
5
7
10
12
14
15
17
18
18
20
21
24
24
26
27
29
29
31
32
2
MARK E. WALKER
1. Introduction
This paper concerns invariants of finitely generated modules over an affine complete intersection ring R. Such a ring is one that can be written as R = Q/(f1 , . . . , fc ),
where Q is a regular ring and f1 , . . . , fc ∈ Q form a regular sequence of elements.
For technical reasons, we will assume Q is a smooth algebra over a field k. More
precisely, this paper concerns invariants of the singularity category, Dsg (R), of R.
This is the triangulated category defined as the Verdier quotient of Db (R), the
bounded derived category of R-modules, by the sub-category of perfect complexes
of R-modules. More intuitively, the singularity category keeps track of just the
“infinite tail ends” of projective resolutions of finitely generated modules. Since R
is Gorenstein, Dsg (R) is equivalent to M CM (R), the stable category of maximal
Cohen-Macaulay modules over R.
Thanks to a Theorem of Orlov [19, 2.1], the singularity category of R is equivalent
to the singularity category of the hypersurface Y of Pc−1
= Proj Q[T1 , . . . , Tc ]
Q
P
c−1
is
cut out by the element W = i fi Ti ∈ Γ(PQ , O(1)). Since the scheme Pc−1
Q
regular Dsg (Y ) may in turn be given by the homotopy category of “twisted matrix
factorizations” associated to the triple (Pc−1
Q , O(1), W ) (see [21], [15], [18], [23],
[2]). In general, if X is a scheme (typically smooth over a base field k), L is a
line bundle on X, and W ∈ Γ(X, L) is regular global section of L, a twisted matrix
factorization for (X, L, W ) consists of a pair of locally free coherent sheaves E0 , E1
on X and morphisms d1 : E1 → E0 , d0 : E0 → E1 ⊗OX L such that each composite
d0 ◦ d1 and (d1 ⊗ idL ) ◦ d0 is multiplication by W . In the special case where
X = Spec(Q) for a local ring Q, so that L is necessarily the trivial line bundle
and W is an element of Q, a twisted matrix factorization is just a classical matrix
factorization as defined by Eisenbud [10].
Putting these facts together, we get an equivalence
X
fi Ti )
Dsg (R) ∼
,
O(1),
= hmf (Pc−1
Q
i
between the singularity category
Pof R and the homotopy category of twisted matrix
factorizations for (Pc−1
,
O(1),
i fi Ti ). The invariants we attach to objects of
Q
Dsg (R) for a complete intersection ring R are actually defined, more generally, in
terms of such twisted matrix factorizations. In detail, for a triple (X, L, W ) as
above, let us assume also that X is smooth over a base scheme S and L is pulledback from S. In the main case of interest, namely X = Pc−1
for a smooth k-algebra
Q
c−1
Q, the base scheme S is taken to be Pk . In this situation, we define a certain
hoc
ad hoc Hochschild homology group HHad
(X/S, L, W ) associated to (X, L, W ).
0
Moreover, to each twisted matrix factorization E for (X, L, W ) we attach a class
hoc
ch(E) ∈ HHad
(X/S, L, W ),
0
called the ad hoc Chern character of E. We use the term “ad hoc” since we offer no
justification that these invariants are the “correct” objects with these names given
by a more theoretical framework. But, if X is affine and L is the trivial bundle,
then our ad hoc Hochschild homology groups and Chern characters coincide with
the usual notions found in [9], [8], [22], [7], [25], [17], [26], [20]. In particular, in
this case, our definition of the Chern character of a matrix factorization is given by
the “Kapustin-Li formula” [14].
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
3
When the relative dimension n of X → S is even, we obtain from ch(E) a “top
Chern class”
ctop (E) ∈ H 0 (X, JW ( n2 ))
where JW is the “Jacobian sheaf” associated to W . By definition, JW is the
dW ∧−
n−1
coherent sheaf on X given as the cokernel of the map ΩX/S
⊗ L−1 −−−−→ ΩnX/S
define by multiplication by the element dW of Γ(X, Ω1X/S ⊗OX L). The top Chern
class contains less information than the Chern character in general, but in the affine
case with L trivial and under certain other assumptions, the two invariants coincide.
The main technical result of this paper concerns the vanishing of the top Chern
class:
Theorem 1.1. Let k be a field of characteristic 0 and Q a smooth k-algebra of
even dimension. If f1 , . . . , fc ∈ Q is a regular sequence of elements with c ≥ 2 such
that the singular locus of R = Q/(f1 , . . . , fc ) is zero-dimensional, then
ctop (E) = 0 ∈ H 0 (X, JW ( n2 ))
P
for all twisted matrix factorizations E of (Pc−1
Q , O(1),
i fi Ti ).
As an application of this Theorem, we prove the vanishing of the invariants ηc
and hc for c ≥ 2 in certain cases. These invariants are defined for a pair of modules
M and N over a complete intersection R = Q/(f1 , . . . , fc ) having the property
that ExtiR (M, N ) and ToriR (M, N ) are of finite length for i ≫ 0. For example, if
the non-regular locus of R is zero-dimensional, then every pair of finitely generated
modules has this property. In general, for such a pair of modules, the even and odd
sequences of lengths of ExtiR (M, N ) and ToriR (M, N ) are governed, eventually, by
polynomials of degree at most c − 1: there exist polynomials
pev (M, N )(t) = ac−1 tc−1 + lower order terms
podd (M, N )(t) = bc−1 tc−1 + lower order terms
with integer coefficients such that
2i+1
lengthR Ext2i
(M, N ) = podd (M, N )(i)
R (M, N ) = pev (M, N )(i) and lengthR ExtR
for i ≫ 0, and similarly for the Tor modules. Following Celikbas and Dao [4,
3.3], we define the invariant hc (M, N ) in terms of the leading coefficients of these
polynomials:
ac−1 − bc−1
hc (M, N ) :=
c2c
The invariant ηc (M, N ) is defined analogously, using Tor modules instead of Ext
modules; see [5].
If c = 1 (i.e., R is a hypersurface) the invariants h1 and η1 coincide (up to a
factor of 12 ) with the Herbrand difference, defined originally by Buchweitz [1], and
Hochster’s θ invariant, defined originally by Hochster [12]. In this case, ExtiR (M, N )
and TorR
i (M, N ) are finite length modules for i ≫ 0 and the sequence of their
lengths are eventually two periodic, and h1 and η1 are determined by the formulas
R
R
R
2hR
1 (M, N ) = h (M, n) = length Ext2i (M, N ) − length Ext2i+1 (M, N ), i ≫ 0
and
R
2η1R (M, N ) = θR (M, n) = length TorR
2i (M, N ) − length Tor2i+1 (M, N ), i ≫ 0.
We can now state the main application of Theorem 1.1:
4
MARK E. WALKER
Theorem 1.2. Let k be a field of characteristic 0 and Q a smooth k-algebra. If
f1 , . . . , fc ∈ Q is a regular sequence of elements with c ≥ 2 such that the singular locus of R = Q/(f1 , . . . , fc ) is zero-dimensional, then hR
c (M, N ) = 0 and
ηcR (M, N ) = 0 for all finitely generated R-modules M and N .
Acknowledgements. I am grateful to Jesse Burke, Olgur Celikbas, Hailong Dao,
Daniel Murfet, and Roger Wiegand for conversations about the topics of this paper.
2. Chern classes for generalized matrix factorizations
In this section, we develop ad hoc notions of Hochschild homology and Chern
classes for twisted matrix factorizations. The connection with the singularity category for complete intersection rings will be explained more carefully later.
Assumptions 2.1. Throughout this section, we assume
• S is a Noetherian scheme.
• p : X → S is a smooth morphism of relative dimension n.
• LS is a locally free coherent sheaf of rank one on S. Let LX := p∗ LS , the
pullback of LS along p.
• W ∈ Γ(X, LX ) is global section of LX .
Recall that the smoothness assumption means that p : X → S is flat and of
finite type and that Ω1X/S is a locally free coherent sheaf on X of rank n.
Assumptions 2.2. When discussing ad hoc Hochschild homology and Chern characters (see below), we will also be assuming:
• p is affine and
• n! is invertible in Γ(X, OX ).
It is useful to visualize Assumptions 2.1 by a diagram. Let q : LS → SL
denote the
geometric line bundle whose sheaf of sections is LS ; that is, LS = Spec( i≥0 L−i
S ).
W
Then we may interpret W as a morphism X −→ LS fitting into the commutative
triangle
X❈
❈❈
❈❈W
p
❈❈
❈!
S o q LS .
The data (p : X → S, L, W ) determines a closed subscheme Y of X, defined by
the vanishing of W . More formally, if z : S ֒→ LS denotes the zero section of the
line bundle LS , then Y is defined by the pull-back square
/X
Y
W
S
z
/ LS .
We often suppress the subscripts and write L to mean LS or LX . For brevity,
we write F (n) to mean F ⊗OX L⊗n
X if F is a coherent sheaf or a complex of such
on X. In our primary application, L will in fact by the standard very ample line
bundle O(1) on projective space, and so this notation is reasonable in this case.
The case of primary interest for us is described in the following example:
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
5
Example 2.3. Suppose k is a field, Q is a smooth k-algebra of dimension n, and
f1 , . . . , fc is regular sequence of elements of Q. Let
X = Pc−1
= Proj Q[T1 , . . . , Tc ],
Q
S = Pc−1
= Proj k[T1 , . . . , Tc ]
k
and define
p:X →S
to be the map induced by k ֒→ Q. Finally, set LS := OS (1) and
X
fi Ti ∈ Γ(X, LX ).
W :=
i
Then the triple (p : X → S, L, W ) satisfies all the hypotheses in Assumptions
2.1, and moreover p is affine. The subscheme cut out by W is
Y = Proj Q[T1 , . . . , Tc ]/W
in this case.
2.1. The Jacobian complex. We construct a complex of locally free sheaves on X
that arises from the data (p : X → S, LS , W ). Regarding W as a map W : X → LS
as above, we obtain an induced map
W ∗ Ω1LS /S → Ω1X/S
on cotangent bundles. There is a canonical isomorphism Ω1LS /S ∼
= q ∗ L−1
S and hence
−1
∗ 1
∼
an isomorphism L−1
→
Ω1X/S , and
.
Using
this,
we
obtain
a
map
L
W
Ω
=
X
X
LS /S
upon tensoring with L we arrive at the map:
dW : OX → Ω1X/S ⊗OX LX =: Ω1X/S (1).
Note that
Λ·X/S (Ω1X/S ⊗OX L) ∼
=
M
ΩqX/S (q)
q
so that we may regard the right-hand side as a sheaf of graded-commutative gradedOX -algebras. Since dW is a global section of this sheaf lying in degree one, we may
form the Jacobian complex with differential given as repeated multiplication by dW :
dW
dW
dW
dW
(2.4)
Ω·dW := OX −−→ Ω1X/S (1) −−→ Ω2X/S (2) −−→ · · · −−→ ΩnX/S (n) .
We index this complex so that ΩjX/S (j) lies in cohomological degree j.
Example 2.5. Let us consider the case where X and S are affine and the line
bundle L is the trivial one. That is, suppose k is a Noetherian ring, A is a smooth
k-algebra, and f ∈ A, and set S = Spec(k), X = Spec(A), L = OS , and W = f ∈
Γ(X, OX ) = A. (The map p : X → S is given by the structural map k → A.)
Then the Jacobian complex is formed by repeated multiplication by the degree
one element df ∈ Ω1A/k :
df ∧−
df ∧−
df ∧−
df ∧−
Ω·df = A −−−→ Ω1A/k −−−→ Ω2A/k −−−→ · · · −−−→ ΩnA/k .
If we further specialize to the case A = k[x1 , . . . , xn ], then we may identify Ω1A/k
with An by using the basis dx1 , . . . , dxn . Then Ω·df is the A-linear dual of the
∂f
∂f
Koszul complex on the sequence ∂x
, . . . , ∂x
of partial derivatives of f .
1
n
6
MARK E. WALKER
∂f
∂f
, . . . , ∂x
form an
If k is a field and f has only isolated singularities, then ∂x
1
n
A-regular sequence. In this case, the Jacobian complex has cohomology only in
degree n and
k[x1 , . . . , xn ]
H n (Ω·df ) ∼
dx1 ∧ · · · ∧ dxn .
= ∂f
∂f
i
h ∂x1 , . . . , ∂x
n
Example 2.6. With the notation of Example 2.3, the Jacobian complex is the
complex of coherent sheaves on Pc−1
associated to the following complex of graded
Q
modules over the graded ring Q[T ] := Q[T1 , . . . , Tc ]:
dW
dW
dW
0 → Q[T ] −−→ Ω1Q/k [T ](1) −−→ · · · −−→ ΩnQ/k [T ](n) → 0
Here, dW = df1 T1 +· · ·+dfc T1 , a degree one element in the graded module Ω1Q/k [T ].
P ∂f
Further specializing to Q = k[x1 , . . . , xn ], we have dfi = j ∂x
dxj and
j
∂f
∂fc
1
· · · ∂x
T1
∂x1
1
∂f1
∂f
∂x2 · · · ∂xc2
T2
·
.
dW =
.. ..
..
.
.
.
∂f1
∂xn
···
∂fc
∂xn
Tc
It is convenient to think of Ω·dW as representing a “family” of complexes of the
type occurring in Example 2.5 indexed by the Q points of Pc−1
Q . That is, given
a tuple a = (a1 , . . . , ac ) of elements of Q that generated the unit ideal, we have
an associated Q-point ia : Spec(Q) ֒→ Pc−1
of Pc−1
Q
Q . Pulling back along ia (and
∗
identifying ia O(1) with the trivial bundle in the canonical way) gives the Jacobian
P
complex of Example 2.5 for (Spec(Q) → Spec(k), O, i ai fi ).
We are especially interested in the cokernel (up to a twist) of the last map in
the Jacobian complex (2.4), and thus give it its own name:
Definition 2.7. The Jacobi sheaf associated to (p : X → S, L, W ), written J (X/S, L, W )
or just JW for short, is the coherent sheaf on X defined as
dW ∧−
n−1
JW = J (X/S, L, W ) := coker ΩX/S
(−1) −−−−→ ΩnX/S .
In other words, JW := Hn (Ω·dW )(−n), where Hn denote the cohomology of a
complex in the abelian category of coherent sheaves on X.
Example 2.8. If S = Spec(k), k is a field, X = Ank = Spec(k[x1 , . . . , xn ]), L = OX
and f a polynomial, then
k[x1 , . . . , xn ]
dx1 ∧ · · · ∧ dxn .
Jf =
∂f
∂f
,
.
.
.
,
∂x1
∂xn
Example 2.9. With the notation of 2.3, JW is the coherent sheaf associated to
the graded Q[T ]-module
dW
n−1
coker ΩQ/k
[T ](−1) −−→ ΩnQ/k [T ] .
If we further specialize to the case Q = k[x1 , . . . , xn ] = k[x], then using dx1 , . . . , dxn
as a basis of Ω1Q/k gives us that JW is isomorphic to the coherent sheaf associated
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
to
⊕n
coker k[x, T ](−1)
(T1 ,...,Tc )
∂fi
∂xi
7
!
t
−−−−−−−−−−−→ k[x, T ] .
2.2. Matrix factorizations. We recall the theory of “twisted matrix factorizations” from [21] and [3].
Suppose X is any Noetherian scheme, L is a line bundle on X and W ∈ Γ(X, L)
any global section of it. (The base scheme S is not needed for this subsection, and
we write L for LX .) A twisted matrix factorization for (X, L, W ) consists of the
data E = (E0 , E1 , d0 , d1 ) where E0 and E1 are locally free coherent sheaves on X and
d0 : E0 → E1 (1) and d1 : E1 → E0 are morphisms such that d0 ◦ d1 and d1 (1) ◦ d0
are each given by multiplication by W . (Recall that E1 (1) denotes E1 ⊗OX L.) We
visualize E as a diagram of the form
d1
d0
E0 (1)
E1 −→
E = E0 −→
or as the “twisted periodic” sequence
d0 (−1)
d1 (−1)
d
d0 (1)
d
d1 (1)
1
0
E0 (1) −−−→ E1 (1) −−−→ · · · .
E1 −→
· · · −−−−→ E1 (−1) −−−−→ E0 −→
Note that the composition of any two adjacent arrows in this sequence is multiplication by W (and hence this is not a complex unless W = 0).
A strict morphism of matrix factorizations for (X, L, W ), say from E to E′ =
′
(E0 , E1′ , d′0 , d′1 ) consists of a pair of morphisms E0 → E0′ and E1 → E1′ such that the
evident pair of squares both commute. Matrix factorizations and strict morphisms
form an exact category, which we write as mf (X, L, W ), for which a sequence of
morphisms is declared exact if it is so in each degree.
If V ∈ Γ(X, L) is another global section, there is a tensor product pairing
− ⊗mf − : mf (X, L, W ) × mf (X, L, V ) → mf (X, L, W + V ).
In the special case L = OX and V = W = 0, this tensor product is the usual
tensor product of Z/2-graded complexes of OX -modules, and the general case is
the natural “twisted” generalization of this. We refer the reader to [2, §7] for a
more precise definition.
There is also a sort of internal Hom construction. Given W, V ∈ Γ(X, L) and
objects E ∈ mf (X, L, W ) and F ∈ mf (X, L, V ), there is an object Hommf (E, F) ∈
mf (X, L, V − W ). Again, if L = OX and W = V = 0, this is the usual internal
Hom-complex of a pair of Z/2-graded complexes of OX -modules, and we refer the
reader to [2, 2.3] for the general definition. If V = W , then Hommf (E, F) belongs
to mf (X, L, 0) and hence is actually a complex. In general, we refer to an object
of mf (X, L, 0) as a twisted two-periodic complex, since it is a complex of the form
d0 (−1)
d
d
d1 (1)
d0 (1)
0
1
E1 (1) −−−→ E0 (1) −−−→ · · · .
E0 −→
· · · −−−−→ E1 −→
In other words, such an object is equivalent to the data of a (homologically indexed)
complex E· of locally free coherent sheaves and a specified isomorphism E· (1) ∼
= E· [2].
(We use the convention that E· [2]i = Ei−2 .)
There is also a notion of duality. Given E ∈ mf (X, L, W ), we define E∗ ∈
mf (X, L, −W ) as Hommf (E, OX ) where here OX denotes the object of mf (X, L, 0)
given by the twisted two-periodic complex
· · · → 0 → OX → 0 → L → 0 → L⊗2 → · · · .
8
MARK E. WALKER
(In the notation introduced below, E∗ = Hommf (E, Fold OX ).) See [2, 7.5] for a
more explicit formula.
These constructions are related by the natural isomorphisms (cf. [2, §7])
Hommf (E, F) ∼
= E∗ ⊗mf F
and
Hommf (E ⊗mf G, F) ∼
= Hommf (E, Hommf (G, F)).
The construction of Hommf (E, F) is such that we have a natural identification
Homstrict (E, F) = Z0 Γ(X, Hommf (E, F)).
Here, Γ(X, Hommf (E, F)) is the complex of abelian groups obtained by regarding
Hommf (E, F) as an unbounded complex of coherent sheaves on X and applying
the global sections functor Γ(X, −) degree-wise, obtaining a complex of Γ(X, OX )modules, and Z0 denotes the cycles lying in degree 0 in this complex.
Let P(X) denote the category of complexes of locally free coherent sheaves on
X and let P b (X) denote the full subcategory of bounded complexes. There is a
“folding” functor,
Fold : P b (X) → mf (X, L, 0),
defined by taking direct sums of the even and odd terms in a complex. In detail,
given a complex (P, dP ) ∈ P b (X), let
M
M
P2i+1 (i).
P2i (i) and Fold(P )1 =
Fold(P )0 =
i
i
The required map Fold(P )1 → Fold(P )0 maps the summand P2i+1 (i) to the summand P2i (i) via dP (i) and Fold(P )0 → Fold(P )1 (1) maps P2i (i) to P2i−1 (i) also
via dP (i).
There is also an “unfolding” functor
Unfold : mf (X, L, 0) → P(X)
which forgets the two periodicity. In detail,
Unfold(E1 → E0 → E1 (1))2i = E0 (−i) and Unfold(E1 → E0 → E1 (1))2i+1 = E1 (−i)
with the evident maps.
These functors are related by the following adjointness properties: Given E ∈
mf (X, L, 0) and P ∈ P b (X), there are natural isomorphisms
∼ HomP(X) (P, Unfold E)
Hommf (X,L,0) (Fold P, E) =
and
HomP(X) (Unfold E, P) ∼
= Hommf (X,L,0) (E, Fold P).
(These functors do not technically form an adjoint pair, since Unfold takes values
in P(X), not P b (X).)
Example 2.10. If E is a locally free coherent sheaf on X, we write E[0] for the
object in P(X) with E in degree 0 and 0’s elsewhere. Then
Fold(E[0]) = (0 → E → 0)
= (· · · → E(−1) → 0 → E → 0 → E(1) → · · · )
indexed so that E lies in degree 0. Letting E[j] = E[0][j], we have
Fold(E[2i]) = (0 → E(i) → 0)
= (· · · → E(i − 1) → 0 → E(i) → 0 → E(i + 1) → · · · )
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
9
with E(i) in degree 0, and
Fold(E[2i + 1]) = (E(i) → 0 → E(i + 1))
= (· · · → E(i − 1) → 0 → E(i) → 0 → E(i + 1) → · · · )
with E(i) in homological degree 1.
If P ∈ P b (X) is a bounded complex and E ∈ mf (X, L, W ) is a twisted matrix
factorization, we define their tensor product to be
P ⊗ E := Fold(P ) ⊗mf E ∈ mf (X, L, W ).
Example 2.11. We will use this construction, in particular, for the complex P =
Ω1X/S (1)[−1] consisting of Ω1X/S (1) concentrated in homological degree −1. Since
Ω1 lies in odd degree, the sign conventions give us
− id ⊗d1 (1)
− id ⊗d0
Ω1X/S ⊗ E1 (1) −−−−−−−→ Ω1X/S ⊗ E0 (1) .
Ω1X/S (1)[−1] ⊗ E = Ω1X/S ⊗ E0 −−−−−→
dW ∧−
We will also apply this construction to the complex P := OX −−−−→ Ω1X/S (1)
indexed so that OX lies in degree 0 to form the matrix factorization
dW ∧−
(OX −−−−→ Ω1X/S (1)) ⊗OX E
which is given explicitly as
E1
⊕
Ω1X/S ⊗E0
d1
0
dW − id ⊗d0
−−−−−−−−−−−−−→
E0
⊕
Ω1X/S ⊗E1 (1)
d0
0
dW − id ⊗d1
−−−−−−−−−−−−−→
E1 (1)
⊕
Ω1X/S ⊗E0 (1)
.
Building on the identity
Homstrict (E, F) = Z0 Γ(X, Hommf (E, F))
we obtain a natural notion of homotopy: Two strict morphisms are homotopic if
their difference lies in the image of the boundary map in the complex Γ(X, Hommf (E, F)).
That is, the group of equivalence classes of strict morphisms up to homotopy from
E to F is H0 Γ(X, Hommf (F, E)). We define [mf (X, L, W )]naive , the “naive homotopy category”, to be the category with the same objects as mf (X, L, W ) and with
morphisms given by strict maps modulo homotopy; that is,
Hom[mf ]naive (E, F) := H0 Γ(X, Hommf (F, E)).
Just as the homotopy category of chain complexes is a triangulated category, so
too is the category [mf (X, L, W )]naive ; see [21, 1.3] or [3, 2.5] for details.
The reason for the pejorative “naive” in this definition is that this notion of
homotopy equivalence does not globalize well: two morphisms can be locally homotopic without being globally so. Equivalently, an objects can be locally contractible
with being globally so. To rectify this, we define the homotopy category of matrix
factorizations for (X, L, W ), written hmf (X, L, W ), to be the Verdier quotient of
[mf (X, L, W )]naive by the thick subcategory consisting of objects that are locally
contractible. (The notion appears to be originally due to Orlov, but see also [21,
3.13] where this category is referred to as the “derived category of matrix factorizations”.)
We list the properties of hmf (X, L, W ) needed in the rest of this paper:
10
MARK E. WALKER
(1) There is a functor mf (X, L, W ) → hmf (X, L, W ), and it sends objects
that are locally contractible to the trivial object.
(2) If α : E → F is strict morphism and it is locally null-homotopic, then α is
sent to the 0 map in hmf (X, L, W ).
(3) When W = 0, hmf (X, L, 0) is the Verdier quotient of [mf (X, L, 0]naive obtained by inverting quasi-isomorphisms of twisted two-periodic complexes.
(4) At least in certain cases, the hom sets of hmf (X, L, W ) admit a more
explicit description. In general, for objects E, F ∈ mf (X, L, W ) there is a
natural map
Homhmf (E, F) → H0 (X, Unfold Hommf (E, F))
where H0 (X, −) denotes sheaf hyper-cohomology [3, 3.5]. This map is an
isomorphism when
(a) W = 0 or
(b) X is projective over an affine base scheme and L = OX (1) is the
standard line bundle; see [3, 4.2].
2.3. Connections for matrix factorizations and the Atiyah class. See Appendix A for recollections on the notion of a connection for a locally free coherent
sheaf on a scheme. Much of the material in this subsection represents a generalization to the non-affine case of constructions found in [26], which were in turn
inspired by constructions in [9].
Definition 2.12. With (p : X → S, L, W ) as in Assumptions 2.1, given E ∈
mf (X, L, W ), a connection on E relative to p, written as ∇E or just ∇, is a pair of
connections relative to p
∇1 : E1 → Ω1X/S ⊗OX E1
and
∇0 : E0 → Ω1X/S ⊗OX E0 .
There is no condition relating ∇E and d1 , d0 .
Since each ∇i is p∗ OS -linear and L is pulled back from S, we have induced
connections
∇1 (j) : E1 (j) → Ω1X/S ⊗OX E1 (j)
and
∇0 (j) : E0 (j) → Ω1X/S ⊗OX E0 (j)
for all j.
Example 2.13. Suppose S = Spec k and X = Spec Q are affine where Q is a
smooth k-algebra of dimension n, L = OX so that W ∈ Q, and thecomponents of
A
E are free Q-modules. Upon choosing bases, we may represent E as Qr −===
− Qr
B
where A, B are r × r matrices with entries in Q such that AB = BA = W Ir . The
choice of basis leads to an associated “trivial” connection
⊕r
∇ = d : Qr → Ω1Q/k ⊗Q Qr = Ω1Q/k
given by applying exterior differentiation to the components of a vector.
As noted in Appendix A, if p is affine, then every vector bundle on X admits a
connection relative to p and so every matrix factorization admits a connection in
this case.
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
11
Definition 2.14. Given a connection ∇ relative to p for a twisted matrix factorization E ∈ mf (X, L, W ), the associated Atiyah class is defined to be the map
AtE,∇ : E → Ω1X/S (1)[−1] ⊗ E
given as the “commutator” [∇, δE ]. In detail, it is given by the pair of maps
∇1 (1) ◦ δ − (id ⊗δ) ◦ ∇0 : E0 → Ω1 (1) ⊗ E1
and
∇0 ◦ δ − (id ⊗δ) ◦ ∇1 : E1 → Ω1 ⊗ E0 .
AtE,∇ is not a morphism of matrix factorizations in general, but by Lemma A.6
we have that AtE,∇ is OX -linear.
Example 2.15. Keep the notations and assumptions of Example 2.13, so that
A
r
E = Qr −===
−Q .
B
Then
1
Ω (1)[−1] ⊗ E =
1
Ω1Q/k
⊕r
⊕r
−B
1
−==
==− ΩQ/k
−A
and the map At : E → Ω (1)[−1] ⊗ E obtained by choosing the trivial connections
is represented by the following diagram:
A
Qr
dA
⊕r
1
ΩQ/k
/ Qr
B
dB
−B
⊕r
1
/ Ω
Q/k
/ Qr
dA
−A
⊕r
1
/ Ω
Q/k
Here dA and dB denote the r×r matrices with entires in Ω1Q/k obtained by applying
d entry-wise to A and B.
This is not a map of matrix factorizations since the squares do not commute.
Indeed, the differences of the compositions around these squares are (dB)A+B(dA)
and (dA)B + A(dB). Since AB = BA = W , these expression both equal dW , a
fact which is relevant for the next construction.
Definition 2.16. Given a connection ∇ for a matrix factorization E , define
dW ∧−
ΨE,∇ : E → OX −−−−→ Ω1X/S (1) ⊗OX E
dW ∧−
as idE + At∇ . (Here, OX −−−−→ Ω1X/S (1) is the complex with OX in degree 0.) In
other words, ΨE,∇ is the morphism whose composition with the canonical projection
dW ∧−
OX −−−−→ Ω1X/S (1) ⊗OX E −−։ E
is the identity and whose composition with the canonical projection
dW ∧−
OX −−−−→ Ω1X/S (1) ⊗OX E −−։ Ω1 (1)[−1] ⊗OX E
is AtE,∇ .
Lemma 2.17. The map ΨE,∇ is a (strict) morphism of matrix factorizations, and
it is independent in the naive homotopy category [mf (X, L, W )]naive (and hence the
homotopy category too) of the choice of connection ∇.
12
MARK E. WALKER
Proof. The proofs found in [26], which deal with the case where X is affine and L
is trivial, apply nearly verbatim. The homotopy relating ΨE,∇ and ΨE,∇′ for two
different connections ∇ and ∇′ on E is given by the map
∇ − ∇′ : E → Ω1X/S ⊗ E,
which is OX -linear by Lemma A.6.
Example 2.18. Continuing with Examples 2.13 and 2.15, the map Ψ in this case
is represented by the diagram
A
Qr
1
dA
Qr
⊕ r
Ω1Q/k
A
dW
B
/ Qr
0
−B
/
1
dB
/ Qr
Qr
⊕ r
Ω1Q/k
B
dW
0
−A
/
1
dA
Qr
⊕ r
Ω1Q/k .
This diagram commutes, confirming that Ψ is indeed a morphism of matrix factorizations.
2.4. Ad hoc Hochschild homology. For an integer j ≥ 0, we define the twisted
two-periodic complex
(j)
ΩdW :=
(j−2)dW
(j−1)dW
jdW
2dW
dW
−→ ΩjX/S (j)) .
Fold OX −−→ Ω1X/S (1) −−−−−→ Ω2X/S (2) −−−−−→ · · · −−→ Ωj−1
X/S (j − 1) −
(j)
Explicitly, ΩdW is the twisted periodic complex
1
1
ΩX/S
ΩX/S (1)
OX
⊕
⊕
⊕
3
3
2
ΩX/S (1)
ΩX/S (2)
ΩX/S (1)
· · · −→ ⊕ −→ 4 ⊕ −→ ⊕ −→ · · ·
Ω5 (2)
Ω5 (3)
Ω
(2)
X/S
X/S
X/S
⊕
⊕
⊕
.
..
..
.
.
.
.
with
Li=⌊ 2j ⌋
i=0
Ω2i
X/S (i) in degree 0 and
Li=⌊ j−1
2 ⌋
i=0
Ω2i+1
X/S (i) in degree 1.
(j)
The coefficients appearing in the maps of ΩdW are necessarily to make the following statement hold true. (We omit its straight-forward proof.)
Lemma 2.19. The pairings
ΩiX/S (i) ⊗OX ΩlX/S (l) → Ωi+l
X/S (i + l)
defined by exterior product induce a morphism
(j)
(l)
(j+l)
ΩdW ⊗mf ΩdW → ΩdW .
in mf (X, L, 0).
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
13
If j! is invertible in Γ(X, OX ) (for example, if X is a scheme over a field k with
char(k) = 0 or char(k) > j), then there is an isomorphism
dW
dW
dW
(j)
j
ΩdW ∼
= Fold OX −−→ Ω1X/S (1) −−→ · · · −−→ ΩX/S (j))
1
n(n−1)···(n−i)
given by collection of isomorphisms ΩiX/S (i) −−−−−−−−−→ ΩiX/S (i), 0 ≤ i ≤ j. In
particular, if n! is invertible in Γ(X, OX ), then there is an isomorphism
(n)
ΩdW ∼
= Fold Ω·dW .
(Recall that n is the relative dimension of X of S, and hence is the rank of Ω1X/S .
It follows that Ωm
X/S = 0 for m > n.)
Definition 2.20. Let p : X → S, L, W be as in Assumptions 2.1 and assume
also that n! is invertible in Γ(X, OX ). We define the (degree 0) ad hoc Hochschild
homology of mf (X, L, W ) relative to S to be
hoc
(X/S, L, W ) := H0 (X, Unfold Fold Ω·dW ).
HHad
0
hoc
As we will see in the next section, HHad
(X/S, L, W ) is the target of what we
0
term the “ad hoc Chern character” of a twisted matrix factorization belonging to
mf (X, L, W ).
We use the adjective “ad hoc” for two reasons. The first is that we offer here no
justification that this definition deserves to be called “Hochschild homology”. See,
however, [22, 24] for the affine case and [20] for the non-affine case with L = OX .
The second reason is that whereas the support of every object of mf (X, L, W )
is contained in the singular locus of the subscheme Y cut out by W (as will be
justified in the next subsection), the twisted two-periodic complex Fold Ω·dW is not
so supported in general. A more natural definition of Hochshild homology would
thus be given by H0Z (X, Unfold Fold Ω·dW ), where Z is the singular locus of Y and
H0Z refers to hypercomology with supports (i.e., local cohomology). Since this more
sensible definition of Hochschild homology is not necessary for our purposes and
hoc
only adds complications to what we do, we will stick with using HHad
.
0
Recall that the Jacobi sheaf is defined as
dW
n−1
JW = J (X/S, L, W ) = coker ΩX/S
(−1) −−→ ΩnX/S
so that there is a canonical map
Ω·dW → Hn (Ω·dW )[−n] = JW (n)[−n].
From it we obtain the map
Fold(Ω·dW ) → Fold(JW (n)[−n]) ∼
= Fold(JW ( n2 )).
Applying Unfold and using the canonical map Unfold Fold(JW ( n2 )) → JW ( n2 ) results in a map Unfold Fold Ω·dW → JW ( n2 ). Finally, applying H0 (X, −) yields the
map
(2.21)
hoc
HHad
(X/S, L, W ) → H 0 (X, JW ( n2 )).
0
This map will play an important role in the rest of this paper. Let us observe that
in certain situations, it is an isomorphism:
14
MARK E. WALKER
Proposition 2.22. Assume S = Spec k for a field k, X = Spec(Q) for a smooth
k-algebra Q, and hence LX = OX and W ∈ Q. If the morphism of smooth varieties
W : X → A1 has only isolated critical points and n is even, then the canonical map
HHad
0
hoc
(X/S, W ) → H 0 (X, JW ( n2 )) =
ΩnQ/k
n−1
dW ∧ ΩQ/k
is an isomorphism.
Proof. These conditions ensure that the complex of Q-modules Ω·dW is exact except
on the far right where it has homology J (n). The result follows since X is affine.
Example 2.23. Assume S = Spec(k) for a field k, X = Ank = Spec k[x1 , . . . , xn ],
and the morphism f : Ank → A11 associated to a given polynomial f ∈ k[x1 , . . . , xn ]
has only isolated critical points. Then
hoc
HHad
(X/S, f ) ∼
= H 0 (X, Jf ( n2 )) =
0
ΩnQ/k
df ∧
n−1
ΩQ/k
=
k[x1 , . . . , xn ]
dx1 ∧ · · · ∧ dxn .
∂f
∂f
( ∂x
, . . . , ∂x
)
1
n
2.5. Supports. We fix some notation. For a Noetherian scheme Y , the non-regular
locus of Y is
Nonreg(Y ) := {y ∈ Y | the local ring OY,y is not a regular local ring}.
Under mild additional hypotheses (e.g., Y is excellent), Nonreg(Y ) is a closed subset
of Y .
Assume g : Y → S is a morphism of finite type. Recall that g is smooth near
y ∈ Y if and only if it is flat of relative dimension n near y and the stalk of Ω1Y /S
at y is a free OY,y -module of rank n. We define singular locus of g : Y → S to be
the subset
Sing(g) = Sing(Y /S) = {y ∈ Y | g is not smooth near y}.
At least for a flat morphism g : Y → S of finite type, the singular locus of g is a
closed subset of Y by the Jacobi criterion. When S = Spec(k) for a field k and
there is no danger of confusion, we write Sing(Y ) instead of Sing(Y / Spec(k)). If
Y is finite type over a field k, then Nonreg(Y ) ⊆ Sing(Y ) = Sing(Y / Spec(k)), and
equality holds if k is perfect.
Proposition 2.24. Assume X is a Noetherian scheme, L is a line bundle on X,
and W is a regular global section of L. Let Y ⊆ X be the closed subscheme cut
out by W . Then for every E, F ∈ mf (X, L, W ), the twisted two-periodic complex
Hommf (E, F) is supported Nonreg(Y ).
Proof. For x ∈ X, we have
Hommf (E, F)x ∼
= Hommf (Ex , Fx )
where the hom complex on the right is for the category mf (Spec(OX,x ), Lx , Wx ).
By choosing a trivialization OX ∼
= Lx , we may identify this category with mf (Q, f )
for a regular local ring Q and non-zero-divisor f . The assertion thus becomes that
if f is non-zero-divisor in regular local ring Q and Q/f is also regular, then the
complex Hommf (E, F) is acyclic for all objects E, F ∈ mf (Q, f ). This holds since the
∗
d Q/f (coker(E), coker(F));
cohomology modules of the complex Hommf (E, F) are Ext
see Theorem 3.2 below.
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
15
Example 2.25. Suppose k is a field, Q is a smooth k-algebra, m is a maximal
ideal of Q, and f1 , . . . , fc ∈ m form a regular sequence such that the non-regular
locus of the ring R := Q/(f1 , . . . , fc ) is {m}. Let X = Pc−1
= Proj Q[T1 , . . . , Tc ],
Q
P
L = OX (1), and W = i fi Ti . Then
c−1
Nonreg(Y ) = Pc−1
Q/m ⊆ PQ .
2.6. Chern classes of matrix factorizations. If E ∈ mf (X, L, W ) is a matrix
factorization that admits a connection (for example, if p is affine), we define strict
morphisms
(j)
(j)
ΨE,∇ : E → ΩdW ⊗ E, for j ≥ 0,
recursively, by letting
(1)
Ψ(0) = idE , ΨE,∇ = ΨE,∇ ,
(j)
and, for j ≥ 2, defining ΨE,∇ as the composition of
(j−1)
idΩ(1) ⊗ΨE,∇
(1)
(j−1)
E −−−−−−−−−→ ΩdW ⊗ ΩdW
∧⊗id
(j)
E
⊗ E −−−−→
ΩdW ⊗ E.
(j)
By Lemma 2.17, ΨE,∇ is independent up to homotopy of the choice of ∇.
Example 2.26. Continuing with Examples 2.13, 2.15 and 2.18, the map Ψ(∧j)
arising from the trivial connection is in degree zero given by
j factors
}|
{
z
(1 + dA) · · · (1 + dA)(1 + dB)
for j even and by
j factors
}|
{
z
(1 + dB) · · · (1 + dA)(1 + dB)
for j odd, where the product is taken in the ring Matr×r (Ω·Q/k ).
By the duality enjoyed by the category mf (X, L, Q), we obtain a morphism of
objects of mf (X, L, 0) of the form
(j)
(j)
Ψ̃E : Endmf (E) = E∗ ⊗mf E → ΩdW .
Moreover, as a morphism in the category [mf (X, L, 0)]naive , it is independent of
the choice of connection. Take j = n and assume n! is invertible, so that we have
an isomorphism
(n)
ΩdW ∼
= Fold Ω·dW .
We obtain the morphism
Endmf (E) → Fold Ω·dW
in mf (X, L, 0) and hence a map of unbounded complexes
(2.27)
chE : Unfold Endmf (E) → Unfold Fold Ω·dW .
id
E
The identity morphism on E determines a morphism OX −−→
Unfold Endmf (E),
0
that is, an element of H (X, Unfold Endmf (E)).
Definition 2.28. Under Assumptions 2.1 and 2.2, define the ad hoc Chern character of E ∈ mf (X, L, W ) relative to p : X → S to be
hoc
ch(E) = chE (idE ) ∈ HHad
(X/S, L, W ).
0
16
MARK E. WALKER
hoc
Recall that when n is even, we have the map HHad
(X/S, L, W ) → H 0 (X, JW ( n2 ))
0
defined in (2.21), and hence we obtain an invariant of E in H 0 (X, JW ( n2 )). Since
this invariant will be of crucial importance in the rest of this paper, we give it a
name:
Definition 2.29. Under Assumptions 2.1 and 2.2 and with n even, define the ad
hoc top Chern class of E ∈ mf (X, L, W ) relative to p : X → S to be the element
ctop (E) ∈ H 0 (X, JW ( n2 )) given as the image of ch(E) under (2.21).
Example 2.30. With the notation and assumptions of Example 2.13, we have
dW
−−→ Ω2i+1
M ker Ω2i
Q/k
Q/k
hoc
hoc
.
HHad
(X/S, L, W ) = HHad
(Spec(Q)/ Spec(k), O, W ) =
0
0
2i−1 dW
2i
i im ΩQ/k −−→ ΩQ/k
and, when n is even, H 0 (X, JW ( n2 )) is the last summand:
H 0 (X, JW ( n2 )) =
ΩnQ/k
n−1
dW ∧ ΩQ/k
.
We have
n factors
}|
{
z
2
tr(dAdB · · · dB).
n!
Here dA and dB are r × r matrices with entries in Ω1 − Q/k and the product is
occurring in the ring Matr×r (Ω·Q/k ), and the map tr : Matr×r (ΩnQ/k ) → ΩnQ/k is
the usual trace map.
Our formula for ctop (E) coincides with the “Kapustin-Li” formula [14] found in
many other places, at least up to a sign: In the work of Polishchuck-Vaintrob [22,
n
Cor 3.2.4], for example, there is an additional a factor of (−1)( 2 ) in their formula
for the Chern character.
ctop (E) =
Since the top Chern class of E ∈ mf (X, L, W ) is defined as the image of idE
under the composition
ch
E
Unfold Endmf (E) −−→
Unfold Fold Ω·dW → JW ( n2 )
of morphisms of complexes of coherent sheaves on X, the following result is an
immediate consequence of Proposition 2.24. Recall that for a quasi-coherent sheaf
F on X and a closed subset Z of X, HZ0 (X, F ) denotes the kernel of H 0 (X, F ) →
H 0 (X \ Z, F ).
Proposition 2.31. For any E ∈ mf (X, L, W ), its top Chern class is supported on
Nonreg(Y ):
0
n
ctop (E) ∈ HNonreg(Y
) (X, JW ( 2 )).
Remark 2.32. It is also follows from what we have established so far that the Chern
character of any E is supported on Nonreg(Y ) in the sense that it lifts canonically
to an element of
H0Nonreg(Y ) (X, Unfold Fold Ω·dW ),
where in general H0Z denote local hyper-cohomology of a complex of coherent
sheaves.
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
17
2.7. Functorality of the Chern character and the top Chern class. The
goal of this subsection is to prove Chern character is functorial in S. Throughout,
we suppose Assumptions 2.1 and 2.2 hold.
Let i : S ′ → S be a morphism of Noetherian schemes and write L′ , X ′ , W ′ for
the evident pull-backs: X ′ = X ×S S ′ , L′S = i∗ (LS ), and W ′ = i∗ (W ). The typical
application will occur when i is the inclusion of a closed point of S.
There is a functor
(2.33)
i∗ : mf (X, L, W ) → mf (X ′ , L′ , W ′ )
induced by pullback along the map X ′ → X induced from i. There is a canonical
isomorphism i∗ Ω·dW ∼
= Ω·dW ′ and hence an hence an induced map
hoc
hoc
i∗ : HHad
(X/S, L, W ) → HHad
(X ′ /S ′ , L′ , W ′ ).
0
0
Proposition 2.34. For E ∈ mf (X, L, W ), we have
hoc
i∗ (ch(E)) = ch(i∗ (E)) ∈ HHad
(X ′ /S ′ , L′ , W ′ ).
0
Proof. Recall that the Chern character of E is determined by a map of matrix
factorizations
E → E ⊗ Ω·dW
which is itself determined by a choice of connection Ej → Ej ⊗OX Ω1X/S for j = 0, 1.
Since our connections are OS -linear, we have the pull-back connection
i∗ Ej → i∗ (Ej ) ⊗OX ′ Ω1X ′ /S ′ , j = 0, 1.
which we use to define the map
i∗ E → i∗ E ⊗ Ω·dW ′ .
Using also that i∗ End(E) = End(i∗ E), it follows that the square
End(E)
/ Fold Ω·
dW
Ri∗ End(i∗ E)
/ Ri∗ Fold Ω· ′
dW
commutes. Applying H0 (X, Unfold(−)) gives the commutative square
H0 (X, Unfold End(E))
hoc
/ HHad
(X/S, L, W )
0
i∗
H0 (X ′ , Unfold End(i∗ E))
i∗
hoc
/ HHad
(X ′ /S ′ , L′ , W ′ ).
0
The result follows from the fact that i∗ : H0 (X, Unfold End(E)) → H0 (X ′ , Unfold End(i∗ E))
sends idE to idi∗ E .
Recall from (2.21) that, when n is even, there is a natural map
hoc
HHad
(X/S, L, W ) → Γ(X, JX/S,L,W ( n2 ))
0
Since i∗ JX/S,L,W ∼
= JX ′ /S ′ ,L′ ,W ′ , we obtain a map
i∗ : Γ(X, JX/S,L,W ( n2 )) → Γ(X ′ , JX ′ /S ′ ,L′ ,W ′ ( n2 )).
18
MARK E. WALKER
Corollary 2.35. When n is even, for any E ∈ mf (X, L, W ), we have
i∗ (ctop (E)) = ctop (i∗ (E)) ∈ Γ(X, JX ′ /S ′ ,L′ ,W ′ ( n2 )).
Proof. This follows from Proposition 2.34 and the fact that
hoc
HHad
(X/S, L, W )
0
/ Γ(X, JX/S,L,W ( n2 ))
i∗
i∗
/ Γ(X ′ , JX ′ /S ′ ,L′ ,W ′ ( n2 ))
hoc
HHad
(X ′ /S ′ , L′ , W ′ )
0
commutes.
3. Relationship with affine complete intersections
We present some previously known results that relate twisted matrix factorizations with the singularity category of affine complete intersections. We start with
some basic background.
For any Noetherian scheme X, we write Db (X) for the bounded derived category
of coherent sheaves on X and Perf(X) for the full subcategory of perfect complexes
— i.e., those complexes of coherent sheaves on X that are locally quasi-isomorphic
to bounded complexes of finitely generated free modules. We define Dsg (X) to be
the Verdier quotient Db (X)/ Perf(X). For a Noetherian ring R, we write Dsg (R)
for Dsg (Spec(R)).
The category Dsg (X) is triangulated. For a pair of finitely generated R-modules
M and N , we define their stable Ext-modules as
i
d (M, N ) := HomD (R) (M, N [i]),
Ext
R
sg
where on the right we are interpreting M and N [i] as determining complexes (and
hence objects of Dsg (R)) consisting of one non-zero component, lying in degrees 0
and −i, respectively. Note that since Db (R) → Dsg (R) is a triangulated functor by
construction and the usual ExtR -modules are given by HomDb (R) (M, N [i]), there
is a canonical map
i
d R (M, N )
ExtiR (M, N ) → Ext
that is natural in both variables.
Proposition 3.1. [1, 1.3] For all finitely generated R-modules M and N , the natural map
i
d (M, N )
Exti (M, N ) → Ext
R
R
is an isomorphism for i ≫ 0.
3.1. Euler characteristics and the Herbrand difference for hypersurfaces.
Throughout this subsection, assume Q is a regular ring and f ∈ Q is a non-zeroα
divisor, and let R = Q/f . Given an object E = (E1 −==
− E0 ) ∈ mf (Q, f ), define
β
α
→ E0 ). The Q-module coker(E) is annihilated by f and
coker(E) to be coker(E1 −
we will regard it as an R-module.
Theorem 3.2 (Buchweitz). [1] If R = Q/f where Q is a regular ring and f is a
non-zero-divisor, there is an equivalence of triangulated categories
∼
=
hmf (Q, f ) −
→ Dsg (R)
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
19
induced by sending a matrix factorization E to coker(E), regarded as an object of
Dsg (R).
In particular, we have an isomorphisms
i
and
d R (coker(E), coker(F)), for all i,
Extihmf (E, F) := Homhmf (E, F[i]) ∼
= Ext
Extihmf (E, F) ∼
= ExtiR (coker(E), coker(F)), for i ≫ 0.
Since the category hmf (Qp , f ) is trivial for all f ∈ p such that Rp is regular,
Homihmf (E, F) is an R-module supported on Nonreg(R), for all i. If we assume
Nonreg(R) is a finite set of maximal ideals, then Homihmf (E, F) has finite length as
an R-module, allowing us to make the following definition.
Definition 3.3. Assume Q is a regular ring and f ∈ Q is a non-zero-divisor such
that Nonreg(R) is a finite set of maximal ideals, where R := Q/f . For E, F ∈
mf (Q, f ), define the Euler characteristic of (E, F) to be
χ(E, E) = length Hom0hmf (E, F) − length Hom1hmf (E, F).
In light of Theorem 3.2, if M and N are the cokernels of E and F, then
χ(E, F) = hR (M, N ),
(3.4)
where the right-hand side is the Herbrand difference of the R-modules M and N ,
defined as
0
1
d R (M, N ) − length Ext
d R (M, N )
hR (M, N ) = length Ext
2i+1
(M, N ), i ≫ 0.
= length Ext2i
R (M, N ) − length ExtR
Lemma 3.5. Suppose φ : Q → Q′ is a flat ring homomorphism between regular
rings, f is a non-zero-divisor in Q, and f ′ := φ(f ) is a non-zero-divisor of Q′ .
Assume Nonreg(Q/f ) = {m} and Nonreg(Q′ /f ′ ) = {m′ } for maximal ideals m, m′
satisfying the condition that mQ′ is m′ -primary.
Then, for all E, F ∈ mf (Q, f ),
χmf (Q,f ) (E, F) = λ · χmf (′ Q,f ′ ) (E ⊗Q Q′ , F ⊗Q Q′ ),
where λ := lengthQ′ (Q′ /mQ′ ).
Proof. Let R = Q/f and R′ = Q′ /f ′ . The induced map R → R′ is flat and hence
for any pair of R-modules M and N , we have an isomorphism
ExtiR (M, N ) ⊗R R′ ∼
= ExtiR′ (M ⊗R R′ , N ⊗R R′ )
of R′ -modules. For i ≫ 0, ExtiR (M, N ) is supported on {m} and ExtiR′ (M ⊗R
R′ , N ⊗R R′ ) is supported on {m′ }. It follows that
lengthR′ ExtiR′ (M ⊗R R′ , N ⊗R R′ ) = lengthR′ (ExtiR (M, N )⊗R R′ ) = λ lengthR ExtiR (M, N ).
′
Hence hR (M ′ , N ′ ) = λhR (M, N ) and the result follows from (3.4).
20
MARK E. WALKER
3.2. Matrix factorizations for complete intersections. Using a Theorem of
Orlov [19, 2.1], one may generalize Theorem 3.2 to complete intersection rings. The
precise statement is the next Theorem. Versions of it are found in [21], [15], [18],
[23]; the one given here is from [2, 2.11].
Theorem 3.6. Assume Q is a regular ring of finite Krull dimension and f1 , . . . , fc
is a regular sequence of elements of Q. Let R = Q/(f1 , . . . , fc ), define X = Pc−1
=
Q
P
Proj Q[T1 , . . . , Tc ] and set W = i fi Ti ∈ Γ(X, O(1)). There is an equivalence of
triangulated categories
∼
=
Dsg (R) −
→ hmf (X, O(1), W ).
The isomorphism of Theorem 3.6 has a certain naturality property that we need.
Suppose a1 , . . . , ac is a sequence of element of Q that generate the unit ideal, and
let i : Spec(Q) ֒→ X be the associated closed immersion. Then we have a functor
X
ai fi ).
i∗ : hmf (X, O(1), W ) → hmf (Q,
i
∗
∗
Technically, the right hand side should
P be hmf (Spec(Q), i O(1), i (W )), but it is
canonically isomorphic to hmf (Q, i aP
i fi ) because there is a canonical isomorphism i∗ O(1) ∼
i∗ (W ) to i ai fi .
= Q that sends P
Also, the quotient map Q/( i ai fi ) −−։ R has finite projective dimension, since,
locally on Q, it is given by modding out by a regular sequence of length c − 1. We
thus an induced functor
X
ai f i )
res : Dsg (R) → Dsg (Q/
i
on singularity categories, given by restriction of scalars.
Proposition 3.7. With the notation above, the square
∼
=
Dsg (R)
/ hmf (Pc−1 , O(1), W )
Q
res
Dsg (Q/
P
i
i∗
∼
=
ai f i )
/ hmf (Q,
commutes up to natural isomorphism.
P
i
ai f i )
Proof. To prove this, we need to describe the equivalence of Theorem 3.6 explicitly.
Let Y ⊆ Pc−1
denote the closed subscheme cut out by W . Then there are a pair of
Q
equivalences
(3.8)
∼
=
∼
=
Dsg (R) −
→ Dsg (Y ) ←
− hmf (X, O(1), W ),
∼
=
which give the equivalence Dsg (R) −
→ hmf (X, O(1), W ) of the Theorem.
The left-hand equivalence of (3.8) was established by Orlov [19, 2.1] (see also
[2, A.4]), and it is induced by the functor sending a bounded complex of finitely
generated R-modules M · to β∗ π ∗ (M · ), where π : Pc−1
→ Spec(R) is the evident
R
projection and β : Pc−1
֒→
Y
is
the
evident
closed
immersion.
R
The right-hand equivalence of (3.8) is given by [3, 6.3], and it sends a twisted
α
β
→ E1 (1)) to coker(α) (which may be regarded
→ E0 −
matrix factorization E = (E1 −
as a coherent sheaf on Y .)
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
For brevity, set g =
the diagram
(3.9)
P
i
ai fi . The Proposition will follow once we establish that
β∗ ◦π ∗
∼
=
/ Dsg (Y ) o
s
ss
ss
k∗
s
ss ∗
yss Lj
Dsg (Q/g) o
∼
Dsg (R)
21
∼
=
hmf (Pc−1
Q , O(1), W )
i∗
hmf (Q, g)
=
commutes up to natural isomorphism, where k : Spec(R) ֒→ Spec(Q/g) is the
canonical closed immersion and j : Spec(Q/g) → Y is the restriction of i. (The
closed immersion j is locally a complete intersection and thus of finite flat dimension. It follows that it induces a morphism on singularity categories, which we write
Lj ∗ .)
The right-hand square of (3.9) commutes since for a matrix factorization E =
(E1 → E0 → E1 (1)) we have
j ∗ coker(E1 → E0 ) ∼
= coker(i∗ E1 → i∗ E0 )
and
j ∗ coker(E1 → E0 ) ∼
= Lj ∗ coker(E1 → E0 ).
The first isomorphism is evident. The second holds since j has finite flat dimension
and coker(E1 → E0 ) is an “infinite syzygy”; that is, coker(E1 → E0 ) has a right
resolution by locally free sheaves on Y , namely
γ ∗ E1 (1) → γ ∗ E0 (1) → γ ∗ E1 (2) → · · · ,
where γ : Y ֒→ X is the canonical closed immersion.
To show the left-hand triangle commutes, we consider the Cartesian square of
closed immersions
c−1 /
PR
O
β
j
iR
O
Spec(R) /
/Y
O
k
O
/ Spec(Q/g),
where iR denote the restriction of i to Spec(R). This square is Tor-independent
Y
(i.e., TorO
i (β∗ OPc−1 , j∗ OSpec(Q/g) ) = 0 for i > 0), from which it follows that
R
Lj ∗ ◦ β∗ = k∗ ◦ Li∗R .
Since Li∗R ◦ π ∗ is the identity map, the left-hand triangle of (3.9) commutes.
4. Some needed results on the geometry of complete intersections
Let k be a field and Q a smooth k-algebra of dimension n. Set V = Spec(Q), a
smooth affine k-variety. Given f1 , . . . , fc ∈ Q, we have the associated morphism
f := (f1 , . . . , fc ) : V → Ack
of smooth k-varieties, induced by the k-algebra map k[t1 , . . . , tc ] → Q sending ti to
fi . Define the Jacobian of f , written Jf , to be the map
Jf : Qc → Ω1Q/k
22
MARK E. WALKER
given by (df1 , . . . , dfc ). (More formally, Jf is the map f ∗ Ω1Ac /k → Ω1V /k induced by
k
f , but we identify f ∗ Ω1Ac /k with Qc by using the basis dt1 , . . . , dtc of Ω1Ac /k . Also,
k
k
Jf is really the dual of what is often called the Jacobian of f .)
For example, if Q = k[x1 , . . . , xn ], then Jf is given by the n× c matrix (∂fi /∂xj ),
using the basis dx1 , . . . , dxn of Ω1Q/k .
For a point x ∈ V , let κ(x) denote its residue field and define
Jf (x) = Jf ⊗Q κ(x) : κ(x)c → Ω1Q/k ⊗Q κ(x) ∼
= κ(x)n
to be the map on finite dimensional κ(x)-vector spaces induced by Jf . Define
(4.1)
Vj := {x ∈ V | rank Jf (κ(x)) ≤ j} ⊆ V,
so that we have a filtration
∅ = V−1 ⊆ V0 ⊆ · · · ⊆ Vc = V.
Note that the set Vc−1 is the singular locus of the map f .
For example, if Q = k[x1 , . . . , xn ], then Vj is defined by the vanishing of the
(j + 1) × (j + 1) minors of the matrix (∂fi /∂xj ).
Example 4.2. Let k be a field of characteristic not equal to 2, let Q = k[x1 , . . . , xn ]
so that V = Ank , let c = 2, and define f1 = 21 (x21 + · · · + x2n ) and f2 = 12 (a1 x21 +
· · · + an x2n ), where a1 , . . . , an ∈ k are such that ai 6= aj for all i 6= j. Thus f is
a morphism between two affine spaces: f : Ank → A2k . For the usual basis, the
Jacobian is the matrix
x1 a1 x1
x2 a2 x2
Jf = .
.. .
..
.
xn
an xn
Then V0 is just the origin in V = Ank and V1 is the union of the coordinate axes.
In the previous example, we have dim(Vj ) ≤ j for all j < c. This will turn out
to be a necessary assumption for many of our result.
Let us summarize the notations and assumptions we will need in much of the
rest of this paper:
Assumptions 4.3. Unless otherwise indicated, from now on we assume:
(1) k is field.
(2) V = Spec(Q) is smooth affine k-variety of dimension n.
(3) f1 , . . . , fc ∈ Q is a regular sequence of elements. We write
f : V → Ack
for the flat morphism given by (f1 , . . . , fc ). (Note that c ≤ n.)
(4) The singular locus of f −1 (0) = Spec(Q/(f1 , . . . , fc )), is zero-dimensional
— say Sing(f −1 (0)) = {v1 , . . . , vm } where the vi ’s are closed points of V .
(5) dim(Vj ) ≤ j for all j < c, where Vj is defined in (4.1).
Observe that the singular locus of f −1 (0) may be identified with Vc−1 ∩ f −1 (0).
For the results in this paper, we will be allowed to shrink about the singular
locus of f −1 (0). In this situation, if char(k) = 0, the last assumption in above list
is unnecessary, as we now prove:
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
23
Lemma 4.4. If all of Assumptions 4.3 hold except possibly (5) and char(k) = 0,
then for a suitably small affine open neighborhood V ′ of the set {v1 , . . . , vm }, we
have dim(Vj′ ) ≤ i for all j ≤ c − 1.
Proof. Since the induced map Sing(f ) → Ac has finite type, the set
B := {y ∈ Sing(f ) | Sing(f ) ∩ dim(f −1 (f (y)) > 0}
is a closed subset of Sing(f ) by the upper-semi-continuity of fiber dimensions
[EGAIV, 13.1.3]. Since Sing(f −1 (0)) = f −1 (0) ∩ Sing(f ) is a finite set of closed
points, it follows that {v1 , . . . , vm } ∩ B = ∅. Thus V ′ := V \ B is an open neigh′
is quasi-finite over Ac .
borhood of the vi ’s and Sing(f |V ′ ) = Sing(f ) ∩ V ′ = Vc−1
′
Shrinking V further, we may assume it is affine.
′
′ ) is
→ f (Vc−1
By [11, III.10.6], we have dim(f (Vj′ )) ≤ j for all j. Since Vc−1
quasi-finite, so is Vj′ → f (Vj′ ) for all j ≤ c − 1, and hence dim(Vj′ ) ≤ i for all
j ≤ c − 1.
Example 4.5. Suppose char(k) = p > 0, Q = k[x1 , . . . , xc ], and let f1 (x) =
xp1 , · · · , fc (x) = xpc . Then Jf : Qc → Ω1Q/k ∼
= Qc is the zero map, and yet
p
p
k[x1 , . . . , xc ]/(x1 , . . . , xc ) has an isolated singularity at (x1 , . . . , xc ). This shows
that the characteristic 0 hypothesis is necessary in Lemma 4.4.
Given the set-up in Assumptions 4.3, we let
S = Pc−1
= Proj k[T1 , . . . , Tc ]
k
and
X = Pc−1
×k V = Proj Q[T1 , . . . , Tc ],
k
and we define
W = f1 T1 + · · · + fc Tc ∈ Γ(X, O(1)).
As before, we have a map
dW : OX → Ω1X/S (1),
or, equivalently, a global section dW ∈ Γ(X, Ω1X/S (1)). Recall that Ω1S/X (1) is the
coherent sheaf associated to the graded module Ω1Q/k [T1 , . . . , Tc ](1), and dW may
be given explicitly as the degree one element
dW = df1 T1 + · · · + dfc Tc ∈ Ω1Q/k [T1 , . . . , Tc ],
or, in other words,
(4.6)
T1
T2
dW = Jf · . .
..
Tc
Proposition 4.7. With Assumptions 4.3, dW is a regular section of Ω1X/S (1).
Proof. Since X = Pc−1
×k V is smooth of dimension n + c − 1 and Ω1X/S is locally
k
free of rank n, it suffices to prove the subscheme Z cut out by dW has dimension
at most c − 1. Using (4.6), we see that the fiber of Z → V over a point x ∈ V is a
linear subscheme of Pc−1
κ(x) of dimension c − 1 − rank(Jf (x)). That is, if x ∈ Vj \ Vj−1 ,
then the fiber over Z → Y over y has dimension c − 1 − j. (This includes the case
j = c in the sense that the fibers over Vc \ Vc−1 are empty.) Since dim(Vj ) ≤ j if
j < c by assumption, it follows that dim(Z) ≤ c − 1.
24
MARK E. WALKER
Corollary 4.8. If Assumptions 4.3 hold, then the Jacobian complex Ω·dW is exact
everywhere except in the right-most position, and hence determines a resolution of
JW (n) by locally free coherent sheaves.
We will need the following result in the next section.
Proposition 4.9. Under Assumptions 4.3, there is an open dense subset U of
Pc−1
such that for every k-rational point [a1 : · · · : ac ] ∈ U , the singular locus of
k
the morphism
X
ai fi : V → A1k
i
has dimension 0.
Proof. Let Z be the closed subvariety of Pc−1
×k V defined by the vanishing of
k
dW (see (4.6)). The fiber of the projection map π : Z → Pc−1
over a k-rational
k
point
[a
:
·
·
·
:
a
]
may
be
identified
with
the
singular
locus
of
the morphism
1
c
P
1
a
f
:
V
→
A
.
The
claim
is
thus
that
there
is
a
open
dense
subset
U of Pc−1
over
i
i
k
i
k
which the fibers of π have dimension zero. Such a U exists because dim(Z) ≤ c − 1,
as we established in the proof of Proposition 4.7. (U may be taken to be the
complement of π(B) where B is the closed subset {z ∈ Z | dim π −1 π(z) ≥ 1} of
Z.)
5. The vanishing of Chern characters
In this section, we prove that the top Chern class of a twisted matrix factorization
vanishes in certain situations. We combine this with a Theorem of PolishchukVaintrob to establish the vanishing of hc and ηc under suitable hypotheses.
5.1. The vanishing of the top Chern class. The following theorem forms the
key technical result of this paper.
Theorem 5.1. If Assumptions 4.3 hold and in addition
Pn is even, char(p) ∤ n!,
c−1
and c ≥ 2, then ctop (E) = 0 for any E ∈ mf (PQ
, O(1), i fi Ti ).
Proof. The key point is that, under Assumptions 4.3 with c ≥ 2, we have
(5.2)
HP0c−1 ×{v1 ,...,vm } (Pc−1
×k V, JW (i)) = 0
k
for all i < n.
To prove this, since
HP0c−1 ×{v1 ,...,vm } (Pc−1 ×k V, JW (i)) =
m
M
HP0c−1 ×{vi } (Pc−1 ×k V, JW (i))
i=1
and since we may replace V = Spec(Q) by any open neighborhood of vi without
affecting the Q-module HP0c−1 ×{vi } (Pc−1 ×k V, JW (i)), we may assume m = 1 and
that there exists a regular sequence of elements x1 , . . . , xn ∈ m that generate the
maximal ideal m of Q corresponding to v = v1 .
Write C := C(x1 , . . . , xn ) for the “augmented Cech complex”
M 1
M 1
1
Q
→
→ ··· → Q
,
Q
Q→
xi
xi xj
x1 · · · xn
i,j
i
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
25
with Q in cohomological
degree 0. Equivalently C = C(x1 ) ⊗Q · · · ⊗Q C(xn ), where
h i
c−1
1
, we have
C(xi ) = (Q ֒→ Q xi ). For any coherent sheaf F on PQ
HP0c−1 ×{v} (Pc−1 ×k V, F ) = Γ(Pc−1 ×k V, H0 (F ⊗Q C)).
(5.3)
The complex C is exact in all degrees except in degree n, and we set E = H n (C).
Explicitly,
i
h
1
Q x1 ···x
n
h
i.
E=
1
Σi Q x1 ···xˆi ···xn
(The localization of E at m is an injective hull of the residue field, but we do not
∼
need this fact.) Thus we have a quasi-isomorphism C −
→ E[−n].
By Corollary 4.8, there is a quasi-isomorphism
∼
Ω·dW (−n)[n] −
→ JW .
Since C is a complex of flat modules, the map
∼
→ JW ⊗Q C
Ω·dW (−n)[n] ⊗Q C −
∼
is also a quasi-isomorphism. Combining this with the quasi-isomorphism C −
→
E[−n] gives an isomorphism in the derived category
Ω· (−n) ⊗Q E ∼
= JW ⊗ C.
dW
For any i we obtain an isomorphism
H0 (JW (i) ⊗ C) ∼
= H0 (Ω·dW (i − n) ⊗Q E)
of quasi-coherent sheaves.
We thus obtain from (5.3) the isomorphism
Pc−1 ×k V, JW (i) ∼
(5.4)
Γ c−1
= Γ Pc−1 ×k V, H0 (Ω·
Pk
×{v}
k
dW (i
k
But H0 (Ω·dW (i − n) ⊗Q E) is the kernel of
− n) ⊗Q E)
dW ∧−
OPc−1 (i − n) ⊗Q E −−−−→ Ω1Pc−1 /Pc−1 (i − n + 1) ⊗Q E.
Q
Q
k
Since c − 1 ≥ 1, the coherent sheaf OPc−1 (i − n) has no global section when i < n,
Q
and hence
Γ Pc−1
×k V, H0 (Ω·dW (i − n) ⊗Q E) = 0,
k
which establishes (5.2).
P
Let Y be the closed subscheme of X = Pc−1
× V cut out by W =
i fi Ti .
k
c−1
We claim Sing(Y
/k)
⊆
P
×
{v
,
.
.
.
,
v
}.
Indeed,
Sing(Y
/k)
is
defined
by
the
1
m
k
P
1
d(f
)T
=
0
and
f
=
0,
i
=
1,
.
.
.
c,
where
d(f
)
∈
Ω
.
The
equations
i i
i
i
i
Q/k
first equation cuts out a subvariety contained in Pc−1
×
V
,
where,
recall,
Vc−1
k
c−1
k
denotes the singular locus of f : V → Ack . The equations fi = 0, i = 1, . . . c, cut out
Pc−1
×k f −1 (0). But Vc−1 ∩ f −1 (0) = Sing(f −1 (0)) = {v1 , . . . , vm }.
k
Since Nonreg(Y ) ⊆ Sing(Y /k), we deduce from (5.2) that
c−1
0
×k V, JW (i)) = 0.
HNonreg(Y
) (Pk
The vanishing of ctop (E) now follows from Proposition 2.31.
Corollary 5.5. If
• k is a field of characteristic 0,
26
MARK E. WALKER
• Q is a smooth k-algebra of even dimension,
• f1 , . . . , fc ∈ Q forms a regular sequence of elements with c ≥ 2, and
• the singular locus of Q/(f1 , . . . , fc ), is zero-dimensional,
P
then ctop (E) = 0 for all E ∈ mf (Pc−1
Q , O(1),
i fi Ti ).
Proof. Let {v1 , . . . , vm } be the singular locus of Q/(f1 , . . . , fc ) and let V ′ = Spec(Q′ )
be an affine open neighborhood of it such that dim(Vi′ ) ≤ i for all i ≤ c − 1. The
existence of such a V ′ is given by Lemma 4.4. The result follows from the (proof
of the) Theorem, since the map
X
X
fi Ti )
fi Ti ) → HP0c−1 ×{v ,...,v } (Pc−1
HP0c−1 ×{v ,...,v } (Pc−1
Q′ , O,
Q , O,
k
1
m
i
k
1
m
i
is an isomorphism.
5.2. The Polishchuk-Vaintrob Riemann-Roch Theorem. We recall a Theorem of Polishchuk-Vaintrob that relates the Euler characteristic of (affine) hypersurfaces with isolated singularities to Chern characters. This beautiful theorem
should be regarded as a form of a “Riemann-Roch” theorem. Since PolishchukVaintrob work in a somewhat different setting that we do, we begin by recalling
their notion of a Chern character.
For a field k and integer n, define Q̂ := k[[x1 , . . . , xn ]], a power series ring in
n variables, and let f ∈ Q̂ be such that the only non-regular point of Q̂/f is its
maximal ideal. Define Ω̂·Q̂/k to be the exterior algebra on the free Q̂-module
Ω̂1Q̂/k = Q̂dx1 ⊕ · · · ⊕ Q̂dxn .
In other words,
Ω̂·Q̂/k = Ω·k[x1 ,...,xn ]/k ⊗k[x1 ,...,xn ] Q̂.
Finally, define
Note that
n−1 df ∧−
Jˆ(Q̂/k, f ) = coker Ω̂Q̂/k
−−−→ Ω̂nQ̂/k .
k[[x1 , . . . , xn ]]
dx1 ∧ . . . ∧ dxn .
Jˆ(Q̂/k, f ) ∼
= ∂f
∂f
h ∂x1 , . . . , ∂x
i
n
Definition 5.6. With the notation above, given a matrix factorization E ∈ mf (Q̂, f ),
A
upon choosing bases, it may be written as E = (Q̂r −===
− Q̂r ) for r × r matrices
B
A, B with entries in Q̂. When n is even, we define the Polishchuk-Vaintrob Chern
character of E to be
n factors
}|
{
z
n
chPV (E) = (−1)( 2 ) tr dA ∧ dB ∧ · · · ∧ dB ∈ Jˆ(Q̂/k, f ).
Theorem 5.7 (Polishchuk-Vaintrob). Assume k is a field of characteristic 0 and
f ∈ Q̂ = k[[x1 , . . . , xn ]] is a power series such that R := Q̂/f has in an isolated
singularity (i.e., Rp is regular for all p 6= m).
If n is odd, then χ(E, F) = 0 for all E, F ∈ mf (Q̂, f ).
If n is even, there is a pairing h−, −i on Jˆ(Q̂, f ), namely the residue pairing,
such that for all E, F ∈ mf (Q̂, f ), we have
χ(E, F) = hchPV (E), chPV (F)i.
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
27
In particular, χ(E, F) = 0 if either chPV (E) = 0 or chPV (F) = 0.
It is easy to deduce from their Theorem the following slight generalization of it,
which we will need to prove Theorem 5.11 below.
Corollary 5.8. Assume k is a field of characteristic 0, Q is a smooth k-algebra of
dimension n, f ∈ Q is a non-zero-divisor such that the singular locus of R := Q/f
is zero-dimensional. Given E ∈ mf (Q, f ), if either n is odd or n is even and
n−1
ctop (E) = 0 in J (Q/k, f ) = ΩnQ/k /df ∧ ΩQ/k
, then
χ(E, F) = 0
for all F ∈ mf (Q, f ).
Proof. We start be reducing to the case where R has just one singular point m and
it is k-rational. We have
X
χmf (Q,f ) (E, F) =
χmf (Qm ,f ) (Em , Fm )
m
where the sum ranges over the singular points of R. If k is an algebraic closure of
k, then for each such m we have
∼ Hom∗
Hom∗hmf (Qm ,f ) (Em , Fm ) ⊗k k =
(Em ⊗k k, Fm ⊗k k)
hmf (Qm ⊗k k,f ⊗1)
and hence
χmf (Qm ⊗k k,f ⊗1) (Em ⊗k k, Fm ⊗k k) = [Q/m : k] · χmf (Qm ,f ) (Em , Fm ).
It thus suffices to prove the Corollary for a suitably small affine open neighborhood
of each maximal ideal m of Q ⊗k k at which Q ⊗k k/f is singular.
In this case, a choice of a regular sequence of generators x1 , . . . , xn of the
maximal ideal of Qm allows us to identify the completion of Q along m with
Q̂ = k[[x1 , . . . , xn ]]. By Lemma 3.5,
χmf (Q,f ) (E, F) = χmf (Q̂,f ) (Ê, F̂)
where Ê = E⊗Q Q̂ and F̂ = F⊗Q Q̂. Moreover, under the evident map J (Qm /k, f ) →
Jˆ(Q̂/k, f ), ctop (E) is sent to a non-zero multiple of chPV (Ê) by Example 2.30. The
result now follows from the Polishchuk-Vaintrob Theorem.
5.3. Vanishing of the higher h and η Invariants. We apply our vanishing
result for ctop to prove that the invariants hc and ηc vanish for codimension c ≥ 2
isolated singularities in characteristic 0, under some mild additional hypotheses.
First, we use the following lemma to prove the two invariants are closely related.
I thank Hailong Dao for providing me with its proof.
Lemma 5.9 (Dao). Let R = Q/(f1 , . . . , fc ) for a regular ring Q and a regular
sequence of elements f1 , . . . , fc ∈ Q with c ≥ 1. Assume M and N are finitely geni
erated R-modules and that M is MCM. If TorR
i (M, N ) and ExtR (HomR (M, R), N )
have finite length for i ≫ 0, then
ηcR (M, N ) = (−1)c hR
c (HomR (M, R), N ).
In particular, if the non-regular locus of R is zero dimensional and hc vanishes
for all pairs of finitely generated R-modules, then so does ηc .
Remark 5.10. The conditions that TorR
i (M, N ) has finite length for i ≫ 0 and that
ExtiR (HomR (M, R), N ) has finite length for i ≫ 0 are equivalent; see [13, 4.2].
28
MARK E. WALKER
Proof. We proceed by induction on c. The case c = 1 follows from the isomorphism
[6, 4.2]
i
R
d R (HomR (M, R), N ) ∼
d −i−1 (M, N ).
Ext
= Tor
Let S = Q/(f1 , . . . , fc−1 ) so that R = S/(fc ) with fc a non-zero-divisor of S.
By [5, 4.3] and [4, 3.4], TorSi (M, N ) and ExtiS (HomR (M, R), N ) have finite length
for i ≫ 0 and
1 S
ηcR (M, N ) = ηc−1
(M, N )
2c
1 S
hR
h (HomR (M, R), N ).
c (HomR (M, R), N ) =
2c c−1
Let S n −−։ M be a surjection of S-modules with kernel M ′ so that we have the
short exact sequence
0 → M ′ → S n → M → 0.
Observe that M ′ is an MCM S-module. Since fc ·M = 0 and fc is a non-zero-divisor
∼
=
of S, we have HomS (M, S n ) = 0. Using also the isomorphism HomR (M, R) −
→
1
ExtS (M, S) of S-modules coming from the long exact sequence of Ext modules
fc
associated to the sequence 0 → S −→ S → R → 0, we obtain the short exact
sequence
0 → S n → HomS (M ′ , S) → HomR (M, R) → 0.
In particular, it follows from these short exact sequences that TorSi (M ′ , N ) and
ExtiS (HomS (M ′ , S), N ) have finite length for i ≫ 0. Moreover, since hSc−1 vanishes
on free S-modules and is additive for short exact sequences, we have
hSc−1 (HomS (M ′ , S), N ) = hSc−1 (HomR (M, R), N ).
Using the induction hypotheses, we get
1
1 S
S
h (HomS (M ′ , S), N ) = (−1)c−1 ηc−1
(M ′ , N ).
hR
c (HomR (M, R), N ) =
2c c−1
2c
S
S
Finally, ηc−1
(M ′ , N ) = −ηc−1
(M, N ), since M ′ is a first syzygy of M . Combining
these equations gives
ηcR (M, N ) = (−1)c hR
c−1 (HomR (M, R), N ).
The final assertion holds since ηc is completely determined by its values on MCM
modules.
Theorem 5.11. Assume
• k is field of characteristic 0,
• Q is a smooth k-algebra,
• f1 , . . . , fc ∈ Q form a regular sequence of elements,
• the singular locus of R := Q/(f1 , . . . , fc ) is zero-dimensional, and
• c ≥ 2.
R
Then hR
c (M, N ) = 0 and ηc (M, N ) = 0 for all finitely generated R-modules M and
N.
Remark 5.12. Since k is a perfect field, the hypotheses that Q is a smooth k
algebra is equivalent to Q being a finitely generated and regular k-algebra. Likewise,
Sing(R/k) = Nonreg(R).
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
29
Proof. By Lemma 5.9, it suffices to prove hR
c (M, N ) = 0. The value of this invariant
is unchanged if we semi-localize at Nonreg(R). So, upon replacing Spec(Q) with
a suitably small affine open neighborhood of Nonreg(R), Lemma 4.4 allows us to
assume all of Assumptions 4.3 hold.
Let EM , EN ∈ mf (Pc−1
Q , O(1), W ) be twisted matrix factorizations corresponding to the classes of M, N in Dsg (R) under the equivalence of Theorem 3.6. Now
choose a1 , . . . , ac ∈ k as in Proposition
P 4.9 so that the singular locus of Q/g is zero
dimensional, where we define g := i ai fi .
A key fact we use is that hR
c is related to the classical Herbrand difference of the
hypersurface Q/g: By [4, 3.4], for all finitely generated R-modules M and N , we
have
1
Q/g
hR
(M, N ) = hQ/g (M, N ),
c (M, N ) = h1
2
and thus it suffices to prove hQ/g (M, N ) = 0. By Proposition 3.7 the affine matrix
factorizations i∗ EM , i∗ EN ∈ hmf (Q, g) represent the classes of M, N in Dsg (Q/g),
where i : Spec(Q) ֒→ Pc−1
is the closed immersion associated to the k-rational
Q
point [a1 : · · · , ac ] of Pc−1
.
It
thus follows from (3.4) that
k
hQ/g (M, N ) = χ(i∗ EM , i∗ EN ).
By Corollary 5.8, it suffices to prove ctop (i∗ EM ) = 0 in J (Qm /k) (when n is even).
But ctop (i∗ EM ) = i∗ ctop (EM ) by Corollary 2.35 and ctop (EM ) = 0 by Corollary
5.5.
Appendix A. Relative Connections
We record here some well known facts concerning connections for locally free
coherent sheaves. Throughout, S is a Noetherian, separated scheme and p : X → S
is a smooth morphism; i.e., p is separated, flat and of finite type and Ω1X/S is locally
free.
Definition A.1. For a vector bundle (i.e., locally free coherent sheaf) E on X, a
connection on E relative to p is a map of sheaves of abelian groups
∇ : E → Ω1X/S ⊗OX E
on X satisfying the Leibnitz rule on sections: given an open subset U ⊆ X and
elements f ∈ Γ(U, OX ) and e ∈ Γ(U, E), we have
∇(f · e) = df ⊗ e + f ∇(e) in Γ(U, Ω1X/S ⊗OX E),
where d : OX → Ω1X/S denotes exterior differentiation relative to p.
Note that
imply that ∇ is OS -linear — more precisely, p∗ (∇) :
the hypotheses
p∗ E → p∗ Ω1X/S ⊗OX E is a morphism of quasi-coherent sheaves on S.
A.1. The classical Atiyah class. Let ∆ : X → X ×S X be the diagonal map,
which, since X → S is separated, is a closed immersion, and let I denote the
sheaf of ideals cutting out ∆(X). Since p is smooth, I is locally generated by
a regular sequence. Recall that I/I 2 ∼
= ∆∗ Ω1X/S . Consider the coherent sheaf
P̃X/S := OX×S X /I 2 on X ×S X. Observe that P̃X/S is supported on ∆(X), so
that (πi )∗ P̃X/S is a coherent sheaf on X, for i = 1, 2, where πi : X ×S X → X
denotes projection onto the i-th factor.
30
MARK E. WALKER
The two push-forwards (πi )∗ P̃X/S , i = 1, 2 are canonically isomorphic as sheaves
of abelian groups, but have different structures as OX -modules. We write PX/S = P
for the sheaf of abelian groups (π1 )∗ P̃ = (π2 )∗ P̃ regarded as a OX − OX -bimodule
where the left OX -module structure is given by identifying it with (π1 )∗ P̃X/S the
right OX -module structure is given by identifying it with (π2 )∗ P̃X/S .
Locally on an affine open subset U = Spec(Q) of X lying over an affine open
−·−
subset V = Spec(A) of S, we have PU/V = (Q⊗A Q)/I 2 , where I = ker(Q⊗A Q −−−→
Q) and the left and right Q-module structures are given in the obvious way.
There is an isomorphism of coherent sheaves on X × X
∼
∆∗ Ω1
= I/I 2
X/S
given locally on generators by dg 7→ g ⊗ 1 − 1 ⊗ g. From this we obtain the short
exact sequence
0 → Ω1X/S → PX/S → OX → 0.
(A.2)
This may be thought of as a sequence of OX − OX -bimodules, but for Ω1X/S and
OX the two structures coincide.
Locally on open subsets U and V as above, we have Ω1Q/A ∼
= I/I 2 , and (A.2)
takes the form
0 → I/I 2 → (Q ⊗A Q)/I 2 → Q → 0.
Viewing (A.2) as either a sequence of left or right modules, it is a split exact
sequence of locally free coherent sheaves on X. For example, a splitting of PX/S →
OX as right modules may be given as follows: Recall that as a right module,
PX/S = (π2 )∗ P and so a map of right modules OX → PX/S is given by a map
π2∗ OX → PX/S . Now, π2∗ OX = OX×S X , and the map we use is the canonical
surjection. We refer to this splitting as the canonical right splitting of (A.2).
Locally on subsets U and V as above, the canonical right splitting of (Q ⊗A
Q)/I 2 −−։ Q is given by q 7→ 1 ⊗ q.
Given a locally free coherent sheaf E on X, we tensor (A.2) on the right by E to
obtain the short exact sequence
π
i
→E →0
→ PX/S ⊗OX E −
0 → Ω1X/S ⊗OX E −
(A.3)
of OX − OX -bimodules. Taking section on affine open subsets U and V as before,
letting E = Γ(U, E), this sequence has the form
0 → Ω1Q/A ⊗Q E → (Q ⊗A E)/I 2 · E → E → 0.
Since (A.2) is split exact as a sequence of right modules and tensor product preserves
split exact sequences, (A.3) is split exact as a sequence of right OX -modules, and
the canonical right splitting of (A.2) determines a canonical right splitting of (A.3),
which we write as
can : E → PX/S ⊗OX E.
The map can is given locally on sections by e 7→ 1 ⊗ e.
In general, (A.3) need not split as a sequence of left modules. Viewed as a
sequence of left modules, (A.3) determines an element of
Ext1 (E, Ω1 ⊗O E) ∼
= H 1 (X, Ω1 ⊗O EndO (E)),
OX
X/S
X
X/S
X
X
sometimes called the “Atiyah class” of E relative to p. To distinguish this class
from what we have called the Atiyah class of a matrix factorization in the body of
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
31
this paper, we will call this class the classical Atiyah class of the vector bundle E,
and we write it as
(E) ∈ Ext1OX (E, Ω1X/S ).
Atclassical
X/S
The sequence (A.3) splits as a sequence of left modules if and only if Atclassical
= 0.
X/S
Lemma A.4. If p : X → S is affine, then the classical Atiyah class of any vector
bundle E on X vanishes, and hence (A.3) splits as a sequence of left modules.
Proof. Since p is affine, p∗ is exact. Applying p∗ to (A.3) results in a sequence
of OS − OS bimodules (which are quasi-coherent for both actions). But since
p ◦ π1 = p ◦ π2 these two actions coincide. Moreover, since (A.3) splits as right
modules, so does its push-forward along p∗ .
It thus suffices to prove the following general fact: If
F := (0 → F ′ → F → F ′′ → 0)
is a short exact sequence of vector bundles on X such that p∗ (F ) splits as a sequence of quasi-coherent sheaves on S, then F splits. To prove this, observe that F
determines a class in H 1 (X, HomOX (F ′′ , F ′ )) and it is split if and only if this class
vanishes. We may identify H 1 (X, HomOX (F ′′ , F ′ )) with H 1 (S, p∗ HomOX (F ′′ , F ′ ))
since p is affine. Moreover, the class of F ∈ H 1 (S, p∗ HomOX (F ′′ , F ′ )) is the image
of the class of p∗ (F ) ∈ H 1 (S, HomOS (p∗ F ′′ , p∗ F ′ )) under the map induced by the
canonical map
HomOS (p∗ F ′′ , p∗ F ′ ) → p∗ HomOX (F ′′ , F ′ ).
But by our assumption the class of p∗ (F ) vanishes since p∗ F splits.
A.2. The vanishing of the classical Atiyah class and connections. Suppose
σ : E → PX/S ⊗OX E is a splitting of the map π in (A.3) as left modules and recall
can : E → PX/S ⊗OX E is the splitting of π as a morphism of right modules given
locally by e 7→ 1 ⊗ e. Since σ and can are splittings of the same map regarded as
a map of sheaves of abelian groups, the difference σ − can factors as i ◦ ∇σ for a
unique map of sheaves of abelian groups
∇σ : E → Ω1X/S ⊗OX P.
Lemma A.5. The map ∇σ is a connection on E relative to p.
Proof. The property of being a connection may be verified locally, in which case
the result is well known.
In more detail, restricting to an affine open U = Spec(Q) of X lying over an
affine open V = Spec(A) of S, we assume E is a projective Q-module and that we
are given a splitting σ of the map of left Q-modules (Q ⊗A E)/I 2 · E −−։ E. The
map ∇σ = (σ − can) lands in I/I 2 ⊗Q Q = Ω1Q/A ⊗Q E, and for a ∈ A, e ∈ E we
have
∇σ (ae) = σ(ae) − 1 ⊗ ae
= aσ(e) − 1 ⊗ ae
= aσ(e) − a ⊗ e + a ⊗ e − 1 ⊗ ae
= a(σ(e) − 1 ⊗ e) + (a ⊗ 1 − 1 ⊗ a) ⊗ e
= a∇σ (e) + da ⊗ e,
since da is identified with a ⊗ 1 − 1 ⊗ a under Ω1Q/A ∼
= I/I 2 .
32
MARK E. WALKER
Lemma A.6. Suppose E, E ′ are locally free coherent sheaves on X and ∇, ∇′ are
connections for each relative to p. If g : E → E ′ is a morphisms of coherent sheaves,
then the map
∇′ ◦ g − (id ⊗g) ◦ ∇ : E → Ω1X/S ⊗OX E ′
is a morphism of coherent sheaves.
Proof. Given an open set U and elements f ∈ Γ(U, OX ), e ∈ Γ(U, E), the displayed
map sends f · e ∈ Γ(U, E) to
∇′ (g(f e)) − (id ⊗g)(df ⊗ e − f ∇(e) = ∇′ (f g(e)) − df ⊗ g(e) − f (id ⊗g)(∇(e))
= df ⊗ g(e) + f ∇′ (g(e)) − df ⊗ g(e) − f (id ⊗g)(∇(e))
= f ∇′ (g(e)) − f (id ⊗g)(∇(e))
= f ((∇′ ◦ g − (id ⊗g) ◦ ∇) (e)) .
Proposition A.7. For a vector bundle E on X, the function σ 7→ ∇σ determines a
bijection between the set of splittings of the map π in (A.3) as a map of left modules
and the set of connections on E relative to p. In particular, E admits a connection
relative to p if and only if Atclassical
(E) = 0.
X/S
Proof. From Lemma A.6 with g being the identity map, the difference of two connections on E is OX -linear. By choosing any one splitting σ0 of (A.3) and its
associated connection ∇0 = ∇σ0 , the inverse of
σ 7→ ∇σ
is given by
∇ 7→ ∇ − ∇σ0 + σ0 .
References
1. R. O. Buchweitz, Maximal Cohen-Macaulay modules and Tate cohomology over Gorenstein
rings, unpublished, available at http://hdl.handle.net/1807/16682, 1987.
2. Jesse Burke and Mark E. Walker, Matrix factorizations in higher codimension, ArXiv e-prints
(2012).
3. Jesse Burke and Mark E. Walker, Matrix factorizations over projective schemes, Homology
Homotopy Appl. 14 (2012), no. 2, 37–61. MR 3007084
4. Olgur Celikbas and Hailong Dao, Asymptotic behavior of Ext functors for modules of finite
complete intersection dimension, Math. Z. 269 (2011), no. 3-4, 1005–1020. MR 2860275
5. Hailong Dao, Asymptotic behavior of Tor over complete intersections and applications, ArXiv
e-prints (2007).
6. Hailong Dao, Some observations on local and projective hypersurfaces, Math. Res. Lett. 15
(2008), no. 2, 207–219. MR 2385635 (2009c:13032)
7. Tobias Dyckerhoff, Compact generators in categories of matrix factorizations, Duke Math. J.
159 (2011), no. 2, 223–274. MR 2824483 (2012h:18014)
8. Tobias Dyckerhoff and Daniel Murfet, The Kapustin-Li formula revisited, Adv. Math. 231
(2012), no. 3-4, 1858–1885. MR 2964627
, Pushing forward matrix factorizations, Duke Math. J. 162 (2013), no. 7, 1249–1311.
9.
MR 3079249
10. David Eisenbud, Homological algebra on a complete intersection, with an application to group
representations, Trans. Amer. Math. Soc. 260 (1980), no. 1, 35–64. MR 570778 (82d:13013)
11. Robin Hartshorne, Algebraic geometry, Springer-Verlag, New York, 1977, Graduate Texts in
Mathematics, No. 52. MR 0463157 (57 #3116)
CHERN CHARACTERS FOR TWISTED MATRIX FACTORIZATIONS
33
12. Melvin Hochster, The dimension of an intersection in an ambient hypersurface, Algebraic
geometry (Chicago, Ill., 1980), Lecture Notes in Math., vol. 862, Springer, Berlin, 1981,
pp. 93–106.
13. Craig Huneke and David A. Jorgensen, Symmetry in the vanishing of Ext over Gorenstein
rings, Math. Scand. 93 (2003), no. 2, 161–184. MR 2009580 (2004k:13039)
14. Anton Kapustin and Yi Li, D-branes in Landau-Ginzburg models and algebraic geometry, J.
High Energy Phys. (2003), no. 12, 005, 44 pp. (electronic). MR 2041170 (2005b:81179b)
15. Kevin H. Lin and Daniel Pomerleano, Global matrix factorizations, Math. Res. Lett. 20 (2013),
no. 1, 91–106. MR 3126725
16. W. Frank Moore, Greg Piepmeyer, Sandra Spiroff, and Mark E. Walker, The vanishing of
a higher codimension analogue of Hochster’s theta invariant, Math. Z. 273 (2013), no. 3-4,
907–920. MR 3030683
17. Daniel Murfet, Residues and duality for singularity categories of isolated Gorenstein singularities, Compos. Math. 149 (2013), no. 12, 2071–2100. MR 3143706
18. Dmitri Orlov, Matrix factorizations for nonaffine LG-models, Math. Ann., To appear.
, Triangulated categories of singularities, and equivalences between Landau-Ginzburg
19.
models, Mat. Sb. 197 (2006), no. 12, 117–132. MR 2437083 (2009g:14013)
20. D. Platt, Chern Character for Global Matrix Factorizations, ArXiv e-prints (2012).
21. Alexander Polishchuk and Arkady Vaintrob, Matrix factorizations and singularity categories
for stacks, Ann. Inst. Fourier (Grenoble) 61 (2011), no. 7, 2609–2642. MR 3112502
22.
, Chern characters and Hirzebruch-Riemann-Roch formula for matrix factorizations,
Duke Math. J. 161 (2012), no. 10, 1863–1926. MR 2954619
23. Leonid Positselski, Coherent analogues of matrix factorizations and relative singularity categories, arXiv:1102.0261.
24. A. Preygel, Thom-Sebastiani and Duality for Matrix Factorizations, ArXiv e-prints (2011).
25. Ed Segal, The closed state space of affine Landau-Ginzburg B-models, J. Noncommut. Geom.
7 (2013), no. 3, 857–883. MR 3108698
26. Yu Xuan, Geometric study of the category of matrix factorizations, (2013), unpublished,
available at http://digitalcommons.unl.edu/mathstudent/45.
Department of Mathematics, University of Nebraska, Lincoln, NE 68588
E-mail address: [email protected]
| 0math.AC
|
arXiv:1706.02850v1 [cs.CV] 9 Jun 2017
Weakly supervised training of deep convolutional neural networks
for overhead pedestrian localization in depth fields
Alessandro Corbetta
Department of Applied Physics
Eindhoven University of Technology
Vlado Menkovski
Department of Mathematics and Computer Science
Eindhoven University of Technology
[email protected]
[email protected]
Federico Toschi
Department of Applied Physics and
Department of Mathematics and Computer Science
Eindhoven University of Technology
[email protected]
1. Introduction
Abstract
Depth field data encodes the distance between each
recorded point and the camera plane. This allows for
highly-accurate crowd dynamics analyses in real-world scenarios, i.e. outside of laboratory environments [1]. With
this technology, for the first time, the ability to analyze between several thousands to few millions actual pedestrian
trajectories has been achieved [3, 6, 7]. This enabled new
statistical insights, unbiased by artificial laboratory conditions (e.g. need for participants to wear tracking hats or
vests, and dynamics regulated by the experimenter instructions) [2, 5, 13, 16]. Furthermore, depth measurements do
naturally protect the privacy of the pedestrians, since individuals remain unrecognizable. This is a requirement for
real-life measurements, and a challenge for methods that
use imaging rather than depth field data [8].
Pedestrian positioning from depth field data, acquired
from sensors such as Microsoft Kinects [9], requires addressing two key tasks, namely: background subtraction
and head localization. For a number of common cases
these two tasks have a straight-forward solution. Since
the camera takes a birds eye view, the background can be
simply subtracted by removing all points beyond a depth
threshold. The head localization can be approached similarly, in fact the points closest to the camera are part of the
pedestrian heads. So far, these tasks have been tackled via
hand-crafted approaches that rely on expert-tweaked depthcloud clustering algorithms [13] (CL). These approaches
segment the different objects mainly based on the assumption that a cluster of neighboring points forms a pedestrian.
This assumption typically holds when the dynamics on the
scene involve low pedestrian densities (about 1.5 ped./m2
Overhead depth map measurements capture sufficient
amount of information to enable human experts to track
pedestrians accurately. However, fully automating this process using image analysis algorithms can be challenging.
Even though hand-crafted image analysis algorithms are
successful in many common cases, they fail frequently when
there are complex interactions of multiple objects in the
image. Many of the assumptions underpinning the handcrafted solutions do not hold in these cases and the multitude of exceptions are hard to model precisely. Deep Learning (DL) algorithms, on the other hand, do not require hand
crafted solutions and are the current state-of-the-art in object localization in images. However, they require exceeding amount of annotations to produce successful models. In
the case of object localization these annotations are difficult
and time consuming to produce. In this work we present
an approach for developing pedestrian localization models
using DL algorithms with efficient weak supervision from
an expert. We circumvent the need for annotation of large
corpus of data by annotating only small amount of patches
and relying on synthetic data augmentation as a vehicle for
injecting expert knowledge in the model training. This approach of weak supervision through expert selection of representative patches, suitable transformations and synthetic
data augmentations enables us to successfully develop DL
models for pedestrian localization efficiently.
1
(A)
(B)
The major advantage that DL methods have over ML image
analysis is that they incorporate automatic feature extraction (or representation learning) as part of the model training. However, one of the main disadvantages of these approaches is the difficulty in incorporating expert knowledge
about the problem and hence the requirement of significant
amount of annotations [15].
Therefore, to develop DL models for pedestrian localization, an expert needs to produce a large amount of handannotated images. These images and their annotations can
then be used to train a DL model such as a Deep Convolution Neural Network (CNN) to produce the target model.
As the number of annotations can be quite high, this becomes very labor intensive and diminishes the advantages
of using CNNs.
In this work we address this problem by proposing a
method for efficient collection of expert annotations for
pedestrian tracking using depth field images.
Our contribution is twofold:
Figure 1. (A) Example scenario of under-performing localization
employing a clustering (CL) approach. The comparison with the
ground truth in (B) shows some typical mistakes, e.g. the inability
to disentangle close subjects, especially when they are small in
size (as infants, cf. top and left of the scene). Furthermore, typical
CL approaches cannot distinguish shapes, thus the localization at
the bottom includes a static obstacle.
max [7]), assorted homogeneously (i.e. composed of adults
of similar size, with no elements such as strollers, carts, and
bikes) and developing in simple geometric settings (corridors in [7, 13, 16]).
However, such designed approaches do not generalize
well and are sensitive to a number of special case scenarios.
This can be as simple as a raised hand that is interpreted as
a head, but it becomes a larger issue when pedestrian density increases and CL algorithms experience difficulties in
disentangling the individuals. Furthermore, these problems
only increase when there are other (i.e. non-pedestrian) objects in the scene, such as moving doors, trolleys and obstacles. Figure 1 contains a sample depth-field map picturing 9
pedestrians, 3 of which are infant (notice the smaller size),
and 1 static object. The CL based localization in Fig. 1(A)
shows some typical mistakes (cf. ground truth in Fig. 1(B)):
twice a couple formed by an adult and a child (top and left
side of the scene) is detected as a unique individual. Based
on the shape of the objects, an expert instead recognizes
that the likely scenario involves two individuals standing
close to each other. The CL algorithms also fails to disentangle a pedestrian at the bottom from a static obstacle in
their neighborhood (even though they are not connected).
This follows from the small size of the object and the lack
of ability of the CL approach to classify shapes. Detailing
every exception and crafting appropriate rules to deal with
such complexity becomes difficult and requires significant
effort that does not transfer well across different measurement scenarios and implementations.
On the other hand, Machine Learning (ML) methods for
image analysis rely on training data rather than design of
rules and exceptions. Given sufficient data and good features, these models tend to perform well and are robust to
special cases. The success of these approaches has been particularly demonstrated with recent developments in Deep
Learning (DL). DL models obtain excellent performance
and are currently state-of-the-art in image localization [11].
• we design a CNN model that can detect a large number of objects in a high density scenario suitable for
pedestrian tracking with overhead depth field images;
• we develop a ’soft’ supervision procedure that provides training data for the model by selecting a number of patches in the original data, designing suitable
transformations to the patches and generation of realistic synthetic data for the CNN model.
The obtained model can be used for real time pedestrian
detection in depth field maps, possibly on large areas exploiting Graphical Processing Units (GPU).
This paper is structured as follows: in Sect. 2 we provide some selected background on depth map based crowd
recording setups. On this basis, in Sect. 3 we describe our
synthetic data generation procedure as well as our neural
network and its training method. In Sect. 5, we examine the
detection performance. A final discussion section closes the
paper.
2. Depth-field measurements
Overhead depth-field maps for pedestrian dynamics
analyses A typical depth-field measurement apparatus for
pedestrian dynamics include sensors placed overhead and
aligned with the vertical axis. Bird eye view, in fact, avoid
mutual occlusions thus eases localization tasks. Moreover,
to bypass the limited range of commercial devices such as
Microsoft KinectTM , sensors are arranged in grids to enlarge
the measurement areas. This requires a merge of the individual sensor signal. Different stitching approaches have
been considered: for instance, in [7], depth images are unified into large depth frames to then undergo detection and
tracking algorithms. In [13], the tracking information from
2
different sensors are merged a posteriori. In the next paragraph, we review some elements of the former approach
as it supports the depth maps “combination algebra” we
employ to generate synthetic annotated depth maps (see
Sect. 3.2)
it has detected in the image. The model can also associate
a class to each object. For our application we only provide
detection of the objects, since we do not need to detect different types of objects. The model overlays a grid over the
image, and produces a binary detection decision for each
cell in the grid. It also produces an offset and the width and
height of the bounding box for that object. The bounding
box size is not limited to the cell size. For a S × S regular
grid the model produces:
Depth-field maps combination algebra In [7], depthfield maps from neighboring Kinect sensors are merged enabling to track pedestrians over a relatively large area. This
requires two operations: (1) the depth-field measured from
each sensor is converted from a perspective view into an axonometric view. Overhead axonometric views of pedestrians are translation invariant, which means that a pedestrian
is represented with the same depth patch regardless whether
they are at the center or at the edges of the sensor view (cf.
patches in Fig. 1). (2) The axonometric view enables seamless combination of depth-field maps from neighboring sensors. Given two neighboring sensors returning the depth
maps d1 = d1 (x, y) and d2 = d2 (x, y) we combine them
as
zi = (xi , yi , wi , hi , ni , pi ) with i = 1, . . . , S 2 ,
where xi , yi are the Cartesian coordinates of the bounding
box at the i-th tile, whose width and height are, respectively,
wi and hi . pi denotes the probability that zi is actually a
bounding box. Namely, for ground truth data, pi = 0 whenever zi plays the role of a placeholder, conversely pi = 1
states that the i-th is a non void bounding box. Finally,
ni = 1 − pi and is kept for extensibility to multi-type object
detection.
The network we employ is composed of a first section
aimed at feature extraction. This is followed by a densely
connected section that combines local features into bounding boxes estimates. The input are depth images (thus single
channel) at resolution 160 × 120 (VGA resolution downsampled by a factor 4 in each direction). Feature extraction
occurs through two stacked layer blocks, each of which contains two convolutional layers with small filter size (3 × 3)
and a final max pool layer. This architecture is closely related to the VGG network [14]. The convolutional layers
and pooling layers are followed by fully connected layers
that end with the output layer consisting of linear outputs
for the bounding box parameters and softmax for the detection probabilities. The diagram of the network is given in
Fig. 2.
As non-linear function approximators, DL models are
trained through a non-convex optimization procedure that
minimizes a defined loss function. Due the complex multipart output of our model, we needed to define multi-part
loss function L:
c((d1 , τ1 ), (d2 , τ2 )) = minimum{(d1 ◦ τ1 ), (d2 ◦ τ2 )}. (1)
In words, the depth field maps are first rigidly translated (by
a composition with the motions τ1 and τ2 ), to register with
the relative positions of the sensors in the physical space.
Then the component-wise minimum is extracted. This is
used to retain for each location the value of smallest distance, actually observed from an aerial view. The operation
in Eq. (1) is commutative and associative, so it can be extended to an arbitrary number of cameras (or depth patches).
3. Method
3.1. CNN model for object localization
Neural networks, more particularly CNNs have demonstrated particular success in image analysis [12]. The major advantage is their hierarchical structure that allows the
models to build complex features and form efficient representation of the input data. In this method we aim to leverage this advantage to improve the detection and localization
of pedestrians from depth maps. We expect that efficient
features and sufficient supervision will produce models that
better disentangle multiple nearby objects and make more
accurate distinction between pedestrian and other objects in
the scene.
The architecture of the proposed CNN model is closely
related to the YOLO object localization approach [10]. This
approach offers computationally efficient localization. This
opens the possibility for real-time analysis or large number
of objects, which are advantageous properties for large scale
pedestrian tracking.
The model processes the whole image in a single pass
and produces a set of bounding boxes for each object that
2
L=
S
X
gt 2
gt
λH H̃(zi , zgt
i ) + λL2 pi L̃2 (zi , zi ),
(2)
i=1
Respectively, it holds
• the “gt” denotes a ground truth (synthetic) bounding
box data;
• H̃ is the categorical cross-entropy function restricted
to the (ni , pi ) components of the bounding box vector
zi ;
• L̃2 is the ordinary Euclidean distance among the spatial components (xi , yi , wi , hi ) of the bounding box
3
Inputs
1x120x160
Feature
maps
32x120x160
Feature
maps
32x120x160
Feature
maps
32x60x80
Feature
maps
64x60x80
Feature
maps
64x60x80
Feature
maps
64x30x40
Hidden
units
76800
Hidden
units
512
Outputs
S²x6
x y wh n p
x y wh n p
...
Convolution
3x3 kernel
Act:ReLu
Convolution
3x3 kernel
Act:ReLu
Max-pooling
2x2 kernel
Convolution
3x3 kernel
Act:ReLu
Convolution
3x3 kernel
Act:ReLu
Max-pooling
2x2 kernel
Flatten
x y wh n p
Fully
connected
Act:4xReLu+SoftMax
Fully
connected
Act:ReLu
Figure 2. Diagram of the neural network employed. Kernel sizes and activation functions (Act) are below each layer.
vector. Notably, the function is multiplied by pgt
i ,
which acts as a switch, turning off the loss for the location parameters when there is no object present in the
ground truth;
• λH and λL2 are weighting factors for the linear combination of the two metrics.
3.2. Weak supervision through synthetic data
DL methods rely on training data to develop models. Beyond this and the network architecture, most of the options
for adding expert knowledge are indirect. One common way
to guide the training is to augment or add synthetically generated training data. This allows for adding properties to the
model such as invariances to translation, rotation, mirroring
and skewing. One can even go further and add noise or synthetic generation of data that relies on understanding of the
domain. This way the model is exposed to larger range of
variances from the input space and can achieve better generalization with smaller amount of natural data available. We
use this opportunity to deal with the difficulty in providing
annotations for our problem. We achieve this by selecting
patches from the original data that correspond to pedestrians. The patches provide for both a example of a pedestrian
and an annotation of the bounding box around the pedestrian. We then use these patches and other patches of nonpedestrian objects to build a synthetic image for training.
We further inject expert knowledge about the the real
data by deciding on how the synthetic images are composed
and by applying carefully designed transformations to the
selected patches. In this manner we achieve an outstanding
amount of training data will little effort from the expert.
More specifically the approach relies on a human expert
to identify few hundred bounding boxes among those annotated correctly by a clustering-based algorithm (cf. Sect 1).
Here we employ a random selection of real depth maps from
existing crowd tracking experiments. These are typically result of the combined output of multi-sensor setups. This a
set of “overhead human patches” Pu .
Figure 3. A selection of patches defined by the expert and used for
the synthetic data generation. The last 12 items are depth artifacts
or objects to be ignored by the localization algorithm.
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
avg. precision (A)
CNN 5
CL 5
CNN 10
CL 10
20
40
pedestrians
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
avg. recall (B)
1.0
IoU coefficient (C)
0.8
0.6
0.4
0.2
20
40
pedestrians
0.0
CNN 5 CL 5 CNN 10CL 10
Figure 4. Comparison between CNN and CL localization approaches in case S = Sa = 5 and S = Sa = 10. We include
average precision (A) and average recall (B), both conditioned to
the number of pedestrians observed, and intersection over union
coefficient (C).
Secondly, we ask the human expert to manually extract patches that are not pedestrians. These can be of two
types: 1. objects and architectonic elements (such as bags,
strollers, carts, tables, and doors), Po ; 2. depth artifacts from
sensor errors (noisy “stain-looking” patches in the depth
4
(1)
(A)
CL (B)
CNN (C)
(3)
(A)
CL (B)
CNN (C)
(2)
(A)
CL (B)
Sample Experiment Pedestrians
1
2
3
S=5
S=10
9
18
15
Precision
CL CNN
1
1
0.94
1
0.75 0.82
CNN (C)
Recall
CL
CNN
0.77
1
0.72 0.94
0.60 0.93
Figure 5. Localization algorithms in action in our two experiments. Comparison of synthetic data (A), clustering-based algorithm (B), and
our CNN (C). Case (1) is analogous to Fig. 1. Precison and recall data for the three samples are in the table. A higher recall for the CNN
approach can be observed as side by side subjects and static objects are typically correctly recognized.
field, counting from few pixels to few dozens), Pn .
We combine via Eq. (1) elements randomly extracted
from augmented versions of Pu , Po and Pn , say P̂u , P̂o
and P̂n . We begin with an empty depth map at VGA resolution (native output resolution of single Kinect sensors), on
which we overlay a S×S regular grid (cf. network output in
Sect. 3.1). For each grid tile, we choose, with probability q
whether to place an additional pedestrian patch chosen randomly from P̂u . We assign to the patch centroid a position
on the tile surface with uniform probability. Thus, the total
number of pedestrians in the depth maps follows a binomial
distribution with n = S 2 and probability q. As a second
step, we extract N ∼ Poisson(λ) patches from P̂o ∪ P̂n ,
that we place at random in the final map.
We further apply random transformation to the patches
including: 90o rotations, flipping, pixel removal (i.e. pixels
are replaced with the floor depth) and addition (i.e. pixels are replaced with the median value of the image), and
rigid depth translation. Finally, we add Gaussian noise to
the produced image as another layer of regularization.
tn denote, respectively, false positives and true negatives).
We consider the CNN output zi to hold a bounding box prediction if pi > 0.5. We further score the accuracy of each
true positive computing the intersection over union (IoU) of
predicted and actual bounding box. We keep the number of
pedestrians as a parameter in the analysis, as we expect it to
be a major determinant of performance degradation.
Data and CNN training We employ pedestrian patches,
as in Fig. 3, extracted from past measurement setups in
which the sensor was located approximately 4 meters above
the ground. Similarly to a single Kinect operating in these
conditions, synthetic depth-field maps cover an area of
about1 LM × Lm = 2.9 m × 2.2 m = 6.4 m2 .
During the training phase, we expose the two networks
to synthetic data featuring a cell occupation probability q
respectively of qa = 0.5 and qb = 0.2. Hence, in case
S = Sa , the network can detect up to 25 pedestrians and,
due to the grid constraint, the minimum distance admitted
between the centroids of first neighbors is, in the worst case,
LM /Sa = 0.58 m. During the training the network is exposed to an average of n = 12.5 pedestrians per depth
image (average density: 1.9 ped/m2 ). In case S = Sb
the network can detect up to 100 pedestrians and the minimum distance between the centroid of first neighbors is
0.28 m (this potentially encompasses people walking hand
in hand). During the training the network is exposed to
depth images including an average of n = 20 pedestrians
(average density: 3.1 ped/m2 ). We implemented and trained
our networks through the Keras library [4] with tensorflow
GPU backend. We trained the two networks for a total of
300 epochs, each including 64.000 random training depth
maps and 6.400 random validation depth maps (batch size:
64).
Clustering algorithm and test data The CL algorithm
4. Experiment design
We conduct two experiments aiming at comparing the
performance of our CNN with a CL algorithm. The experiments differ in the grid size adopted for the CNN, respectively S = Sa = 5 and S = Sb = 10. Each of our
CNN undergoes a training phase against synthetic data (cf.
Sect 3.2). Then, we expose both the CNN and the considered CL algorithm to further synthetic data (more general
and not seen during the training) and we measure their performance.
Performance evaluation To grade the performance of
the algorithms we consider their output on a cell basis. We
evaluate the algorithms precision (i.e. tp/(tp + f p), cf. explanation below) and recall (i.e. tp/(tp + f n)). We account
cell measurement as a true positive, tp, if the cell is correctly
predicted to hold the centroid of a bounding box (let f p and
1 Calculated considering the characteristic field of view of Kinect sensors [9] plus the assumption that individuals of height larger or equal than
1.4 m have to be fully resolved, even at the edges of the camera sight cone
5
we compare with is similar to what employed in [7, 13].
First, foreground blobs are randomly sampled. Hence, the
samples undergo a complete-linkage hierarchical clustering. The clustering tree is cut at a cutoff height comparable
with the average human shoulder size. Finally, the cluster
larger than a threshold are retained as pedestrians.
We compare, employing synthetic data, our two CNNs
with one CL algorithm, which has fixed parameters. Synthesized test depth data include between 1 and 20 pedestrians (roughly uniformly distributed), for comparison with
the first CNN, and between 1 and 35 for comparison with
the second CNN.
ples, we developed a procedure producing synthetic training
data as a means for efficient delivery of ’soft’ supervision
from an expert.
References
[1] M. Boltes and A. Seyfried. Collecting pedestrian trajectories.
Neurocomputing, 100:127–133, 2013.
[2] D. Brščić, T. Kanda, T. Ikeda, and T. Miyashita. Person
tracking in large public spaces using 3-d range sensors. IEEE
Trans. Human-Mach. Syst., 43(6):522–534, 2013.
[3] D. Brščić, F. Zanlungo, and T. Kanda. Density and velocity
patterns during one year of pedestrian tracking. Transportation Research Procedia, 2:77 – 86, 2014.
[4] F. Chollet. Keras. https://github.com/fchollet/keras, 2015.
[5] A. Corbetta, L. Bruno, A. Muntean, and F. Toschi. High
statistics measurements of pedestrian dynamics. Transportation Research Procedia, 2:96–104, 2014.
[6] A. Corbetta, C. Lee, R. Benzi, A. Muntean, and F. Toschi.
Fluctuations around mean walking behaviours in diluted
pedestrian flows. Phys. Rev. E, 95:032316, 2017.
[7] A. Corbetta, J. Meeusen, C. Lee, and F. Toschi. Continuous
measurements of real-life bidirectional pedestrian flows on
a wide walkway. In Pedestrian and Evacuation Dynamics
2016, pages 18–24. University of Science and Technology
of China press, 2016.
[8] A. Johansson, D. Helbing, and P. Shukla. Specification of
the social force pedestrian model by evolutionary adjustment to video tracking data. Advances in complex systems,
10(supp02):271–288, 2007.
[9] Microsoft Corp. Kinect for Xbox 360, available online:
http://www.xbox.com/en-us/kinect/, 2011. Redmond, WA,
USA.
[10] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You
only look once: Unified, real-time object detection. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.
[11] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. International Journal of Computer
Vision (IJCV), 115(3):211–252, 2015.
[12] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
et al. Imagenet large scale visual recognition challenge.
International Journal of Computer Vision, 115(3):211–252,
2015.
[13] S. Seer, N. Brändle, and C. Ratti. Kinects and human kinetics: A new approach for studying pedestrian behavior. Transport. Res. C-Emer., 48:212–228, 2014.
[14] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014.
[15] F. Yu, A. Seff, Y. Zhang, S. Song, T. Funkhouser, and
J. Xiao. Lsun: Construction of a large-scale image dataset
using deep learning with humans in the loop. arXiv preprint
arXiv:1506.03365, 2015.
5. Results and discussion
In Fig. 4 we report the results of the experiments. Our
CNN shows a precision equal (case S = 5) or higher (case
S = 10) than the CL approach (Fig. 4(A)), and significantly
better results for recall (Fig. 4(B)). For those cells in which a
bounding box has been correctly localized, we evaluate the
localization accuracy by measuring the IoU coefficient. In
this case, the CNN performance is comparable with the CL
approach. In Fig. 5 we include samples of synthetic depth
field maps from our experiments for visual inspection.
Following our expectations, the CNN delivers higher
localization performance than the CL approach. In fact,
the CNN succeeds in disentangling neighboring pedestrians
and in avoiding non-pedestrian elements (cf. Fig. 5). This
results in the higher recall performance. Notably, we leveraged expert knowledge efficiently to extract the patches
combined in the synthetic depth maps. As the localization quality is determined by the examples rather than handcrafted procedures, we expect the method to generalize and
transfer across different real-life measurement setups. Possibly with the only effort of enriching the patch library with
object characteristic to specific location.
6. Consclusions
We targeted a high performance pedestrian localization
tool for overhead depth field data capable or running in
real-time setting. Depth data is often employed in reallife pedestrian dynamics research, and, so far, hierarchical
clustering-based approaches have been mostly used for localization. Here we aimed at bypassing the typical shortcomings of such approaches, leveraging the generalization
power of deep learning models.
We presented a convolutional neural network approach
showing significantly higher recall performance than clustering methods. This was possible as the network learns
the typical shape of individuals. As such it can disentangle neighboring subjects and diminish false positive outputs
(e.g. objects, depth artifacts). To bypass the difficulties imposed by the need for a large number of annotated exam6
[16] F. Zanlungo, T. Ikeda, and T. Kanda. Potential for the dynamics of pedestrians in a socially interacting group. Phys.
Rev. E, 89:012811, Jan 2014.
7
| 9cs.NE
|
Optimization of OFDM radar waveforms using
genetic algorithms
arXiv:1405.4894v1 [cs.NE] 25 Apr 2014
Gabriel Lellouch and Amit Kumar Mishra
University of Cape Town, South Africa,
[email protected]
[email protected],
Abstract. Keywords: OFDM Radar, Genetic Algorithm, NSGA-II,
PSLR, ISLR, PMEPR
In this paper, we present our investigations on the use of single objective
and multiobjective genetic algorithms based optimisation algorithms to improve
the design of OFDM pulses for radar. We discuss these optimization procedures
in the scope of a waveform design intended for two different radar processing
solutions. Lastly, we show how the encoding solution is suited to permit the
optimizations of waveform for OFDM radar related challenges such as enhanced
detection.
1
Introduction
With the advent of powerful digital hardware, software defined radio and radar
have become an active area of research and development [1]. This in turn has
given rise to many new research directions in the radar community, which were
previously not comprehensible. One such direction is the recently investigated
OFDM radar [3], radars which use OFDM waveforms instead of the classic linear
frequency modulated (LFM) waveform.
OFDM is a special form of multicarrier modulation (MCM), where a single
data stream is transmitted over a number of lower rate subcarriers. Alphabets
such as BPSK, QPSK, etc., are commonly used to code the information. The
resulting complex symbols, also called phase codes (-1,+1 in BPSK) are modulate
the subcarriers [16]. When signal gets to the receiver, a demodulation stage
retrieves the transmitted phase codes and eventually the binary message. In
radar, the priority is to detect the presence of targets and possibly estimate some
of their features through the following measurable quantities: range, Doppler,
azimuth and elevation. It is thus of utmost interest for the radar designer to
understand the possibilities offered by the OFDM structure and optimize it to
suit its needs.
It can be quickly verified that non-coded OFDM pulses will not be suitable
in radar systems that operate with the conventional matched filter processing,
since they will give rise to high sidelobes. Another drawback of non-coded OFDM
pulses is their varying time domain signal. Strong variations are detrimental since
these may saturate the signal and cause distortions. Therefore, the OFDM pulse
needs to be tailored before it becomes a suitable radar waveform.
In this work, we argue that the emerging evolutionary algorithms are particularly adapted to solve the pulse design problem when the pulse is an OFDM
signal. Although a number of techniques have been proposed to mitigate the peak
to mean envelope power ratio (PMEPR) and the peak-to-sidelobe ratio (PSLR)
[14], their flexibility is somewhat limited. For example, the Newman phasing
technique gives very low PMEPR for the single OFDM symbol case, and this
for most numbers of subcarriers N . However, as soon as some subcarriers are
suppressed the PMEPR deteriorates.
It may be remarked here that it is not the scope of this paper to investigate
the use of newer evolutionary techniques, rather we show how we can integrate
some simple and easy-to-implement memetic computing techniques in the design
of waveforms for OFDM radar. We will focus firstly on the genetic algorithm
(GA) optimization technique and then on the multiple objective optimization
genetic algorithm (MOO-GA) based technique. Although the former method
offers a straightforward implementation [17], many implementations exist in the
case of the MOO-GA. In this work, we use the well known non-dominated sorting
genetic algorithm II (NSGA-II). It has proved to be much faster than the earlier
version NSGA while providing diversity in the solutions [4].
There are two major novelties of this work. First of all, the use of OFDM
pulses as a radar signal is in itself a new direction. Secondly, the use of GA based
techniques to design OFDM radar pulses is also novel. The rest of the paper is
organized as follows. In section 2 we present the first step of our waveform design and characterize the impact of the processing solution on this design. The
objective here is to fix some of the OFDM parameters to build the frame of the
waveform. We also stress the interconnection of the different OFDM parameters.
In section 3 we discuss the details of the optimization. We review the various
objective functions, discuss the parameters that we use in our optimization problem and give our motivations to use GA based techniques as compared to other
existing methods. In section 4, we review both of our GA based optimization
methods. We present the technical details of the encoding strategy as well as the
population size. We then present and discuss our results in section 5. We start
with the case of single objective optimization before moving on to the multiple
objective optimization case. In section 6, we propose a case study where the single objective GA is integrated into a waveform design procedure for enhancing
the target detection. We give our conclusions in section 5.
2
Waveform design
In this section, we present the successive steps that form our design strategy for
characterizing the fixed parameters of the pulses. We show how they are inferred
from, on the one hand, the processing and on the other hand, the scenario. This
analysis fixes the frame of our pulses. We will show in the following section how
we optimize the rest of our free parameters to compose pulses with improved
radar features.
2.1
Processing related constraints
In the scope of a pulsed OFDM radar waveform, we proposed in [3] two processing alternatives. The first alternative is based on the combination of matched
filtering and Doppler processing whilst the second alternative transforms the received signal in the frequency domain in the same way as OFDM communication
systems operate. After a demodulation stage which suppresses the phase codes,
two orthogonal DFT processing are applied to form a range Doppler image. The
key characteristics of both processing are recalled in Table 1. In the rest of the
paper we refer to the former as the conventional processing while we name the
latter our frequency domain processing.
Table 1. Conventional processing characteristics
Conventional processing
Immune to intersymbol inPros
terference
Doppler sidelobes are phase
codes independent
Cons Range sidelobes are phase
codes dependent
Frequency domain processing
Range and Doppler sidelobes are
phase codes independent
Subject to intersymbol interference
Because the frequency domain processing is subject to inter-symbol interference, we will use it to track targets. Indeed, in tracking configurations we can
assume to have some prior knowledge of the illuminated scene and in particular
the target extent. The rule of thumb is that the return echo from the closest
point scatterer and the return echo from the furthest point scatterer of the target fall within the same time cell. When satisfied, this condition insures that
the orthogonality between the subcarriers is maintained. This issue has been
of utmost interest in the early years of OFDM signalling for communication to
cope with the multipath effect. To that end the concept of cyclic prefix has been
introduced [16]. In our current analysis, rather than inserting a cyclic prefix, we
choose to match the size of the time cell according to the target extent. Since
the conventional processing does not come up with a severe design constraint
we choose to use the constraint of the frequency domain processing as the main
guideline.
Sampling frequency In our analysis, the received signal which we feed into
either of these processing is the complex signal formed from the real and imaginary components respectively in the I and Q channels of the receiver. If the
transmitted OFDM pulse has a bandwidth B, the received complex signal has
the same bandwidth. Because the signal is complex, the Nyquist theorem states
that the sampling frequency can be taken as low as fs = B and the time cell
size is thus inversely proportional to the bandwidth, ts = 1/B. The size of the
range cell is then given by c/2B, where c is the speed of light.
Bandwidth As a result, if we want to design the size of the time cells such
that, despite the superposition of all echoes returning from the different point
scatterers of the target, the orthogonality property of the subcarriers is maintained, we shall adjust the bandwidth to comply with: c/2B ≥ ∆Rt where ∆Rt
is the target range extent. Practically we can add a margin to account for the
target radial velocity and the uncertainty on the target extent and position. Not
only the received echoes shall remain in the same time cell as a result of the
first pulse but also throughout the coherent processing interval. In the end the
bandwidth can be obtained from:
c
(1)
B=
2(∆R + α)
where α is the margin in range.
Therefore, if there is no need for high range resolution we suggest to base
the bandwidth selection on intersymbol interference mitigation considerations
instead. When the radar is in a tracking mode, the target is known and the use
of high resolution range profiles is a fortiori not necessary.
2.2
Scenarios related constraints
Other parameters need to be fixed. The pulse length and the number of subcarriers that will compose the OFDM signal.
Pulse length Pulse compression waveforms have the unique advantage to offer
low peak power transmissions over a long time to provide the same maximum
detection range and the same range resolution as would be obtained from a short
pulse with a high peak power. But since the radar receiver is switched off when
the pulse is being transmitted we cannot afford to have a very long pulse. We
commonly refer to as the eclipsed zone, the window that lies between the radar
and the minimum detection range. Even though we could choose one value for
each target, for simplicity we choose only one for all. If, say, we expect targets
from Rmin = 1.5 km, an upper bound for the pulse length is found to be [15],
tp = 2 · Rmin /c = 10 µs.
Number of subcarriers The orthogonality property is another example of the
unique OFDM structure. It states that the bit duration tb is inversely proportional to the subcarrier spacing ∆f , tb = 1/∆f . In the extreme case where the
pulse is composed of one symbol the maximum number of subcarriers Nmax used
in the pulse can be derived from:
Nmax =
2BRmin
c
(2)
For the same pulse bandwidth and pulse duration a smaller number of subcarriers
can be used if we construct the pulse from several symbols. For example we can
decide to use 250 subcarriers and have 4 symbols in the pulse to maintain the
same duration. The subcarrier spacing is then increased from 100 kHz to 400
kHz.
In light of the above analysis, the fixed parameters for our pulses are summarized in Table 2.
Table 2. Scenarii characteristics for the waveform design
Case 1 Case 2
(walker) (truck)
Range extent (m)
2
10
Margin (m)
1
5
Bandwidth (MHz)
50
10
500
100
Maximum number of subcarriers
3
Optimizing the pulse for radar
Having fixed the frame of our pulses, literally the bandwidth and the maximum
number of subcarriers, we are now ready to concentrate our analysis on the optimization of the OFDM pulse for radar. Firstly, we need to define our objective
functions and secondly, we need to identify the OFDM parameters that we will
use to run our optimization procedure.
3.1
Objective functions
An OFDM symbol is built as a sum of weighted complex sinusoids, where every
sinusoid has a given starting phase. When the OFDM pulse is composed of
several symbols it can be expressed as:
x(t) =
K
N X
X
wn an,k rk (t) · exp(j2πn∆f t).
(3)
n=1 k=1
N is the number of subcarriers, wn corresponds to the weight applied on subcarrier n and the phase code an,k is attributed to subcarrier n in symbol k. K
is the total number of symbols in the pulse. The function rk (t) refers to the
rectangular window for every symbol:
1 (k − 1)tb ≤ t ≤ ktb
rk (t) =
0 elsewhere
When using OFDM for radar two important aspects must be considered.
Sidelobe level Firstly, would the processing be based on a correlation function like in our conventional processing where matched filtering is applied in
range, the sidelobes at the output shall be maintained as low as possible. This
consideration is true for any signal and a fortiori for our OFDM signal. When
both functions are equal, the correlation function becomes the autocorrelation
function. The output of R(τ ) is then given by:
Z ∞
R(τ ) =
x(t)x∗ (t − τ )dt.
(4)
−∞
In our analysis, we rather use the discrete form R[m]:
R[m] =
NX
K−1
x[p]x∗ [p − m],
(5)
p=0
where m takes integer values between −N K + 1 and N K − 1. x[p] represent the
discrete values of the OFDM pulse taken at the discrete instants ptb /N =, where
p takes integer numbers from 0 to N K − 1. If the pulse is composed of only one
symbol, then p takes values from 0 to N − 1, just like n the subcarrier index.
In the end, x[p] = x(ptb /N ). In Eq. 5, we assume that x[p] = 0 for all forbidden
values of p, that is p < 0 and p > N K − 1.
To cope with practical applications we commonly distinguish two objective
functions. The first function is the peak sidelobe level ratio (PSLR). It returns
the ratio between the highest sidelobe and the peak.
max|R[m]|
P SLR =
m
|R[0]|
, m 6= 0
(6)
The second function is the integrated sidelobe level ratio (ISLR). It returns the
ratio between the cumulation of the sidelobes and the peak.
P
|R[m]|
ISLR = m
, m 6= 0
(7)
|R[0]|
The relative importance between these two figures of merit depends on the application as well as the environment. For example, if the radar operates in the
presence of distributed clutter, it will be important to work with low ISLR in
order to keep the weak targets visible. In that case, high ISLR can be interpreted
as an increase of the noise floor. Conversely if the application requires detection
of targets in the presence of strong discrete clutter, the PSLR is more critical
and must be kept low to prevent from deceptively considering one sidelobe as
another small target.
Peak to mean envelope power ratio This second aspect is specific to OFDM
signals. Its impact on the radar performance is less straightforward, nonetheless
it is as relevant as the previous PSLR and ISLR. It characterizes the variations in
time of the envelope. Too strong variations can be detrimental to the radar since
the signal may saturate. A common expression for this peak-to-mean envelope
power ratio (PMEPR) is given by [7]:
max|x[n]|2
,
P M EP R = 1 nP
|x[n]|2
N
(8)
An important comment to make at this stage concerns the finesse of the
sampling that we consider in each of our objective functions. In Eq. 6, Eq. 7 and
Eq. 8, we have assumed to work at the critical sampling rate fs = B such that
the sampling period is ts = tb /N , as a result of the relationships that govern the
OFDM structure. In Fig. 1 we stress the impact of oversampling. Because of the
quick temporal variations of the OFDM signal the PMEPR will not be the same
at the critical sampling rate than it will be with an oversampling factor of 20. In
[2], we stressed that the autocorrelation function of multicarrier signals presents
strong variations between consecutive critically sampled instants, which is not
the case of phase coded signal with a single carrier. Practically, the relevance to
include or not oversampling in this analysis shall be evaluated with regard to
the implementation strategy in terms of the baseband OFDM signal sampling
frequency. In this work, we thus choose to oversample the OFDM signal before
evaluating our objective functions. In Fig. 1(b), we show how both PSLR and
ISLR calculations exclude the values around the main peak. Also discussed in
[2], the total peak extent is equal to twice the Rayleigh resolution, which in time
is 2/B.
(a) Time domain OFDM pulse
(b) Autocorrelation output
Fig. 1. Pulse amplitude versus time of an uncoded OFDM pulse and autocorrelation
function versus range. N = 3 and K = 3. Both functions are plotted at the critical
sampling rate fs = B and with oversampling fs = 20B.
Optimization strategy Thus, in the process of designing an OFDM pulse,
extra care shall be employed so that the signal results in tolerable PMEPR as
well as reasonable sidelobe level. Note however that in this statement we have
assumed that the processing was based on a correlation function. In our case, this
is valid when we intend to use our conventional processing. However, if rather,
we decide to use our frequency domain processing, we have recalled in Table 1
that both the range and Doppler sidelobes were phase codes independent. In that
case, the optimization can focus on the PMEPR only. In a nutshell, this analysis
indicates that with our conventional processing we are interested in optimizing 3
objectives, while with our frequency domain processing, we would have a single
objective. We thus have to come up with two solutions, a multi-objective solution
and a single objective one.
Parameters for optimization We have seen in Eq.3 that the parameters still
available after our initial design in section 2 were the weights wn and the phase
codes an,k . We decide to leave the weights and focus primarily on the phase
codes. To justify this approach, we point out that the scattering centres of a
target resonate variably at different frequencies, therefore it may be desirable to
leave the weights for an optimization that would account for the target model
as we will see in section 6. For now, we assume equal weights such that the
pulse has normalized energy. The phase codes are also assumed to have constant
modulus. Each an,k is thus a number on the unit complex circle.
Existing optimization methods The problem of minimizing the PSLR and
the PMEPR has been extensively researched since the emergence of the multicarrier concept. Levanon et al. [14] have reported some the major contributions
in that field. Two strategies have popped up. The first strategy assumed identical
sequences, such that all subcarriers are assigned the same phase code. Optimizing
the PMEPR of the pulse results in optimizing the PMEPR of a single symbol.
Newman, Schroeder and Narahashi have suggested different phasing methods
to decrease the PMEPR. In their concept the weights wn are complex values
with quadratic dependence on n. PMEPR as low as 2 (3dB) can be obtained
for any relevant value of N (up to 65,000). The second design strategy is based
on modulating all N subcarriers with consecutive ordered cyclic shifts (COCS)
of an ideal chirplike sequence (CLS) of length K. For example, a OFDM pulse
based on COCS of P4 codes can give PMEPRs below 2 and PSLR below -15 dB
for a large range of N (between 0 and 70).
Motivation for genetic algorithm optimization methods Although the
aforementioned techniques produce excellent results in terms of PMEPR alone or
PMEPR and PSLR combined, we suggest to introduce genetic algorithm based
methods in this field for three main reasons. Firstly, it will diversify the solutions
and increase the potential number of good codes. Not necessarily the best but
those good enough for the application. Secondly, we can tune the objective functions to put emphasis on some criteria. For example, if we wish to have extremely
low sidelobes close to the main peak in the autocorrelation function and can tolerate higher values further away, the objective function can be modified with no
harm. Another example is the case of a banned sub-band. Assume that in the
presence of a jammer we ban the use of one or few subcarriers. The energy on
these subcarriers could be picked up and the presence of our transmission would
be revealed. If we rely on any of the previous strategies, the loss of one or more
subcarriers can potentially destroy the PMEPR, the PSLR or both. Another
interesting example is the use of sparse spectrum OFDM pulses to mitigate the
range ambiguity in SAR OFDM [9]. In both cases, the genetic algorithm based
optimization can come up with sets of phase codes that will improve both the
PMEPR and the PSLR in either of these configurations. Thirdly, we can also
optimize the ISLR which is not in the focus of the other methods. Not only we
can optimize it solely but we can optimize it together with the PSLR and the
PMEPR by means of multi-objective optimization (MOO) techniques.
4
Optimization techniques
In this section, we present the genetic algorithm based techniques that we use
to find our phase codes.
Problem encoding The first step in the implementation of any genetic algorithm is to generate an initial population. Following the canonical genetic
algorithm guideline [17], this implies encoding each element of the population
into a binary string. Note however that techniques based on real numbers have
also been developed [6]. The MOO-GA that we discuss in section 4.1 applies
one of them. For now, we simply encode one phase code (value between 0 and
2π) into a string of q genes. When we are dealing with a pulse composed of N
subcarriers and K symbols, we end up with N K strings of q genes each. Stacking these strings together we create one element of the population, which is then
formed from Q = N Kq genes. This element is called a chromosome. If we require
to use for example binary phase shift keying (BPSK), then q = 1, quadrature
phase shift keying (QPSK), then q = 2. In the more general case where we have
no restriction we can consider the largest value authorized by our system. In our
case we use q = 18. The resolution in angle is then ∆θ = 2π/2q ' 0.024 mrad.
Note that the larger q the slower the algorithm. This is caused by the increase
of the search space as discussed hereafter. With the values of N and K that we
consider in this paper, the chromosome length can be as large as 9000, (N = 500
and K = 1). The search space S ”reduces” to the binary strings of length Q.
Population size To understand what the population size L shall be, we followed
the guideline given in [13]. The starting point is to say that every point in the
search space shall be reachable from the initial population by crossover only.
This can happen only if there is at least one instance of every gene at each locus
in the entire population. On the assumption that every gene is generated with
random probability (P(1)=1/2 and P(0)=1/2) the probability that at least one
gene is present at each locus is given by:
P = (1 − (1/2)L )Q
(9)
With Q = 9000 we see that L ' 23 when we take P = 99.9% and L ' 26 when
we take P = 99.99%.
4.1
Genetic algorithm
The genetic algorithm implemented in this work is a two-stage process. Goldberg
defined this class of genetic algorithms as simple genetic algorithms (SGA) [17].
It starts with the current population. Then selection is applied to form the
intermediate population. Next, recombination and mutation are applied to form
the next population. The process of going from the current population to the next
population represents one generation of the execution of the genetic algorithm.
We mentioned earlier that we make use of this algorithm for the single objective
optimization, where the objective is the PMEPR. The steps are:
1. Initialize the current population by generating L chromosomes. Each chromosome has its Q genes set to 0 or 1 with equal probability.
2. For each chromosome, convert the N K binary sequences into real numbers
and create a set of N K phase codes: exp(jan,k ).
3. For any of the L sets, compute the oversampled complex OFDM signal,
obtained as a result of applying an IDFT1 on the phase codes [8].
4. Evaluate the PMEPR according to Eq. 8 and attach this number to the
corresponding chromosome.
5. Form the intermediate population by discarding the weakest element (highest
PMEPR) and duplicating the strongest element (lowest PMEPR).
6. Prepare for recombination by associating chromosomes by pair. Each pair
shall be composed of two distinct chromosomes.
7. For each pair, apply a one point crossover [17]. This process results in the
generation of L offspings.
8. Apply mutation every two generations. When the generation number is odd,
select at random Lmut offsprings and for each of them apply mutation on
one of their gene, again, selected at random.
9. Feed the intermediate population into the current population.
Steps 2 to 9 are carried out until the stopping criteria is met. The latter is
characterized by two elements. A threshold on the population mean fitness that
guarantees satisfactory solutions in the current population as well as a threshold
on the standard deviation which assesses the convergence of the entire population
towards a minimum solution.
1
If there are several symbols in the pulse the IDFT is applied on each vector of phase
codes corresponding to one symbol and so on. At the end the outputs are stacked
together to form the pulse
4.2
Multi-objective genetic algorithm
Our multi-objective problem could be solved as a single optimization problem by
formulating an objective function in the form s = αP M EP R+βP SLR+γISLR
and then optimizing with respect to s with our GA. The drawback of that solution is judicious selection of the weights. If multiple solutions are required, the
problem has to be run repeatedly for different sets of weights. To overcome this
difficulty, many multi-objective evolutionary optimization algorithms have been
developed, which produce a set of non dominated solutions in a single run. In this
work, we make use of the well-known non-dominated sorting genetic algorithm
NSGA-II [4]. The principal breakthrough of this algorithm is the convergence towards the true Pareto optimal set with a good spread or diversity of the solutions
[5]. When selecting the best elements not only the respective fitness functions
are evaluated but also the crowding distance, which tells whether this element
is in a high density zone or conversely in an low density zone. At equal fronts,
we select the isolated element in order to maintain diversity in the solutions. In
light of this preparation we describe the steps of this algorithm, which we use
with either two or three of our objective functions. In comparison to our GA, the
NSGA-II uses real numbers throughout. Genetic operations such as cross-over
and mutation employ a method that simulates the equivalent binary processes.
A thorough review of this technique is described in [6].
1. Initialize the population with L sets of N K phases, all real numbers between
0 and 2π and calculate the phase codes exp(jan,k ).
2. For any of the L sets, compute the oversampled complex OFDM signal,
obtained as a result of applying an IDFT on the phase codes and calculate
the autocorrelation function.
3. Evaluate the objective functions following Eq. 6, Eq. 7 and Eq. 8.
4. Sort the chromosomes according to non-dominated sort and form the fronts
(≡ rank).
5. Evaluate the crowding distance of each chromosome.
6. Select the parent chromosomes using binary tournament selection. In binary tournaments, two chromosomes are randomly chosen and the strongest
in terms of rank is selected to be in the parent population. If individuals
with the same rank are encountered, their crowding distance is compared. A
lower rank and higher crowding distance is the selection criteria. The parent
population has a size of L/2.
7. Apply genetic operations such as cross-over and mutation on the selected
chromosomes to produce the off-springs.
8. Combine the off-spring population with the parent population and select the
best L chromosomes for the next generation, again based on the rank and if
needed on the crowding distance.
Steps 2 to 8 are carried out until the stopping criteria is met, which occurs in
this case when the total number of generations has been covered.
5
Simulation results
In this section we present our simulation results. We start with the results obtained with our GA based algorithm.
5.1
Single objective: PMEPR
In our simulations, we considered a population of L = 22 chromosomes, and
the number of chromosomes selected for mutation every two generations was
Lmut = 5. Despite the simplicity of our genetic algorithm, we are able to retrieve
phase code sequences with optimal PMEPR properties. Following our previous
comment, we address the interesting case of sparse spectrum. In Fig. 3, we
compare the PMEPR of three sets of phase codes for two different configurations.
In the first configuration, all subcarriers are enabled. We see in Fig. 2(a) that
our GA solution with 10 subcarriers outperforms the Newman solution. We also
show the relative gain as compared to the uncoded case. The latter results in
the highest PMEPR. Possibly not harmful with 10 subcarriers it cannot be
tolerated with 500 subcarriers. In Fig. 2(b) we present our GA solution when
we disabled two subcarriers and compare now the PMEPR of this solution with
the Newman and uncoded cases when the same subcarriers have been disabled.
Our GA solution outperforms the Newman solution.
(a) Full band
(b) 80% of the band
Fig. 2. Time domain signal of the OFDM symbol for different coding. When the coverage of the band is sparse our GA solution outperforms significantly the Newman phase
coding.
In Table 3, we evaluated the PMEPR for higher numbers of subcarriers. We
chose to use N = 500 and N = 100 to comply with our design parameters
summarized in Table 2. We observe that when we face a sparse spectrum we
have no difficulty to find solutions that outperform the Newman phases. The
results given in the last column were produced while using only two genes per
phase, as with QPSK. This aspect is attractive in particular if we were to use
a communication system for radar with a predefined alphabet, such as QPSK,
etc.
Table 3. Simulation results for the PMEPR (in dB) with different levels of sparsity
Full
N=100 70%
50%
Full
N=500 70%
50%
No coding Newman GA solution GA solution (QPSK)
100
1.8
3.3
3.4
70
4.2
2.9
3.1
50
4.1
3.2
3.3
500
1.8
3.9
4.6
350
4.9
3.9
4.6
250
5.0
4.5
4.5
In Fig. 3, we show the convergence of our GA for both cases of ”no modulation” and QPSK modulation. We observe what we could intuitively guess.
Because of the shorter chromosome size, the QPSK tends to converge in a more
chaotic manner than our ”no modulation” case. In the latter, each phase code
is represented by a binary string composed of 18 bits. As a result, from one
generation to the next, the genetic operations will transform the phase code set
(equivalent chromosome) in a less radical fashion than with QPSK. The convergence appears thus smoother.
(a) No modulation (18 genes)
(b) QPSK (2 genes)
Fig. 3. Convergence of the SGA when using either 18 genes or 2 genes to describe one
phase. The population contains 22 elements.
5.2
Multiple obectives: PMEPR and PSLR
One of the benefit of the NSGA-II as compared to the prior version, NSGA is its
fast computing time as a result of the selection strategy based on chromosome
rank and crowding distance. On top, the use of binary simulated crossover and
mutation improves the algorithm complexity. The large chromosome sizes from
our GA are simply suppressed as we work throughout the algorithm with sets
of real numbers for the phase codes. In Figs. 4, we present the improvement
obtained with the NSGA-II when we intend to find a set of phase codes for our
two design cases given in Table 2. In both cases, we have taken a smaller number
of subcarriers, but the time bandwidth product remains the same, N K = 100
and N K = 500 respectively. For comparison, we plotted 100 realizations of the
initial random population. With a population size equal to L = 40, our clouds
comprise 4000 points. After 10000 generations our set of optimal solutions is
considerably improved.
(a) N =25, K=4
(b) N =125, K=4
Fig. 4. Comparison between random phases and phases resulting from the NSGA-II
optimization.
In Fig. 5, we propose to compare three cases, all having equal time bandwidth
product, like in our design. We see that when the same number of generations
is considered, the design with the smallest number of subcarrier will provide the
best pareto front. Again, this results confirms an intuitive guess.
5.3
Multiple objectives: PMEPR, PSLR and ISLR
As opposed to the single optimization case, which we presented in [2], we can
feed our three objective functions of interest into our MOO-GA to provide us
with a 3D map of solutions, related to a particular design. In light of the intended
application and the design constraints the most suitable solution can be selected.
Fig. 6 shows the 3D pareto fronts of the design cases evaluated in [2].
(a) N =25, K=4
(b) N =125, K=4
Fig. 5. Convergence of the NSGA-II optimization.
Fig. 6. Pareto front after 10000 generations for two pulse design.
6
Case study: GA for PMEPR optimization in a target
detection enhancement procedure
In this section we show how our single objective GA optimization can be integrated in a procedure intended to enhance the target detection. Our OFDM
pulse is tailored in two steps. Firstly, we find the best weights to enhance the
detection. Secondly, we run our GA to find satisfactory sets of phase codes. Our
procedure is presented in Fig. 7
Fig. 7. OFDM pulse optimization in two steps.
6.1
SNR as the design metric
Unlike traditional notions where imaging-based metrics for waveform design are
best for target detection and classification, which lead to our objectives of low
PSLR and low ISLR, another approach relies on the well-known SNR metric.
In that case, the optimum waveform will not necessarily have good PSLR nor
ISLR properties.
6.2
Transmitted signal receiver-filter pair
It has been shown [10] that the signal at the output of the matched filter Vs (f ),
in a radar receiver, is given by the product of the target reflectivity spectrum
ς(f ), the waveform spectrum X(f ) and the filter transfer function H(f ):
Vs (f ) = ς(f + fc )X(f )H(f ).
(10)
The frequency domain operations undergone by the waveform are sumarized
in Fig. 8, where Vtx (f ) is the Fourier transform of the transmitted signal vtx (t):
vtx (t) = x(t) exp(j2πfc t),
(11)
fc is the carrier frequency and x(t) is the OFDM pulse as given in Eq. 3. In this
analysis, we consider that the pulse is composed of a single OFDM symbol.
Fig. 8. Model of the received signal in the frequency domain.
Optimum receiver-filter pair In the case of an extended target, unlike the
standard matched filtering problem applicable when the target is modelled as
a point scatterer, the optimal receiver filter shall be matched to the waveform
scattered by the target, not the transmitted target itself [11]. When this is the
case, the maximum SNR at time t0 is given by the following equation [11]:
Z ∞
S
|ς(f + fc )X(f )|2
=
df.
(12)
N t0
Snn (f )
−∞
While previous works have researched solutions of this SNR optimization
problem in the time domain, to find appropriate x(t), we argue that we can
benefit from the OFDM structure and find directly X(f ). Assuming our OFDM
signal to be zero outside its bandwidth2 , and considering that the stationary
additive Gaussian noise is white, with one-sided power spectral density Snn (f ) =
N0 , we can modify Eq. 12 into:
Z B
1
S
=
|X(f )ς(f + fc )|2 df,
(13)
N t0
N0 0
If we know the target reflectivity spectrum, we can rewrite Eq. 13 in discrete
terms:
N −1
1 X
S
=
|X[n]|2 · |ς[n]|2 ,
(14)
N t0
N0 n=0
an
where X[n] = w√nN
is our discrete OFDM spectrum, X[n] = X(n∆f ) and n
takes values from 0 to N − 1. Following this expression, when weights are equal
to 1 the signal is normalized (unit energy). Assuming that we transmit at all
subcarriers, our optimization problem reduces in finding the weights wn such
that:
arg max
wn
N
−1
X
wn2
2
· |ς[n]| , s.t.
n=0
wn 6= 0
PN −1 2
n=0 wn = N
(15)
A solution to this problem is obtained when wn ∝ |ς[n]|.
6.3
Simulation setup and results
In our analysis we choose to work at X-band. We consider a 2 GHz bandwidth
B centred around 10 GHz, fc = 9 GHz. Before we elaborate on the reflectivity
spectrum of our target within this band we stress the need to normalize ς[n] and
describe our normalization strategy.
Target reflectivity spectrum normalization For relevant comparisons, we
normalize the discrete target reflectivity spectrum following the strategy suggested in chapter 14 in [12]. Essentially, when a flat spectrum OFDM pulse
of unit energy interacts with the normalized reflectivity spectrum ςnorm [n], the
2
The OFDM spectrum of the baseband pulse is considered to spread between 0 and
B rather -B/2 to B/2
frequency-domain reflected signal shall have unit average power. The discrete
elements
of our flat spectrum OFDM pulse of unit energy are given by X[n] =
√
an / N . Hence, we find for ςnorm [n]:
N2
· |ς[n]|2 ,
|ςnorm [n]|2 = PN −1
2
|ς[n]|
n=0
6.4
(16)
Complex target model
In our simulation we construct a synthetic target from P = 50 point scatterers
√
with equal unit reflectivity ςi = σi = 1 and located at ranges Ri from the
radar. The individual point scatterers are assumed to be perfectly conducting
spheres, large enough to have a reflectivity constant within the band of interest.
It can be shown [15] that the compound target reflectivity spectrum ς(f ) is equal
to:
P
X
√
Ri
ς(f ) =
σi exp(−j4πf ).
(17)
c
i=1
The point scatterers are randomly distributed within a rectangle, 5 meters wide
and 10 meters long, whose center is 10 km away from the radar along the x axis.
Fig. 9(a) shows the position of the scatterers, while Fig. 9(b) gives the target
reflectivity power spectrum. As expected, we observe strong variations of the
power spectrum within the frequency band.
(a) Synthetic target
(b) Reflectivity power spectrum
Fig. 9. Complex target made of P = 50 point scatterers.
7
SNR and PMEPR improvements
Following the above methodology, in the first step, the N weights are derived
from the normalized target reflectivity spectrum, so that the intermediate OFDM
spectrum X1 (f ) has enhanced detection capability. In the second step, X1 (f ) is
fed into our GA optimization to find a set of N phase codes that will improve
the PMEPR. In the end, as shown in Fig 7 our pulse Xopt (f ) has enhanced
detection capabilities for the target of interest and a reasonably low PMEPR.
(a) Weights
(b) PMEPR improvement
Fig. 10. Use of the single objective GA to optimize the PMEPR based when an optimized set of weights is applied on the subcarriers.
In table 4, we show the gain in dB of the weights optimization method.
Table 4. Result for the detection enhancement
Average power (dB)
8
Flat spectrum Optimal weights
0
2.4
Conclusion
In this paper, we showed that GA based techniques are suitable to optimize or
improve the design of OFDM pulses for radar. We inspected the possibility to
incorporate these optimization techniques in the more general waveform design in
regards of two processing solutions. In our first conventional processing solution,
primarily, the PSLR and the PMEPR need to be optimized. The former is a
figure-of-merit for the detectability of small targets and the latter relates to the
severity of the distortion. The MOO-GA based algorithm NGSA-II has been used
to optimize this multi-objective problem. We observed that with the PMEPR
and the PSLR for the objective functions a substantial improvement is achieved
as compared to the case where random coding is applied. We also showed that the
more symbols in the pulse the better the optimization in PMEPR and PSLR, for
the same time bandwidth product. In our second, frequency domain processing
solution, we stressed that the main focus should be on the minimization of
the PMEPR. Therefore, the single objective GA can be used. We showed that
this evolutionary technique can produce solutions that outperform the most
robust methods. We also demonstrated that in some relevant cases, when the
OFDM spectrum is sparse, our GA solution gives outstanding results. Finally,
we presented a case study to attest the relevance of our GA optimization method
where the optimization relies on the set of phase codes of the OFDM pulse. In
this case study, we showed that the problem of finding a suitable radar pulse
for enhancing the detection of a known target could be solved in two steps. In
the first step, the weights of the subcarriers are selected in light of the target
reflectivity spectrum. In the second step, our GA based optimization method is
applied on this set of weights to search for the most appropriate phase codes
that would minimize the PMEPR. For completeness we may conclude saying
that when both PSLR and ISLR need to be optimized, we suggest to use a
MOO-GA based on two objectives rather than three, and use the PMEPR as a
constraint.
References
1. Langman, A. and Hazarika, O. and Mishra, A.K. and Inggs, M. White RHINO: A
Low Cost Whitespace Communications and Radar Hardware Platform. In Radioelektronika Conference, 2013, April. 2013.
2. Lellouch, G. and Mishra, A.K. Multi-carrier based Radar signal optimization using
genetic algorithm. In Proceedings of the SocProS 2013, December. 2013.
3. Lellouch, G. and Mishra, A.K. and Inggs, M. Processing alternatives in OFDM
Radar. Submitted to IEEE International Radar Conference, 2014, Oct. 2014.
4. Deb, K. and Pratap, A. and Agarwal, S. and Meyarivan, T. A fast and elitist
multiobjective genetic algorithm: NSGA-II. In IEEE Transactions on Evolutionary
Computation, vol. 6, number 2, pp. 182-197, 2002.
5. Seshadri, A. A fast elitist multiobjective genetic algorithm: NSGA-II. In MATLAB
Central, 2006.
6. Agrawal, R.B. and Deb, K. Simulated binary crossover for continuous search space.
In Citeseer, 1994.
7. Jones, A.E. and Wilkinson, T.A. and Barton S.K. Block Coding Scheme for Reduction of Peak to Mean Envelope Power Ratio of Multicarrier Transmission Schemes.
In Electronics Letters,, vol. 30, issue 25, pp. 20982099, Dec. 1994.
8. Lellouch, G. and Tran, P. and Pribic, R. and Van Genderen, P. OFDM waveforms
for frequency agility and opportunities for Doppler processing in radar. In IEEE
Radar Conference, 2008, May. 2008, pp. 1-6.
9. Riché, V. and Meric, S. and Baudais, J-Y. and Pottier, E. Optimization of OFDM
SAR signals for Range ambiguity suppression. In Radar Conference (EuRAD), 2012.
2012, pp. 278-281.
10. Wilkinson, A.J. and Lord, R.T. and Inggs, M.R. Stepped-frequency processing
by reconstruction of target reflectivity spectrum. In Communications and Signal
Processing, 1998. COMSIG’98.. 1998, pp. 101-104.
11. Bell, M.R. Information theory and radar waveform design. In IEEE Transactions
on Information Theory, vol. 39, number 5, pp. 1578-1597, 1993.
12. Gini, F. and De Maio, A. and Patton, L. Waveform design and diversity for
advanced radar systems. Institution of engineering and technology, 2012.
13. Reeves, C. Handbook of metaheuristics. 55–82 Springer, 2003.
14. Levanon, N. and Mozeson, E. Radar signals. Wiley-Interscience, 2004.
15. Richards, M.A. and Scheer, J. and Holm, W.A. Principles of modern radar: basic
principles. SciTech Pub., 2010.
16. Hara, S. and Prasad, R. Multicarrier techniques for 4G mobile communications.
Artech House Publishers, 2003.
17. Whitley D. A Genetic Algorithm Tutorial. Colorado State University.
| 9cs.NE
|
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
1
Covert Attacks in Cyber-Physical
Control Systems
Alan Oliveira de Sá∗† , Luiz F. Rust da Costa Carmo∗‡ , Raphael C. S. Machado‡
of Mathematics / NCE - Tércio Pacitti Institute,
Federal University of Rio de Janeiro, 21.941-901, RJ – Brazil
† Admiral Wandenkolk Instruction Center – Brazilian Navy,
Ilha das Enxadas, Baı́a de Guanabara – Rio de Janeiro – RJ – Brazil
[email protected]
† National Institute of Metrology, Quality and Technology (Inmetro)
Av. Nossa Senhora das Graças, 50, Xerém, Duque de Caxias, 25.250-020, RJ – Brazil
{lfrust,rcmachado}@inmetro.gov.br
Index Terms—Security, Cyber-Physical Systems, Networked
Control Systems.
I. I NTRODUCTION
T
HE integration of the systems used to control physical
processes via communication networks aims to assign to
such systems better operational and management capabilities,
as well as reduce its costs. Motivated by these advantages,
there is a trend to have an increasing number of industrial process and critical infrastructure systems driven by Networked
Control Systems (NCS) [1]–[4], also referred as NetworkBased Control Systems (NBCS) [5], [6]. A NCS, shown in
Figure 1, consists of a physical plant, described by its transfer
function G(z), a controller, which runs a control function
C(z), and a communication network that interconnect both
devices through a forward stream and a feedback stream. The
forward stream connects the output of the controller to the
plant’s actuators. The feedback stream connects the output of
the plant’s sensors to the controller’s input.
At the same time it brings several advantages, the integration
of controllers and physical plants in a closed loop through a
communication network also exposes such control systems to
new threats, typical of the cyber domain. In this context, studies have been conduced aiming to characterize vulnerabilities
and propose security solutions for the NCSs.
Controller
forward stream
Plant
r(k) +
e(k) Control
function
_
C(z )
u(k)
u’(k)
Physical
Process
G(z )
Sensors
Abstract—The advantages of using communication networks
to interconnect controllers and physical plants motivate the
increasing number of Networked Control Systems, in industrial
and critical infrastructure facilities. However, this integration also
exposes such control systems to new threats, typical of the cyber
domain. In this context, studies have been conduced, aiming to
explore vulnerabilities and propose security solutions for cyberphysical systems. In this paper, it is proposed a covert attack for
service degradation, which is planned based on the intelligence
gathered by another attack, herein proposed, referred as System
Identification attack. The simulation results demonstrate that the
joint operation of the two attacks is capable to affect, in a covert
and accurate way, the physical behavior of a system.
Actuators
arXiv:1609.09537v3 [cs.CR] 3 Apr 2017
∗ Institute
Network
y’(k)
y(k)
feedback stream
Figure 1. Networked Control Systems (NCS)
One possible way to attack a NCS is by hacking its software,
i.e. changing the configuration or even the code executed by
the controller, following a strategy similar to that used by the
Stuxnet worm [7]. Another possible way for an attacker to
negatively affect a NCS is by interfering on its communication
process. Basically, an attacker may interfere in the forward
and/or the feedback streams by three different means: inducing
a jitter, causing data loss due to packet drop outs, or even
injecting false data in the communication process.
In this work, it is developed an attack where false data
is injected in the communication process of a NCS. It is
demonstrated the possibility to degrade the service performed
by a plant, through interventions that produce subtle changes
on its physical behavior. Such interventions aim to reduce the
efficiency of the plant or even cause it damage in mid/long
therm. It is worth mentioning that an uncontrolled intervention
in a NCS may lead the plant to an immediate breakdown,
or even significantly change its behavior, which may cause
the disclosure of the attack and the eventual failure of the
operation. Thus, the changes driven by the attack herein
proposed are dimensioned so that the modifications in the
plant’s behavior are physically difficult to be perceived. That
is why the present attack is classified as physically covert.
To ensure that the attack to a NCS is physically covert,
the attacker must plan his offensive based on an accurate
Copyright (c) 2017 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from
the IEEE by sending a request to [email protected]. The final version of this paper is available at http://dx.doi.org/10.1109/TII.2017.2676005.
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
knowledge about the system dynamics, otherwise the consequences of the attack may be unpredictable. One possible way
to obtain such knowledge is through conventional intelligence
operations, performed to collect information regarding to the
design and the dynamics of the NCS. Another way to gather
information about the targeted system is through what we
refer in this paper as a Cyber-Physical Intelligence attack.
To this end, it is also proposed in this paper the System
Identification attack, which aims to collect information about
the plant’s transfer function G(z) and the controller’s control
function C(z) of a NCS. This attack is proposed based on the
Backtracking Search Optimization algorithm (BSA) [8].
This work motivated the formalization of a number of
concepts related to covertness and intelligence in the context of
the cyber-physical security. Thus, an additional contribution of
this paper is the proposition of a terminology that encompasses
a whole new class of attacks on cyber-physical systems. The
proposed taxonomy establishes a new approach regarding to
the covertness of attacks on cyber-physical systems, which
must be analyzed from two aspects simultaneously: the physical and the cybernetic aspects.
It is worth mentioning that the objective of this work is not
to facilitate covert attacks for service degradation in cyberphysical control systems. The purpose of this work is to
demonstrate the degree of accuracy that may be achieved
in this kind of attack, especially when supported by System
Identification attacks and, therefore, encourage the research
for countermeasures to such attacks. The rest of this paper
is organized as follows: First, in Section II, some related
works are presented. Later, in Section III, it is proposed a
taxonomy regarding to the cyber-physical attacks that may
happen in the control loop of a NCS. In Section IV, it is
described a System Identification attack. Then, in Section V,
it is defined a cover attack for service degradation. After
that, in Section VI, there are reported the results obtained by
simulations of covert attacks for service degradation, supported
by System Identification attacks. Finally, in Section VII, some
conclusions and possible future works are presented.
In [4], the authors present a Supervisory Control and
Data Acquisition (SCADA) testbed using TrueTime, a MATLAB/Simulink based tool. They demonstrate an attack where a
malicious agent sends false signals to the controller and to the
actuator of a NCS. In that paper, the false signals are randomly
generated aiming to make a DC motor lose its stability. This
kind of attack does not require a previous knowledge about
the plant and controller of the NCS. On the other hand, the
desired physical effect and the covertness of the attack can
not be ensured due to the unpredictable consequences of the
application of random false signals to a system which the
model is unknown.
More recently, in [10], the authors give a general framework
for the analysis of a wide variety of methods of attack in
NCSs. In their classification, it is stated that covert attacks
in NCSs require high level of knowledge about the targeted
system. Exemples of covert attacks are provided in [11],
[12]. In these works the attacks are reformed by a man-inthe-middle (MitM), where the attacker needs to know the
model of the plant under attack and also inject false data in
both the forward and the feedback streams. The covertness
of the attacks described in [11], [12], which depends on the
difference between the actual model of the plant and the model
used by the attacker, is analyzed from the perspective of the
signals arriving to the controller, without addressing if the
physical effects on the plant are noticeable, or if they are covet
when faced by a human observer.
In [10]–[12], where it is required a previous knowledge
about the models of the NCS under attack, it is not described
how the knowledge about the system is obtained by the
attacker. It is just stated that a model is previously known
to subsidize the design of the attack. The joint operation,
herein proposed, of a covert attack for service degradation,
supported by a System Identification attack, aims to fill this
hiatus, demonstrating how the data of a NCS can be obtained
and how a covert attack can take advantage from it. The
Table I presents a synthesis of the characteristics of the attacks
referred in this section.
Table I
S YNTHESIS OF THE RELATED ATTACKS
II. R ELATED W ORKS
The possibility of cyber-physical attacks became a reality
after the launch of the Stuxnet worm [7] and has been
motivating researches concerning the security of NCSs. In this
section, some works related to this subject are presented.
In [6] the authors propose two queueing models to evaluate
the impact of delay jitter and packet loss in a NCS under
attack. The attack is not designed taking into account a
previous knowledge about the models of the controller and the
physical plant. Thus, to affect the plant’s behavior, the attacker
arbitrarily floods the network with an additional traffic, causing
jitter and packet loss. In this tactics, the excess of packets in
the network can reduce the covertness of the attack, allowing
the adoption of countermeasures, such as packet filtering [6]
or blocking the malicious traffic on its origin [9]. Additionally,
the arbitrary intervention in a system which the model is
unknown may lead the plant to an extreme physical behavior,
which is not desired if it is intended a covert attack.
2
Attack
Stuxnet worm [7]
Long, et al. [6]
Farooqui, et al. [4]
Smith [11], [12]
Method
Modifications in
the PLC code
Inducing jitter
and packet loss
Data injection
Data injection
System
knowledge
How the knowledge
is obtained
Yes
None
Experiments in
a real system
N/A
None
Yes
N/A
Not described
III. TAXONOMY
In this section it is presented a taxonomy concerning the
possible attacks on cyber-physical control systems. In Section
III-A, the attacks are briefly described and classified according
to the way they act in the NCS. In Section III-B, it is proposed
a new approach for the analysis of the covertness of attacks
in cyber-physichal systems.
A. Classification of the Attacks
The attacks to cyber-physichal control systems may take
place on its devices – i.e. the controller, and the plant’s sensors
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
^LJƐƚĞŵ<ŶŽǁůĞĚŐĞ
ĐĐĞƐƐƚŽƚŚĞĚĂƚĂ
ĐĐĞƐƐƚŽƚŚĞĐŽŶƚƌŽů ůŽŽƉ
Ž^ͲŽŶƚƌŽůůĞĚ
:ŝƚƚĞƌ
Ž^ͲƌďŝƚƌĂƌLJ
:ŝƚƚĞƌ
Ž^
Ž^ͲƌďŝƚƌĂƌLJ
ĂƚĂůŽƐƐ
ZĞƋƵŝƌĞŵĞŶƚƐ
and actuators – and/or on its communication system, affecting
the forward and the feedback streams. As a premise, we must
consider that the service intended to be attacked/protected in
such system is the work performed by the physical process,
controlled by the NCS.
Considering the aforementioned definition of service in a
NCS, the attacks may be classified within three different
categories, as shown in Figure 2:
• Denial-of-Service (DoS) [13]: in a NCS, the DoS attacks
comprehends all kind of cyber-physical attacks that deny
the operation of the physical process, interrupting the
execution of the work, or service, that the controlled
plant is intended to do. The attack results, for example, in
behaviors that may shut the plant down or even destroy
it in a short therm.
• Service Degradation (SD): the SD attacks consist of
malicious interventions that are done in the control loop
in order to reduce the efficiency of the service, i.e. the
efficiency of the physical process, or even reduce the
mean time between failure (MTBF) of the plant in mid
therm or long therm.
• Cyber-physical Intelligence (CPI): the concept of Cyberphysical Intelligence, herein proposed, is different from
the concept of where cyber-physical systems are integrated with intelligent systems [14]. In the present
taxonomy, the CPI attacks comprehend actions that are
performed in the control loop of a NCS in order to gather
information about the system’s operation and/or its design. This attacks are intended to acquire the intelligence
necessary to plan covert and controlled attacks, or even
to subsidize data for replay attacks [7].
Ž^ͲƌďŝƚƌĂƌLJ
ĂƚĂ/ŶũĞĐƚŝŽŶ
Ž^ͲŽŶƚƌŽůůĞĚ
ĂƚĂůŽƐƐ
Ž^ͲŽŶƚƌŽůůĞĚ
ĂƚĂ/ŶũĞĐƚŝŽŶ
^ͲŽŶƚƌŽůůĞĚ
:ŝƚƚĞƌ
^ĞƌǀŝĐĞ
ĞŐƌĂĚĂƚŝŽŶ
^ͲŽŶƚƌŽůůĞĚ
ĂƚĂůŽƐƐ
^ͲŽŶƚƌŽůůĞĚ
ĂƚĂ/ŶũĞĐƚŝŽŶ
LJďĞƌͲƉŚLJƐŝĐĂů
ŝŶƚĞůůŝŐĞŶĐĞ
ĂǀĞƐĚƌŽƉƉŝŶŐ
;>ŽǁĞƌͿ
^LJƐƚĞŵ
/ĚĞŶƚŝĨŝĐĂƚŝŽŶ
ŽŵƉůĞdžŝƚLJŽĨƚŚĞĂƚƚĂĐŬ
;,ŝŐŚĞƌͿ
Figure 2. Classification and requirements of the cyber-physical attacks that
act in the control loop of a NCS.
In Figure 2, six kinds of DoS attacks are presented, with
their respective requirements. From this six DoS attacks, the
less complex are the arbitrary ones:
3
DoS-Arbitrary Jitter: in this kind of attack, the delay
of the forward and/or the feedback stream is arbitrarily
changed, without a previous knowledge about the models
of the NCS, in order to lead the system to an instability
or to a condition that causes the interruption of the
physical process. This attack only requires access to the
control loop, once it may be performed by just consuming
the resources of the system, such as the bandwidth of
communication links, or the computational resources of
the equipments that are in the control loop.
• DoS-Arbitrary Data Loss: in this kind of attack, the
attacker prevents the data from reaching the actuator
and/or the controller of the NCS. The communication
channel is jammed arbitrarily, without a previous knowledge about the models of the NCS, leading the system to
an instability or to a condition that causes the interruption
of the physical process. It is worth mentioning that some
DoS-Arbitrary Jitter attack may result in a DoS-Arbitrary
Data Loss attack, if an eventual higher delay cause packet
drop outs. As the DoS-Arbitrary Jitter attack, this attack
only requires access to the control loop of the NCS.
• DoS-Arbitrary Data Injection: in such attacks, the attacker
sends arbitrary false data to the controller, as it was sent
by the sensors, and/or to the actuators, as it was sent by
the controller. The false data is injected in the NCS closed
loop without a previous knowledge about the models of
the NCS. This attack is more complex than the DoSArbitrary Jitter and the DoS-Arbitrary Data Loss attacks,
given that it requires access to the data that flows in the
control loop of the NCS.
The attacks classified as DoS-controlled – DoS-Controlled
Jitter, DoS-Controlled Data Loss, and DoS-Controlled Data
Injection – shown in Figure 2 interfere in the control loop
of a NCS by the same means that their respective DoSArbitrary attacks. The difference between a DoS-Controlled
attack and a DoS-Arbitrary attack is that, in the former, the
interference caused by the attacker is precisely planned and
executed, in order to achieve the exact desired behavior that
leads the physical service to an interruption, in a more efficient
way. Thus, to achieve such efficiency, a DoS-Controlled attack
require an accurate knowledge about the NCS models, i.e. the
plant and the controller transfer functions, which have to be
analyzed to plan the attack.
Regarding to the SD attacks, we must consider the three different kinds of attack shown in Figure 2: SD-Controlled Jitter,
SD-Controlled Data Loss, and SD-Controlled Data Injection.
The difference between a SD-Controlled attack and a DoSControlled attack is that the former is not intended to interrupt
the physical process in a short therm. It aims to keep the process running with reduced efficiency, sometimes also targeting
a gradual physical deterioration of the controlled devices. To
succeed, the SD-Controlled attacks need to be planned based
on an accurate knowledge about the dynamics and the design
of the NCS. Otherwise, the attack can eventually interrupt the
physical process, due to unpredicted reasons, evolving to a
DoS attack.
The system knowledge required to both DoS-Controlled and
SD-Controlled attacks, can be gathered through CPI attacks,
•
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
as shown in Figure 2. The first, and simpler, CPI attack
is the eavesdropping attack [15], [16], which consists of
simply capturing the data transmitted through the forward and
feedback streams of the NCS. The second CPI attack, herein
proposed, is the System Identification attack, which aims to
obtain information about the control function of the controller
and the transfer function of the plant, by analyzing the signals
that flow in the network between the controller and the plant.
The CPI attacks by themselves do not impact on the NCS,
but they are an useful tool to plan efficient and accurate DoSControlled and SD-Controlled attacks.
B. Cybernetic vs. Physical Covertness
The covertness of an attack regards to its capacity to not be
perceived or detected. In the case of cyber-physical attacks on
NCSs, the covertness must be simultaneously analyzed in two
different domains: the cyber domain; and the physical domain.
In this sense, it is presented in this section the definition of
what is a cybernetically covert attack and what is a physically
covert attack:
•
•
Cybernetically covert attacks: are the attacks that have
low probability to be detected by algorithms that monitor
the softwares, communications and data of the NCS, or
by systems that monitor the dynamics of the plant.
Physically covert attacks: are attacks that cause physical
effects that can not be easily noticed or identified by
a human observer. The attack slightly modifies some
behaviors of the system in a way that it physically affects
the plant, but the effect is not easily perceptible or it can
eventually be understood as a consequence of another
root cause, other than an attack.
IV. T HE S YSTEM I DENTIFICATION ATTACK
The System Identification attack, herein proposed, is intended to assess the coefficients of the plant’s transfer function
G(z) and the controller’s control function C(z). Both functions are linear time-invariant (LTI). The attack uses the BSA
metaheuristic, proposed in [8] and briefly described in [17],
to minimize the fitness function presented in this section.
The BSA is an evolutionary algorithm that uses the information obtained by past generations – or iterations – to perform
the search for solutions for optimization problems. The algorithm has two parameters that are empirically adjusted: the size
of its population P ; and η, described in [17], that establishes
the amplitude of the movements of the individuals of P . The
parameter η must be adjusted aiming to assign to the algorithm
both good exploration and exploitation capabilities.
If the input i(k) and the output o(k) of a device of the
NCS is known, the model of such device can be assessed
by applying the known i(k) in an estimated model, which
must be adjusted until its estimated output ô(k) converge to
o(k). In this sense, the BSA is used to iteratively adjust the
estimated model, by minimizing a specific fitness function,
until the estimated model converge to the actual model of the
real device, that can be a controller or a plant of the NCS.
4
To establish the fitness function, firstly, it must be considered a generic LTI system, whose transfer function Q(z) is
represented by (1):
Q(z) =
O(z)
an z n + an−1 z n−1 + ... + a1 z 1 + a0
, (1)
= m
I(z)
z + bm−1 z m−1 + ... + b1 z 1 + b0
wherein I(z) is the input of the system, O(z) is the output
of the system, n and m are the order of the numerator
and the denominator, respectively, and [an , an−1 , ...a1 , a0 ] and
[bm−1 , bm−2 , ...b1 , b0 ] are the coefficients of the numerator
and the denominator, respectively, that are intended to be
found by this System Identification attack. Also, it must be
considered that i(k) and o(k) represent the sampled input and
output of the system, respectively, where I(z) = Z[i(k)],
O(z) = Z[o(k)], k is the number of the sample and Z
represents the Z-transform operation.
In this System Identification attack, i(k) and o(k) are firstly
captured by an eavesdropping attack [15], [16], for exemple,
during a monitoring period T . To deal with the eventual loss
of samples, that may not be received by the attacker during
T , the algorithm holds the value of the last received sample,
according with (2), wherein x(k) can either be i(k) or o(k):
if the sample k is lost;
x(k − 1)
(2)
x(k) =
x(k)
otherwise.
Then, after acquiring i(k) and o(k), the captured i(k)
is applied to the input of an estimated model, that
is described by a transfer function whose coefficients
[an,j , an−1,j , ...a1,j , a0,j , bm−1,j , bm−2,j , ...b1,j , b0,j ] are the
coordinates of an individual j of the BSA. The application
of i(k) to the input of the estimated model results in an
output signal ôj (k). After obtaining ôj (k), the fitness fj
of the individual j is computed comparing the output o(k),
captured from the attacked device, with the output ôj (k) of
the estimated model, according with (3):
fj =
N
P
(o(k) − ôj (k))2
k=0
,
(3)
N
wherein N is the number of samples that exist during the
monitoring period T . Note that, if the attacker do not lose
any sample of i(k) and o(k) during T , then min fj = 0
when [an,j , an−1,j , ...a1,j , a0,j , bm−1,j , bm−2,j , ...b1,j , b0,j ] =
[an , an−1 , ...a1 , a0 , bm−1 , bm−2 , ...b1 , b0 ], i.e. when the estimated model converges to the actual model of the attacked
device.
It is possible to establish an analogy between this System
Identification attack and the Known Plaintext cryptanalytic
attack [18], wherein i(k) and o(k) correspond to the plaintext
and ciphertext, respectively, the form of the generic transfer
function Q(z) corresponds to the encryption algorithm and the
actual coefficients of Q(z) corresponds to the secret key.
V. T HE C OVERT ATTACK
FOR
S ERVICE D EGRADATION
Based on the taxonomy presented in Section III-A, the
attack described in this section is classified as a SD-Controlled
Data Injection attack. Its purpose is to reduce the MTBF of the
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
plant and/or reduce the efficiency of the physical process that
the plant performs, by inserting false data in the control loop.
At the same time, the attacker desires that the attack meets
the requirement of being physically covert, as the definition
presented in Section III-B.
One way to degrade a physical service is through the
induction of an overshoot during the transient response of
a plant. The overshoots, or peaks occurred when the system
exceeds the targeted value during the transient response, can
cause stress and possibly damage physical systems such as
mechanical, chemical and electromechanical systems [19],
[20]. Additionally, once they occur in a short period of time,
the overshoots are difficult to be noticed by a human observer.
Another way to degrade the service of a plant is causing a
constant steady state error on it, i.e. producing a constant error
when t → ∞. A low proportion steady state error, besides
being difficult to be perceived by a human observer, may
reduce the efficiency of the physical process or, occasionally,
stress and damage the system in mid/long therm.
In the present attack, to achieve either of the two mentioned
effects, i.e. an overshoot or a constant steady state error, the
attacker interfere in the NCS’s communication process by
injecting false data into the system in a controlled way. To do
so, the attacker act as a MitM that executes an attack function
M (z), as presented in Figure 3, wherein U ′ (z) = M (z)U (z),
U (z) = Z[u(k)] and U ′ (z) = Z[u′ (k)]. The function
M (z) is designed based on the models of the plant and the
controller, both obtained through the System Identification
attack, described in Section IV. The effectiveness of the attack,
therefore, depends on the design of M (z), which in turn
depends on the accuracy of the System Identification attack.
It is worth mentioning that, in Figure 3, although the MitM
is placed in the forward stream, it is also possible perform an
attack by interfering in the feedback stream of the NCS. The
MitM may act in wired or wireless networks, such as in [21].
Controller
forward stream
e(k) Control
function
_
C(z )
u(k)
M(z )
u’(k)
MitM
Physical
process
G(z )
Sensors
r(k) +
Actuators
Plant
Network
y’(k)
y(k)
feedback stream
Figure 3. MitM attack
VI. R ESULTS
In this section, there are presented the results obtained
through simulations that combines the System Identification
attack with physically covert SD-Controlled attacks. First, in
Section VI-A, the model of the attacked system is described.
Then, in Section VI-B, there are presented the results obtained
by the System Identification attack. After that, in Section VI-C,
5
there are presented the results achieved by simulations of physically covert SD-Controlled Data Injection attacks, planned
based on the data gathered by the System Identification attack.
A. The Model of the Attacked System
The attacked NCS has the same architecture of the NCS
shown in Figure 1, and consists of a Proportional-Integral (PI)
controller that controls the rotational speed of a DC motor. The
PI control function C(z) and the DC motor transfer function
G(z) are the same as in [6]. The equations of this NCS are
represented in (4):
C(z) =
c1 z − c2
z−1
G(z) =
z2
g1 z + g2
− g3 z + g4
(4)
wherein c1 = 0, 1701, c2 = −0, 1673, g1 = 0, 3379, g2 =
0, 2793, g3 = −1, 5462 and g4 = 0, 5646. The sample rate of
the system is 50 samples/s and the set point r(k) is an unitary
step function. The network delay is not taken into account in
the simulations of this paper.
B. Results of the System Identification Attack
In this Section, the performance of the System Identification
attack is evaluated through a set of simulations performed
in MATLAB. The SIMULINK tool is used to compute the
output ôj of the estimated models, whose coefficients are the
coordinates of an individual j of the BSA.
The structure of the equations represented in (4) are previously known by the attacker once that, as a premise, it is
known that the target is a NCS that controls a DC motor using
a PI controller. In these simulations, the goal of the System
Identification attack is to discover g1 , g2 , g3 , g4 , c1 and c2 ,
also taking into account scenarios in which the attacker occasionally loses samples of the forward and feedback streams.
Each time that the DC motor is turned on, the forward
and the feedback streams are captured by the attacker during
a period T = 2s. All initial conditions are considered 0,
by the time that the motor is turned on. The coefficients of
G(z), [g1 , g2 , g3 , g4 ], and the coefficients of C(z), [c1 , c2 ],
are computed separately considering that, albeit the closed
loop, G(z) and C(z) are independent transfer functions. To
assess [g1 , g2 , g3 , g4 ], the attacker considers the forward stream
as the input and the feedback stream as the output of the
estimated plant. In the opposite way, to assess [c1 , c2 ], the
attacker considers the feedback stream as the input and the
forward stream as the output of the estimated controller.
To simulate the loss of samples, it is considered four
different rates l of sample loss: 0, 0.05, 0.1 and 0.2. Thus,
a sample is lost by the attacker every time that l < P, where
P ∼ U (0, 1) and U is the uniform distribution. There are
executed 100 different simulations for each rate of sample loss.
In the BSA, the population is set to 100 individuals and
η, empirically adjusted, is 1. To assess the coefficients of the
controller [c1 , c2 ], the algorithm is executed for 600 iterations.
To assess the coefficients of the plant [g1 , g2 , g3 , g4 ], the
number of iterations is increased to 800, due to the higher
number of dimensions of the search space in this case. The
limits of each dimension of the search space are [−10, 10].
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
Figure 4 shows the means of 100 estimated values of g1 ,
g2 , g3 , g4 , c1 and c2 , with a Confidence Interval (CI) of 95%,
considering different rates of sample loss. The actual values of
the coefficients of C(z) and G(z) are also depicted in Figure 4.
Note that the scales of Figures 4(a), 4(b), 4(c) and 4(d) are
different from the scales of Figures 4(f) and 4(f), due to the
smaller amplitude of the CI of c1 and c2 . In Addition, some
statistics of the obtained results are presented in Table II.
According with Table II the distributions of g1 , g2 , g3 and g4
have a high skewness, while the distributions of c1 and c2 have
a moderate skewness. Regarding the kurtosis, the distributions
of all coefficients of G(z) and C(z) are leptokurtic. However,
analyzing Table II, it is not possible to state a clear general
pattern of an increasing/decreasing skewness or kurtosis, in
face of the growth of sample loss.
In Figure 4, it is possible to verify that, in all cases, the ICs
tend to grow with the increase of the sample loss. The same
thing occur with the standard deviations shown in Table II.
Regarding to the coefficients of G(z), Figure 4 shows that the
difference between the mean and the actual value of g1 , g2 ,
g3 and g4 also tends to raise with the increase of sample loss.
It is worth mentioning that the performance of the algorithm
when computing g3 and g4 is better then when computing
g1 and g2 , regarding the means and their CIs. This behavior
results from the higher sensitivity that the output of G(z) has
to the variation of its poles than to the variations of its zeros.
It means that, in this problem, fj grows faster for errors in g3
and g4 than for errors in g1 and g2 , making the BSA population
converge more accurately in dimensions g3 and g4 .
In Figure 4 it is also possible to note that the accuracy
obtained with the coefficients of C(z) is better than the
accuracy of the coefficients of G(z), for all rates of sample
loss. The means of c1 and c2 are closer to their actual values,
with a smaller CI. In fact, the optimization process is more
effective when computing the coefficients of C(z) due to its
smaller search space, that has only two dimensions instead of
the four dimensions of the G(z) problem.
As an additional metric to evaluate the performance of the
algorithm, it is computed |Eg | = |Gr −Ge | and |Ec | = |Cr −Ce |,
that synthesize the error of the estimated coefficients of G(z)
and C(z), respectively, for each solution found. Gr and Ge are
vectors that contain the actual and the estimated coefficients
of G(z), respectively. Similarly, Cr and Ce are vectors that
contain the actual and the estimated coefficients of C(z),
respectively. The histograms of |Eg | and |Ec | are presented
in Figure 5, considering the mentioned rates of sample loss.
The histograms graphically show that |Eg | and |Ec |, which
correspond to the modulus of the error of the estimated
coefficients of G(z) and C(z), respectively, tend to present
higher values as the loss of samples grows. It can also be
confirmed by the increase of the standard deviation of the
coefficients of G(z) and C(z) presented in Table II. However,
according with Figure 5, the mode of this errors remain close
to zero for all considered rates of sample loss.
C. Results of the Service Degradation Attacks
In this section, there are presented the results obtained
through simulations of SD-Controlled Data Injection attacks,
6
performed by a MitM acting in the control link of the NCS, as
shown in Figure (3). The attacks were simulated in MATLAB,
aiming to evaluate their accuracy when planned based on the
results provided in Section VI-B, obtained by the System
Identification attack. There were performed two sets of attack.
The first one, aims to cause an overshoot of 50% in the
rotational speed of the motor. The second one, aims to cause a
stationary error of −10% in the rotational speed of the motor
when it is on the steady state.
In the attack aiming the overshoot, the function executed by
the attacker is M (z) = Ko . Performing a root locus analysis
considering the obtained models, the attacker adjusts Ko to
make the system underdamped, with a peak of rotational speed
50% higher than its steady state speed. The values of Ko are
adjusted considering the average of the coefficients estimated
in Section III-B. Table III shows the values of Ko , estimated
considering different rates of sample loss during the System
Identification attack, as well as the overshoots obtained with
the respective Ko in the real model. In Figure 6 it is possible
to compare the response of the system without attack, with the
response of the system with an attack aiming the overshoot of
50%. It is also possible to note that the attack to the actual
model presents, in the time domain, a response quite similar
to the attack estimated with the model obtained by the System
Identification attack. This can be verified not only in the case
where the system is identified with 0% of sample loss, but also
in the worst considered case, i.e. with 20% of sample loss. It
is worth mentioning that all responses presented in Figure 6
converge to 1 rad/s.
In the attack where objective is to cause a stationary error
of −10% on the rotational speed of the motor, the attacker
executes (5):
M (z) =
KEss (z − 1)
,
z − 0.94
(5)
wherein KEss is adjusted based on the data obtained with
the System Identification attack. The pole of M (z) is added
aiming to allow a stationary error in the system. The zero
of M (z) is intended to format the root locus in order to
guarantee the existence of a stable KEss that leads the system
to a stationary error of −10%. Table III shows the KEss
resultant from different rates of sample loss during the System
Identification attack, as well as the stationary errors obtained
with the respective KEss in the real model.
According with the data presented in Table III, it is possible
to state that the SD-Controlled Data Injection attack, designed
based on the data gathered by the System Identification attack,
is capable to modify, in an accurate way, the response of the
physical system, considering all the evaluated rates of sample
loss. In the worst case, i.e. with 20% of sample loss, it is
obtained an overshoot of 45.94% and a stationary error of
−9.8%, quite close to the desired values of 50% and −10%,
respectively. Such accuracy allows the attacker to keep his
offensive under control, leading the system to a behavior that
is predefined as physically covert and capable to degrade the
service performed by the plant under attack.
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
0.4
Actual
0.35
0.3
0.25
g2
0.3
0.25
0.2
0.15
−1,5
Actual
−1,45
0.2
0.1
0.1
−1,3
0.05
−1,25
0
5
0
20
10
0
5
−1,2
20
10
(a) g1 of G(z)
0.18
0.5
0.17
0.45
0.165
−0,18
0.4
−0,175
Actual
−0,17
0.16
0.155
0.35
−0,16
−0,155
0.3
0.15
−0,15
0.25
0.145
−0,145
5
20
10
Actual
−0,165
c2
c1
0.175
20
10
(c) g3 of G(z)
Actual
0
5
Loss of samples (%)
(b) g2 of G(z)
0.55
0.2
0
Loss of samples (%)
Loss of samples (%)
g4
−1,4
−1,35
0.15
0.05
0
Actual
−1,55
g3
0.35
g1
−1,6
0.4
0
Loss of samples (%)
5
20
10
0
5
Loss of samples (%)
(d) g4 of G(z)
20
10
Loss of samples (%)
(e) c1 of C(z)
(f) c2 of C(z)
Figure 4. Mean, with a CI of 95%, of the estimated coefficients of G(z) and C(z), in face of different rates of sample loss.
Table II
S TATISTICS OF THE RESULTS OBTAINED WITH DIFFERENT RATES OF SAMPLE LOSS
Loss of
samples
0%
5%
10%
20%
Loss of
samples
Mean
g1
g2
0.32793
0.31835
0.30473
0.26963
0.29652
0.29689
0.30461
0.33352
g1
g2
g3
g4
-1.54121
0.55983
-1.54251
0.56085
-1.54110
0.55925
-1.53119
0.54916
Skewness(*)
g3
g4
c1
c2
g1
g2
0.16991
0.16997
0.16999
0.16989
-0.16712
-0.16719
-0.16724
-0.16716
0.03097
0.07572
0.08781
0.14120
0.04288
0.11523
0.13483
0.22378
c1
c2
0%
-1.21214
1.23278
1.75298
-1.73202
-0.64331
0.79458
5%
-2.34607
1.64875
1.35284
-1.41346
-0.42288
0.36037
10%
-2.52938
1.97711
1.18018
-1.26045
-0.23379
0.13377
20%
-3.24122
1.75186
1.68335
-1.71055
-0.40055
0.37927
nd
(*) Computed in accordance with the Pearson’s 2
Coefficient of Skewness.
(a) Distribution of |Eg |
Figure 5. Histograms of |Eg | and |Ec | in face of different rates of sample loss
g1
g2
0.18846
0.08094
0.16833
0.21292
0.19433
0.10527
0.17123
0.21127
Standard deviation
g3
g4
c1
c2
0.00986
0.00944
0.03322
0.03194
0.04076
0.03922
0.08596
0.08313
Kurtosis
g3
g4
0.00167
0.00287
0.00397
0.00596
0.00178
0.00287
0.00399
0.00598
c1
c2
0.21259
0.09412
0.25041
0.25054
0.15119
0.02540
0.24361
0.23883
0.16472
0.03118
0.23429
0.24441
0.21218
0.09802
0.24811
0.24932
(b) Distribution of |Ec |
7
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
1.5
1.5
Without attack
Estimated attack
Actual attack
Rotational speed (rad/s)
Rotational speed (rad/s)
Without attack
Estimated attack
Actual attack
1
0.5
0
8
0
0.2
0.4
0.6
0.8
1
Time (s)
1
0.5
0
0
0.2
0.4
0.6
0.8
1
Time (s)
(a) Attack based on the data obtained without loss of samples
(b) Attack based on the data obtained with 20% of sample loss
Figure 6. Response of the system to attacks planned to cause an overshoot of 50% in the rotational speed of the motor.
Table III
VALUES OF Ko , KEss
AND THE RESULTS OBTAINED WITH THE ATTACKS
Sample loss in the System Identification attack
0%
5%
10 %
20 %
Ko
Overshoot in the real model
KEss
Stationary error in the real model
4.0451
48.90 %
5.7471
−10%
4.0745
49.43 %
5.7803
−10%
4.0828
49.57 %
5.8140
−9.9%
3.796
45.94 %
5.8823
−9.8%
VII. C ONCLUSIONS
This work proposes a physically covert attack for service
degradation, which the performance depends on the knowledge
about the model of the plant under attack and its controller.
To obtain such knowledge, it is proposed a System Identification attack, based on the BSA algorithm. The effectiveness
of the System Identification attack is demonstrated and its
performance is statistically analyzed in face of different rates
of sample loss. The results achieved by the physically covert
attacks for service degradation, designed based on the data
gathered by the System Identification attack, demonstrate the
high degree of accuracy that may be achieved with the joint
operation of the two attacks. In the worst case, i.e. with 20%
of sample loss during the System Identification attack, the
attacker attained an overshoot of 45.94% and a stationary
error of −9.8%, quite close to the desired values of 50% and
−10%, respectively. In both physically covert interventions,
the accuracy of the attacks ensure that they will not evolve to
unwanted behaviors, physically perceivable.
As future work, it is encouraged the research of techniques
capable to avoid, or complicate, physically convert attacks
planned with the data obtained by System Identification attacks. In this sense, we plan to further investigate countermeasures capable to make it difficult to obtain information
about cyber-physical control systems, which is essential for
planning covert and controlled attacks.
R EFERENCES
[1] Y. Tipsuwan, M.-Y. Chow, and R. Vanijjirattikhan, “An implementation
of a networked pi controller over ip network,” in Industrial Electronics
Society, 2003. IECON’03. The 29th Annual Conference of the IEEE,
vol. 3. IEEE, 2003, pp. 2805–2810.
[2] R. A. Gupta and M.-Y. Chow, “Networked control system: overview and
research trends,” Industrial Electronics, IEEE Transactions on, vol. 57,
no. 7, pp. 2527–2535, 2010.
[3] L. Zhang, L. Xie, W. Li, and Z. Wang, “Security solutions for networked
control systems based on des algorithm and improved grey prediction
model,” International Journal of Computer Network and Information
Security, vol. 6, no. 1, p. 78, 2013.
[4] A. A. Farooqui, S. S. H. Zaidi, A. Y. Memon, and S. Qazi, “Cyber
security backdrop: A scada testbed,” in Computing, Communications and
IT Applications Conference (ComComAp), 2014 IEEE. IEEE, 2014, pp.
98–103.
[5] M.-Y. Chow and Y. Tipsuwan, “Network-based control systems: a
tutorial,” in Industrial Electronics Society, 2001. IECON’01. The 27th
Annual Conference of the IEEE, vol. 3. IEEE, 2001, pp. 1593–1602.
[6] M. Long, C.-H. Wu, and J. Y. Hung, “Denial of service attacks
on network-based control systems: impact and mitigation,” Industrial
Informatics, IEEE Transactions on, vol. 1, no. 2, pp. 85–96, 2005.
[7] R. Langner, “Stuxnet: Dissecting a cyberwarfare weapon,” Security &
Privacy, IEEE, vol. 9, no. 3, pp. 49–51, 2011.
[8] P. Civicioglu, “Backtracking search optimization algorithm for numerical
optimization problems,” Applied Mathematics and Computation, vol.
219, no. 15, pp. 8121–8144, 2013.
[9] A. C. Snoeren, C. Partridge, L. A. Sanchez, C. E. Jones, F. Tchakountio,
B. Schwartz, S. T. Kent, and W. T. Strayer, “Single-packet ip traceback,”
IEEE/ACM Transactions on Networking (ToN), vol. 10, no. 6, pp. 721–
734, 2002.
[10] A. Teixeira, I. Shames, H. Sandberg, and K. H. Johansson, “A secure control framework for resource-limited adversaries,” Automatica,
vol. 51, pp. 135–148, 2015.
[11] R. Smith, “A decoupled feedback structure for covertly appropriating
networked control systems,” in Proceedings of the 18th IFAC World
Congress 2011, vol. 18, no. 1. IFAC-PapersOnLine, 2011.
[12] R. S. Smith, “Covert misappropriation of networked control systems:
Presenting a feedback structure,” Control Systems, IEEE, vol. 35, no. 1,
pp. 82–92, 2015.
[13] A. Hussain, J. Heidemann, and C. Papadopoulos, “A framework for
classifying denial of service attacks,” in Proceedings of the 2003
conference on Applications, technologies, architectures, and protocols
for computer communications. ACM, 2003, pp. 99–110.
[14] C. Ramos, Z. Vale, and L. Faria, “Cyber-physical intelligence in the context of power systems,” in Future Generation Information Technology.
Springer, 2011, pp. 19–29.
[15] S. Khatri, P. Sharma, P. Chaudhary, and A. Bijalwan, “A taxonomy
of physical layer attacks in manet,” International Journal of Computer
Applications, vol. 117, no. 22, 2015.
[16] Y. Zou and G. Wang, “Intercept behavior analysis of industrial wireless
sensor networks in the presence of eavesdropping attack,” IEEE Transactions on Industrial Informatics, vol. 12, no. 2, pp. 780–787, 2016.
[17] A. O. de Sá, N. Nedjah, and L. de Macedo Mourelle, “Distributed
efficient localization in swarm robotic systems using swarm intelligence
algorithms,” Neurocomputing, vol. 172, pp. 322–336, 2016.
FINAL VERSION OF THIS PAPER AVAILABLE AT IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (HTTP://DX.DOI.ORG/10.1109/TII.2017.2676005)
[18] W. Stallings, Cryptography and network security: principles and practices. Pearson Education India, 2006.
[19] M. El-Sharkawi and C. Huang, “Variable structure tracking of dc
motor for high performance applications,” Energy Conversion, IEEE
Transactions on, vol. 4, no. 4, pp. 643–650, 1989.
[20] T. Tran, Q. P. Ha, and H. T. Nguyen, “Robust non-overshoot time
responses using cascade sliding mode-pid control,” Journal of Advanced
Computational Intelligence and Intelligent Informatics, 2007.
[21] H. Hwang, G. Jung, K. Sohn, and S. Park, “A study on mitm (man in
the middle) vulnerability in wireless network using 802.1 x and eap,” in
Information Science and Security, 2008. ICISS. International Conference
on. IEEE, 2008, pp. 164–170.
9
| 3cs.SY
|
Even faster sorting of (not only) integers
Marek Kokot, Sebastian Deorowicz, and Maciej Dlugosz
arXiv:1703.00687v1 [cs.DS] 2 Mar 2017
Institute of Informatics, Silesian University of Technology,
Akademicka 16, 44-100 Gliwice, Poland
{marek.kokot,sebastian.deorowicz,maciej.dlugosz}@polsl.pl
Abstract. In this paper we introduce RADULS2, the fastest parallel
sorter based on radix algorithm. It is optimized to process huge amounts
of data making use of modern multicore CPUs. The main novelties include: extremely optimized algorithm for handling tiny arrays (up to
about a hundred of records) that could appear even billions times as
subproblems to handle and improved processing of larger subarrays with
better use of non-temporal memory stores.
1
Introduction
Sorting is one of the fundamental tasks in computer science. Despite the fact
that the problem was discussed many times from both theoretical and practical
perspective, it remains open and improved solutions are still proposed. The main
reason of its popularity is a wide range of applications. Frequently sorting is an
important component of other algorithms. Ordered sequences may often be faster
processed or stored in compacted form. Sorting algorithms may be divided into
two groups: comparison based and non-comparison based. In the first group, the
elements of input sequence are compared in pairs, which implies the lower bound
of the number of comparisons Ω(n log n) [12]. In the second group, no comparisons are made, thus the lower bound may be relaxed to Ω(n). Unfortunately,
non-comparison attempt requires additional assumptions about the data, which
makes this group of algorithms less general. Despite the algorithm category, the
technical details of implementation have crucial meaning in practice.
One of such aspects is memory access pattern. In modern computer architectures, processors are equipped with small amount of fast cache memory, which
minimizes the negative impact of high latency of the main memory access. The
cache is usually divided into a couple of levels: L1, L2, L3. When a processor
core needs to load a particular memory location, it is first checked if the required
address is already in the cache (checking starts from the lowest level and continues to the highest one). If the required address is absent in the cache it must
be loaded from the main memory. Such a situation is called a cache miss [5].
Algorithms should be constructed to avoid cache misses whenever possible.
The second important factor on the way to fast algorithms is parallelization.
Nowadays even desktop computers and laptops are equipped with multicore processors, which may be used to increase performance. In practice only for simple
algorithms optimal speedup is easily achievable. For the complex ones it is often
2
Marek Kokot, Sebastian Deorowicz, Maciej Dlugosz
hard to keep the desired speedup when the number of involved cores increases.
Combining both cache-friendly memory access and parallelization introduces
new issues that need to be addressed. In a multicore system each core has its
own L1 cache, while L2 may be assigned to a single core or may be shared among
a group of cores. L3 is typically common to all cores in a chip. The data in the
cache is organized in lines (typically of 64 bytes). When two cores operate on a
separate part of the same cache line and at least one of them performs writing,
the slow synchronization occurs. Such a situation is called false sharing [10] and
similarly as cache misses should be avoided.
The other important factor is understanding the cost of branch mispredictions [8]. Modern processors are equipped with long pipelines [19]. For each
conditional instruction processor predicts how it will be evaluated and chooses
the next instructions to load into pipeline. If the actual result of comparison is
different than predicted then branch misprediction occurs and the whole pipeline
must be cleaned, which introduces long delay. In some cases it is possible to avoid
branch mispredictions using conditional assembly instructions, e.g., CMOVcc [6].
Radix sort [4] is a non-comparison technique and has the time complexity
O(kn), where n is the number of elements to be sorted and k is the length of keys
in digits. The digit is a fixed-size part of a key, e.g., in case of sorting strings a
single character. In general radix sort works in k phases, each phase is related to
sorting keys according to the current digit. Digits are usually from a small alphabet and counting sort may be applied per single phase. There are two categories
of radix sort depending on the order in which digits are processed. The first variant, LSD (least significant digit), starts sorting whole data from the least significant digit and continues to the most significant one. It is required to use a stable
sort per each phase to ensure that the final result is valid. The second variant,
MSD (most significant digit), works in the opposite direction and starts sorting
the data according to the most significant digit. This results in partitioning data
into bins that may be further processed independently in the recursive manner.
The MSD radix has an important advantage over the LSD variant. The bins
created after each phase tend to shrink (except bins with all keys equal). After a
few phases the bins become tiny and switching to other algorithm improves the
overall performance. A common approach here is to use some comparison-based
sorting algorithm, e.g., insertion sort [15]. Typically, the time of handling such
tiny bins could be from 10 to 50 percent of total time of radix sort.
Sorting tiny arrays is a special case, which should be considered separately.
Often algorithms with poor time complexity perform well for tiny arrays. Therefore many complex sorting algorithms, like introspective sort [16], uses them as
subroutines. Yet tiny array sorting seems to be not fully investigated.
Recently we proposed RADULS [13], the fastest existing radix sort. In this
paper we present its significantly improved successor. The main novelties in
RADULS2 are motivated by the following investigations. Firstly, we considered
the special case of tiny array sorting. We proposed a new hybrid algorithm, which
is significantly faster than the hybrid of insertion sort, Shell sort, and introspective sort used in RADULS. It also outperforms insertion sort used in common
Even faster sorting of (not only) integers
3
std::sort function implementations in the C++ standard libraries. Secondly, we
optimized the last phase of radix sort, in which the subarrays are small, but not
so small to satisfy the tiny bin criterion. Thirdly, we improved the processing of
small bins to take the most of non-temporal memory stores present at modern
CPUs.
The paper is organized as follows. In Section 2 we give a short description
of existing radix sorters and algorithms that are capable to sort tiny arrays
efficiently. Our algorithm is described in Section 3. The experimental results are
presented in Section 4. In the last section we conclude the paper.
2
2.1
Related works
Radix sorters
Probably the fastest version of LSD radix sort was proposed by Satish et al. [17].
The main idea is to divide the scatter phase into two steps to make it cachefriendly. Before saving to the main memory, each record is stored into a cache
buffer. When the buffer for some digit value becomes full, it is transfered using
a non-temporal memory store to the main memory. The size of buffer per single
digit value is a multiple of 64 B (size of a single memory transfer). This approach
increases bandwidth utilization and reduces number of cache misses.
In [2] an efficient MSD radix sort was proposed. Its in-place nature results in
difficulty of parallelization. The algorithm performs series of element swaps to
sort them according to the current digit. Unfortunately, there are high dependencies between the operations, which make them hard to perform in parallel.
The authors introduced speculative permutation and repair phases to address
this problem. Another issue occurring is significant imbalance of bins sizes. The
authors handles this by a distribution-adaptive load balancing.
In [13] we proposed variant of MSD radix sort. The main idea is to perform
appropriate optimization depending on the current phase and size of currently
processed bin. For instance, if the current bin is big enough, a buffered version
of scattering inspired by Satish et al. is used. When the current bin is tiny (say
contains tens of elements), some comparison-based sorting algorithm is chosen.
2.2
Tiny array sorters
Sorting tiny arrays is not as widely discussed in the literature as the case of big
ones. For most applications it is irrelevant which sorting algorithm is chosen if
the number of records is low. Still many sorters stops their default procedure and
switch to a simpler algorithm when the number of remaining records is below
given threshold. Such a situation may occur billions of times for large inputs.
Thus, improving tiny sorting procedure will also improve overall running time
of given sorting algorithm. For example introspective sort [16], which is usually
a part of standard C++ library is a hybrid of quicksort [9] and heapsort [21],
yet for tiny ranges it switches to insertion sort.
4
Marek Kokot, Sebastian Deorowicz, Maciej Dlugosz
In the case of sufficiently tiny arrays all the data are usually in L1 cache,
so there are no cache misses, and the cost of memory accesses is low. Thus, the
dominant factors that have impact on the running time are: (i ) simplicity of
the algorithm, that usually is reflected in the small number of jump operations,
(ii ) number of comparisons, and the cost of single comparison, (iii ) number of
assignments. As the assignment is usually a simple (branchless) operation it is
faster than comparison, which, especially for longer than 8 B keys, are hard to
be made without jumps.
As said above, insertion sort is usually taken as a “tiny sorter”. Shell sort [18]
being its generalization is also an obvious candidate. Sorting networks [1] are constructed with series of so-called comparison elements or comparators. Each such
element performs a conditional swap of its arguments. The number of such operations, as well as the indices of the compared records, are fixed for a given input
size, i.e., they are independent on the ordering of input elements. This allows the
compiler to hard-code the indices in the code, so the addressing is faster than
when the index must be loaded from a variable. There are two criteria of sortingnetwork optimality. The first one is related to its depth defined as the number
of layers of comparators (a layer is a set of comparators which may perform
operations independently). The second criteria is sorting network size defined as
the total numer of comparators. Optimal-depth sorting network are known only
up to 17 records [7], optimal-size networks are known only up to 10 records [3].
In the enumeration sorting [12] the final position of each element is determined by comparing it against each other. The number of comparisons is
O(n2 ), but the number of element moves is only O(n). Even if this algorithm
requires high number of comparisons, they may be implemented branchless with
SETcc assembly instruction. It allows casting boolean comparison result to integer value, which may be further used to update final position of the element.
3
3.1
Our algorithm
General idea
Our algorithm is an MSD radix sort with 1 B digits. It follows the same general
scheme as its predecessor [13]:
1. Parallel processing of the whole array according to the first digit in a cacheefficient way.
2. Serial processing of small bins obtained in the previous stages, in a cacheefficient way. Several bins are processed by separate threads in the same time.
3. Serial processing of tiny bins (containing up to tens or hundreds of records)
using a hybrid of insertion sort, Shell sort, introspective sort. Several threads
process different bins in parallel.
The main improvements in this paper are in the small and tiny bin processing
subroutines. These phases constitute the majority of RADULS processing time,
so the impact of the proposed novelties on the total sorting time is remarkable.
Even faster sorting of (not only) integers
3.2
5
Small bin processing
Small bins are processed in a decreasing order of their sizes. Each thread handles
a single bin at a time. There are two cases here. In the first one, the bin is small
enough to fit in a half of L2 cache, so cache misses are unlikely. In such a situation
simple counting and scatter approach is chosen.
Larger bins will likely do not fit into cache, so many cache misses can be expected. Thus, a buffered version similar to the first phase (motivated by Satish et
al. [17]) is chosen. Nevertheless, there are some traps here. In the first pass it can
be assured that the input array is aligned to 64 B which allows to use efficient
non-temporal memory storage operations. In the following stages, the beginning
address of subarray representing bin is likely unaligned. Thus, non-temporal
memory storage operations used in RADULS are inefficient here. Depending on
the CPU architecture they are a few times slower than in the case of aligned
address. (Usually, the difference is larger for AMD Opteron than Intel Xeon.)
To solve this problem in RADULS2 we implemented special alignment of
subarrays. The beginning of the aligned bin address can be localed in other bin.
Nevertheless, the transfers from buffers (located in cache) to the main memory
are made in such a way that other bins are not overwritten. Such solution resulted
in about two times faster scatter step than in RADULS.
3.3
Tiny bin processing
Taking into account the results of our experiments on “tiny sorters” (discussed
in Section 4) we constructed a hybrid algorithm that is able to handle tiny arrays
efficiently. The algorithm exploits advantages of sorting networks (SN) [1] and
enumeration sort (ES) [12] for the tiniest arrays, while for larger ones it uses our
modification of BlockQuicksort (BQS) [6] in which we replaced insertion sort for
the tiniest arrays by sorting network or enumeration sort.
For the smallest records (i.e., containing just 8 B key) the hybrid switches at
n = 64 elements from SN to BQS with SN for the tiniest arrays. The sorting
networks for n > 64 seem to be of almost the same speed as BQS, but the source
code for SN becomes long (and also the code must be prepared for each array
size), which leads to long compilation and large executable.
If the data field exists the situation becomes more complex as conditional
swaps of records in the array takes longer time, and often cannot be made without
jumps (or at least compilers tend to prepare the code with branches). Therefore,
ES becomes more attractive, as it makes just n (branchless) assignments. Even
the fact that ES makes O(n2 ) comparisons seems to be acceptable (at least for
tiniest arrays) if we take into account that the comparisons are branchless. Therefore, for 16 B or 24 B data fields the hybrid employs ES when the number of elements is not larger than 24 and BQS with ES for the tiniest subarrays otherwise.
The most complex situation is the case of 8 B data field. For n ≤ 16 the
hybrid employs the ES, but then, up to 32 records, the SN is used. It seems
that the quadratic number of comparisons of ES becomes the problem for larger
arrays. For n > 32 BQS with SN is used.
6
Marek Kokot, Sebastian Deorowicz, Maciej Dlugosz
The situation of 16 B keys and no additional data is different as it is hard
to implement the comparison of elements in a branchless way. Therefore, the
proposed hybrid uses ES according to the most significant 8 B word (to make
use of branchless comparisons) and then fix the ordering using IS. For n > 24 we,
however, employ BQS with the above ES+IS as a dispatch for tiniest subarrays.
In our general radix procedure we analyze each newly created bin to make
a decision about its further processing. There are bins that do not meet the
requirements of being tiny, but our general radix procedure introduce observable
overhead due to unnecessary analysis of empty subbins. To address this problem
we treat bins of size at most 256 in a special manner. During the histogram
construction we store the identifiers of subbins containing at least 2 elements
that need further processing. Thus, the other subbins can be easily omitted.
4
4.1
Experimental results
Tiny array sorting algorithms
All tiny sorters were implemented in the C++14 programing language. A couple
of SSE instructions were used to improve the performance of sorting network in
case of 16 B records. For compilation we used GCC 6.3.0 and Visual C++ 2015.
The experiments were run on four platforms:
–
–
–
–
Laptop with Intel Core i7-4700MQ CPU clocked at 2.4 GHz (Windows),
Workstation with Intel Core i7-5820K CPU clocked at 3.30 GHz (Linux),
Workstation with two Intel Xeon E5-2670v3 CPUs clocked at 2.3 GHz (Linux),
Server with four AMD Opteron 6320 CPUs clocked at 2.8 GHz (Linux).
The results of sorting tiny arrays of size from 2 to 64 records with the
following algorithms: insertion sort (IS), BlockQuicksort (BQS), enumeration
sort (ES), introspective sort implemented as a part of standard C++ library
(std::sort), sorting network (SN), our hybrid algorithm (Hybrid), enumeration
sort followed by insertion sort (ES+IS) are shown in Fig. 1. The running times
are divided by the size of the sorted array.
The fastest method for records with 8 B key and no data field or 8 B data
field is SN (except for the smallest number of records—up to 10). Yet it is worth
mentioning, that our implementation of sorting network for 16 B records uses
SSE to allow elements swapping by branchless code. Without such optimization
SN performed much worse as compilers tend to produce the code with branching
instructions. The cost of swapping the elements is easy observable in case of 32 B
records—in such a situation ES clearly outperforms SN.
The case of 16 B keys and no additional data is quite interesting. The comparison of elements cannot be performed in a single step, thereby maintaining
branchless code is harder. Our implementation of SN suffers in such situation, yet
it is observable that Visual C++ is able to perform better optimizations, which
leads to much lower runtime than in case of GCC. Our two-stage technique helps
to remove branch mispredictions overhead, in fact our hybrid algorithm in such
situation is SN according to 8 B key and 8 B data field followed by IS.
Even faster sorting of (not only) integers
key=8B, record=8B, Windows, Intel i7
40
BQS
Time [ns]
30
ES
SN
std::sort
Hybrid
10
10
20
40
60
key=8B, record=16B, Linux, Intel i7
40
BQS
30
Time [ns]
30
20
0
0
ES
SN
std::sort
Hybrid
10
20
40
60
0
0
20
40
60
40
40
Time [ns]
60
key=8B, record=32B, Linux, Xeon
key=8B, record=32B, Linux, Opteron
20
20
BQS
0
IS
ES
SN
std::sort
Hybrid
20
40
60
0
0
40
BQS
IS
ES
ES+IS
SN
Hybrid
20
40
60
key=16B, record=16B, Linux, Intel i7
key=16B, record=16B, Windows, Intel i7
Time [ns]
40
30
10
0
20
key=8B, record=16B, Linux, Xeon
40
20
0
0
IS
20
0
key=8B, record=8B, Linux, Xeon
40
IS
20
0
7
40
std::sort
20
20
0
0
20
40
Array size [elements]
60
0
0
20
40
60
Array size [elements]
Fig. 1: Running times (per element) for tiny sorters. The examined algorithms:
BlockQuicksort (BQS), insertion sort (IS), enumeration sort (ES), sorting networks (SN), introspective sort from C++ (std::sort), enumeration sort according
to most significant 8 B word followed by insertion sort (ES+IS), our hybrid algorithm (Hybrid). The legends are the same for both charts of each row.
8
Marek Kokot, Sebastian Deorowicz, Maciej Dlugosz
The sudden rise of running time of SN above 32 records is an effect of much
higher number of comparison elements. For such values we used Batcher’s OddEven algorithm [1] to generate networks. For smaller values we was able to use
more efficient networks. In case of std::sort similar behavior may be explained
by a threshold of switching to insertion sort.
4.2
RADULS 2
The results of experiments for RADULS 2 and the fastest competitors are given
in Fig. 2. Similarly as in [13] we picked for comparison the following algorithms:
– TBB—the parallel comparison sort implemented in the Intel Threading
Building Blocks [11] (2017 Update 3 release),
– MCSTL—the parallel hybrid sort [20], included in GNU’s libstdc++ library,
– Satish—the buffered LSD radix sort introduced by Satish et al. [17] with the
buffer size for a specific digit equal to 4 times the cache line size,
– PARADIS—the state-of-the-art in-place radix sort algorithm by Cho et al. [2],
– RADULS—the state-of-the-art radix sort algorithm [13].
The main experiments are for 16 threads for 16 B records (8 B key and 8 B
data). As we can see RADULS2 is the fastest sorting algorithm in all the experiments. For 1 G elements it is 23% (uniform distribution) and 28% (Zipf
distribution) faster than RADULS when run at Xeon-based workstation. At
Opteron-based server the advantage is 18%. The scalability experiments show
that RADULS2 scales almost in the same way as RADULS.
We also evaluated the behavior of RADULS2 for records of different sizes
(Fig. 3). RADULS2 was noticeably faster than its predecessor, however, the
difference becomes smaller for longer records. It is also interesting to compare
the relative throughput for different platforms. Opterons were clocked higher
(2.8 GHz) than Xeons (2.3 GHz) but sorted the data about two times slower.
RADULS is used in KMC3 [14], our k-mer counter for genomic data. The
excellent performance of KMC3 is in significant part due to fast sorter. When
sorting KMC set from [13] the gain in throughput was 17% (413 M 16 B-key
records) and 10% (1887 M 16 B-key records).
5
Conclusions
In this paper we examined sorting algorithms for tiny arrays, as well as we proved
the relevance of this topic by applying proposed hybrid solution to RADULS—
our radix sort algorithm. We also proposed other improvements, which—together
with “tiny sorters”—constitute new version, i.e., RADULS2. Our experiments
confirm the significance of hardware-aware programming, especially considering
cache memory and avoiding branch mispredictions.
Acknowledgement
The work was supported by the Polish National Science Center upon decision
DEC-2015/17/B/ST6/01890.
Even faster sorting of (not only) integers
Throughput [M elements/s]
Uniform random (Xeon)
300
200
200
TBB
100
MCSTL
Satish
Paradis
RADULS
RADULS2
100
0
1
2
3
Array size [G elements]
0
4
Uniform random (Opteron)
200
Throughput [M elements/s]
Zipf theta = 0.75 random (Xeon)
300
0
9
0
1
2
3
Array size [G elements]
4
Zipf theta = 0.75 random (Xeon)
1,000
150
100
100
50
10
0
0
1
2
3
Array size [G elements]
4
1
2
3
4
6 8
No. threads
12 16
24
Fig. 2: Experimental comparison of best sorting algorithms. The right bottom
chart shows how the algorithm speed ups for growing number of threads. The
remaining charts present throughput for 16 threads when sorting 2 G records of
size 16 B (8 B key and 8 B data).
Throughput [M elements/s]
Uniform random
600
400
200
0
(8,8)
(16,8)
(16, 16)
(24,8)
(32,8)
OPTERON
(8,8)
(16,8)
(16, 16)
(Record size, Key size)
TBB
MCSTL
Satish
RADULS
(24,8)
(32,8)
XEON
RADULS2
Fig. 3: Experimental comparison of algorithms when sorting 2 G records of various sizes using 16 threads.
10
Marek Kokot, Sebastian Deorowicz, Maciej Dlugosz
References
1. Batcher, K.: Sorting networks and their applications. Proceedings of the AFIPS
Spring Joint Computing Conference 32, 307–314 (1968)
2. Cho, M., Brand, D., Bordawekar, R., Finkler, U., Kulandaisamy, V., Puri, R.:
PARADIS: an efficient parallel algorithm for in-place radix sort. In: Proceedings
of the VLDB Endowment. pp. 1518–1529 (2015)
3. Codish, M., Cruz-Filipe, L., Frank, M., Schneider-Kamp, P.: Sorting nine inputs
requires twenty-five comparisons. Journal of Computer and System Sciences 82(3),
551–563 (2016)
4. Cormen, T., Leiserson, C., Rivest, R., Stein, C.: Introduction to algorithms, Third
Edition. MIT press (2009)
5. Drepper, U.: What every programmer should know about memory. Red Hat, Inc
11 (2007)
6. Edelkamp, S., Weiß, A.: Blockquicksort: How branch mispredictions dont affect
quicksort. arXiv preprint arXiv:1604.06697 (2016)
7. Ehlers, T., Müller, M.: New bounds on optimal sorting networks. Lecture Notes in
Computer Science 9136, 167–176 (2015)
8. Fog, A.: The microarchitecture of Intel, AMD and VIA CPUs. an optimization
guide for assembly programmers and compiler makers. (2016), http://www.agner.
org/optimize/microarchitecture.pdf
9. Hoare, C.: Quicksort. The Computer Journal 5(1), 10–16 (1962)
10. Intel: Intel Guide for Developing Multithreaded Application. Intel (2011), http:
//www.intel.com/software/threading-guide
11. Intel:
Threading
Building
Blocks
(2016),
https://www.
threadingbuildingblocks.org/
12. Knuth, D.: The Art of Computer Programming. Addison-Wesley (1968)
13. Kokot, M., Deorowicz, S., Debudaj-Grabysz, A.: Sorting data on ultra-large scale
with RADULS. arXiv preprint arXiv:1612.02557 (2016)
14. Kokot, M., Dlugosz, M., Deorowicz, S.: KMC 3: counting and manipulating k-mer
statistics. arXiv preprint arXiv:1701.08022 (2017)
15. McIlroy, P., K., B., McIlroy, M.: Engineering radix sort. Computing systems 6(1),
5–27 (1993)
16. Musser, D.: Introspective sorting and selection algorithms. Software: Practice and
Experience 27(8), 983–993 (1997)
17. Satish, N., Kim, C., Chhugani, J., Nguyen, A., Lee, V., Kim, D., Dubey, P.: Fast
sort on CPUs and GPUs: A case for bandwidth oblivious SIMD Sort. In: Proceedings of the 2010 ACM SIGMOD International Conference on Management of Data.
pp. 351–362 (2010)
18. Shell, D.: A high-speed sorting procedure. Communications of the ACM 2(7), 30–32
(1959)
19. Shen, J., Lipasti, M.: Modern processor design: fundamentals of superscalar processors. Waveland Press (2013)
20. Singler, J., Sanders, P., Putze, F.: MCSTL: The multi-core standard template
library. Lecture Notes in Computer Science 4641, 682–694 (2007)
21. Williams, J.: Algorithm 232: Heapsort. Communications of the ACM 7(6), 347–348
(1964)
| 8cs.DS
|
Copyright ©2018 IEEE
Accepted to 2018 IEEE International Symposium on Biomedical Imaging
UNSUPERVISED DISCOVERY OF TOXOPLASMA GONDII MOTILITY PHENOTYPES
Mojtaba S. Fazli1 , Stephen A. Vella2 , Silvia N.J. Moreno2 , Shannon Quinn1
1
Department of Computer Science, The University of Georgia, Athens, GA, USA
arXiv:1801.02591v2 [cs.CV] 11 Jan 2018
2
Department of Cellular Biology, The University of Georgia, Athens, GA, USA
ABSTRACT
Toxoplasma gondii is a parasitic protozoan that causes disseminated toxoplasmosis, a disease that afflicts roughly a
third of the worlds population. Its virulence is predicated
on its motility and ability to enter and exit nucleated cells;
therefore, studies elucidating its mechanism of motility and in
particular, its motility patterns in the context of its lytic cycle,
are critical to the eventual development of therapeutic strategies. Here, we present an end-to-end computational pipeline
for identifying T. gondii motility phenotypes in a completely
unsupervised, data-driven way. We track the parasites before
and after addition of extracellular Ca2+ to study its effects
on the parasite motility patterns and use this information to
parameterize the motion and group it according to similarity
of spatiotemporal dynamics.
We linked fluctuations of fluorescence (cytosolic Ca2+ concentration) to motility patterns. Using our algorithm, we
extracted and quantified motility tracks, parameterized the
motion trajectories, and use spectral clustering to identify
distinct motility phenotypes. In particular, we repeat this
procedure for T. gondii parasites both before and after addition of extracellular Ca2+ , which results in Ca2+ influx and
stimulation of motility [5]. This not only allows us to identify
distinct T. gondii motility phenotypes in a completely unsupervised, data-driven way, but also enriches our knowledge of
these phenotypes with respect to the conditions under which
they manifest. Ideally, this work will help pave the way to
understanding key differences in the biology of T. gondii and
mammalian cells and the potential for future therapeutic drug
discovery.
Index Terms— Toxoplasma gondii, computer vision,
tracking, spectral clustering, autoregressive models
2. METHODS
1. INTRODUCTION
Toxoplasma gondii is the causative agent of toxoplasmosis
and is considered one of the most successful parasitic infections. Approximately 1/3 of the worlds population will
test positive for T. gondii, a parasite capable of infecting
any nucleated cell [1]. Though T. gondii infection is usually mild or asymptomatic within in healthy host, infections
are life-long, and the immunocompromised such as AIDS
patients, fetuses, and organ-transplant recipients are at risk
of developing severe complications from the reactivation of
dormant tissue cysts [1, 2, 3]. Ca2+ is a universal signaling
molecule that regulates numerous cellular processes in both
prokaryotes, unicellular and multicellular eukaryotes [4].
T. gondiis pathogenesis is directly linked to its lytic cycle,
comprised of invasion, replication, egress, and motility, and
progression forward is regulated through Ca2+ signaling [5].
Therefore, the motion of T. gondii is critical to understanding its lytic cycle and, ultimately, in developing potential
therapies. Cytosolic Ca2+ oscillations in T. gondii precede
activation/enhancement of each step of the lytic cycle [5],
making these oscillations a critical avenue through which to
analyze T. gondii pathogenesis. Here we present an unsupervised, data-driven pipeline for identifying T. gondii motility
patterns in response to cytosolic Ca2+ . We used extracellular parasites that express in vivo a derivatized variant of the
Green Fluorescent Protein (GFP), GCaMP6f, whose fluorescence is directly proportional to Ca2+ concentration [6].
2.1. Data Acquisition
T. gondii tachyzoites (RH) strain were maintained as described previously using Dulbeccos modified minimal essential media with 1% FBS. GCaMP6f is genetically encoded
calcium indicator whose fluorescence is sensitive to cytosolic
calcium concentration (Ca2+ ) and allows us to track Ca2+ dynamics in motile cells. Construction of GCaMP6f-expressing
parasites was performed by cloning the coding sequence of
GCaMP6f into a ptubP30GFP (a selection-less plasmid and a
kind gift from Boris Striepen). Plasmids were electroporated
into RH parasites and subcloned via cell sorting. Videos
were acquired using an LSM 710 confocal microscope on
10% FBS (Fetal Bovine Serum) coated glass bottom cover
dishes (MATTEK) [5]. Addition of 2M Thapsigargin and/or
1.8 mM CaCl2 was noted. Images were processed using the
FIJI ImageJ software suite [7] and converted to z-stacks for
processing in OpenCV. For our experiments, we select 10
videos with average number of 14 cells in each video.
2.2. Software
We implemented our pipeline using Python 3 and associated
scientific computing libraries (NumPy, SciPy, scikit-learn,
matplotlib). The core of our tracking algorithm used a combination of tools available in the OpenCV 3.1 computer vision library. The full code for our pipeline is available under
the MIT open source license at https://github.com/
quinngroup/toxoplasma-isbi2018 .
Accepted to 2018 IEEE International Symposium on Biomedical Imaging
Copyright ©2018 IEEE
2.3.3. Parameterization of normalized KLT trajectories
2.3. Computational Pipeline
2.3.1. Tracking Algorithm
We used a Kanade-Lucas-Tomasi (KLT) tracker to develop
trajectories for T. gondii parasites. KLT is an optical flow
approach which uses spatial intensities to conduct a focused search of high-probability object locations in the future
[8][9][10]. It is fast, requiring few matches between sequential images. Here, we identify features extracted from objects
in each frame and process them through the KLT module to
be tracked over the entire video. The resulting tracker has
been shown to be robust in our previous work (Fig. 1) [11].
Performing unsupervised discovery of T. gondii motility phenotypes required featurizing the trajectories of each object in
such a way as to be invariant to the absolute spatial locations,
but sensitive to relative changes over time. Therefore, we used
a simple linear model that encodes a Markov-based transition
prediction method: an autoregressive (AR) model. AR models are a type of stochastic process, used heavily in modeling
linear systems and predicting future observations as a function of some number of previous ones; in this way, it implements the Markov assumption, in which future observations
are independent of all others, conditioned on some number of
immediate preceding observations. The number of prior observations used to make future predictions is referred to as the
order of the system, and follows as:
p
X
xt =
Aj xt−j + C
(1)
j
Fig. 1. (left) Object tracking results using KLT. (right) 3D visualization of a specific cells trajectory. Hue shows the fluorescence of the object, corresponding to Ca2+ concentration.
2.3.2. Preprocessing and normalization of KLT trajectories
KLT trajectories for each parasite object comprised a twodimensional position vector for each frame. In order to standardize these data to be used in downstream analysis, we employed numerous preprocessing steps. First, we discarded any
trajectories below a certain minimum number of frames. This
happened if the KLT tracker lost an object (e.g., due to occlusion, movement outside plane of focus). Second, we split the
trajectories of each object into two sets, capturing the object
trajectory before and after the addition of extracellular Ca2+ ,
to be analyzed distinctly. Third, for each trajectory in the
two sets, we truncated the number of frames to be equivalent
across all objects.
After applying these preprocessing steps, we ended with
a corpus containing 139 trajectories, each capturing the twodimensional movement of distinct T. gondii parasites. Each
trajectory consisted of 300 frames, half before and half after
the addition of extracellular Ca2+ . Our data were represented
in the following structures, X and Y (representing x and y
coordinates of each object in each frame, respectively):
x(0,0) . . .
x(0,149)
..
..
X(139,150) = ...
.
.
x(138,0)
...
y(138,149)
y(0,0)
..
= .
y(138,0)
...
..
.
...
y(0,149)
..
.
y(138,149)
Y(139,150)
where Eq. 1 denotes a p-order system [12]. In our work,
we are interested in the transition matrices {A1, ..., Ap},
which are quantitative encodings of how the system evolves,
and therefore represents a parameterization of the system that
is not sensitive to absolute spatial positions. In our work, we
found a system of order p = 5 worked well, whereby each
139x150x2 object was represented as five 2x2 AR motion
parameters:{A1, A2, A3, A4, A5}. These parameters were
flattened and concatenated into a single 20-element feature
vector, and this operation was repeated for each of the 139
objects.
2.3.4. Establishing pairwise motion similarity
The AR motion parameters are powerful quantitative representations of movement that are invariant to absolute spatial
coordinates. Unfortunately, they are geodesics that do not
span a Euclidean space [13], and therefore are not amenable
to more traditional unsupervised pattern-recognition strategies such as K-means that rely on Euclidean-based pairwise
distance metrics. Instead, we constructed several custom kernel matrices to capture pairwise similarity, with the goal of
using these matrices as graph embeddings to use in spectral
clustering. To compute a pairwise similarity (kernel) matrix,
we stacked the AR parameters learned from the before and
after sequences of T. gondii motility to form two matrices:
ARb(0,0) . . .
ARb(0,19)
..
..
..
ARBef ore(139,20) =
.
.
.
ARb(138,0)
...
ARa(0,0)
..
=
.
ARa(138,0)
...
..
.
...
ARAf ter(139,20)
ARb(138,19)
ARa(0,19)
..
.
ARa(138,19)
We then computed pairwise affinities, or similarities, using
the radial-basis function (RBF) or Gaussian kernel. This kernel matrix A has the form:
Accepted to 2018 IEEE International Symposium on Biomedical Imaging
Copyright ©2018 IEEE
clustering results. The left column indicates trajectories from
before the addition of cytosolic Ca2+ , and the right column of
Aij = Aji = e−γ||xi −xj ||
after Ca2+ addition. Each subfigure is a single object, illuswhere γ is a free parameter that designates the width of the
trated by a line showing the extent of its movement over the
Gaussian kernel. Unfortunately, heat kernels are very sen150-frame duration of the segment.
sitive to the choice of gamma, so the value was chosen by
First, we observe that object trajectories are significantly
empirical cross-validation, and selected =0.1 (Fig. 2). Other
richer and more dynamic after the addition of cytosolic Ca2+
kernels were also tested, including trajectory-based kernels
than before; we saw this in Fig. 1 (right), and is wellsuch as spatial covariance [14], angle covariance (using disunderstood that since Ca2+ regulates the T. gondii lytic cycle,
placement angles instead of position vectors), and Pearson
we can expect more activity when those concentrations are
covariance matrices [15]. However, these did not yield lowhigh [4, 5]. Second, and most relevant to the goal of this
dimensional embeddings of sufficient rank to be useful in
work, is that we can observe qualitatively unmistakable simspectral clustering.
ilarity of trajectory dynamics, especially in the after-Ca2+
clusters. Most of the before-Ca2+ trajectories are largely
2.3.5. Spectral clustering to identify similar motility phenoimmotile (Fig. 3, bottom left), but there are clusters that show
types
similar motion motifs, including full or partial circles (Fig. 3,
Clustering methods are a natural choice for unsupervised
upper left). The after-Ca2+ clusters show similar groupings
learning; they typically impose few assumptions on the data,
of motifs: a clear partial-twirl, sometimes combined with a
defining only a notion of similarity with which to group data
straight trajectory (Fig. 3, lower right), or seemingly rantogether. K-means is a favorite for its speed and simplicdom motion with a long outward arc (Fig. 3, upper right).
ity; unfortunately, in our case, the drawbacks of K-means
Particularly when viewed in context (Fig. 4), we can clearly
make it an impractical option. For one, K-means assumes
Euclidean data; AR motion parameters are geodesics that do
not live in a Euclidean space. For two, K-means assumes
isotropic clusters; it is possible that AR motion parameters
exhibit isotropy in their space, but without a proper distance
metric this cannot be explicitly tested. For these reasons, we
chose spectral clustering. Spectral clustering operates on the
spectrum of the underlying graph of the data [16], imposed
through a similarity computation using a pairwise kernel. Using this affinity kernel, a graph Laplacian is computed, and its
principal eigenvectors are used to embed the original data in
a low-dimensional space where they are separable by simpler
clustering strategies such as K-means. While this requires a
full diagonalization of the Laplacian and can therefore pose
a computational bottleneck with large data, our 139 trajectories is sufficiently small for to be solved by out-of-the-box
spectral clustering implementations, such as scikit-learn [17]
Fig. 3. Clustering results of trajectories before (left column)
and after (right column) addition of cytosolic Ca2+ . Two
sample clusters are shown in each column. Subplots are
aligned to center for easy viewing.
Fig. 2. RBF kernels from AR motion parameters using gamma
= 0.1, from AR motion parameters computed before (left) and
after (right) addition of cytosolic Ca2+ .
3. RESULTS AND DISCUSSION
Using established heuristics from previous work on the T.
gondii parasite and its lytic cycle [5, 11], we chose k = 5
clusters for our analysis. Figure 3 shows a summary of the
identify several distinct T. gondii motility phenotypes such
as helical, twirling, and circularthat are identified through
manual heuristics [1-5].
There are limitations to this approach. First, while the
RBF kernel encodes nonlinear relationships, it still relies on
a Euclidean-style pairwise comparison. Second, despite the
empirical performance of AR parameters in encoding twodimensional spatial movement, this is not an ideal application
of this technique; it is likely that the relative simplicity of
T. gondii motility patterns, and the small number of motion
motifs, are responsible for the relatively good empirical per-
Accepted to 2018 IEEE International Symposium on Biomedical Imaging
Copyright ©2018 IEEE
[3] Jeffrey L Jones, Monica E Parise, and Anthony E Fiore, “Neformance. Third, the combination of small quantity of data
glected parasitic infections in the united states: toxoplasmoand low-resolution temporal models precludes a deeper exsis,” The American journal of tropical medicine and hygiene,
ploration of the manifold of T. gondii motility phenotypes.
vol. 90, no. 5, pp. 794–799, 2014.
Nevertheless, the results strongly suggest such a manifold ex[4] David E Clapham, “Calcium signaling,” Cell, vol. 80, no. 2,
ists, and could be elucidated in future work.
pp. 259–268, 1995.
Fig. 4. Aggregated plotting of trajectories before (A) and after (B) addition of cytosolic Ca2+ . Each object is colored by
its cluster identifier, and placed according to its original absolute spatial location. Here, some trajectories(straight lines
mostly in part B) are artifacts and were excluded from the
final clustering results.
4. CONCLUSION
In this work, we have demonstrated a computational pipeline
for tracking, extracting, parameterizing, and clustering motion trajectories of T. gondii parasites in an initial effort to
elucidate discrete patterns in the parasites lytic cycle. This is
a first step toward a mechanistic understanding of the parasites virulence as a function of its motility patterns. We
have also demonstrated the application of a linear time series
(AR) model to the parameterization of two-dimensional spatial motility, and its effectiveness in clustering for recognizing
gross motion patterns. Given the relatively short tracks of the
T. gondii objects, we felt AR models (i.e., a linear model)
would be sufficient to parameterize motion. This technique
could be applied pretty much as-is to any trajectory analysis
of discrete objects.
In the future, we anticipate refining the elements of this
pipeline to extract finer-grained motility patterns. Specifically, we anticipate deep learning to be the next evolution
of this pipeline: using segmentation and tracking tools such
as Faster RCNN to track the objects, and parameterizing the
motility sequences using recurrent neural networks instead of
linear systems.
5. ACKNOWLEDGMENTS
The authors acknowledge Zhu Hong Li for construction of the
GCaMP6f non-selectable strain. This work was supported in
part by AWS in Education Grant Award. We acknowledge
partial support from the NSF Advances in Biological Informatics (ABI) under award number 1458766.The Toxoplasma
work was funded by the National Institutes of Health (AI110027 and AI-096836 to SNJM).
6. REFERENCES
[1] Geita Saadatnia and Majid Golkar, “A review on human toxoplasmosis,” Scandinavian journal of infectious diseases, vol.
44, no. 11, pp. 805–814, 2012.
[2] Francis Derouin and H Pelloux, “Prevention of toxoplasmosis
in transplant patients,” Clinical Microbiology and Infection,
vol. 14, no. 12, pp. 1089–1101, 2008.
[5] Lucas Borges-Pereira, Alexandre Budu, Ciara A McKnight,
Christina A Moore, Stephen A Vella, Miryam A Hortua Triana, Jing Liu, Celia RS Garcia, Douglas A Pace, and Silvia NJ
Moreno, “Calcium signaling throughout the toxoplasma gondii
lytic cycle a study using genetically encoded calcium indicators,” Journal of Biological Chemistry, vol. 290, no. 45, pp.
26914–26926, 2015.
[6] Tsai-Wen Chen, Trevor J Wardill, Yi Sun, Stefan R Pulver,
Sabine L Renninger, Amy Baohan, Eric R Schreiter, Rex A
Kerr, Michael B Orger, Vivek Jayaraman, et al., “Ultrasensitive fluorescent proteins for imaging neuronal activity,” Nature,
vol. 499, no. 7458, pp. 295–300, 2013.
[7] Johannes Schindelin, Ignacio Arganda-Carreras, Erwin Frise,
Verena Kaynig, Mark Longair, Tobias Pietzsch, Stephan
Preibisch, Curtis Rueden, Stephan Saalfeld, Benjamin Schmid,
et al., “Fiji: an open-source platform for biological-image analysis,” Nature methods, vol. 9, no. 7, pp. 676–682, 2012.
[8] Bruce D Lucas, Takeo Kanade, et al., “An iterative image registration technique with an application to stereo vision,” 1981.
[9] Carlo Tomasi and Takeo Kanade, “Detection and tracking of
point features,” 1991.
[10] Jianbo Shi et al., “Good features to track,” in Computer Vision
and Pattern Recognition, 1994. Proceedings CVPR’94., 1994
IEEE Computer Society Conference on. IEEE, 1994, pp. 593–
600.
[11] Mojtaba Sedigh Fazli, Stephen Andrew Vella, Silvia NJ
Moreno, and Shannon Quinn, “Computational motility tracking of calcium dynamics in toxoplasma gondii,” arXiv preprint
arXiv:1708.01871, 2017.
[12] P. BOURKE,
“Auto-regression analysis (ar),”
http://local.wasp.uwa.edu.au/pbourkke/other/ar/, 1998.
[13] Franco Woolfe and Andrew Fitzgibbon, “Shift-invariant dynamic texture recognition,” Computer Vision–ECCV 2006, pp.
549–562, 2006.
[14] Andrew Gordon Wilson, “Covariance kernels for fast automatic pattern discovery and extrapolation with gaussian processes,” University of Cambridge, 2014.
[15] Bruce Weaver and Karl L Wuensch, “Spss and sas programs
for comparing pearson correlations and ols regression coefficients,” Behavior Research Methods, vol. 45, no. 3, pp. 880–
895, 2013.
[16] Ulrike Von Luxburg, “A tutorial on spectral clustering,” Statistics and computing, vol. 17, no. 4, pp. 395–416, 2007.
[17] Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al.,
“Scikit-learn: Machine learning in python,” Journal of Machine Learning Research, vol. 12, no. Oct, pp. 2825–2830,
2011.
| 1cs.CV
|
arXiv:1711.03002v1 [math.AG] 8 Nov 2017
BERKOVICH LOG DISCREPANCIES IN POSITIVE
CHARACTERISTIC
ERIC CANTON
Abstract. We introduce and study a log discrepancy function on the Berkovich
analytifications of schemes in positive characteristic that shares many properties with the analogue in characteristic zero. We prove several theorems of
Jonsson and Mustaţă regarding log canonical thresholds of graded sequences
of ideals, using birational F -singularity techniques in place of log resolutions.
Strong F -regularity and sharp F -purity of Cartier subalgebras are detected
using log discrepancies of semivaluations in the same way Kawamata log terminal and log canonical singularities are defined using divisorial log discrepancies. We generalize a transformation formula for finite morphisms by Schwede
and Tucker to all valuations. Additionally, we prove that the minimal log discrepancy is constructibly lower-semicontinuous on F -finite schemes in positive
characteristic, a theorem proven for normal varieties in characteristic zero by
Ambro.
Contents
1. Introduction
2. Berkovich spaces
3. Log discrepancies of Cartier prealgebras
4. Connections with F -singularities
5. Lower-semicontinuity
6. Transformation of log discrepancies under finite morphisms
7. Log canonical thresholds of graded sequences of ideals
References
1
4
7
15
16
19
22
33
1. Introduction
We study singularities of Noetherian schemes X → Spec(k) over a field k of characteristic p > 0 using log discrepancies associated to OX -linear maps F∗e OX → OX .
These log discrepancies are assigned to valuations centered on X and trivial on k,
and more generally to such valuations on the function fields of integral subschemes
of X; these semivaluations are the points of the Berkovich space Berk(X). Several
groups of authors have used log discrepancies to study graded sequences of ideals on
X using the structure of the space of valuations Val(X) for X of characteristic zero:
Jonsson and Mustaţă first showed in [JM12] that when X is regular and excellent
over Q, one can realize ValQ (X) as an inverse limit
ValQ (X) = lim QM (Y, D)
←−
(Y,D)
1
2
ERIC CANTON
over the cofiltered collection of log-smooth pairs (Y, D) admitting a proper birational morphism to X. They use this description to extend the definition of log
discrepancy along a divisor over X to every rank-one valuation of X, a technique
extended to singular varieties by Boucksom, de Fernex, Favre, and Urbinati in
[Bou+15].
Consider now a connected normal scheme X → Spec(k) of characteristic p > 0
with fraction field L and finite absolute Frobenius endomorphism F : X → X.
Long have researchers known (or implicitly used, see [MR85; RR85; HW02; Sch09])
that to a nonzero OX -linear map ψ : F∗e OX → OX one can associate an effective
Q-divisor ∆ψ such that KX + ∆ψ is Q-Cartier with index not divisible by p. If
π : Y → X is a proper birational morphism from a normal Y , the transformation
formula
KY + ∆ψ,Y = π ∗ (KX + ∆ψ )
allows one to associate a new OY -linear map ψY : F∗e OY (d(pe − 1)∆ψY e) → OY .
The coefficient on irreducible divisors E ⊂ Y appearing in ∆ψ,Y determine the log
discrepancies A(E, X, ∆ψ ), studied heavily in the minimal model program, see e.g.
[KM98; Kol97; K192; Kol13]. These coefficients are determined algebraically as
e e (a/b)) := (1/b)ψ(F e bpe −1 a)
follows: for a/b ∈ L with a, b ∈ OX , the equation ψ(F
∗
∗
uniquely determines an extension of ψ to
L.
The
log
discrepancy
of
∆
along
E⊂Y
ψ
e e b) =
ordE (b) for b ∈ L such that ψ(F
is equal to the supremum of the values e1
p −1
∗
1. Taking this as our starting point, we define the log discrepancy of maps like ψ,
and, more generally, of graded OX -subsheaves of ⊕HomL (F∗e L, L) closed under selfcomposition, along a semivaluation with compatible home. This definition recovers
many properties usually established using resolutions of singularities for normal
excellent Q-schemes. Our methods do not rely upon such resolutions.
Should we learn that log resolutions exist in positive characteristic, one then
has available the definition of Jonsson and Mustaţă [JM12; Bou+15]. Our main
result is that these definitions will agree, establishing our approach as an alternative formulation making essential use of Cartier subalgebras, sheaves familiar from
birational F -singularity literature.
Main Theorem (Theorem 3.23). Let X be a normal F -finite scheme with Cartier
algebra CX , and suppose log resolutions exist in characteristic p > 0. Let ∆ ≥ 0
be a Q-divisor on X such that KX + ∆ ∼Q 0. For valuations w ∈ Val(X), denote
by AX (w; ∆) the log discrepancy defined as in [JM12] and by A(w; CX · ∆) our log
discrepancy from §3. Then AX (w; ∆) = A(w; CX · ∆) for all v ∈ Val(X).
Essentially by definition, these log discrepancies detect F -singularities. We include a brief study of this relationship, proving that sharp F -purity and strong
F -regularity are characterized using log discrepancies of semivaluations in the same
way that log canonical and Kawamata log terminal singularities are characterized
using divisorial valuations.
Theorem (Theorem 4.4). Let X be integral with trivial valuation trivX : L× → 0.
The Cartier algebra CX is
(a) sharply F -pure if and only if A(ζ; CX ) ≥ 0 for all ζ ∈ Berk(X).
(b) strongly F -regular if and only if A(ζ; CX ) > 0 for all ζ ∈ Berk(X), ζ 6=
trivX .
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
3
A consequence of the proof of this theorem is that when CX is F -pure, its F -pure
centers are exactly those x ∈ X with mldB (x; CX ) = 0.
Valuative measurement of sections of OX mapped to one by φ ∈ CX provides
a numerical analogy between F -splitting techniques and the different ideal of field
extensions from algebraic number theory. Indeed, this circle of ideas is already
illustrated by Schwede and Tucker in [ST14], where they show the ramification
divisor of a dominant, generically finite étale morphism π : Y → X is associated
with the trace of the corresponding finite separable extension of fraction fields.
Inspired by this paper, and drawing heavily from the techniques there, we prove
the log-discrepant version of their transformation formula for finite morphisms.
Theorem (Theorem 6.4). Let π : Y → X be a finite dominant morphism between
F -finite integral schemes. For every nonzero k(X)-linear map T : k(Y ) → k(X)
and every ψ ∈ CX,e , there exists a corresponding T-transpose ψT : F∗e k(Y ) →
k(Y ) [ST14, Definition 5.3] and a T-log ramification function ρ(−; T) on Val(Y )
(Definition 6.2). For every w ∈ Val(Y ) with A(w; ψT ) < ∞ extending v ∈ Val(X)
we have
A(v; ψ) = A(w; ψT ) + ρ(w; T).
The added generality over their setting of normal schemes is afforded by not using
divisors.
One important consequence of the description of Val(X) in [JM12] is that log
discrepancy is lower-semicontinuous on Val(X); we prove this semicontinuity on
the entire Berkovich space. In many ways, it is this shared property that affords
us the similarity with the log discrepancies on valuation spaces in characteristic
zero. As a demonstration of this assertion, we prove the main theorems from
[JM12] regarding valuations computing the log canonical threshold lctB (a? ) of a
multiplicatively graded sequence of ideals a? on a regular F -finite scheme. To state
these results, we write jt for the ideal of those sections f ∈ OX with
v(f ) + A(v; at? ) > 0
for all valuations v centered on X; this ideal is well-known as the multiplier ideal.
Theorem (Theorem 7.18; cf [JM12], Theorem A). Let λ = lctB (a? ) < ∞. For
any generic point x of an irreducible component of V(jλ : q) there exists a valuation
with center x computing lctB (a? ), i.e. such that λ = A(v)/v(a? ).
These computing valuations are obtained here and in [JM12] using a compactness
argument, and their properties (e.g. Abhyankar) do not seem to be revealed from
the proof. Jonsson and Mustaţă conjecture that these valuations must be quasimonomial, a condition equivalent to Abhyankar for excellent schemes in characteristic zero. We state the analogous conjectures here using locally quasi-monomial
valuations, a classification that works better in positive characteristics thanks to
Knaf and Kuhlmann’s local monomialization of Abhyankar valuations [KK05].
Conjecture 1 (Conjecture 7.19; cf. [JM12], Conjecture B). Let X be a regular F finite scheme and let a? be a graded sequence of ideals on X such that lct(a? ) < ∞.
◦ Weak version: there exists a locally quasi-monomial valuation that computes lctB (a? ).
◦ Strong version: any valuation computing lctB (a? ) is locally quasi-monomial.
4
ERIC CANTON
Following [JM12], we reduce this conjecture to what is hopefully a more approachable form on affine spaces.
Conjecture 2 (Conjecture 7.20; cf. [JM12], Conjecture C). Consider Ank , where
k is an algebraically closed field of positive characteristic and n ≥ 1. Let a? be
a graded sequence of ideals vanishing only at a closed point x ∈ Ank , and suppose
lctqB (a? ) < ∞, where 0 6= q ⊆ OAn is an ideal.
◦ Weak version: there exists a locally quasi-monomial valuation computing
lctqB (a? ) and centered at x.
◦ Strong version: any valuation of transcendence degree 0 over Ank , centered
at x, and computing lctqB (a? ) must be locally quasi-monomial.
Theorem (Theorem 7.23; cf. [JM12], Theorem D). If Conjecture 2 holds for all
n ≤ d, then Conjecture 1 holds for all X with dim(X) ≤ d.
We preceed all of this in section two with a brief sketch of the theory of Berkovich
spaces as we need it for this paper; we require only the topological properties, and
so do not discuss the sheaf theory. In section three, we define the log discrepancy
and derive some essential elementary properties. Additionally, we prove in section
five that log discrepancies are lower-semicontinuous on Berk(X), and use this to
deduce constructible lower-semicontinuity on X for any F -finite scheme.
2. Berkovich spaces
By a scheme we always mean a separated Noetherian scheme over a field k, and
a point of a scheme refers to any (not necessarily closed) point. If X is a scheme
and Z ⊆ X is an integral subscheme, we denote by k(Z) the associated function
field of Z, and by OX,Z the local ring at the generic point ηZ of Z. For local
sections f ∈ OX,Z , we write f + IZ for the residue of f in k(Z). If F and G are two
OX -modules, we denote by HomX (F , G ) the OX -module (U 7→ HomU (F |U , G |U )).
2.1. Berkovich spaces. The center and home functions. We provide a summary of the Berkovich theory as we need it, providing sketches of proofs when the
ideas involved are necessary for later sections. We refer to [DFN15] and [Con08]
for more detailed developments of the following material.
Definition 2.1. Let X be a scheme.
(a) A (rank one, k-trivial) semivaluation on X is a pair ζ = (w, Z) consisting
of an integral closed subscheme Z ⊆ X and a valuation w : k(Z) → R∞ :=
R ∪ {+∞}. We require w to satisfy the conditions:
◦ w(k × ) = 0
◦ w is non-negative on some local ring OZ,z ⊆ k(Z).
If f ∈ OX is a section in a neighborhood of ηZ , we define evf (ζ) = ζ(f ) =:
w(f + IZ ).
(b) The subscheme of ζ = (w, Z) is called the home of ζ (on X), and the home
function hX : Berk(X) → X is given by hX (ζ) = ηZ , where ηZ is the generic
−1
point of Z. Note that h−1
X (ηZ ) ( hX (Z) when Z is not a point.
(c) We denote by Berk(X) the set of all semivaluations on X. If X is integral
with generic point ηX , we also define Val(X) to be the subset of all semivaluations on X with hX (ζ) = ηX , i.e. rank one, k-trivial valuations on k(X)
having center on X.
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
5
(d) Since any subscheme Z ⊆ X is separated, for any semivaluation ζ = (w, Z)
on X, there is a unique z ∈ Z such that OZ,z is dominated by the valuation
ring Aw := w−1 [0, ∞]. This z is called the center of ζ on X, denoted cX (ζ).
This defines the center function cX : Berk(X) → X.
(e) When X is clear from context, we will write OX,c(ζ) for OX,cX (ζ) . We follow
a similar convention with hX (ζ).
Remark 2.2. Let N be the nilradical of X. Integral subschemes of X and Xred are
the same, and so Berk(X) = Berk(Xred ) as sets. Moreover, f ∈ N is sent to +∞
by all ζ ∈ Berk(X), so evf = ev0 . Shortly, we will define a topology on Berk(X),
and we see that Berk(X) and Berk(Xred ) also agree as topological spaces.
We now topologize the set Berk(X) to define the Berkovich space of semivaluations on X. We do this in two steps, beginning with the affine case.
Definition 2.3. Suppose X = Spec(R) is a reduced affine scheme. Topologize
Berk(X) via the map
Y
Berk(X) ,→
[0, ∞]
f ∈R
ζ 7→
Y
ζ(f ).
f ∈R
Here we consider [0, ∞] as a compact space. One checks easily that Berk(X) is a
closed subspace of this product space, so is compact by Tychonoff’s theorem. The
Hausdorff property is inherited under this embedding.
Remark 2.4. It is equivalent to give the set Berk(Spec(R)) the weakest topology
such that evf is continuous for all f ∈ R.
Remark 2.5. The center function is well-known to be anticontinuous in the sense
that for Zariski-open subsets U ⊆ X the preimage c−1
X (U ) ⊆ Berk(X) is closed
[JM12]. To see this, note that f 6∈ p if and only if ζ(f ) = 0 when cX (ζ) = p.
Lemma 2.6. For any affine X = Spec(R), hX : Berk(X) → X is continuous.
Proof. We check that h−1
X (V(I)) is closed for all proper ideals I ⊂ R. If ηZ =
hX (ζ) ∈ V(I) and pZ is prime ideal of R corresponding
to ηZ , then I ⊆ pZ . Thus
T
−1
ζ(I) = {+∞}. We conclude h−1
(V(I))
=
ev
(+∞),
which is closed by
X
f ∈I
f
continuity of each evf .
Remark 2.7. Any morphism f : Spec(S) = W → Spec(R) = X between affine
schemes induces a continuous map f∗ : Berk(W ) → Berk(X) by restricting semivaluations, so it follows that when f is an open immersion the subspace topology
on f∗ Berk(W ) agrees with the one defined directly on Berk(W ) as above.
Definition 2.8 (Berkovich space of a scheme). Let X be a scheme and let U and
V be two affine open subschemes of X. Since X is separated, U ∩ V is again affine,
and the topology on Berk(U ∩ B) is identical to the subspace topology induced
from Berk(U ) or Berk(V ). Thus, there is a unique topology on Berk(X) whose
open subsets are those A such that A ∩ Berk(U ) is open for every affine subscheme.
Remark 2.9. As a word of caution, it is eminently not true that Berk(U ) = h−1
X (U )
for open subsets of X. Indeed: supposing that X is integral for simplicity, every
valuation having center on X lives at the generic point of X, and so Val(X) ⊆
6
ERIC CANTON
h−1
X (U ). However, not every valuation having center on X necessarily has center on
−1
U . We do have Berk(U ) = c−1
X (U ) for U ⊆ X open, and Berk(X \U ) = hX (X \U ).
Remark 2.10. Since the topology of Berk(X) is entirely determined by the topology
of Berk(U ) as U ranges over affine subschemes U ⊂ X, checking flavors of continuity
reduces to checking this continuity for affine X. Thus, cX remains anticontinuous
and hX continuous for any X.
2.2. Z-valuations and the D-space of a scheme. Now we develop some basic
properties of a totally disconnected subset of Berk(X) which we call D(X). This
space is discrete in a kind of way, providing a fairly rigid structure that makes the
center map continuous. This space nevertheless sees a great deal of the information
in Berk(X): indeed, D(X) is still large in the sense that the space R>0 D(X) of
R>0 multiples of semivaluations in D(X) is dense in Berk(X), since it contains all
discrete valuations, which are themselves dense.
Definition 2.11. We will say that a valuation w ∈ Val(X) is a Z-valuation if
im(w) ⊆ Z, and denote by D(X) the set of Z-valuations. The Z-semivaluations on
X are the elements in the closure D(X) of D(X) in Berk(X).
Remark 2.12. Just as for Berk(X), we know that D(U ) = D(X)∩Berk(U ) for open
subschemes U ⊆ X. Let us explain this statement now.
First, note that for any open subscheme U ⊆ X we have Berk(U ) = c−1
X (U ),
and so in particular if U = Spec(R) is an affine open subscheme of X then D(U ) =
D(X) ∩ Berk(U ), as both sets are described as Z-valuations on the function field
L of X that are non-negative on R. Since Berk(U ) is closed in Berk(X), the
closure of D(U ) in Berk(X) must be contained in Berk(U ), and so agrees with
D(U ). This implies that D(U ) ⊆ D(X) ∩ Berk(U ). For the reverse inclusion, let
ζ ∈ D(X) ∩ Berk(U ) and let {vα } ⊂ D(X) be a convergent net with limit ζ. Since
cX (ζ) ∈ U , we know ζ(f ) ≥ 0 for all f ∈ R, and so vα → ζ implies that there
exists α0 such that for α ≥ α0 we have vα (f ) > −1. Since vα takes values only
in Z, we conclude that vα (f ) ≥ 0 for all α ≥ α0 and f ∈ R, which is to say that
{vα }α≥α0 ⊂ D(U ), so ζ ∈ D(U ).
The case when Y ( X is a proper closed subscheme is much more difficult. In
general, neither inclusion needed for this equality is clear to the author.
We now conclude this section by proving that the center function becomes continuous with compact fibers upon restriction to D(X).
Lemma 2.13. The center function cD = cX |D : D(X) → X is continuous.
Proof. We may assume X = Spec(R) is affine. Let f ∈ R and ζ ∈ D(X). Then
cD (ζ) ∈ V(f ) if and only if ζ(f ) > 0. Since ζ ∈ D(X), im(ζ) ⊆ Z, and so ζ(f ) > 0
−1
is equivalent to ζ(f ) ≥ 1. Therefore, c−1
D (V(f )) = evf [1, ∞] ∩ D(X) is closed.
Lemma 2.14. For all x ∈ X, the fiber c−1
D (x) is closed and compact.
Proof. Let U = Spec(R) be an affine neighborhood of x and suppose x ∈ Spec(R)
corresponds to the prime p ⊂ R. Since c−1
D (x) ⊆ D(U ), we may assume that
X = Spec(R). Now, cX (ζ) = x for ζ ∈ D(X) if and only if ζ(g) = 0 for all g ∈ R \ p
and ζ(f ) ≥ 1 for all f ∈ p. These are both closed conditions.
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
7
3. Log discrepancies of Cartier prealgebras
Let X → Spec(k) be an integral scheme of characteristic p > 0 with fraction
field L. The (absolute) Frobenius morphism F : X → X is defined by (f 7→ f p ) for
f ∈ OX . We say that X is F -finite if F is a finite morphism, an assumption we
generally do not need (and thus do not globally make) in this section.
We begin with two classic lemmas that motivate our definitions. Both of these
are special cases of Grothendieck duality.
Lemma 3.1. Suppose X is an F -finite normal scheme and let KX be a canonical
divisor. Then HomX (F∗e OX , OX ) ∼
= F∗e OX ((1 − pe )KX ) as F∗e OX -modules.
Lemma 3.2. Let (R, m) ⊆ (S, n) be a finite local extension of Gorenstein rings.
We have an isomorphism of S-modules HomR (S, R) ∼
= S.
We will use this first lemma to associate φ : F∗e OX → OX with sections of
OX ((1 − pe )KX ), thus to Q-divisors ∆φ on X (and schemes birational to X) having the property that (pe − 1)(KX + ∆φ ) ∼ 0. It is this association which lends
these φ to birational study of singularities. Without log resolutions, we must find
replacements for standard approaches in characteristic zero that make heavy use
of them. However, we can use the second lemma to perform an absolute comparison between φ and ψ ∈ HomL (F∗e L, L) ∼
= F∗e L, since φ = ψ F∗e h for some h. See
Proposition 3.15 and Theorem 6.4 for examples.
Definition 3.3. Let
Q=
M
e≥0
Qe ⊆
M
HomL (F∗e L, L) =: CL ,
e≥0
be a graded quasicoherent OX -sheaf, acting on Qe as an F∗e OX -module via precomposition. We say that Q is a Cartier prealgebra on X if φ ∈ Qe implies
φm := φ ◦ F∗e φm−1 ∈ Qme for all m ≥ 1
One of the main interesting examples of Cartier preaglebras are Cartier subalgebras [Sch10].
Definition 3.4. The Cartier algebra of X is CX = ⊕e HomX (F∗e OX , OX ), viewed
as a graded non-commutative sheaf of rings on X. A Cartier subalgebra on X is a
graded subring of CX .
Remark 3.5. Cartier subalgebras can be thought of as having a similar relationship
to Cartier prealgebras that effective divisors have with arbitrary divisors. To wit,
we can generate a Cartier prealgebra hφiX on X for any φ ∈ HomL (F∗e L, L), and
the principal Cartier subalgebras are those hφiX ⊆ CX .
We temporarily introduce the following ad-hoc definition that allows us to restrict
Cartier prealgebras to compatible subschemes. For subalgebras, we can take all Me
below to be OX , and for individual φ : F∗e L → L on an F -finite scheme there is
always (locally) an effective divisor E such that φ(F∗e OX ) ⊆ OX (E), which is to
say that we can take Me to be OX (−pe E) on an appropriate neighborhood.
Definition 3.6. Let I ⊆ OX be an ideal on X and Q a Cartier prealgebra on X.
We will say that I is Q-compatible at x ∈ X if there exists a neighborhood U of x
and a sequence of line bundles M• = {Me }e≥0 on U such that:
(a) Q|U ⊆ ⊕e HomU (F∗e Me , OU ) and
8
ERIC CANTON
(b) φ(F∗e (IMe )x ) ⊆ Ix for all φ ∈ Qe,x and all e ≥ 0.
We will say that I is Q-compatible when I is Q-compatible for every x ∈ X.
If W is the closed subscheme of X associated to I, then we also say that W is
Q-compatible. We say that M• realizes the Q-compatibility of I at x.
Proposition/Definition 3.7. Suppose Z is an closed integral subscheme of X
that is Q-compatible at x ∈ Z, and suppose M• realizes this compatiblility. Since
φ(F∗e (IZ Me )x ) ⊆ IZ,x , there is a well-defined OZ,x -linear function
e
φkM
Z : F∗ (Me,x |Z ) → OZ,x .
Since Z is integral, and Me,x |Z is a line bundle on Z, we may embed this sheaf into
k(Z). The resulting generic extension φkZ : F∗e k(Z) → k(Z) does not depend on
M• nor the embedding. Applying this procedure to every ψ ∈ Qx yields a Cartier
prealgebra on Z that we will denote by QkZ or QkIZ , depending on the context.
Proof. Fixing a sequence M• realizing the Φ-compatibility of p at x, denote by
Φn kM
Z the generic extension obtained from a fixed embedding of Mne |Z into k(Z).
Suppose
c+p
f +p
n M
ne
=
Φ kZ F∗
g+p
d+p
for c ∈ R and f, g, d ∈ R \ p. Then
ne p
(d + p)Φn kM
Z (F∗ g
en
−1
f + p) = gc + p ∈ k(Z).
en
It follows that d Φn (F∗ne g p −1 f ) − gc ∈ p. In particular, by reversing this argue pen −1
f + p) = gc + p for any sequence L• realizing
ment we see (d + p)Φn kL
Z (F∗ g
compatibility at x and any chosen embedding of each Le .
Definition 3.8 (Log discrepancy of a Cartier prealgebra). We define the log discrepancy of a Cartier prealgebra at a semivaluation in three steps:
(a) Let φ ∈ HomL (F∗e L, L) and w ∈ Val(X). We define the log discrepancy of
φ at w to be
1
n
en
A(w; φ) = sup
w(b)
:
b
∈
L
such
that
φ
(F
b)
=
1
.
∗
en − 1
n≥1 p
(b) Let Q be a Cartier prealgebra on X, w ∈ Val(X), and set x = cX (w). The
log discrepancy of Q at w is defined by
A(w; Q) = sup A(w; φ).
φ∈Qx
(c) Let ζ ∈ Berk(X) with home Z := hX (ζ). If Z is Q-compatible at x = cX (ζ),
we define the log discrepancy of Q at ζ by
A(ζ; Q) = A(ζ; QkZ ).
If Z is not compatible at x then we set A(ζ; Q) = +∞.
The next proposition provides a way of computing log discrepancies at semivaluations without restricting to the home.
Proposition 3.9. Let R = OX,x and ζ ∈ Berk(X) with cX (ζ) = x. Set hX (ζ) = ηZ
and let p ⊂ R be the ideal of Z at x. Let Φ ∈ HomL (F∗e L, L) and define the Cartier
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
9
prealgebra hΦiR := ⊕n≥1 Φn F∗en R on Spec(R). Suppose p is hΦiR -compatible at x.
Then
A(ζ; hΦiR ) = A(ζ; Φ)
1
en
n
en
n
en
= sup
(ζ(r)
−
p
ζ(Φ
(F
r)))
:
r
∈
R
s.t.
Φ
(F
r)
∈
R
\
p
.
∗
∗
en − 1
n≥1 p
Proof. The first equality is clear from the assumption cX (ζ) = x, which implies
ζ(r) ≥ 0 for all r ∈ R and ζ(r) > 0 for r ∈ px . Let ΦkZ : F∗e k(Z) → k(Z) be the
induced mapping associated to any sequence of line bundles realizing compatibility
at p. Suppose
f +p
n
ne
=1
Φ kZ F∗
g+p
for f, g ∈ R \ p. Then
Φn kZ (F∗ne g p
n
en
−1
f + p) = g + p 6= 0 ∈ k(Z),
en
(F∗ne g p −1 f ))
= ζ(g). Therefore, A(ζ; Φ) is at most the supremum on the
so ζ(Φ
second line above. Equality is shown by reversing this process.
The following proposition is used to prove the essential conservation property
(3.18)(b), and is also quite useful when computing log discrepancies in practice. The
usefulness of this proposition is the reason we imposed the power-closed property
on Cartier prealgebras; everything up to this point is true for arbitrary graded
subsheaves Q ⊆ CL .
Proposition 3.10. Let Q be a Cartier prealgebra on X, ζ ∈ Berk(X) with Qcompatible home Z = hX (ζ). Let x = cX (ζ). Then for φ ∈ Qe,x ,
1
n
ne
A(ζ; φ) = lim sup
w(b)
:
b
∈
k(Z)
such
that
φ
(F
b)
=
1
.
∗
pne − 1
n→∞
Similarly, we have an equality
A(ζ; Q) = lim sup sup A(ζ; φ).
e→∞
φ∈Qe,x
Proof. The first equality follows from observing that if φm (F∗me b) = 1 then
φmn (F∗mne b(p
me
mne
−1)/(pme−1 )
(pmne −1)/(pme −1)
) = 1,
mne
and w(b)/(p − 1) = w(b
)/(p
− 1). The second equality is
similar, using that φ ∈ Qe,x implies φn ∈ Qne,x and A(ζ; φ) = A(ζ; φn ).
Example 3.11 (F -finite DVRs; cf. [CMS15]). One can associate divisors to (p−1 )linear maps and deduce log discrepancies using this association. As this process is
well-known to experts, we only sketch the construction and refer to [Sch09, Theorem
3.11] and [BS13, §4] for more detailed explanations. Let R = Aw be an F -finite
DVR associated to a normalized discrete valuation w and E ∈ Spec(R) the closed
point, and fix a local parameter t ∈ R. Then HomR (F∗e R, R) is generated by the
e
projection Φe onto the F∗e tp −1 factor of F∗e R. Since R is Gorenstein and F -finite,
local duality implies that
HomR (F∗e R, R) ∼
= F∗e R
as F∗e R-modules
10
ERIC CANTON
and that our choice of t fixes an isomorphism so that F∗e 1 corresponds to Φe .
Therefore, if φ = Φe F∗e (utc ) ∈ HomR (F∗e R, R) and φ(F∗e b) = 1 then
c + w(b) = w(utc ) + w(b) = pe − 1.
Setting (pe − 1)∆φ := div(utc ) = cE gives a Q-divisor with A(E; ∆φ ) = 1 −
1
pe −1 w(b). This is the key point used by Cascini, Mustaţă, and Schwede (Lemma 3.22),
and their result is crucial for proving that our definition of log discrepancy agrees
with the log discrepancy defined in [JM12] should log resolutions exist. Note that
by F -finiteness, for all φ : F∗e L → L we have that tc φ(F∗e R) ⊆ R for some c ≥ 0 (so
in particular hφiR is a Cartier prealgebra) and that by the above computation we
need only compute the supremum
1
e
e
(w(b) − w(φ(F∗ b)) : φ(F∗ b) 6= 0
sup
pe − 1
to compute A(w; hφiR ).
Example 3.12 (Regular local rings). Let κ be an F -finite field and R = κ[[x1 , . . . , xd ]]
[κ:κp ]
with maximal ideal m, and let {λi }i=1 be a basis for κ over κp . Then R is F -finite,
F∗ R is free over R with basis {λi xa1 1 · · · xadd : 0 ≤ aj ≤ p − 1, 1 ≤ i ≤ [κ : κp ]},
and HomR (F∗ R, R) is generated canonically by the projection Φ onto the xp−1 =
(x1 · · · xd )p−1 -basis element. Now if v ∈ c−1
X (x) ∩ Val(X) and Φ(F∗ b) = 1 for some
b ∈ Frac(R), then b = uxp−1 for some u ∈ R× . It follows that
n
X
A(v; Φ) =
v(xi ).
j=1
We reduce the case of a valuation centered on any regular local ring to this case in
Lemma 7.13(a).
Example 3.13. If w is a discrete valuation whose associated valuation ring R is
not F -finite, then Datta and Smith proved in [DS16, Lemma 4.2.4] (see also [DS17])
that HomR (F∗e R, R) = 0 for all e ≥ 1. I claim that this implies that A(w; φ) = +∞
for any nonzero φ : F∗e L → L. This follows from observing that the set of real
numbers {w(φ(F∗e r)) : φ(F∗e r) 6= 0}r∈R is unbounded below: if it were bounded
below, say by −M = w(t−M ), then tM φ ∈ HomR (F∗e R, R) gives a nonzero map.
Lemma 3.14. Let φ, ψ ∈ HomL (F∗e L, L) and suppose ψ = φ F∗e h for h ∈ L.
For every valuation v on L with A(v; φ) < ∞, A(v; φ) + v(h)/(pe − 1) = A(v; ψ).
Furthermore, A(v; φ) = +∞ implies A(v; ψ) = +∞.
Proof. This follows directly from: φ(F∗e f ) = 1 if and only if ψ(F∗e (f /h)) = 1.
Proposition 3.15 (Birational transformation rule). Let π : Y → X be a proper
birational morphism between regular F -finite schemes, and fix canonical divisors
KX and KY agreeing wherever π is an isomorphism. Setting KY /X = KY − π ∗ KX ,
A(v; CX ) = A(v; CY ) + v(KY /X )
for all v ∈ Val(X) with A(v; CX ) < ∞. Here by v(KY /X ) we mean the valuation of
any representative for this Cartier divisor in a neighborhood of cY (v).
Proof. Suppose x = cX (v) and y = cY (v). From HomZ (F∗ OZ , OZ ) ∼
= F∗ OZ ((1 −
p)KZ ) for Z ∈ {X, Y } we are given generators Φz for CZ,z ∼
= F∗ OZ ((1 − p)KZ )z
such that divZ (ΦZ ) ∩ Spec(OZ,z ) = (1 − p)KZ ∩ Spec(OZ,z ). There exists hS/R ∈ L
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
11
such that Φx = Φy F∗ hy/x since HomL (F∗ L, L) is one-dimensional over F∗ L. As we
have seen,
A(v; Φx ) = A(v; Φy F∗ hy/x )
= A(v; Φy ) − v(hy/x )/(p − 1).
We know that A(v; Φx ) = A(v; CX ) and A(v; Φy ) = A(v; CY ), and so conclude that
v(KY /X ) = v(KY ) − v(π ∗ KX ) = −v(hy/x )/(p − 1).
3.1. Graded sequences of ideals. We now study how Cartier prealgebras can
be twisted by sequences of ideals with multiplicative structures, obtaining formulas
similar to Lemma 3.14. Let {ae }e∈N1 be a collection of ideals on X. We say that
{ae } is multiplicatively graded (or simply graded) if as at ⊆ as+t for all t ∈ N, and
[pt ]
that {ae } is F -graded if a1 = OX and as at ⊆ as+t for all s, t ≥ 0. While it is
standard to use the notation a• for both of these, we have need to use these two
concepts together and so we will always interpret a• as an F -graded sequence and
a? as a multipliciatively graded sequence.
Interesting examples of graded sequences of ideals are obtained as the base loci
of |L ⊗m | for line bundles L on X. Another common source of graded sequences
of ideals are those associated to a semivaluation ζ on X, defined by
as (ζ) = {f ∈ OX : ζ(f ) ≥ s} for s ∈ Z≥0 .
We write a? (ζ) for this graded sequence.
The F -graded condition is precisely what is needed to make new Cartier prealgebras from old. Every Cartier subalgebra on a Gorenstein scheme arises in this way;
see [Bli13; BST12]. One also uses F -graded sequences to study pairs (X, at ) with
a ⊆ OX an ideal and t ≥ 0 a real parameter by studying the F -graded sequence
e
with terms ae := adt(p −1)e [Sch10].
Definition 3.16. We can twist a Cartier prealebra Q by an F -graded ideal sequence a• = {ae ⊆ OX }e∈N on X and obtain a new OX -subsheaf of CL . We do this
by setting
Q · a• = ⊕e≥0 (Qe · F∗e ae )
nX
o
(φi · F∗e ai ) : ai ∈ ae and φi ∈ Qe .
= ⊕e≥0
The F -graded condition ensures that Q · a• ⊆ Q is again a Cartier prealgebra.
A special case of interest is constructed from the data of a Cartier prealgebra Q
on X, an ideal a ⊆ OX , and a real number t ≥ 0. We define
Q · at := Q · {adt(p
Note that {ae = adt(p
e
−1)e
e
−1)e
}e≥1 .
}e≥1 is F -graded.
For ζ ∈ Berk(X) with center x and an ideal sheaf a ⊆ OX , we define
ζ(a) = min{ζ(f ) : f ∈ ax }.
The following lemma allows us to evaluate semivaluations on sequences of ideals in
two ways; the existence of the limits is well-known.
12
ERIC CANTON
Lemma 3.17. (cf. [Ein+06; JM12]) Let ζ ∈ Berk(X) and {ae }e∈N1 be a sequence
of ideals on X, and suppose ζ(ae ) < ∞ for all e. If a• is F -graded, we have the
limit
ζ(ae )
ζ(ae )
ζF (a• ) = lim e
= inf e
.
e→∞ p − 1
e≥1 p − 1
If a? is a graded sequence, then
ζ(a? ) := lim
e→∞
ζ(ae )
ζ(ae )
= inf
.
e≥1
e
e
Proof. See [JM12, Lemma 2.3].
The following lemma is applied throughout the paper.
Lemma 3.18. Let Q and R be Cartier prealgebras on X, ζ ∈ Berk(X) a semivaluation with Q-compatible home Z.
(a) Monotonicity: If Q ⊆ R then A(ζ; Q) ≤ A(ζ; R).
(b) Conservation: Let a• be an F -graded sequence of ideals on X. If ζF (a• ),
A(ζ; Q), and A(ζ; Q · a• ) ∈ R then A(ζ; Q) = A(ζ; Q · a• ) + ζF (a• ).
Proof. Monotonicity follows directly from the definition; we therefore begin with
conservation. Towards this end, Q · a• ⊆ Q implies uniform compatibility of Z
with Q · a• . By passing to a• OZ and QkZ we may assume ζ = w is a valuation
with associated valuation ring Aw . For each e ≥ 0, since X is Noetherian we can
select ge ∈ ae such that ae Aw = ge Aw ⊆ Aw .
Put x = cX (w) and let φ ∈ Qe,x . From Proposition 3.10 we have
1
n
ne
(w(b/g
)
+
w(g
))
:
φ
(F
b)
=
1
A(w; φ) = lim sup
ne
ne
∗
pne − 1
n→∞
= wF (a• ) + lim sup A(w; φn · F∗ne gne ).
n→∞
∈ (Q · a• )ne,x and that ψ(F∗ne b/gne ) = 1. Taking
Note that ψ = φ ·
suprema above over all φ ∈ Qe,x and all ψ ∈ (Q · a• )e,x gives
n
(F∗ne gne )
A(w; Q) ≤ wF (a• ) + A(w; Q · a• ).
We now establish the reverse inequality. Since a• is F -graded, it suffices (by
Lemma 3.17) to show that for all ε > 0 there exists e > 0 such that
A(w; Q · a• ) +
1
wF (ae ) < A(w; Q) + ε.
pe − 1
We find e > 0, ψ ∈ (Q · a• )e,x , and b ∈ L such that ψ(F∗e b) = 1 and
(3.18.1)
A(w; Q · a• ) − ε <
1
w(b).
pe − 1
By definition of Q · a• , there exist φ1 , . . . φn ∈ Qe,x and a1 , . . . , an ∈ ae,x such that
ψ(F∗e b) =
n
X
φi (F∗e ai b) = 1.
1
Let −c = min1≤i≤n {w(φi (F∗e ai b))} ≤ w(
assume that −c = w(φ1 (F∗e a1 b)) so that
φi (F∗e ai b)) = 0. By reindexing, we may
e
φ1 (F∗e uf p a1 b) = 1 for some unit u ∈ A×
w
P
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
13
and some f ∈ Aw such that w(f ) = c. Since a1 ∈ ae , w(ae ) ≤ w(a1 ), and so
A(w; Q · a• ) +
1
1
w(ae ) − ε ≤ A(w; Q · a• ) + e
w(a1 ) − ε
pe − 1
p −1
1
< e
(w(b) + w(a1 ) + cpe )
p −1
≤ A(w; Q).
Here we are using (3.18.1) between the first and second line, and then using that
e
e
φ1 ∈ Qe and φ1 (F∗e uf p a1 b) = 1, so that w(f p a1 b) ≤ (pe − 1)A(w; Q).
Remark 3.19. The assumption that ζF (a• ), A(ζ; Q), and A(ζ; Q · a• ) be finite is
essential. Indeed, for a fairly trivial counterexample without these assumptions
(which demonstrates the main trouble), we can take ζ = trivX , Q = CX , and
ae = 0 for e > 1. Then ζF (a• ) = +∞, A(ζ; Q) = 0, and A(ζ; Q · a• ) = −∞.
Alternatively, suppose X is regular, Z ⊂ X is a proper integral subscheme of
[pe ]
dimension at least one, let pZ be the associated prime ideal, and set ae = (pZ : pZ )
for e > 1. Then pZ is compatible with D = CX · a• , and DkZ = CZ [Fed83]. Let
z ∈ Zreg and ζ = ordz be the valuation corresponding to the exceptional fiber of
Blz (Z) → Z. Then A(ζ; D) = A(ζ; CZ ) = dim(Z) while ζF (a• ) = +∞ = A(ζ; CX ).
Definition 3.20 (Log discrepancies of graded sequences). Let a? be a graded
sequence of ideals and Q a Cartier prealgebra on X. For semivaluations ζ and
a t ∈ [0, ∞) we define the log discrepancy of the sequence of Cartier prealgebras
t/m
Q · am to be the following supremum:
A(ζ; Q · at? ) := sup A(ζ; Q · at/m
m ).
m≥1
If we define
Q · at? =
X
Q · at/m
m
m≥1
then (Q · at? ) is a Cartier prealgebra having log discrepancy A(ζ; Q · at? ).
Corollary 3.21. Let Q be a Cartier prealgebra on X, ζ ∈ Berk(X) whose home Z
is uniformly Q-compatible, and let a? ⊆ OX be a multiplicatively graded sequence
of ideals on X such that ζ(a? ) < ∞, and let t ≥ 0. If both A(ζ; Q) and A(ζ; Q · at )
are finite, then
A(ζ; Q) = A(ζ; Q · at? ) + t ζ(a? ).
We will return to studying asymptotic invariants of graded sequences of ideals in
§7, turning now to prove that our definition is the “correct” one, in the sense that
should log resolutions in characteristic p exist, the log discrepancy function defined
here agrees with the one obtained via the techniques of [JM12] for every valuation.
The first step is to establish the result for divisorial valuations; this result was
shown to the author by Karl Schwede, who proved it with Paolo Cascini and Mircea
Mustaţă. For Q-divisors ∆ ≥ 0 we denote by CX · ∆ the Cartier subalgebra
⊕HomX (F∗e OX (d(pe − 1)∆e), OX ) ⊆ CX .
Lemma 3.22 ([CMS15]). Let ∆ ≥ 0 be a Q-Weil divisor on a normal F -finite
integral scheme X such that (pe − 1)(KX + ∆) ∼ 0 for some integer e > 0. Let
π : Y → X be a proper birational morphism from a normal F -finite integral scheme
14
ERIC CANTON
Y and E ⊂ Y an irreducible divisor. Let AX (E; ∆) be the standard log discrepancy
of ∆ on E, i.e. the coefficent of E in E − ∆Y , with ∆Y defined below:
KY + ∆Y = π ∗ (KX + ∆).
Then AX (E; ∆) = A(ordE ; CX · ∆).
Proof. Fixing a canonical divisor KX on X induces a canonical divisor on every Y
as above by requiring KY to agree with KX wherever π : Y → X is an isomorphism.
The assumption that (pe − 1)(KX + ∆) ∼ 0 now gives the section F∗e 1 of
∼ F e OX ((1 − pe )(KX + ∆))
F e OX =
∗
∼
= HomX (F∗e OX (d(pe − 1)∆e), OX ).
Let ψ : F∗e OX (d(pe − 1)∆e) → OX be the map corresponding to F∗e 1. The formula
(pe − 1)(KY + ∆Y ) = (pe − 1)π ∗ (KX + ∆)
now gives a map ψY : F∗e OY (d(pe − 1)∆Y e) → OY whose associated divisor has
coefficients described as in Example 3.11.
From this, our main theorem follows easily.
Theorem 3.23. Suppose log resolutions of singularities exist in positive characteristic, let ∆ ≥ 0 be a Q-Weil divisor on X such that (pe − 1)(KX + ∆) ∼ 0 for some
e > 0, and let AX (w; ∆) be the log discrepancy of (X, ∆) at w ∈ Val(X) [JM12;
Bou+15]. Then A(w; CX · ∆) = AX (w; ∆).
Proof. We use the notation
P of [JM12]. If (Y, D) → (X, ∆) is a log smooth pair
mapping to (X, ∆), D =
Di , and w ∈ QM (Y, D) then
X
AX (w; ∆) =
w(Di )AX (ordDi ; ∆)
[JM12, Proposition 5.1].
X
=
w(Di )A(ordDi ; CX · ∆).
(Lemma 3.22)
Every log smooth pair (Y, D) over X induces a retraction map
r(Y,D) : Val(X) → QM (Y, D)
[JM12, §3.1, 4.3].
For b ∈ L, we have w(b) ≥ r(Y,D) (w)(b), and furthermore if (Y, D) is a log resolution
for (X, div(b)) then w(b) = r(Y,D) (w)(b) [JM12, Corollary 4.8]. In particular
(‡)
w(b) = sup r(Y,D) (w)(b)
for all b ∈ L.
(Y,D)
Therefore, if x = cX (w), we have
1
e
w(b) : φ(F∗ b) = 1 for some φ ∈ (CX · ∆)e,x
A(w; CX · ∆) = sup
pe − 1
e≥1
1
e
r(Y,D) (w)(b) : φ(F∗ b) = 1
(using (‡))
= sup sup
e
e≥1 (Y,D) p − 1
1
e
= sup sup
r
(w)(b)
:
φ(F
b)
=
1
(Y,D)
∗
pe − 1
(Y,D) e≥1
= sup A(r(Y,D) (w); CX · ∆)
(Y,D)
= sup AX (r(Y,D) (w); ∆) = AX (w; ∆).
(Y,D)
Here we are using that cX (w) = cX (r(Y,D) (w)) for all (Y, D).
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
15
This theorem establishes the approach to log discrepancies put forth in this
section as an alternative to the birational approach by guaranteeing that even
should we learn log resolutions exist in positive characteristic, the functions on
valuation spaces will agree. We note that our extension to Berk(X)\Val(X) is nonstandard, as our definition is not the maximal lsc extension to true semivaluations.
It is this non-standard extension that gives our log discrepancies the ability to
detect F -singularities, as shown in the next section.
4. Connections with F -singularities
In this section, we briefly explore the relationship of our log discrepancies with
two classes of F -singularities. We prove that sharply F -pure and strongly F -regular
Cartier subalgebras are characterized using log discrepancies of semivaluations in
the same way that log canonical and Kawamata log terminal singularities are defined
using standard log discrepancies.
Throughout, we fix a Cartier subalgebra D ⊆ CX on an F -finite scheme X. We
recall the following definitions, citing only the modern incarnations defined in our
setting by Schwede. See also [HR76; HH89].
Definition 4.1 ([Sch08; Sch10]). We say that D is
◦ sharply F -pure if for all x ∈ X the stalk (Dx )e contains a surjective map
for some e > 0.
◦ strongly F -regular if for all x ∈ X and all f ∈ OX,x there exists e > 0 and
φ ∈ (Dx )e such that φ(F∗e f ) = 1.
Compatible prime ideals of D are called centers of F -purity. Note that since
the zero ideal is always compatible, and minimal primes of compatible ideals are
compatible (4.9, 4.10 from [Sch10]), the generic points of each irreducible component
of X are centers of F -purity.
For any x ∈ X, the Dx -compatible ideal of OX,x maximal under inclusion is a
prime called the splitting prime, defined below. If P is the splitting prime of Dx ,
then DkP is a strongly F -regular Cartier subalgebra on V(P ).
Definition 4.2. The splitting prime of D at x ∈ X is the ideal of R := OX,x
P(Dx ) = {f ∈ R : φ(F∗e f ) ∈ m for all φ ∈ Dx , |φ| = e }.
Standard facts about P = P(Dx ) include (see [AE05]):
(a) P is proper if and only if D is F -pure.
(b) As suggested by the name, P is prime whenever it is proper.
(c) When P is proper, no prime p ∈ Spec(R) with P ( p can be D-compatible.
In particular, P = 0 if and only if D is strongly F -regular.
(d) The restriction DkP is strongly F -regular.
Example 4.3. We can use log discrepancies of semivaluations to detect well-studied
types of F -singularities. Let x ∈ X and denote by trivx the trivial semivaluation
k(x)× → {0}. For any Cartier subalgebra D, there are exactly three possiblities:
−∞ iff D is not sharply F -pure at x.
0
iff x is a center of F -purity for D.
A(trivx , D) =
+∞ iff D is sharply F -pure at x and P(Dx ) 6= mx .
16
ERIC CANTON
We now prove our main theorem of this section, giving one way that sharply
F -pure and strongly F -regular singularities should correspond to log canonical and
Kawamata log terminal singularities, respectively. If X = X1 ∪ · · · ∪ Xn is an
irreducible decomposition of X, and triv1 , . . . , trivn are the trivial valuations on the
reductions of the components Xi , we set Berk(X)∗ = Berk(X) \ {triv1 , . . . , trivn }.
Similarly, we set D(X)∗ = D(X) ∩ Berk(X)∗ .
Theorem 4.4. The Cartier subalgebra D is
(a) sharply F -pure if and only if A(ζ; D) ≥ 0 for all ζ ∈ Berk(X)∗ .
(b) stronly F -regular if and only if A(ζ; D) > 0 for all ζ ∈ Berk(X)∗ .
Moreover, the same remains true if we replace Berk(X)∗ with D(X)∗ in either part.
Proof. Since A(ζ; D) depends only on φ ∈ Dc(ζ) and both sharp F -purity and strong
F -regularity are local conditions, we assume Spec(R, m) = (X, cX (ζ)).
Suppose first that D is sharply F -pure and let φ ∈ De be surjective with
φ(F∗e f ) = 1. Then
1
A(ζ; D) ≥ e
w(f ) ≥ 0.
p −1
If D is not sharply F -pure then A(trivx ; D) = −∞.
When D is strongly F -regular, by choosing f ∈ m the same inequality as before
shows that A(ζ; D) > 0. If D is not strongly F -regular, by the first part of this
theorem we may assume D is sharply F -pure. Then the splitting prime P(D) is
proper (see [AE05; Sch10]), and by localizing further we may assume P(D) = m.
Since we are assuming D is sharply F -pure, A(trivx ; D) = 0.
Finally, note that in each case the implication “⇐” is established using only
the trivial semivaluations on residue fields of points. Since trivx ∈ D(X), we can
replace Berk(X) with D(X).
Question 4.5. Suppose CX is sharply F -pure but not strongly F -regular at x ∈ X.
Does there exist a non-trivial ζ ∈ Berk(X) with cX (ζ) = x and A(ζ; CX ) = 0?
5. Lower-semicontinuity
For this section, let X be an F -finite integral scheme of characteristic p > 0. We
show that A(−; Q) is lower-semicontinuous on Berk(X) for any Cartier prealgebra
Q. We then deduce that the minimal log discrepancy function derived from log
discrepancies on Berk(X) is lsc on X considered with the constructible topology.
Definition 5.1. Recall that a function f : Y → R±∞ = [−∞, ∞] on a topological
space Y is lower-semicontinuous (lsc) at y0 ∈ Y if one of the following equivalent
conditions holds:
(a) For every convergent net yα → y0 , f (y0 ) ≤ lim inf α f (yα ).
(b) For every ε > 0 there exists an open neighborhood U ⊆ Y of y0 such that
f (y0 ) − ε < f (y) for all y ∈ U ;
(c) Considering R±∞ with the toplogy of half-open intervals (a, ∞], the mapping f is continuous.
Remark 5.2. Since Berk(X) is not generally first countable, we use nets and not
sequences for questions of convergence and compactness.
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
17
Definition 5.3. The constructible topology on X is minimal topology containing
both Zariski open and Zariski closed subsets. This is equivalent to the topology
with basis consisting of finite unions of locally closed subsets.
The following lemma is a technical generalization of the classical way of producing a new lsc function from a given collection of lsc functions. The author thanks
Kevin Tucker for suggesting this approach, which leads to a much simpler proof of
lower-semicontinuity than the author’s original.
Lemma 5.4 (cf. [Fol13], Proposition 7.11). Let A be a topological space and let G
be a sheaf of lsc functions on A. If f (x) = sup{g(x) : g ∈ Gx }, then f is lsc.
Proof. We use definition 5.1(c). For an open set (r, ∞] ⊆ R±∞ , let G = ∪{Gx :
f (x) ∈ (r, ∞]}. Observe that f −1 (r, ∞] = ∪g∈G g −1 (r, ∞] is open.
Theorem 5.5. For every Cartier prealgebra Q on X, the log discrepancy A(−; Q)
is lsc on Berk(X).
Proof. Recall that Q is quasi-coherent, and Berk(X) has an open cover by Berk(U )
for Spec(R) = U ⊆ X affine open. Quasi-coherence gives some graded module Q ⊆
e and A( ; Q) is constructed as in Lemma 5.4
⊕ HomL (F∗e L, L) such that Q|U = Q,
from the locally constant sheaf associated to U 7→ {A( ; φ) : φ ∈ Q} for U ⊆
Berk(U ). Therefore, it is enough to assume X = Spec(R), fix φ ∈ Q, and prove
that A( ; φ) is lsc.
For any f ∈ L = Frac(R) and n ≥ 1, set s(f, n) = φn (F∗ne f ). Varying f ∈ L, we
obtain an open cover of Berk(X) using the open subsets Uf,n = ev−1
s(f,n) (−∞, ∞) ⊆
Berk(X). On each Uf,n , the function
1
evf (ζ) − pen evs(f,n) (ζ)
E(f, n, ζ) = ne
p −1
is continuous. We then induce a sheaf of continuous functions on Berk(X) by
G(U) = {E(f, n, )|U : U ⊆ Uf,n }.
Since A(v; φ) = sup{E(f, n, ζ) : Gζ }, we apply (5.4) once again and conclude that
A( ; φ) is lsc.
Remark 5.6. If A(v; φ) ∈ R for v ∈ Val(X), and ε > 0, we can be more precise in
demonstrating a basic open subset U ⊆ Val(X) with A(v; φ) − ε < A(w; φ) for all
w ∈ U . Indeed, keeping notation from the previous proof, suppose A(v; φ) − ε/2 <
E(f, n, ζ) for some f ∈ L with φn (F∗ne f ) = 1. One can check that U = {w ∈
Val(X) : |v(f ) − w(f )| < (pen − 1)ε/2} has the desired property.
Being a weak topology, a natural basis for the topology near a given ζ ∈ D(X)
over an affine subscheme Spec(R) of X is given by finite intersections of sets of the
form
{α ∈ D(Spec(R)) : |ζ(f ) − α(f )| < ε}
or
{α ∈ D(Spec(R)) : α(f ) > M },
where M and ε are positive real numbers and f ∈ R. The first type of set is used
in the case ζ(f ) < ∞ and the second when ζ(f ) = ∞. An important property of
D(X) not shared by Berk(X) is that every open set of this form is a finite union of
sets of the form {α ∈ D(X) : α(f ) = z} and {α ∈ D(X) : α(f ) ≥ z} for integers
z. Therefore, any finite intersection of these sets can be written as a finite union
18
ERIC CANTON
of sets of the form {β ∈ D(X) : β(f ) = a} and {β ∈ D(X) : β(g) > M } for some
f, g ∈ R and a, M ∈ Z.
Proposition 5.7. The image of a basic open subset of D(X) under cD is constructible, and the quotient topology induced by cD is the constructible topology.
Proof. We assume that X = Spec(R) and that our basic open subset is of the form
U := {α ∈ D(X) : α(fi ) = ai for 1 ≤ i ≤ s} for some fi ∈ R and ai ∈ Z; the case
of α(fi ) ≥ Mi is very similar. Suppose α ∈ U. The condition that p = cD (α) is
equivalent to p ∈ D(fi ) := V(fi )c for fi such that ai = 0 and p ∈ V(fi ) for ai > 0
(note that by definition α(f ) ≥ 0 for every f ∈ R). Re-numbering so that ai = 0
for 1 ≤ i ≤ t and ai > 0 for t < i ≤ s, we conclude that
!
!
t
s
\
[
\
cD (U) =
D(fi )
V(fi ) ,
i=1
i=t+1
which is constructible. Taking preimages gives
!
t
[
\
−1
−1
cD (cD (U)) =
evfi (0)
s
\
!
ev−1
fi (0, ∞]
,
i=t+1
i=1
showing that the quotient topology is the constructible topology.
We use the following lemma to pass lsc functions from D(X) to X by minimizing
on fibers of cD . These semicontinuity results are typically deduced for excellent
schemes over Q using log resolutions (e.g. [Amb99, Theorem 2.2]); such results
become special corollaries of our lemma.
Lemma 5.8. Let f : Y → Z be a continuous surjective function between topological
spaces. Assume that Y is compact and Z is Hausdorff. Fixing T : Y → R±∞ , define
a function on Z by
Tf (z) = inf T (y).
f (y)=z
Then Tf is lsc on Z whenever T is lsc on Y .
Proof. If Tf (z) = −∞ then lower-semicontinuity is automatic at z, so we may
assume Tf (z) > −∞ for all z. Fix z∗ ∈ Z; we show lower-semicontinuity of Tf
in two steps, treating the case Tf (z∗ ) = +∞ first. In this case, T (y) = +∞ for
all y ∈ f −1 (z∗ ). Now if zν → z∗ is a convergent net in Z, we wish to show that
lim inf ν Tf (zν ) = +∞. Assume we have some subnet zµ with lim inf Tf (zµ ) = t∞ <
∞; there exists a subsequence zn with t∞ − 1/n < Tf (zn ) ≤ t∞ . We select yn ∈
f −1 (zn ) with Tf (zn ) ≤ T (yn ) < Tf (zn ) + 1, so that t∞ ≤ lim inf n T (yn ) ≤ t∞ + 1.
Take a subsequence {yns } ⊆ {yn } so that lim inf n T (yn ) = lims T (yns ). Since f is
continuous, f (yns ) =: zns → z∗ , so every convergent subsequence of yns must have
limit y∗ ∈ f −1 (z∗ ). We know T (y) = +∞ for y ∈ f −1 (z∗ ), and T is lsc on Y , so
lims T (yns ) = +∞. This contradicts t∞ ∈ R.
For the final case when Tf (z∗ ) ∈ R, we follow a similar argument as before with
more careful estimates. Since Tf is not lsc at z∗ , there exists a net zβ → z∗ and
δ > 0 such that
(lim inf Tf (zβ ) = Tf (z∗ ) − δ) < Tf (z∗ ).
β
For all n ≥ 1, there exists βn so that, setting zn := zβn ,
Tf (zn ) < (Tf (z∗ ) − δ) + δ/(2n).
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
19
Since Tf (zn ) is an infemum, there must exist yn ∈ f −1 (zn ) satisfying
T (yn ) < Tf (zn ) + δ/(2n).
Passing to a convergent subsequence ynt → y∗ , which exists by compactness of Y ,
we know that f (y∗ ) = limt f (ynt ) by continuity. Since Z is Hausdorff, we conclude
f (y∗ ) = limt znt = z∗ . Combining these estimates, we see
Tf (z∗ ) ≤ T (y∗ )
≤ lim inf T (ynt )
t→∞
since T is lsc on Y
≤ lim inf (Tf (znt ) + δ/(2nt ))
t→∞
≤ lim inf (Tf (z∗ ) − δ + δ/(nt ))
t→∞
= Tf (z∗ ) − δ < Tf (z∗ ).
This is absurd, and we conclude that Tf is lsc on Z.
Definition 5.9. The Berkovich minimal log discrepancy of Q at x ∈ X is defined
to be
mldB (x; Q) = min A(ζ; Q),
where we minimize over ζ ∈ D(X) with cD (ζ) = x. Compactness of the fibers of
cD (Lemma 2.14) proves this minimum is achieved.
Note that if mldB (x; Q) < ∞ then any ζ achieving this minimum must have a
Q-compatible home Z. Note also that minimizing A(−; Q) over Berk(X) instead
of D(X) produces an uninteresting function: clearly A(λ ζ; Q) = λ A(ζ; Q) for all
λ ∈ R>0 , so the minimum at every x ∈ X is either −∞ or 0, based on if A(−; Q)
takes on negative values or not. It may be interesting to study the minimum of
A(ζ; Q)/ζ(x) with ζ ∈ Berk(X), cX (ζ) = x, and ζ 6= trivx .
Theorem 5.10 (cf. [Amb99]). The Berkovich minimal log discrepancy mldB (−; Q) :
X → R±∞ is lsc in the constructible topology on X. Equivalently, for any x ∈
X and ε > 0, there is a constructible subset G ⊂ X containing x such that
mldB (x0 ; Q) > mldB (x; Q) − ε for all x0 ∈ G.
Proof. The theorem follows directly from Lemma 5.8, since cD : D(X) → X Constr
satisfies the hypotheses and A(−; Q) is lsc on D(X).
Remark 5.11. It is a major open problem in the (characteristic zero) minimal model
program to determine if the (usual) minimal log discrepancy is lsc in the Zariski
topology on the set of closed points of a variety, as this would imply the termination
of arbitrary (undirected) sequences of flips.
Another natural function to consider is the appropriate version of the log canonical threshold of graded sequences of ideals with respect to strongly F -regular Cartier
subalgebras. Many of the proofs found in the final sections of [JM12] can be adapted
to our setting, and so we devote §7 to this function.
6. Transformation of log discrepancies under finite morphisms
In this section we study how log discrepancies of valuations change under finite
morphisms. We restrict ourselves to the valuation spaces in this section. The setup
throughout is the following:
20
ERIC CANTON
6.1. Notation and T-transposes. Suppose L ⊆ M is a finite field extension of
F -finite fields, and denote by q = pe a positive integer power of p. Fix an algebraic
closure of M and denote by by L1/q ⊆ M 1/q the fields of q-th roots of L and M in
this algebraic closure. As a field, M 1/q is isomorphic to M ; as an M -module, M 1/q
is isomorphic to F∗e M .
The following theorem of Schwede and Tucker provides maps completing certain
squares set up by φ : L1/q → L and T : M → L.
Theorem 6.1 (cf. [ST14]). Given nonzero L-linear maps T : M → L and φ :
L1/q → L, there exists a unique M -linear map φT : M 1/q → M the diagram below
commutes:
M 1/q
(†)
T
φT
1/q
L1/q
M
T
φ
L
Schwede and Tucker call φT the T-transpose of φ. When X and Y are normal,
they describe how to assign T a divisor RT on Y , such that
π ∗ ∆φ = ∆φT + RT .
Our goal in this section is to show that a similar formula holds for arbitrary F -finite
fields L and log discrepancies of φ and φT on valuations. We need to define a new
variant of ramification adapted to our setting.
6.2. Log ramification.
Definition 6.2. Let L ⊆ M be a finite extension of fields, T : M → L be nonzero
and L-linear, µ ∈ Val(M ), and λ := µ|L . Define the log ramification of T along µ
to be
ρ(µ; T) = sup {µ(b) − λ(T(b))} = sup µ(b).
b∈M ×
T(b)=1
2
Example 6.3 (cf. 3.4 in [ST14]). Let R = F3 [x ] ⊆ F3 [x] = S. The ramification
divisor of S|R is given by div(x), which is the divisor associated to the trace TrS|R ∈
HomR (S, R) ∼
=S S. One can check that
ρ(ordx ; TrS|R ) = 0
and ρ(ordf ; TrS|R ) = 1 for all irreducible f ∈ S \ {x}.
Theorem 6.4. Let L ⊆ M be a finite extension of fields, T : M → L and φ :
L1/q → L nonzero L-linear maps, and φT the T-transpose of φ. For µ ∈ Val(M )
with A(µ; φT ) < ∞ extending λ ∈ Val(L) we have
A(λ; φ) = A(µ; φT ) + ρ(µ; T).
We first prove the following lemma to lessen the notational burden in the proof
of Theorem 6.4.
Suppose (R, m) ⊆ (S, n) is a finite extension of local Gorenstein rings. As we
saw in Lemma 3.2, there is an isomorphism of S-modules
(6.4.1)
HomR (S, R) = HomR (S, ωR ) ∼
=S ωS = S.
This isomorphism is fundamental: [ST14] depends heavily on this isomorphism
to prove existence and uniqueness of transposes, and we have already used this
isomorphism in Example 3.11.
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
21
Lemma 6.5. Let L ⊆ M , φ, and T be as before. For all m ∈ M × and µ ∈ Val(M )
with A(µ; φT ) < ∞ and ρ(µ; T) < ∞ we have:
A(µ; φT·m ) = A(µ; φT ) + µ(m)
and
ρ(µ; T) = ρ(µ, T · m) + µ(m).
In particular, the expression
A(µ; φT ) + ρ(µ, T) = A(µ; φT·m ) − µ(m) + ρ(µ; T · m) + µ(m)
is invariant under scaling T by m ∈ M × .
Proof. Both statements equations follow directly from the definitions, though before
we see the log discrepancy equation we need to know how to compare φT and φT·m .
Applying (6.4.1) to M ⊆ M 1/q and L ⊆ M 1/q , we see that both HomM (M 1/q , M )
and HomL (M 1/q , L) are one-dimensional M 1/q vector spaces. Composition with T
now gives a nonzero M 1/q -linear isomorphism
) : HomM (M 1/q , M ) → HomL (M 1/q , L).
T◦(
Now recalling (†) we see
(T ◦ φT ) · m1/q = (φ ◦ T1/q ) · m1/q
= φ ◦ (T · m)1/q
= (T · m) ◦ φT·m .
Using M -linearity of φT·m we get
T ◦ φT · m1/q = T ◦ (φT·m · m).
Composition with T being an isomorphism tells us that φT = φT·m · m(q−1)/q =
e
φT·m F∗e m(p −1) . Applying Lemma 3.14 now gives the result.
Proof of Theorem 6.4. First we prove that A(λ; φ) ≤ A(µ; φT ) + ρ(µ; T). Fix ε > 0
and choose ` ∈ L such that φ(`1/q ) = 1 and
1
λ(`) > A(λ; φ) − ε/2.
q−1
(6.5.1)
Since T 6= 0, it is surjective, and so there exists m ∈ M such that T(m) = `. Let
n 0 such that
ρ(µ; T) − µ(m/`) < (q n − 1)ε/2.
(6.5.2)
n
n
Replacing ` by `(q −1)/(q−1) , m by (`(q −1)/(q−1) m/`), and q by q n leaves the
quantities µ(m/`) and λ(`)/(q − 1) unchanged. This replacement also preserves the
properties T(m) = ` and φ(`1/q ) = 1.
Note that T(φT (m1/q )) = φ(T(m)1/q ) = 1, allowing us to conclude that
(6.5.3)
− µ φT (m1/q ) ≥ −ρ(µ; T).
22
ERIC CANTON
Using this, we deduce
1
µ(m) − qµ(φT (m1/q ))
A(µ; φT ) ≥
q−1
1
µ(m) − λ(`) + λ(`) − qµ(φT (m1/q ))
=
q−1
1
=
λ(`) + µ(m/`) − qµ(φT (m1/q ))
q−1
1
≥
(λ(`) + µ(m/`) − qρ(µ; T)) by (6.5.3)
q−1
1
(λ(`) + ρ(µ; T) − (q − 1)ε/2 − qρ(µ; T)) by (6.5.2)
>
q−1
> A(λ; φ) − ε/2 − ρ(µ; T) − ε/2 by (6.5.1).
This establishes the claimed inequality. We proceed to prove the reverse inequality.
Towards this end, note first that, using Lemma 6.5, we may assume ρ(µ; T) = 0.
The situation is then the following:
◦ µ(b) ≤ λ(T(b)) for all b ∈ M , and
◦ for all δ > 0 there exists mδ such that T(mδ ) = 1 and −δ < µ(mδ ) ≤ 0.
Fix ε > 0 and choose x ∈ M such that φT (x1/q ) = 1 and
1
A(µ; φT ) − ε/2 <
µ(x).
q−1
Set δ = (q − 1)ε/(2q) and select mδ as mentioned. Note that φT (mδ x1/q ) = mδ so
that φ(T(mqδ x)1/q ) = T(φT (mδ x1/q )) = 1. Therefore,
1
µ(b) − ε/2
A(µ; φT ) − ε <
q−1
1
=
(µ(mqδ b) − qµ(mδ )) − ε/2
q−1
1
<
µ(mqδ b)
q−1
1
≤
λ(T(mqδ b))
q−1
≤ A(λ; φ).
Thus we have A(λ; φ) = A(µ; φT ) + ρ(µ; T).
0
Example 6.6. We can apply the previous theorem to the case T = ψ : L1/q → L.
0
0
Every λ ∈ Val(L) has a unique extension λ0 to L1/q given by λ0 (b1/q ) = λ(b)/q 0 ,
so we will not distinguish between λ and its extensions.
By definition, we see that ρ(λ; ψ n ) ≤ A(λ; ψ)(q n − 1)/q n for all λ ∈ Val(L)
with equality in the case λ corresponds to an F -finite DVR. Moreover, ρ(λ; ψ n ) ↑n
A(λ; ψ). From this it follows that A(λ; φψn ) ↓n A(λ; φ) − A(λ; ψ).
7. Log canonical thresholds of graded sequences of ideals
In this section, we study log canonical thresholds of graded sequences of ideals in
positive characteristics, re-proving several theorems of Jonsson and Mustaţă along
the way. The proofs found in [JM12] generally go through with only mild modifications, replacing characteristic zero techniques with those F -singularity techniques
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
23
we have already utilized. To avoid constantly passing between points and the associated sheaf of ideals, we will generally make statements like “Let m ∈ X be a
point. . . ” understanding schemes to have underlying topological spaces consisting
of the set of prime ideals of OX (that is, X ∼
= Spec(OX )). Unless explicitly stated
otherwise, all schemes in this section are understood to have characteristic p > 0.
7.1. Notation and LCT of graded sequences. Recall from §3.1 that we notate
a? for sequences of ideals that are (multiplicatively) graded, and write b• for an
F -graded sequence. Let us fix a connected F -finite regular scheme X of dimension
d ≥ 1 and characteristic p > 0, and denote by L the fraction field of X. We also
fix a strongly F -regular Cartier subalgebra D and a nonzero ideal q on X. Denote
by trivX the trivial valuation L× → {0} and by Val(X)∗ := Val(X) \ {trivX }. Any
v ∈ Val(X)∗ defines a graded sequence of nonzero ideals, denoted here by a? (v), by
setting
as (v) = {f ∈ R : v(f ) ≥ s}; s ∈ N0 .
Jonsson and Mustaţă prove the following statement showing that w(a? (v)) compares
the values of w and v, asymptotically.
Lemma 7.1 (cf. Lemma 2.4 [JM12]). Let v ∈ Val(X)∗ be nontrivial and define
a? (v) as above. Then
w(b)
w(a? (v)) = inf
v(b)
for every w ∈ Val(X), where b ranges over ideals on X such that v(b) > 0.
Recall from §3 we define, for any graded sequence a? on X,
X
D · at? =
D · at/m
and A(ζ; D · at? ) = sup A(ζ; D · at/m
m
m ).
m
m≥1
Definition 7.2 (Log canonical threshold of graded sequence). The log canonical
threshold of a? with respect to q, D, and ζ ∈ Berk(X) is
lctqB (ζ; D, a? ) = sup{t ≥ 0 : ζ(q) + A(ζ; D · at? ) > 0}.
Here we consider (+∞) + (−∞) ≤ 0, a situation that only arises when ζ(q) = +∞.
This definition is equivalent to
lctqB (ζ; D, a? ) = sup{t ≥ 0 : ζ(q) + A(ζ; D · at/m
m ) > 0 for some m}.
The log canonical threshold of a? with respect to q and D is
lctqB (D, a? ) =
inf
v∈Val(X)∗
lctqB (v; D, a? ).
We will show, as do Jonsson and Mustaţă, that this infemum is in fact achieved.
Lower-semicontinuity of lctqB (−; D, a? ) is automatic from the second description
t/m
above, given that we know (q 7→ ζ(q)) is continuous and that A(−; D · am ) is lsc
on Berk(X).
Example 7.3. If v ∈ Val(X)∗ has A(v; D) < ∞ and v(a? ) > 0, then conservation
of log discrepancy (Corollary 3.21(b)) implies
A(v; D) + v(q)
,
v(a? )
an expression familiar from the literature regarding log discrepancies on valuation
spaces (e.g. [JM12; Bou+15]).
lctB (v; D, a? ) =
24
ERIC CANTON
Lemma 7.4. Let v ∈ Val(X)∗ with A(v; D) < ∞ and consider graded sequences
a? and b? on X. Suppose lctqB (v; D, a? ) < ∞. Then
lctqB (v; D, b? ) > lctqB (v; D, a? )
Consequently,
lctqB (v; D, b? )
⇔
v(a? ) > v(b? ).
= +∞ if and only if v(b? ) = 0.
Proof. Write λ = lctqB (v; D, a? ). We know that A(v; D · at? ) = A(v; D) − t v(a? ) for
any t ≥ 0, so finiteness of λ implies v(a? ) > 0, hence v(q) + A(v; D · aλ? ) = 0. Using
the corresponding expression for A(v; D · bt? ), we see that
v(q) + A(v; D · bλ? ) = v(q) + A(v; D · aλ? ) + λ(v(a? ) − v(b? ))
= λ(v(a? ) − v(b? )).
Since λ ∈ (0, ∞), v(q) + A(v; D · bλ? ) > 0 if and only if v(a? ) > v(b? ), and the
desired comparison between log canonical thresholds follows.
Now we prove that lctqB (v; D, b? ) = +∞ if and only if v(b? ) = 0. If v(b? ) = 0,
then clearly lctqB (v; D, b? ) = +∞. For the reverse direction, it suffices to prove
(ε)
(ε)
that for any ε > 0 there exists a sequence of ideals a? such that v(a? ) = ε and
(ε)
q
lctB (v; D, a? ) < ∞. Using Lemma 7.1, we see that this is achieved by taking
(ε)
w = (1/ε)v and a? = a? (w).
Example 7.5. If trivX is the trivial valuation on X, then ζ(q) = 0 = A(ζ; D) and
lctqB (D, a? ) = −∞. For ζ 6= trivX , if ζ(as ) = 0 for some s ≥ 1, then (D · at? )c(ζ) =
Dc(ζ) is strongly F -regular, so lctqB (ζ; D · a? ) = +∞. Alternatively, lctqB (ζ; D · a? ) =
+∞ if A(ζ; D · at? ) > −∞ for all t and ζ(q) = +∞, a condition satisfied whenever
ζ 6= trivX and (D · at? )h(ζ) is F -pure for all t > 0 (e.g. non-trivial valuations). On
the other hand, A(ζ; D ·at? ) = −∞ for t > fpt(hX (ζ); D, a? ), the F -pure threshold of
(D·at? ) at hX (ζ) (see [Sch10; Dat17]). We see that lctqB (ζ; D, a? ) = fpt(hX (ζ); D, a? )
for every q when fpt(hX (ζ); D, a? ) < ∞.
7.2. Asymptotic multiplier ideals of Cartier subalgebras. We define ideals
JB (D · at? ) containing information about values, and in particular minima, of
A(−; D · at? ). We model our definition on the valuation-theoretic description of
asymptotic multiplier ideals, see e.g. [Laz04; JM12; Bou+15].
Definition 7.6 (Asymptotic multiplier ideal of (D · at? )). Let t ∈ R≥0 . The asymptotic multiplier ideal of (D · at? ) is
JB (D · at? ) = {f ∈ OX : ∀ v ∈ Val(X)∗
= {f ∈ OX : ∀ v ∈ Val(X)∗
v(f ) + A(v; D · at? ) > 0}
∃ m s.t. v(f ) + A(v; D · at/m
m ) > 0}.
Remark 7.7. It follows directly from the definition that λ = lctqB (D, a? ) is also
the supremum over all t ≥ 0 such that q ⊆ JB (D · at? ). We will see soon that the
subscheme V(JB (D ·aλ? ) : q) defines the locus where minimum-achieving valuations
can be centered.
The following proposition should seem familiar from standard treatments of asymptotic multiplier ideals such as [Laz04]. We do not need this statement in the
sequel, but include it for interest.
Proposition 7.8. For all m, n ≥ 1 and v ∈ Val(X)∗
t/(mn)
A(v; D · at/m
).
m ) ≤ A(v; D · amn
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
25
Consequently, there is an inclusion of multiplier ideals
t/(mn)
JB (D · at/m
).
m ) ⊆ JB (D · amn
t/m
From this we deduce that the collection {JB (D · am )}m≥1 has a unique maximal
element, and this element is JB (D · at? ).
Proof. Note that if a ∈ N1 and b ∈ R≥0 , then dabe ≤ adbe, so
E := D · {and(p
m
e
−1)t/(mn)e
}e≥1 ⊆ D · {ad(p
m
e
−1)t/me
}e≥1 = D · at/m
m .
nd(pe −1)t/(mn)e
d(pe −1)t/(mn)e
Since a? is multiplicatively graded, we know am
⊆ amn
and
t/(mn)
so monotonicity gives A(v; E ) ≤ A(v; D · amn ). Therefore, it suffices to check
t/m
that A(v; D · am ) = A(v; E ). The inequalities
d(pe − 1)t/me ≤ nd(pe − 1)t/(mn)e ≤ d(pe − 1)t/me + n
n+d(pe −1)t/me
), and a direct computation shows
imply that E ⊇ F = ⊕(De · F∗e am
t/m
A(v; F ) = A(v; D · am ).
The inequality between log discrepancies directly implies the containment of
multiplier ideals. The statement regarding the collection of multiplier ideals is
standard and only relies on the Noetherian property; see [Laz04].
7.3. The Conjectures of Jonsson and Mustaţă. In [JM12], Jonsson and Mustaţă conjecture statements, here numbered 7.19 and 7.20, regarding valuations computing the log canonical threshold of a? . We say that v computes this lct when
lctqB (D, a? ) = lctqB (v; D, a? ). They prove that 7.20 implies 7.19, and establish 7.20
when dim(X) ≤ 2 or each am is monomial. In positive characteristics, 7.20 follows
from [DS16] in dimension two; see Example 3.13 and compare with [JM12, §9].
Continuing with our setup of D and q 6= 0 on a connected, regular, F -finite
scheme X, fix a nonzero graded sequence a? . We simplify notation by writing jt for
JB (D · at? ).
If v ∈ Val(X)∗ computes lctqB (D, a? ) =: λ < ∞, then I claim cX (v) ∈ V(jλ : q)
[JM12, §7]. Indeed, if v computes lctqB (D, a? ), then v(q) + A(v; D · aλ? ) = 0. Now
consider f ∈ (jλ : q). Since f q ⊆ jλ , the definition of jλ gives the estimate v(f ) =
v(f ) + v(q) + A(v; D · aλ? ) > 0. Therefore, v(f ) > 0 and cX (v) ∈ V(jλ : q).
Lemma 7.9. Let Y be an integral scheme, m ⊂ OY a nonzero proper ideal, and let
G be a strongly F -regular Cartier subalgebra on Y . For all N ∈ R≥0 , the subspace
VN := {ζ ∈ Berk(Y ) : ζ(m) = 1
and
A(ζ; G ) ≤ N }
is a compact subset of Val(Y )∗ ( Berk(Y ).
Proof. Since (ζ 7→ ζ(m)) is continuous and (ζ 7→ A(ζ; G )) is lsc on Berk(Y ), it
is automatic that VN is closed, hence compact. Moreover, strong F -regularity
implies that no non-zero primes are compatible with G . Thus, if ζ 6∈ Val(Y ) then
A(ζ; G ) = +∞, and if ζ = trivX then ζ(m) = 0. Thus, VN ⊆ Val(Y )∗ .
7.4. Locally quasi-monomial valuations. Let v ∈ Val(X). The value group
of v is Γv = v(L× ) ⊆ R. The rational rank of v is ratrk(v) = dimQ (Γv ⊗ Q).
If (B, m) is a local subring of L dominated by the valuation ring (Av , mv ), then
(B/m) = ` ⊆ κ(v) = Av /mv and we set the transcendence degree of v over B to be
26
ERIC CANTON
tr. degB (v) = tr. deg(κ(v)|`). When v ∈ Val(X) with center x we set tr. degX (v) =
tr. degOX,x (v). A fundamental estimate is Abhyankar’s inequality:
ratrk(v) + tr. degX (v) ≤ dim(OX,x ).
Valuations achieving equality are called Abhyankar valuations, and are the most
geometric and well-behaved of valuations. One easy way to obtain an Abhyankar
valuation is via the following construction, outlined more carefully in §3.1 of [JM12].
Let (R, m, κ) be a regular local ring. For any regular system of parameters (rsop)
r1 , . . . , rd for R, there is an isomorphism R̂ ∼
, . . . , rd ]], and so we may view
= κ[[r1P
f ∈ R ⊆ R̂ as having an expansion of the form f = u∈Nd cu ru , where cu is either
0
zero or a unit in R̂, and r(u1 ,...,ud ) = r1u1 · · · rdud . Jonsson and Mustaţă prove in
Proposition 3.1 of [JM12] that for any α = (α1 , . . . , αd ) ∈ Rd≥0 , there is a unique
Pd
valuation v = valα on R̂ such that v(f ) = min{ j=1 αj uj : cu 6= 0}. Restricting v
to R now gives an Abhyankar valuation; see [JM12], Proposition 3.6(ii), which (as
they mention in the proof) works over any ground field. In the same proposition,
they show we may in fact assume that the αi are rationally independent by blowingup appropriately. Thus, there exists a regular local ring (R0 , m0 ) of dimension
ratrk(v), and so that the residue field of valα is algebraic over R0 /m0 .
Now let Y be a regular F -finite scheme and v ∈ Val(Y ). We say that a divisor H on Y has simple normal crossings (and that H is an SNC divisor) if
for all y ∈ Y there exist r1 , . . . , rn ∈ OY,y forming part of a rsop and satisfying H ∩ Spec(OY,y ) = div(r1 · · · rn ). Fixing H with simple normal crossings,
for any v ∈ Val(Y ) centered at y ∈ Y we define r(Y,H) (v) ∈ Val(Y ) by taking
α = (v(r1 ), . . . , v(rd )) in the previous construction. From the definition of r(Y,H) (v)
and the axioms for a valuation, we see that r(Y,H) (v)(f ) ≤ v(f ) for every f ∈ OY ,
hence cY (r(Y,H) (v)) ⊆ cY (v), viewing centers as identical to the associated prime
ideal of OY . We call r(Y,H) the monomialization retraction associated to (Y, H).
In characteristic zero, every Abhyankar valuation is obtained as the result of
some monomialization retraction, induced from a proper birational Y → X and
some SNC divisor H on Y (see [ELS01] and Proposition 3.7 in [JM12]). For fraction
fields of finite transcendence degree over a field k ⊇ Fp , Knaf and Kuhlmann show in
[KK05] that an Abhyankar valuation whose residue field is separable over k admits
a local monomialization over X, essentially removing the properness assumption on
Y → X. Since valuations arising from monomialization seem to work best when
passing to completions, and we make essential use of this soon, we introduce the
following definition.
Definition 7.10. Let v ∈ Val(X) with center x ∈ X. We say that v is locally quasimonomial if there exists a regular local ring (R, m) dominating and birational to
OX,x , and a rsop for R giving the SNC divisor H on Y = Spec(R) such that
v = r(Y,H) (v).
The following lemmas allow us to pass locally quasi-monomial valuations on a
regular F -finite scheme to an appropriate affine space via completion. The proofs
for these lemmas given in [JM12] are independent of characteristic.
Lemma 7.11 (cf. [JM12, Lemma 3.10]). Let x ∈ X and consider the completion
0
0
[
morphism φ : X 0 = Spec(R0 ) → X, where R0 = O
X,x . If v ∈ Val(X ) has center
0
0
the closed point and v = v |X , then κ(v) = κ(v ) and the value groups of v and v 0
are equal. In particular, tr. degX 0 (v 0 ) = tr. degX (v) and ratrk(v 0 ) = ratrk(v).
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
27
Lemma 7.12 (cf. [JM12, Lemma 3.11]). Let k ⊆ K be an algebraic field extension
and φ : AnK → Ank the corresponding morphism of affine spaces. Suppose that v 0
is a valuation on K(x1 , . . . , xn ) with center 0 ∈ AnK , and let v be its restriction
to k(x1 , . . . , xn ). Then v is centered at 0 ∈ Ank , tr. degAnK (v 0 ) = tr. degAnk (v), and
ratrk(v 0 ) = ratrk(v).
Jonsson and Mustaţă prove the next lemma for the more general setting of a
regular morphism, but the only case we ever apply it to here is that of completion
of F -finite (hence excellent) regular local rings, and so content ourselves to prove
it in this case.
Lemma 7.13. Let x ∈ X and let φ : X 0 = Spec(R0 ) → X be the completion
morphism at x. Set q0 = qR0 , a0? = a? R0 , and extend D = CX · b• to D 0 = CX 0 · b0•
on X 0 , where b0• = b• R0 . Suppose λ = lctqB (D, a? ) < ∞.
(a) For any v 0 ∈ Val(X 0 ) and t ≥ 0, we have A(v 0 |R ; D · at? ) = A(v 0 ; D 0 · a0t
? ).
q0
q
0 0
(b) lctB (D , a? ) ≥ lctB (D, a? ) withpequality in the case m ∈ V(jλ : q).
√
(c) If m ∈ V(jλ : q) then jλ R0 = JB (X 0 ; D 0 · a0λ
? ).
Proof. Let v = v 0 |R . By definition, vF (b• ) = vF (b0• ) and v(a? ) = v 0 (a0? ). Moreover,
F -finiteness and flatness prove that HomR0 (F∗e R0 , R0 ) ∼
= HomR (F∗e R, R) ⊗R R0 . A
t
direct computation then shows that A(v; D · a? ) = A(v 0 ; D 0 · a0t
? ).
Moving on to (b), the inequality is clear from (a). We wish to show equality
under the assumption m ∈ V(jλ : q). We can identify lctqB (D, a? ) as the supremum
0
over t such that 1 ∈ (jt : q), and likewise check 1 ∈ (j0t : q0 ) for lctqB (D 0 , a0? ).
Moreover, we can check this condition locally by checking at primes containing
0
(jλ : q). Of course, 1 ∈ (j0t : q0 ) is equivalent to v 0 (q0 ) + A(v 0 ; D 0 · a0t
? ) > 0 for all v ,
t
and we have already shown that the expression on the left is v(q) + A(v; D · a? ).
Building on ideas from the previous paragraph, suppose m ∈ V(jλ : q). Equality of radicals follows from the statement that v 0 (j0λ ) > 0 if and only if v(jλ ) =
λ
v 0 (jλ R0 ) > 0. These are both equivalent to A(v 0 ; D 0 · a0λ
? ) = A(v; D · a? ) < 0.
Remark 7.14. Jonsson and Mustaţă show that in fact j0t = jt R0 for all t ≥ 0, but
their proof uses that one can base change log resolutions (used to compute jt ) along
regular morphisms. Lacking this technique for computing jt , we do not know if this
holds in our setting.
Lemma 7.15 (cf. [JM12], Lemma 5.3). Let H be an SNC divisor on a regular F finite scheme X and let v ∈ Val(X) with retraction v = r(X,H) (v). Then A(v; CX ) ≤
A(v; CX ) with equality if and only if v = v.
Proof. Denoting p = cX (v) ⊆ cX (v) = m and localizing, we may assume X =
Spec(R, m), and set d = dim(X). Let r1 , . . . , rn ∈ p give a rsop for Rp and extend
this to a full rsop r1 , . . . , rd for R. A by-now-familiar computation gives
A(v; CX ) = v(r1 · · · rd ) =
n
X
j=1
v(rj ) ≤
d
X
v(rj ) ≤ A(v; CX ).
j=1
Strict inequality is achieved as follows. If p ( m then rn+1 6∈ p and v(rn+1 ) > 0,
hence the first inequality above is strict. Suppose then that p = m and let π : Y →
X be the blow-up of m with exceptional divisor E. We know cY (v) ∈ E and that
π ∗ H = π∗−1 (H) + dE is supported on an SNC divisor H 0 . Setting v 0 = r(Y,H 0 ) (v)
28
ERIC CANTON
gives a quasi-monomial v 0 with r(X,H) (v 0 ) = r(X,H) (v) = v, v 0 (KY /X ) = v(KY /X ),
and cY (v 0 ) 6= cY (v). Applying Proposition 3.15, we may assume X = Y , and now
we are back in the previous case.
Theorem 7.16 (cf. [JM12, Theorem 7.8]). Let v ∈ Val(X)∗ with A(v; D) < ∞.
The following assertions are equivalent:
(a) There exists a? on X such that v computes lctqB (D, a? ) < ∞.
(b) For every w ∈ Val(X)∗ such that w(a) ≥ v(a) for all ideals a on X, we have
A(w; D) + w(q) ≥ A(v; D) + v(q).
(c) The valuation v computes lctqB (D, a? (v)) < ∞.
Proof. Our proof is very similar to the one found in [JM12]. Notice, however, that
we do not have part (ii) of their theorem, since the multiplier ideals jt are not known
to be subadditive.
Certainly (c) implies (a). We prove (a) ⇒ (b) ⇒ (c). For the first implication, suppose v computes lctqB (D, a? ) < ∞, implying v(a? ) > 0. If w(a) ≥ v(a)
for all ideals a on X, then Lemma 7.1 shows w(a? ) ≥ v(a? ). Since v minimizes
lctqB (−; D, a? ), we know
A(v; D) + v(q)
A(w; D) + w(q)
≥
w(a? )
v(a? )
implying that
A(w; D) + w(q)
w(a? )
≥
≥1
A(v; D) + v(q)
v(a? )
proving (b).
Supposing (b) holds, we prove (c). We see from the definition that v(a? (v)) = 1,
and so assertion (c) is equivalent to proving
lctqB (w; D, a? (v)) ≥ A(v; D) + v(q).
If w(a? (v)) = 0 then lctqB (w; D, a? (v)) = +∞ and this is trivial. We therefore
assume w(a? (v)) > 0, in which case the desired inequality is
A(w; D) + w(q)
≥ A(v; D) + v(q).
w(a? (v))
(∗)
The left hand side is 0-homogeneous in w, and so we may assume w(a? (v)) = 1.
Since a? (v) is the sequence of valuation ideals for v, if w(a? (v)) = 1 then w(a) ≥ v(a)
for all ideals a. But then (∗) holds by assumption (b).
Still following the approach of Jonsson and Mustaţă, we now prove that we may
modify a1 and q so that they are both locally primary to a chosen minimal prime
of (jλ : q). For s ∈ N1 and an ideal m on X, we define
cj =
(7.16.1)
j
X
ai ms(i−j) .
i=0
The graded structure of a? implies that c? is graded. We see also that aj + msj ⊆ cj
for all j ≥ 1, implying that v(c? ) ≤ min{v(a? ), sv(m)} for all s and v ∈ Val(X).
In fact, this is an equality: if sv(m) ≤ v(a? ) then −v(ai ) ≤ −si v(m) for all i ≥ 1.
Therefore,
s v(m) =
s(j − i) v(m) + v(ai )
sj v(m) − v(ai ) + v(ai )
≤
.
j
j
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
29
We then see that
v(cj )
= min
0≤i≤j
j
v(ai ) + s(j − i)v(m)
j
= sv(m)
giving v(c? ) = sv(m). Otherwise, s v(m) > v(a? ), so there exists j0 such that
sj v(m) > v(aj ) for all j ≥ j0 . This implies that for j ≥ j0 and n ≥ j
v(an )
v(aj ) + v(an−j )
v(aj ) + s(n − j)v(m)
≤
<
< s v(m).
n
n
n
For any j < j0 we have (v(aj ) + s(n − j)v(m))/n → s v(m) as n → ∞. Therefore,
v(c? ) = v(a? ) when v(a? ) < sv(m).
Lemma 7.17 (cf. [JM12]). Assume lctqB (D, a? ) = λ < ∞ and let m be the generic
point of an irreducible component of V(jλ : q). Defining c? as in (7.16.1) with
s 0 gives lctqB (D, a? ) = lctqB (D, c? ). If v ∈ Val(X)∗ computes lctqB (D, c? ), then
v computes lctqB (D, a? ).
p
p
Proof. Note that (jλ : q) ⊆ m. First consider the special case m = (jλ : q) and
n
q
let n ∈ N1 such that mn q ⊆ jλ . Define λ0 = lctm
B (D, a? ) > λ and let s 0 so that
n/s < (λ0 − λ). For 0 < ε 1 we still have n/s < (λ0 (1 − ε) − λ).
We denote by W the set of v ∈ Val(X)∗ with lctqB (v; D, a? ) < ∞. Since mn ⊆
(jλ : q) and v(jλ : q) > 0, v(m) > 0. We also know that v(a? ) > 0 by (7.4).
Therefore,
A(v; D) + v(q)
v∈W min{v(a? ), sv(m)}
A(v; D) + v(q)
≤ inf
,
v∈Wε min{v(a? ), sv(m)}
n
o
λ
where Wε = v ∈ W : lctqB (v; D, a? ) ≤ 1−ε
= Vλ/(1−ε) . Note that when v ∈ Wε
we have
A(v; D) + v(q)
λ0 − lctqB (v; D, a? ) = λ0 −
v(a? )
(1 − ε)λ0 − λ
≥
1−ε
> (1 − ε)λ0 − λ > n/s.
lctqB (D, c? ) = inf
Additionally, the definitions imply that
λ0 −
A(v; D) + v(q)
sv(m)
≤
v(a? )
v(a? )
for all v ∈ Wε ⊆ W . We can thus estimate
lctqB (D, c? )
A(v; D) + v(q)
v(a? )
≤ inf
max 1,
v∈Wε
v(a? )
sv(m)
(
−1 )
A(v; D) + v(q)
n
A(v; D) + v(q)
0
≤ inf
max 1,
λ −
v∈Wε
v(a? )
s
v(a? )
= inf
v∈Wε
Thus,
lctqB (D, c? )
A(v; D) + v(q)
= lctqB (D, a? ).
v(a? )
≤ lctqB (D, a? ). The other inequality follows from monotonicity.
30
ERIC CANTON
Now treating the general case of m a minimal prime of (jλ : q), we may complete
the log canonical thresholds. By Theorem 7.16(c),
p 0 X at mpwithout changing
(jλ : q) = (jλ : q)R0 , so we reduce to the previous case.
The proof also demonstrates that if v computes lctqB (D, c? ) then
lctqB (D, a? ) = lctqB (D, c? )
A(v; D) + v(q)
v(c? )
A(v; D) + v(q)
≥
v(a? )
≥ lctqB (D, a? )
=
so the inequalities are equalities, proving that v also computes lctqB (D, a? ).
We are ready to prove the following existence result using essentially the same
method as Jonsson and Mustaţă.
Theorem 7.18 (cf. [JM12, Theorem 7.3]). Let λ = lctqB (D, a? ) < ∞. For any
generic point m of an irreducible component of V(jλ : q) there exists a valuation
with center m computing lctqB (D, a? ).
p
Proof. Applying Lemma 7.13, we may assume (R, m) is local and m = (jλ : q).
Enlarging a? using the previous proposition, we also assume that ms ⊆ a1 for some
s 0. Since λ < ∞, we can fix M ∈ R with λ < M < ∞, and lctqB (D, a? ) is
unchanged by considering only those v ∈ Val(X)∗ such that v(a? ) > 0 and
lctqB (v; D, a? ) =
A(v; D) + v(q)
≤ M.
v(a? )
Note that here we are using that finiteness of lctqB (v; D, a? ) implies cX (v) = m and
v(a? ) > 0, so 0 < v(a? ) ≤ sv(m). Then A(v; D) ≤ A(v; D) + v(q) ≤ M v(a? ) ≤
M sv(m). Re-scaling v if necessary, we may assume v(m) = 1, which now gives
A(v; D) ≤ N := M s. Thus, v ∈ VN as defined in (7.9). We see that
lctqB (D, a? ) = inf lctqB (w; D, a? ),
w∈VN
and compactness yields v ∈ VN achieving this minimum.
We do not get very much information about the valuation computing lctqB (D, a? )
above. In analogy with the discrete valuation case (where lctqB (v; D, a? ) = +∞ for
non-divisorial discrete valuations), one might expect these valuations to be locally
quasi-monomial. Jonsson and Mustaţă conjecture exactly this, albeit with quasimonomial valuations, since their schemes are excellent over Q.
Conjecture 7.19 (cf. [JM12, Conjecture 7.4]). Suppose lctqB (D, a? ) = λ < ∞.
◦ Weak version: for any generic point m of an irreducible component of
V(jλ : q) there exists a locally quasi-monomial valuation v ∈ Val(X)∗ computing lctqB (D, a? ) with cX (v) = m.
◦ Strong version: any non-trivial valuation computing lctqB (D, a? ) must be
locally quasi-monomial.
Jonsson and Mustaţă reduce this conjecture to the (potentially) simpler case of
affine space.
BERKOVICH LOG DISCREPANCIES IN POSITIVE CHARACTERISTIC
31
Conjecture 7.20 (cf. Conjecture 7.5 in [JM12]). Suppose X = Ank with n ≥ 1 and
k algebraically closed of characteristic p > 0. Consider a graded sequence of ideals
a? on X vanishing only at a closed point x ∈ X and lctqB (D, a? ) < ∞. Then:
◦ Weak version: there is a locally quasi-monomial valuation v computing
lctqB (D, a? ) with cX (v) = x.
◦ Strong version: any valuation with transcendence degree 0 over k computing lctqB (D, a? ) must be locally quasi-monomial.
As expected from [JM12], Conjecture 7.20 may be reduced to Conjecture 7.19.
To make reductions as they do, we need their second “enlarging lemma.”
Lemma 7.21 (cf. [JM12]). Suppose that λ = lctqB (D, a? ) < ∞ and let m ∈ X
with ms ⊆ a1 . If N > λs and r = q + mN then lctrB (D, a? ) = λ. Furthermore,
v ∈ Val(X)∗ computes lctqB (D, a? ) if and only if v computes lctrB (D, a? ).
Proof. As noted previously, q 6⊆ jλ but q ⊆ jt for all t < λ. To prove lctrB (D, a? ) = λ,
it therefore suffices to show that mN ⊆ jλ . This, of course, follows from our choice
of N > λs: we want to check that
(∗)
N v(m) + A(v; D, aλ? ) > 0 for every valuation v ∈ Val(X)∗ .
If v(m) = 0, then 0 ≤ v(a? ) ≤ s v(m) = 0 and A(v; D · aλ? ) = A(v; D) > 0 by
strong F -regularity. We therefore assume v(m) > 0. Rescaling v does not change
the truth of (∗), and so we assume v(m) = 1. Our assumed inclusion ms ⊆ a1 gives
λ v(a? ) ≤ sλ < N . We know A(v; D, aλ? ) = A(v; D) − λ v(a? ) ≥ 0, and we conclude
that (∗) holds.
We now prove that v computes lctqB (D, a? ) if and only if it computes lctrB (D, a? ).
The inclusion q ⊆ r implies that v(r) ≤ v(q), so if v ∈ Val(X)∗ computes λ =
lctqB (D, a? ) = lctrB (D, a? ) then
λ ≤ lctrB (v; D, a? ) =
A(v; D) + v(r)
A(v; D) + v(q)
≤
=λ
v(a? )
v(a? )
and v also computes lctrB (D, a? ). To finish the proof, it suffices to prove that
v(q) = v(r) whenever lctrB (v; D, a? ) = λ. Recalling that v(a? ) ≤ s v(m), we have
v(m) ≥
v(a? )
A(v; D) + v(r)
v(r)
=
>
.
s
λs
N
This implies that N v(m) > v(r) = min{v(q), v(mN )} and completes the proof.
Lemma 7.22. Let κ ⊆ K be an algebraic extension of F -finite fields and let AnK →
Anκ be the associated morphism. Extend D, a? , and q on Anκ to D 0 , a0? , q0 on AnK as
before. For every valuation v 0 ∈ Val(AnK ) centered at 0 with restriction v to Anκ ,
0
lctqB (v; D, a? ) = lctqB (v 0 ; D 0 , a0? ).
Proof. The proof is very similar to 7.13. Specifically, we know A(v; D) = A(v 0 ; D)
and v(a? ) = v 0 (a0? ), so A(v; D · at? ) = A(v 0 , D 0 · a0t
? ) for all t ≥ 0. As before,
v(q) = v 0 (q0 ), so the log canonical thresholds are equal.
Theorem 7.23 (cf. Theorem 7.5 in [JM12]). If the weak (resp. strong) version
of Conjecture 7.20 holds for every n ≤ N then the weak (resp. strong) version of
Conjecture 7.19 holds for all X with dim(X) ≤ N .
32
ERIC CANTON
Proof. We begin with the weak versions. Suppose λ = lctqB (D, a? ) < ∞ and let m
be a minimal prime of (jλ : q). If v ∈ Val(X)∗ computes lctqB (D, a? ) and is centered
at m, the continuous extension to the m-adic completion R0 of OX,x = R has the
same rational rank and transcendence degree by Lemma 7.11, and so v is locally
quasi-monomial if and only if this extension is, too. Applying Lemmas 7.17, 7.21,
and 7.13, we may assume R = κ[[x1 , . . . , xn ]], m = rad(R), and that ms ⊂ a1 ∩ q.
We now wish to apply Lemma 7.13 “in reverse” to reduce to the case of An .
Write S = κ[y1 , . . . , yn ] and identify the n = (y1 , . . . , yn )-adic completion of S with
R. Fixing a generating set for q, we approximate each generator g of q by g̃ ∈ S
with g − g̃ ∈ mN , where N 0 so that v(g) = v(g̃) for every valuation v centered at m. Performing similar approximations for every a ∈ a? , we obtain new
ideals q̃0 = q̃R and ã0? extended from S and with v(q̃) = v(q), v(ã? ) = v(a? ), and
nN ⊆ ã1 ∩ q̃. We have an isomorphism HomS (F∗e S, S) ⊗S R ∼
= HomR (F∗e R, R) for
all e ≥ 0 thanks to F -finiteness and flatness of completion. Thus, in every degree
De ⊆ HomR (F∗e R, R) may be identified as Ee ⊗S R for some Ee ⊆ HomS (F∗e S, S).
Now we apply Lemma 7.13 to ⊕e Ee = E , q̃, and ã? on S and conclude that
lctqB (D, a? ) = lctq̃B (E , ã? ) and that every valuation computing lctqB (D, a? ) is the
continuous extension of one on S centered at the closed point 0 ∈ Anκ . We are
assuming that Conjecture 7.20 holds on Anκ , thus have a locally quasi-monomial
valuation centered at 0 computing lctq̃B (E , ã? ), yielding such a valuation on Spec(R).
Let us now proceed to the implication between strong versions. Our strategy will
be similar, the crux being twice applying Lemma 7.13 to reduce to the case of affine
space. Suppose v ∈ Val(X)∗ computes lctqB (D, a? ) = λ < ∞. As we have seen,
m = cX (v) ∈ V(jλ : q). We wish to show that v is quasi-monomial. The dimension
formula (Theorem 15.6 in [Mat89]) implies that tr. degX (v) is the maximum of
dim(OX,cX (v) ) − dim(OY,cY (v) ) over all proper birational Y → X with Y regular.
Thus, by passing to some Y → X (and using Proposition 3.15), we may assume
tr. degX (v) = 0. We apply Lemma 7.13 with m = cX (v) and reduce to the case
assume X = Spec(κ[[x1 , . . . , xd ]]) with κ an F -finite field, v is a valuation centered
at the closed point of X, computing lctqB (D, a? ), and such that κ(v) is algebraic
over κ. Applying Theorem 7.16, we assume that a? = a? (v), implying ms ⊆ a1 for
some s 0. Now applying Lemma 7.21 we assume that mN ⊆ q for some N > λs.
Applying (7.13) again in reverse, followed by (7.22) to assume κ = κ, we conclude
v = v 0 |Y must be locally quasi-monomial, so v 0 is, too.
7.5. The Monomial Case. We now prove a special case of Conjecture 7.19 under
the assumption each as is generated by monomials.
Proposition 7.24 (cf. [JM12], Proposition 8.1). Let a? be a graded sequence of
monomial ideals on X = Ank = Spec(k[x1 , . . . , xn ]). For any nonzero ideal q on
X, there exists a monomial valuation computing λ = lctqB (CX , a? ) < ∞, and any
valuation computing lctqB (CX , a? ) is monomial.
Proof. Let H = div(x1 · · · xn ) and v ∈ Val(X)∗ . The retraction v := r(X,H) (v)
agrees with v on monomials, so v(a? ) = v(a? ) and v(q) ≤ v(q). Lemma 7.15 now
shows
lctqB (v; CX , a? ) =
A(v; CX ) + v(q)
A(v; CX ) + v(q)
≤
= lctqB (v; CX , a? )
v(a? )
v(a? )
REFERENCES
33
for every q, with a strict inequality when v 6= v. Thus we see (using Theorem 7.18)
there must be a monomial valuation computing lctqB (CX , a? ), and that any computing valuation is monomial.
Acknowledgements: This paper is part of the author’s Ph.D. thesis. I thank
Tommaso de Fernex, Mircea Mustaţă, Karl Schwede, and Kevin Tucker for numerous helpful discussions. This project grew out of discussions with Rankeya Datta,
Felipe Pérez, and Karl at the 2015 AMS MRC in Commutative Algebra. I am
indebted to Paolo Cascini, Mircea, and Karl for sharing their result on log discrepancies of divisorial valuations with me, upon which my definition of log discrepancy
is modeled. I am especially grateful to my advisor, Wenliang Zhang, for organizing
and supporting (DMS #1606414) my visit to the University of Utah during the
Spring 2016 semester, during which I established §1 − 3 and §5, and for his careful
reading of drafts of this document. Finally, I want to thank the University of Utah
mathematics department for being such wonderful hosts during my stay.
References
Ian M. Aberbach and Florian Enescu. “The structure of F -pure rings”.
In: Math. Z. 250.4 (2005), pp. 791–806. issn: 0025-5874.
[Amb99] Florin Ambro. “On minimal log discrepancies”. In: Mathematical Research Letters 6.5 (July 1999), pp. 573–580.
[Bli13]
Manuel Blickle. “Test ideals via algebras of p−e -linear maps”. In: J.
Algebraic Geom. 22.1 (2013), pp. 49–83. issn: 1056-3911. url: https:
//doi.org/10.1090/S1056-3911-2012-00576-1.
[Bou+15] S. Boucksom et al. “Valuation spaces and multiplier ideals on singular
varieties”. In: London Math. Soc. Lecture Note Ser. 417 (2015), pp. 29–
51.
[BS13]
Manuel Blickle and Karl Schwede. “p−1 -linear maps in algebra and geometry”. In: Commutative algebra. Springer, New York, 2013, pp. 123–
205. url: https://doi.org/10.1007/978-1-4614-5292-8_5.
[BST12]
Manuel Blickle, Karl Schwede, and Kevin Tucker. “F -signature of pairs
and the asymptotic behavior of Frobenius splittings”. In: Adv. Math.
231.6 (2012), pp. 3232–3258. issn: 0001-8708. doi: 10.1016/j.aim.
2012.09.007. url: http://dx.doi.org/10.1016/j.aim.2012.09.
007.
[CMS15] Paolo Cascini, Mircea Mustaţă, and Karl Schwede. Private correspondence. 2015.
[Con08]
Brian Conrad. “Several approaches to non-archimedean geometry”. In:
AMS University Lecture Series 45 (2008).
[Dat17]
R. Datta. “Uniform approximation of Abhyankar valuation ideals in
function field of prime characteristic”. In: ArXiv e-prints (May 2017).
arXiv: 1705.00447 [math.AG].
[DFN15] Antoine Ducros, Charles Favre, and Johannes Nicaise. Berkovich Spaces
and Applications. Vol. 2119. Lecture Notes in Mathematics. Springer,
2015. isbn: 9783319110288. doi: 10.1007/978-3-319-11029-5.
[DS16]
Rankeya Datta and Karen E. Smith. “Frobenius and valuation rings”.
In: Algebra and Number Theory 10.5 (2016), pp. 1057–1090. doi: 10.
2140/ant.2016.10.1057.
[AE05]
34
[DS17]
[Ein+06]
[ELS01]
[Fed83]
[Fol13]
[HH89]
[HR76]
[HW02]
[JM12]
[K192]
[KK05]
[KM98]
[Kol13]
REFERENCES
Rankeya Datta and Karen E. Smith. “Correction to the article “Frobenius and valuation rings” [ MR3531362]”. In: Algebra Number Theory
11.4 (2017), pp. 1003–1007. issn: 1937-0652. url: https://doi.org/
10.2140/ant.2017.11.1003.
Lawrence Ein et al. “Asymptotic invariants of base loci”. In: Ann. Inst.
Fourier (Grenoble) 56.6 (2006), pp. 1701–1734. issn: 0373-0956. url:
http://aif.cedram.org/item?id=AIF_2006__56_6_1701_0.
Lawrence Ein, Robert Lazarsfeld, and Karen E. Smith. “Uniform bounds
and symbolic powers on smooth varieties”. In: Invent. Math. 144.2 (2001),
pp. 241–252. issn: 0020-9910.
Richard Fedder. “F -purity and rational singularity”. In: Trans. Amer.
Math. Soc. 278.2 (1983), pp. 461–480. issn: 0002-9947.
G.B. Folland. Real Analysis: Modern Techniques and Their Applications. Pure and Applied Mathematics: A Wiley Series of Texts, Monographs and Tracts. Wiley, 2013. isbn: 9781118626399. url: https://
books.google.com/books?id=wI4fAwAAQBAJ.
Melvin Hochster and Craig Huneke. “Tight closure and strong F -regularity”.
In: Mém. Soc. Math. France (N.S.) 38 (1989). Colloque en l’honneur de
Pierre Samuel (Orsay, 1987), pp. 119–133. issn: 0037-9484.
Melvin Hochster and Joel L. Roberts. “The purity of the Frobenius
and local cohomology”. In: Advances in Math. 21.2 (1976), pp. 117–172.
issn: 0001-8708.
Nobuo Hara and Kei-Ichi Watanabe. “F-regular and F-pure rings vs.
log terminal and log canonical singularities”. In: J. Algebraic Geom.
11.2 (2002), pp. 363–392. issn: 1056-3911.
Mattias Jonsson and Mircea Mustaţă. “Valuations and asymptotic invariants for sequences of ideals”. In: Ann. Inst. Fourier (Grenoble) 62.6
(2012), 2145–2209 (2013). issn: 0373-0956. doi: 10.5802/aif.2746.
url: http://dx.doi.org/10.5802/aif.2746.
János Kollár and 14 coauthors. Flips and abundance for algebraic threefolds. Papers from the Second Summer Seminar on Algebraic Geometry
held at the University of Utah, Salt Lake City, Utah, August 1991,
Astérisque No. 211 (1992). Paris: Société Mathématique de France,
1992, pp. 1–258.
Hagen Knaf and Franz-Viktor Kuhlmann. “Abhyankar places admit local uniformization in any characteristic”. eng. In: Annales scientifiques
de l’École Normale Supérieure 38.6 (2005), pp. 833–846. url: http :
//eudml.org/doc/82676.
János Kollár and Shigefumi Mori. Birational geometry of algebraic varieties. Vol. 134. Cambridge Tracts in Mathematics. With the collaboration of C. H. Clemens and A. Corti, Translated from the 1998 Japanese
original. Cambridge: Cambridge University Press, 1998, pp. viii+254.
isbn: 0-521-63277-3.
János Kollár. Singularities of the minimal model program. Vol. 200.
Cambridge Tracts in Mathematics. With a collaboration of Sándor
Kovács. Cambridge: Cambridge University Press, 2013, pp. x+370. isbn:
978-1-107-03534-8.
REFERENCES
[Kol97]
[Laz04]
[Mat89]
[MR85]
[RR85]
[Sch08]
[Sch09]
[Sch10]
[ST14]
35
János Kollár. “Singularities of pairs”. In: Algebraic geometry—Santa
Cruz 1995. Vol. 62. Proc. Sympos. Pure Math. Providence, RI: Amer.
Math. Soc., 1997, pp. 221–287.
Robert Lazarsfeld. Positivity in algebraic geometry. II. Vol. 49. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Positivity
for vector bundles, and multiplier ideals. Berlin: Springer-Verlag, 2004,
pp. xviii+385. isbn: 3-540-22534-X.
Hideyuki Matsumura. Commutative ring theory. Second Ed. Vol. 8.
Cambridge Studies in Advanced Mathematics. Translated from the Japanese by M. Reid. Cambridge: Cambridge University Press, 1989, pp. xiv+320.
isbn: 0-521-36764-6.
V. B. Mehta and A. Ramanathan. “Frobenius splitting and cohomology
vanishing for Schubert varieties”. In: Ann. of Math. (2) 122.1 (1985),
pp. 27–40. issn: 0003-486X.
S. Ramanan and A. Ramanathan. “Projective normality of flag varieties
and Schubert varieties”. In: Invent. Math. 79.2 (1985), pp. 217–224. issn:
0020-9910.
Karl Schwede. “Generalized test ideals, sharp F -purity, and sharp test
elements”. In: Math. Res. Lett. 15.6 (2008), pp. 1251–1261. issn: 10732780.
Karl Schwede. “F -adjunction”. In: Algebra Number Theory 3.8 (2009),
pp. 907–950. issn: 1937-0652. doi: 10.2140/ant.2009.3.907. url:
http://dx.doi.org/10.2140/ant.2009.3.907.
Karl Schwede. “Centers of F -purity”. In: Math. Z. 265.3 (2010), pp. 687–
714. issn: 0025-5874. doi: 10.1007/s00209-009-0536-5. url: http:
//dx.doi.org/10.1007/s00209-009-0536-5.
Karl Schwede and Kevin Tucker. “On the behavior of test ideals under
finite morphisms”. In: J. Algebraic Geom. 23.3 (2014), pp. 399–443.
issn: 1056-3911. url: https://doi.org/10.1090/S1056-3911-201300610-4.
| 0math.AC
|
Information and estimation
in Fokker-Planck channels
Andre Wibisono∗ , Varun Jog∗, and Po-Ling Loh∗†
arXiv:1702.03656v1 [cs.IT] 13 Feb 2017
Departments of Electrical & Computer Engineering∗ and Statistics†
University of Wisconsin - Madison
Madison, WI 53706
Email: [email protected], [email protected], [email protected]
Abstract—We study the relationship between information- and
estimation-theoretic quantities in time-evolving systems. We focus
on the Fokker-Planck channel defined by a general stochastic
differential equation, and show that the time derivatives of
entropy, KL divergence, and mutual information are characterized by estimation-theoretic quantities involving an appropriate
generalization of the Fisher information. Our results vastly
extend De Bruijn’s identity and the classical I-MMSE relation.
I. I NTRODUCTION
Information theory and statistical estimation are closely
intertwined. Various identities and inequalities arise from
fundamental concepts such as mutual information, Fisher
information, and estimation error, and close parallels between
the fields provide an avenue for devising and deriving new
results. As a canonical example, the information-estimation
result known as the moment-entropy inequality [1] states that
among all continuous random variables with a fixed variance,
Gaussian random variables maximize entropy. Furthermore,
Stam’s inequality [2] states that for a fixed Fisher information,
Gaussian random variables minimize entropy. The celebrated
Cramér-Rao bound from statistics [3], which establishes a
lower bound on the variance of an estimator in terms of the
Fisher information, follows from the aforementioned facts.
In the past decade, significant effort has been devoted to uncovering new relationships between information-theoretic and
estimation-theoretic quantities, beginning with the I-MMSE
identity of Guo et al. [4] for additive Gaussian noise channels.
The identity, provided in Theorem 1 below, states that the
derivative of the mutual information between the channel input
and output with respect to the signal to noise ratio (snr) is
proportional to the minimum mean square error (mmse) in
estimating the input from the output. As derived in Stam [2],
the result is equivalent to De Bruijn’s identity (cf. equation
(2)). The restatement of De Bruijn’s identity in terms of the
mmse spawned a host of additional information-estimation
results, including an extension to non-Gaussian additive noise
[5], a generalization to the mismatched estimation setting
[6], and several pointwise information-estimation relations [7].
Different I-MMSE type relations were also obtained for the
Poisson channel [8], [9] and Lévy channel [10].
In information theory, a “channel” is a conditional distribution relating input symbols to output symbols. Whereas
this coding-theoretic model is very useful for communication
channels, however, it possesses certain drawbacks. Many realworld examples such as weather systems and financial markets
are best explained as systems evolving in time according to
random and deterministic influences. Although it is possible
to view an evolving system as a communication channel,
where the current state is the channel input and the state at a
future time is the channel output, such an interpretation lacks
insight about the path of the system. Nonetheless, informationtheoretic ideas are still useful in characterizing the behavior
of time-evolving systems. For instance, one might characterize
how much information about the future is contained in the
present state using quantities such as entropy, KL divergence,
and mutual information. This has been discussed extensively
in the climate science literature [11], [12], [13], [14].
One of the simplest and most useful ways of modeling
evolving systems is via continuous-time Markov chains with a
continuous state space, which may be analyzed using stochastic differential equations (SDEs). In particular, the probability
distribution of such systems evolves according to a partial
differential equation known as the Fokker-Planck equation. We
focus on an information-theoretic analysis of time-evolving
systems described by SDEs, and study the rate of change
of various fundamental quantities as a function of time. We
show that these rates are conveniently expressed in terms of
a generalized Fisher information, so our results may be interpreted as generalizations of De Bruijn’s identity for the SDE
or Fokker-Planck setting. Notably, we obtain a clean identity
expressing the time derivative of mutual information in terms
of the mutual Fisher information, allowing us to derive new IMMSE relations. Our results are readily specialized to specific
stochastic processes, including Brownian motion, OrnsteinUhlenbeck processes, and geometric Brownian motion.
The remainder of the paper is organized as follows: In
Section II, we review existing results and present the family
of SDEs to be analyzed. In Section III, we develop our main
results concerning the evolution of entropy, KL divergence,
and mutual information in terms of estimation-theoretic quantities. We conclude in Section IV with a discussion of open
questions. Proofs are contained in the Appendix.
II. BACKGROUND
AND PROBLEM SETUP
We begin by presenting formal statements of De Bruijn’s
identity and the I-MMSE relation, followed by a detailed
characterization of the SDE framework discussed in our paper.
A. DeBruijn’s identity and I-MMSE
Consider the additive Gaussian noise channel
√
Xt = X0 + tZ
(1)
where Z ∼ N (0, 1) is independent of X0 , and t > 0 is a time
parameter—in this case equal to the variance of the noise—
that controls how much randomness is added to the system.
As t increases, we expect the output Xt to be more random.
De Bruijn’s identity [2] confirms this intuition and asserts that
1
d
H(Xt ) = J(Xt ),
dt
2
(2)
R
where H(Xt ) = − R pt (x) log pt (x)dx is the Shannon entropy
with pt denoting the density of Xt , and
"
2 # Z ′ 2
∂
pt (x)
J(Xt ) = E
log pt (Xt )
=
dx > 0 (3)
∂x
R pt (x)
is the (nonparametric) Fisher information.
Another common parameterization of the channel (1) is
√
Ysnr = snrX + Z
where snr > 0 is the signal to noise ratio and Z ∼ N (0, 1)
is independent of X. Guo et al. [4] established the following
I-MMSE relation, which states that the mutual information
I(snr) = I(X; Ysnr ) increases at a rate given by the mmse, and
showed their result is equivalent to De Bruijn’s identity (2).
A convenient way to study Xt is via its density pt . When
Xt follows the SDE (5), the density function p(x, t) = pt (x)
satisfies a partial differential equation called the Fokker-Planck
equation. This result is classical and technically requires that
a(x, t) and σ(x, t) satisfy appropriate regularity and growth
conditions (e.g., smoothness and Lipschitz properties [15]).
Lemma 1. [Proof in Appendix A-A] The density p(x, t) of the
process described by equation (5) satisfies
∂p(x, t)
∂
1 ∂2
(b(x, t)p(x, t)), (6)
= − (a(x, t)p(x, t)) +
∂t
∂x
2 ∂x2
where b(x, t) = σ(x, t)2 .
Note that Lemma 1 implies the channel (5) is linear in the
space of input distributions, which means a mixture of inputs
produces a mixture of outputs. Also note that when b(x, t) is
independent of x, our model falls under the PDE framework
of Toranzo et al. [16]; however, our SDE framework (5) is
somewhat more general. Consider the following examples:
1) Brownian motion: As discussed above, the choice a ≡ 0
d
and
√ σ ≡ 1 produces the solution Xt = X0 + Wt = X0 +
tZ ∼ N (X0 , t), which is precisely the Gaussian channel (1).
The Fokker-Planck equation is the classical heat equation
∂p(x, t)
1 ∂ 2 p(x, t)
.
=
∂t
2 ∂x2
Any starting point X0 = x0 gives rise to the explicit solution
(x − x0 )2
1
exp −
.
(7)
p(x, t) = √
2t
2πt
2) Ornstein-Uhlenbeck process: The SDE is
Theorem 1 (Guo et al. [4]). We have
dXt = −αXt dt + dWt ,
d
1
I(snr) = mmse(X | Ysnr ),
dsnr
2
2
where mmse(X | Ysnr ) = E[(X − E[X | Ysnr ]) ] denotes the
minimum mean square error for estimating X from Ysnr .
In terms√
of the time parameterization (1), by setting snr =
and Xt = tY1/t we see that Theorem 1 is equivalent to
1
d
I(X0 ; Xt ) = − 2 mmse(X0 | Xt ).
dt
2t
B. SDEs and Fokker-Planck equation
1
t
(4)
with α > 0. This corresponds to a(x, t) = −αx and
σ(x, t) = b(x, t) ≡ 1. The Ornstein-Uhlenbeck process is
mean-reverting and arises in stochastic modeling of interest
rates and particle velocities.
The explicit
q solution is Xt =
R t αs
d −αt
1
−αt
X0 + 0 e dWs = e X0 + 2α
e
(1 − e−2αt )Z, so
1
) as t → ∞. The Fokker-Planck equation is
Xt → N (0, 2α
Consider a general channel that outputs a real-valued
stochastic process (Xt )t≥0 following the SDE
∂
1 ∂ 2 p(x, t)
∂p(x, t)
,
(8)
= α (xp(x, t)) +
∂t
∂x
2 ∂x2
which may also be solved explicitly [17]. Note that if α → 0,
we recover Brownian motion (7).
3) Geometric Brownian motion: The SDE is
dXt = a(Xt , t)dt + σ(Xt , t)dWt ,
dXt = µXt dt + σXt dWt ,
(5)
where (Wt )t≥0 is standard Brownian motion, and a(x, t) and
σ(x, t) > 0 are arbitrary real-valued smooth functions. The
choice a ≡ 0 and σ ≡ 1 generates the Gaussian channel (1),
but the SDE framework (5) is considerably more general. The
drift a(x, t)dt is the deterministic part of the dynamics, and the
diffusion σ(x, t)dWt introduces randomness by incrementally
adding Gaussian noise: For small δ > 0, we have
√
Xt+δ ≈ Xt + a(Xt , t)δ + σ(Xt , t) δZ
where Z ∼ N (0, 1) is independent of Xt .
where µ ∈ R and σ > 0. This corresponds to a(x, t) = µx,
σ(x, t) = σx, and b(x, t) = σ 2 x2 . Geometric Brownian
motion is used to model asset prices in financial mathematics,
notably in Black-Scholes option pricing [18]. The explicit
σ2
solution
is Xt = X0 exp((µ
− 2 )t + σWt ), so log Xt ∼
2
N log X0 + (µ − σ2 )t, σ 2 t . The Fokker-Planck equation is
∂
σ2 ∂ 2 2
∂p(x, t)
(x p(x, t)),
= −µ (xp(x, t)) +
∂t
∂x
2 ∂x2
which also has an explicit solution [17].
(9)
III. M AIN
B. From information to estimation
RESULTS
We now generalize the information-estimation relations
for the Gaussian channel (1) to the SDE channel (5). We
express the time derivatives of fundamental informationtheoretic quantities—entropy, relative entropy, and mutual
information—in terms of a generalized Fisher information,
producing analogs of De Bruijn’s identity (2). Although entropy is not always increasing, relative entropy and mutual
information decrease at a rate given by the relative and mutual
Fisher information, respectively. We further interpret the Fisher
information via a generalized Bayesian Cramer-Rao lower
bound and express the mutual Fisher information as the mmse
of estimating a function of the input from the output, thus
producing generalizations of the I-MMSE relation (4).
A. Generalized Fisher information
Let b : R → (0, ∞) be a positive function. We define the
Fisher information with respect to b to be
2 Z
∂
p′ (x)2
log p(X)
= b(x)
dx, (10)
Jb (X) = E b(X)
∂x
p(x)
R
where p denotes the distribution of the random variable X.
When b ≡ 1, we obtain the usual Fisher information (3). Our
generalized Fisher information differs from other notions in
literature in that we are still motivated by Shannon entropy.
In contrast, Toranzo et al. [16] consider φ-Fisher information
to study φ-entropy, while Lutwak et al. [19] and Bercher [20]
consider q-Fisher information to study q-entropy, which includes Rényi and Tsallis entropies. For distributions p and q,
we define the relative Fisher information with respect to b:
Jb (p k q) =
Z
p(x)b(x)
R
∂
p(x)
log
∂x
q(x)
2
dx,
(11)
Alternatively, Jb (p k q) may be viewed as the Bregman divergence of Jb . Recall that the KL divergence, or relative entropy,
may be written as the Bregman divergence of Shannon entropy.
Lemma 2. [Proof in Appendix A-B]
We define the statistical Fisher information with respect to b
as the weighted average of the pointwise Fisher information:
Z
pY (y) b(y) Φ(X | Y = y) dx.
(14)
Φb (X | Y ) =
R
The following key result provides a bridge between information and estimation:
Theorem 2. [Proof in Appendix B-A] The mutual Fisher
information is equal to the statistical Fisher information:
Jb (X; Y ) = Φb (X | Y ).
We now derive two theorems illustrating the intimate connections between the statistical Fisher information and quantities in estimation theory.
1) Estimation-theoretic lower bound: The pointwise Fisher
information (13) has a natural estimation interpretation via the
Cramer-Rao lower bound; similarly, the statistical Fisher information (14) provides a lower bound on the estimation error
when we have a prior on the parameters. The following result
provides a weighted version of van Trees’ inequality [21]:
Theorem 3. [Proof in Appendix B-B] Consider a parameterized family of distributions {pY |X ( · | x)} with prior pX (x).
For any estimator T (y) of x,
1
1
1
2
E
(T (Y ) − X) ≥
=
.
b(X)
Φb (Y |X) + Jb (X)
Jb (X|Y )
Thus, the conditional Fisher information Jb (X | Y ) is inversely proportional to the hardness of estimating X from Y .
2) mmse relation: We define the mmse of Y given X with
respect to b as
mmseb (Y | X) = min E[b(X)(T (X) − Y )2 ],
T
Jb (p k q) = Jb (p) − Jb (q) − h∇Jb (q), p − qi.
In particular, Jb (p k q) ≥ 0 shows that Jb (p) is convex in p.
Analogous to the the mutual information, which measures
the reduction in conditional entropy, we define the mutual
Fisher information of random variables X and Y with respect
to b:
Jb (X; Y ) = Jb (Y | X) − Jb (Y ),
In fact, the mutual Fisher information (12) is equal to a
natural generalization of the statistical Fisher information, a
central estimation-theoretic quantity. Recall that the pointwise
statistical Fisher information Φ(y) = Φ(X | Y = y) of a
parameterized family of distributions {pX|Y ( · | y)} is
Z
∂
2
Φ(X|Y = y) = pX|Y (x|y)
log pX|Y (x|y) dx. (13)
∂y
R
(12)
R
where Jb (Y | X) = R pX (x)Jb (Y | X = x) dy is the conditional Fisher information. Note that Jb (X; Y ) is the difference
between E[Jb (pY |X (· | X))] and Jb (E[pY |X (· | X)]) = Jb (Y ).
Since Jb is convex, this implies Jb (X; Y ) ≥ 0.
where the minimization is over all estimators T (X). Note
that the minimizer corresponds to the conditional expectation
T (X) = E[Y |X], regardless of b. For a parameterized family
of distributions {pY |X ( · | x)}, consider the pointwise score
function
∂
log pY |X (y | x).
ϕ(x, y) =
∂y
Given a prior pX , by Bayes rule we can define the other conditional distribution pX|Y (x|y) ∝ pX (x)pY |X (y|x). Observe
that for every fixed y, if X ∼ pX|Y (· |y), then ϕ(X, y) is an
unbiased estimator of the nonparametric score function:
E[ϕ(X, y) | Y = y] =
∂
log pY (y).
∂y
This fact leads to the following result:
Theorem 4. [Proof in Appendix B-C]
Jb (X; Y ) = mmseb (ϕ(X, Y ) | Y ).
C. De Bruijn’s identity
We now describe our generalizations of De Bruijn’s identity.
In the statements below, we write bt (x) = b(x, t) = σ(x, t)2 .
1) Time derivative of entropy: Our first result relates the
rate of change of Shannon entropy to the Fisher information:
Theorem 5. [Proof in Appendix C-A] Let Xt be the output
of the SDE (5). Then
d
1
1 ∂2
∂
b(X
,
t)
.
H(Xt ) = Jbt (Xt ) + E
a(Xt , t) −
t
dt
2
∂x
2 ∂x2
Note that in the case of heat equation, when a ≡ 0 and
b ≡ 1, this result recovers the classical De Bruijn’s identity (2).
However, in the general case, the entropy does not necessarily
always increase. This may seem odd, but as the results below
show, we obtain monotonicity by considering relative entropy.
derivative of KL divergence: Let K(p k q) =
R 2) Time p(x)
p(x)
log
q(x) dx denote the KL divergence. The following
R
result establishes that the relative entropy between any two
solutions is always decreasing, with a rate given by the relative
Fisher information:
Theorem 6. [Proof in Appendix C-B] Let Xt , Yt denote the
output random variables of the channel (5) with distributions
pt , qt . Then
1
d
K(pt k qt ) = − Jbt (pt k qt ).
dt
2
Thus, the KL divergence is a contracting map along any
two trajectories pt and qt , implying the existence of at most
one fixed point of the channel; i.e., a stationary probability
distribution p∞ satisfying equation (6). However, such a
distribution p∞ may not always exist, as in the case of the
heat equation.
3) Time derivative of mutual information: Recall that the
mutual information satisfies
I(X; Y ) = H(Y ) − H(Y | X),
R
where H(Y | X) = R pX (x)H(Y | X = x) dx is the conditional Shannon entropy. Consider the time derivative of
I(X0 ; Xt ), where Xt is the output of the channel (5) with input X0 . Theorem 5 expresses the time derivative of H(Xt ) in
terms of the Fisher information Jbt (Xt ); since the channel (5)
is linear, we also obtain a formula for the time derivative of
the conditional entropy H(Xt |X0 ) in terms of Jbt (Xt |X0 ).
Recalling definition (12), this yields the following result:
Theorem 7. [Proof in Appendix C-C] The output Xt of the
SDE channel (5) with input X0 satisfies
d
1
I(X0 ; Xt ) = − Jbt (X0 ; Xt ).
dt
2
D. Special cases
We specialize our results to the examples in Section II-B.
For more details, see Appendix D.
1) Brownian motion: Since b ≡ 1, the function Jb is the
usual nonparametric Fisher information J, and Theorem 5
yields De Bruijn’s identity (2). We calculate the conditional
√
d
Fisher information using the coupling Xt = X0 + tZ:
Z
√
1
pX0 (x0 ) J(x0 + tZ) dx0 = .
J(Xt | X0 ) =
t
R
The mutual Fisher information is Φ(X0 |Xt ) = 1t −J(Xt ) ≥ 0,
which implies J(Xt ) ≤ 1t . Theorem 7 says that
1
d
1
J(Xt ) −
.
I(X0 ; Xt ) =
dt
2
t
Using the formula (7), we may compute the score function:
ϕ(x0 , xt ) =
∂
1
log pXt |X0 (xt |x0 ) = − (xt − x0 ).
∂xt
t
By Theorem 4, we then obtain
1
1
mmse((Xt − X0 )| Xt ) = 2 mmse(X0 | Xt ),
t2
t
which, with Theorem 7, recovers the I-MMSE identity (4).
2) Ornstein-Uhlenbeck process: Again, we have Jb = J.
Applying Theorem 5 yields
Φ(X0 | Xt ) =
1
d
H(Xt ) = J(Xt ) − α.
dt
2
Note that if α ≤ 0, entropy always increases—but if α > 0,
which is the regime of interest,
q entropy need not be monotonic.
d −αt
1
(e2αt − 1)Z), we may com(X0 + 2α
Using Xt = e
pute the conditional Fisher information J(Xt | X0 ) = 1−e2α
−2αt
and the mutual Fisher information
2α
Φ(X0 | Xt ) =
− J(Xt ).
1 − e−2αt
Since Φ(X0 | Xt ) ≥ 0, this yields the bound J(Xt ) ≤
2α
1−e−2αt , which monotonically decreases to 2α as t → ∞.
Using the explicit solution to equation (8), we obtain
2α(xt − e−αt x0 )
.
1 − e−2αt
By Theorems 4 and 7, we then deduce the I-MMSE relation
ϕ(x0 , xt ) = −
d
2α2 e−2αt
mmse(X0 | Xt ).
I(X0 ; Xt ) = −
dt
(1 − e−2αt )2
3) Geometric Brownian motion: Since b(x, t) = σ 2 x2 , we
have Jb 6= J. Applying Theorem 5 yields
d
1
1
H(Xt ) = Jb (Xt ) + µ − σ 2 .
dt
2
2
Thus, the entropy may not increase monotonically. Using the
explicit solution to equation (9), we may compute
!
2
1 log xt − log x0 − (µ − σ2 )t
+1 .
ϕ(x0 , xt ) = −
xt
σ2 t
Then by Theorem 4, we derive the I-MMSE relation
1
d
I(X0 ; Xt ) = − 2 2 mmse(log X0 | Xt ).
dt
2σ t
E. Multivariate extension
Our results extend without difficulty to the multivariate
setting where Xt is a stochastic process in Rd evolving
according to the SDE (5), where a(x, t) ∈ Rd is a drift vector,
σ(x, t) ∈ Rd×d is a covariance matrix, and Wt is standard
Brownian motion in Rd . The weight matrix is given by
b(x, t) = σ(x, t)σ(x, t)⊤ ,
which is assumed to be uniformly positive definite.
We define the generalized Fisher information (10) with
respect to a positive definite matrix b(x):
i Z k∇p(x)k2
h
b(x)
dx,
Jb (X) = E k∇ log p(X)k2b(X) =
p(x)
d
R
where kvk2b(x) = v ⊤ b(x)v = Tr(b(x)vv ⊤ ) is the Mahalanobis
inner product of v ∈ Rd . The relative (11) and mutual Fisher
information (12) are defined similarly, and the statistical Fisher
information (14) is defined as
Z
Φb (X | Y ) =
pY (y) Tr(b(y)Φ(X | Y = y)) dy,
Rd
matrix
Rwhere Φ(X | Y ∂= y) is the usual ∂Fisher information
⊤
p
(x|y)(
log
p
(x|y))(
log
p
(x|y))
dx.
d
X|Y
X|Y
X|Y
∂y
∂y
R
With these definitions, our results hold unchanged (see
Appendix E for more details).
IV. D ISCUSSION
AND FUTURE WORK
We have established information-estimation identities for
time-evolving systems. Our results extend the classical De
Bruijn’s identity, which concerns the rate of entropy growth
in a Brownian motion process, to the time derivatives of
entropy, KL divergence, and mutual information for processes
described by general SDEs. The predictability of such systems relies on the information contained in the current state
regarding future states. At a high level, the current state
contains progressively less information about future states
of the system; we derive the specific rates of change in
information in terms of quantities arising from estimation.
Theorem 7 relates the derivative of I(X0 ; Xt ) to the statistical Fisher information, which by Theorem 3 is inversely
proportional to the difficulty of estimating X0 from Xt . This
difficulty should increase with time, suggesting that I(X0 ; Xt )
may decrease in a convex manner. Costa [22] showed that the
entropy of Brownian motion is a convex function, and Chen
et al. [23] showed that the first four derivatives of entropy
alternate in sign. We conjecture a similar property for higherorder derivatives of mutual information in general SDEs.
We suspect that the relationship between Fisher information
and estimation may be generalized to Bregman divergences.
2
Our results are based on the function fx (y) = 12 b(x)y
p , corresponding to Gaussian randomness generated via b(x)dWt .
It may be interesting to investigate stochastic processes corresponding to general convex functions. Finally, one could
explore the connection to optimal transport, which interprets
the Fokker-Planck equation (6) as the gradient flow of relative
entropy in the space of probability densities with respect to
the Wasserstein metric. Villani [24] then views Theorem 6 as
an analog of De Bruijn’s identity. Note that such a viewpoint
focuses on a PDE rather than SDE formulation.
R EFERENCES
[1] T. M. Cover and J. A. Thomas, Elements of information theory. John
Wiley & Sons, 2012.
[2] A. Stam, “Some inequalities satisfied by the quantities of information of
Fisher and Shannon,” Information and Control, vol. 2, no. 2, pp. 101–
112, 1959.
[3] E. L. Lehmann and G. Casella, Theory of point estimation. Springer
Science & Business Media, 2006.
[4] D. Guo, S. Shamai, and S. Verdú, “Mutual information and minimum
mean-square error in Gaussian channels,” IEEE Transactions on Information Theory, vol. 51, no. 4, pp. 1261–1282, 2005.
[5] D. P. Palomar and S. Verdú, “Representation of mutual information via
input estimates,” IEEE Transactions on Information Theory, vol. 53,
no. 2, pp. 453–470, 2007.
[6] S. Verdú, “Mismatched estimation and relative entropy,” IEEE Transactions on Information Theory, vol. 56, no. 8, pp. 3712–3720, 2010.
[7] K. Venkat and T. Weissman, “Pointwise relations between information
and estimation in Gaussian noise,” IEEE Transactions on Information
Theory, vol. 58, no. 10, pp. 6264–6281, 2012.
[8] R. Atar and T. Weissman, “Mutual information, relative entropy, and
estimation in the poisson channel,” IEEE Transactions on Information
theory, vol. 58, no. 3, pp. 1302–1318, 2012.
[9] J. Jiao, K. Venkat, and T. Weissman, “Pointwise relations between
information and estimation in the poisson channel,” in 2013 IEEE
International Symposium on Information Theory, pp. 449–453, IEEE,
2013.
[10] J. Jiao, K. Venkat, and T. Weissman, “Relations between information
and estimation in scalar Lévy channels,” in 2014 IEEE International
Symposium on Information Theory, pp. 2212–2216, IEEE, 2014.
[11] L.-Y. Leung and G. R. North, “Information theory and climate prediction,” Journal of Climate, vol. 3, no. 1, pp. 5–14, 1990.
[12] T. Schneider and S. M. Griffies, “A conceptual framework for predictability studies,” Journal of Climate, vol. 12, no. 10, pp. 3133–3155,
1999.
[13] R. Kleeman, “Measuring dynamical prediction utility using relative
entropy,” Journal of the Atmospheric Sciences, vol. 59, no. 13, pp. 2057–
2072, 2002.
[14] T. DelSole, “Predictability and information theory. part I: Measures of
predictability,” Journal of the Atmospheric Sciences, vol. 61, no. 20,
pp. 2425–2440, 2004.
[15] M. C. Mackey, Time’s Arrow: The Origins of Thermodynamics Behavior.
Springer-Verlag, 1992.
[16] I. V. Toranzo, S. Zozor, and J.-M. Brossier, “Generalization of the
de Bruijn’s identity to general φ-entropies and φ-Fisher informations,”
arXiv preprint arXiv:1611.09400, 2016.
[17] S. E. Shreve, Stochastic calculus for finance II: Continuous-time models,
vol. 11. Springer Science & Business Media, 2004.
[18] F. Black and M. Scholes, “The pricing of options and corporate
liabilities,” The Journal of Political Economy, vol. 81, no. 3, pp. 637–
654, 1973.
[19] E. Lutwak, D. Yang, and G. Zhang, “Representation of mutual information via input estimates,” IEEE Transactions on Information Theory,
vol. 53, no. 2, pp. 453–470, 2007.
[20] J.-F. Bercher, “Some properties of generalized fisher information in
the context of nonextensive thermostatistics,” Physica A: Statistical
Mechanics and its Applications, vol. 392, no. 15, pp. 3140 – 3154,
2013.
[21] R. D. Gill and B. Y. Levit, “Applications of the van Trees inequality: A
Bayesian Cramér-Rao bound,” Bernoulli, pp. 59–79, 1995.
[22] M. Costa, “A new entropy power inequality,” IEEE Transactions on
Information Theory, vol. 31, no. 6, pp. 751–760, 1985.
[23] F. Cheng and Y. Geng, “Higher order derivatives in Costa’s entropy
power inequality,” IEEE Transactions on Information Theory, vol. 61,
no. 11, pp. 5892–5905, 2015.
[24] C. Villani, Topics in optimal transportation. No. 58, American Mathematical Soc., 2003.
A PPENDIX A
P ROOFS OF LEMMAS
In this Appendix, we provide proofs of the technical lemmas
stated in our paper. Throughout, we assume that all densities
and functions are smooth and rapidly decreasing, so we may
differentiate under the integral sign and apply integration
by parts with all boundary terms zero. These are standard
assumptions in parabolic partial differential equations, and
may be ensured by constraining the functions a and σ in
the channel definition (5) to be smooth and Lipschitz (see
Mackey [15] and the references cited therein for technical
details). These assumptions hold for all our examples.
A. Proof of Lemma 1
The Fokker-Planck equation is a standard result; we follow
the outline of Mackey [15, §11].
Let h : R → R be a smooth, compactly supported function,
and consider the expectation
Z
p(x, t)h(x)dx.
E(t) = E[h(Xt )] =
R
d
We compute the time derivative Ė(t) = dt
E(t) in two ways:
First, by differentiating under the integral sign,
Z
Z
∂p(x, t)
d
h(x)dx. (15)
p(x, t)h(x)dx =
Ė(t) =
dt R
∂t
R
Second, we compute Ė(t) as the limit of the difference
1
δ (E(t+ δ)− E(δ)) as δ → 0. From the channel definition (5),
for small δ, we have
√
Xt+δ = Xt + a(Xt , t)δ + σ(Xt , t) δZ,
where Z ∼ N (0, 1) is independent of Xt . By the second-order
Taylor expansion of h, and ignoring terms of order smaller
than δ, we then have
√
h(Xt+δ ) = h(Xt ) + δh′ (Xt )σ(Xt , t)Z
1
+ δ h′ (Xt )a(Xt , t) + h′′ (Xt )σ(Xt , t)2 Z 2 .
2
We take the expectation of both sides. Since Z and Xt are
independent,
E[h′ (Xt )σ(Xt , t)Z] = E[h′ (Xt )σ(Xt , t)] E[Z] = 0,
so the middle term vanishes. Using E[Z 2 ] = 1, we obtain
E[h(Xt+δ )] = E[h(Xt )]
h
i
1
+ δE h′ (Xt )a(Xt , t) + h′′ (Xt )b(Xt , t) ,
2
which shows that
i
h
1
Ė(t) = E h′ (Xt )a(Xt , t) + h′′ (Xt )b(Xt , t) .
2
Using integration by parts twice, this implies that
Z
1 ′′
′
Ė(t) =
p(x, t) h (x)a(x, t) + h (x)b(x, t) dx
2
R
Z
∂
1 ∂2
h(x) −
(b(x, t)p(x, t)) dx,
(a(x, t)p(x, t)) +
=
2
∂x
2 ∂x
R
(16)
where all boundary terms vanish, since h has compact support.
Comparing equations (15) and (16), and using the fact that h
is arbitrary, we conclude that
∂
1 ∂2
∂p(x, t)
(b(x, t)p(x, t)),
= − (a(x, t)p(x, t)) +
∂t
∂x
2 ∂x2
which is the Fokker-Planck equation (6).
Observe that the Fokker-Planck equation (6) conserves
mass. If p(x, t) satisfies (6), then
Z
Z
d
∂p(x, t)
dx
p(x, t)dx =
dt R
∂t
R
Z
Z
1
∂
∂2
(b(x, t)p(x, t))dx
=−
(a(x, t)p(x, t))dx +
2 R ∂x2
R ∂x
=0+0=0
where each integral above is zero by integration by parts. Thus,
if we start with p(x, 0) which is a probability density, then the
solution p(x, t) stays a probability density at each time t ≥ 0.
B. Proof of Lemma 2
We first compute the gradient ∇Jb (p) with respect to its
argument p, which is a probability density function p : R → R
representing the distribution of a random variable. Recall that
∇Jb (p) satisfies
1
h∇Jb (p), vi = lim (Jb (p + δv) − Jb (p)),
(17)
δ→0 δ
R
for all v : R → R with R v(x)dx =
R 0 (so q + δv is still a
probability density), where hf, gi = R f (x)g(x)dx.
For ease of notation, we suppress dependence on x in what
follows, and use ′ to denote the derivative. We also assume that
p is smooth and rapidly decreasing, so all boundary terms are
zero in the application of integration by parts. We first have
the following lemma:
′2
′ ′
Lemma 3. ∇Jb (p) = b pp2 − 2 (bpp ) .
R
Proof. Let h : R → R with R p(x)h(x)dx = 0, and let δ > 0
be small. Let v = ph, so
log(p + δv) = log p + log(1 + δh) = log p + δh + o(δ).
Squaring the derivative yields
(log(p + δv))′2 = (log p)′2 + 2δh′ (log p)′ + o(δ).
We now multiply the above equation by b(p+δv) = bp+δbph
and integrate over x, to obtain
Z
Jb (p + δv) = Jb (p)+δ (2bph′ (log p)′ + bph(log p)′2 )+o(δ)
Z
= Jb (p) + δ h(−2(bp(log p)′ )′ + bp(log p)′2 ) + o(δ),
where we have used integration by parts in the second equality.
Since v = ph and h is arbitrary, equation (17) implies that
−2(bp(log p)′ )′ + bp(log p)′2
p
′ 2
′ ′
p
−2(bp )
+b
,
=
p
p
∇Jb (p) =
as desired.
B. Proof of Theorem 3
Lemma 2 follows from Lemma 3 by a direct computation.
The Bregman divergence of Jb from q to p is
We follow the outline of Gill and Levit [21, §2]. For every
y, we have
Z
x=+∞
∂
pXY (x, y) dx = pXY (x, y)
= 0.
∂x
x=−∞
DJb (p, q) = Jb (p) − Jb (q) − h∇Jb (q), p − qi
′2
Z
Z
Z
p′2
(bq ′ )′
q ′2
q
= b
− b
− (p − q) b 2 − 2
p
q
q
q
Z
Z
Z
Z
′2
′2
p ′ ′
q
p
− bp 2 + 2
(bq ) − 2 (bq ′ )′ .
= b
p
q
q
The last term above is zero (assuming q is also smooth and
rapidly decreasing) because it is the integral of Ra boundary
term. By integration by parts, the third term is 2 pq (bq ′ )′ =
R
′
′
−2 bq ′ ( pq − pq
q2 ). Therefore, we have
Z
Z
Z
′
p′2
q ′2
′q
DJb (p, q) = b
− 2 bp + bp 2
p
q
q
Z
= bp(log p − log q)′2 = Jb (p k q),
as desired.
A PPENDIX B
P ROOFS OF INFORMATION - ESTIMATION THEOREMS
In this Appendix, we prove the theorems appearing in
Section III-B, linking information to estimation.
Moreover, by integration by parts,
Z
Z
∂
pXY (x, y) dx = − pXY (x, y)dx = −pY (y).
x
∂x
Therefore,
Z Z
(T (y) − x)
∂
pXY (x, y) dx dy = 1,
∂x
which we may also write as
i
h
∂
log pXY (X, Y ) = 1,
E (T (Y ) − X)
∂x
where (X, Y ) ∼ pXY . By the Cauchy-Schwarz inequality, we
then have
"
#2
∂
(T (Y ) − X) p
p
b(X)
1=E
log pXY (X, Y )
∂x
b(X)
∂
2
(T (Y ) − X)2
E b(X)
log pXY (X, Y ) . (19)
≤E
b(X)
∂x
A. Proof of Theorem 2
By Bayes’ rule, we have the identity
∂
∂
∂
log pX|Y (x|y) =
log pY |X (y|x)−
log pY (y). (18)
∂y
∂y
∂y
Using the fact that
Observe that
Z
Z
∂
∂
∂
∂
∂
log pXY (x, y) =
log pX (x) +
log pY |X (y|x),
log pX|Y (x|y)dx =
pX|Y (x|y)dx = 0.
pX|Y (x|y)
∂x
∂x
∂x
∂y
∂y
Therefore, equation (18) implies
we also have
Z
∂
∂
2
∂
pX|Y (x|y)
log pY |X (y|x) dx =
log pY (y).
log pXY (X, Y )
E b(X)
∂y
∂y
∂x
2
∂
In particular, we have the following mmse-like relation:
Z
log pX (X)
= E b(X)
∂
2
∂x
Φ(X | Y = y) = pX|Y (x|y)
log pX|Y (x|y) dx
2
∂
∂y
Z
2
2
∂
∂
+ E b(X)
log pY |X (Y |X)
∂x
log pY |X (y|x) dx −
log pY (y) .
= pX|Y (x|y)
∂
∂
∂y
∂y
+
2E
b(X)
log
p
(X)
log
p
(Y
|
X)
X
Y |X
Multiplying by b(y) and integrating with respect to pY (y), we
∂x
∂x
obtain
= Jb (X) + Φb (Y | X)
(20)
Z
Φ(X | Y ) =
b(y) pY (y) Φ(X | Y = y) dy
Z Z
∂
2
=
b(y) pXY (x, y)
log pY |X (y|x) dx dy
∂y
Z
∂
2
− b(y) pY (y)
log pY (y) dy
∂y
Z
= pX (x) Jb (Y | X = x) dx − Jb (Y )
= Jb (X; Y ),
as desired.
where the last expectation above is zero because it is equal to
Z Z
∂
∂
b(x) pX (x) pY |X (y|x) dy dx
∂x Z
Z∂x
∂
∂
= b(x) pX (x)
pY |X (y|x) dy dx = 0,
∂x
∂x
R
R ∂
∂
∂
pY |X (y|x)dy = ∂x
1 = 0.
pY |X (y|x)dy = ∂x
since ∂x
Combining inequality (19) with equation (20) yields the desired result.
C. Proof of Theorem 4
Since the minimizer in the mmse definition is the conditional expectation, we may write
mmseb (ϕ(X, Y ) | Y )
= E[b(Y )(ϕ(X, Y ) − E[ϕ(X, Y )|Y ])2 ]
= E[b(Y )ϕ(X, Y )2 ] − E[b(Y )(E[ϕ(X, Y ) | Y ])2 ]. (21)
The first term in equation (21) is equal to
Z
Z
2
∂
log pY |X (y|x) dy dx
pX (x) b(y)pY |X (y|x)
∂y
Z
= pX (x)Jb (Y | X = x) dx = Jb (Y | X).
∂
log pY (Y ), the
On the other hand, since E[ϕ(X, Y ) | Y ] = ∂y
second term in equation (21) is equal to
Z
2
∂
log pY (y) dy = Jb (Y ).
b(y)pY (y)
∂y
Therefore, equation (21) implies that
mmseb (ϕ(X, Y ) | Y ) = Jb (Y | X) − Jb (Y ) = Jb (X; Y ),
as desired.
A PPENDIX C
G ENERALIZATIONS OF D E B RUIJN ’ S
IDENTITY
In this Appendix, we prove the theorems in Section III-C.
A. Proof of Theorem 5
We assume that p = p(x, t) is smooth and rapidly decreasing, so the boundary terms become zero when we apply
integration by parts, and we may differentiate under the
integral sign.
We first write
Z
Z
∂p
∂
d
H(Xt ) = −
log p dx − p log p dx.
dt
∂t
∂t
R
R
∂
The second integral above is equal to ∂p
p dx =
∂t dx = ∂t
∂
1
=
0.
We
then
substitute
the
Fokker-Planck
equation
(6)
∂t
to obtain
Z
Z 2
∂(ap)
∂ (bp)
1
d
log p dx. (22)
H(Xt ) =
log p dx −
dt
∂x
2
∂x2
By two applications of integration by parts, the first integral
in equation (22) is equal to
Z
Z
Z
∂ log p
∂p
∂(ap)
log p dx = − ap
dx = − a dx
∂x
∂x
∂x
Z
∂a
∂
=
p dx = E
a(Xt , t) . (23)
∂x
∂x
For the second integral in equation (22), we will use the
identity
∂ 2 log p
1 ∂ 2 p ∂ log p 2
.
=
−
2
∂x
p ∂x2
∂x
By integration by parts, we then have
Z 2
Z
∂ (bp)
∂ 2 log p
log
p
dx
=
bp
dx
∂x2
∂x2
Z
Z
∂ log p 2
∂ 2p
dx
= b 2 dx − bp
∂x
∂x
Z 2
Z
∂ b
∂ log p 2
=
dx
p dx − bp
2
∂x
∂x
2
∂
=E
(24)
b(Xt , t) − Jbt (Xt ).
∂x2
Combining equations (22), (23), and (24) yields the desired
conclusion.
B. Proof of Theorem 6
For ease of notation, we write p ≡ pt and q ≡ qt . By
differentiating under the integral sign and using the chain rule,
we have
Z
d
d
p
K(pt k qt ) =
p log dx
dt
dt
q
Z
Z
Z
p
∂ log p
∂ log q
∂p
log dx + p
dx − p
dx
=
∂t
q
∂t
∂t
Z
Z
∂p
p ∂q
p
=
log dx + 0 −
dx,
(25)
∂t
q
q ∂t
R ∂ log p
where
the
middle
integral
above
is
zero
because
p ∂t dx =
R ∂p
R
∂
∂
p dx = ∂t 1 = 0. We apply the Fokker-Planck
∂t dx = ∂t
equation (6) and use integration by parts to write the first
integral in equation (25) as
Z
Z
∂(ap) 1 ∂ 2 (bp)
p
p
∂p
−
log dx
log dx =
+
∂t
q
∂x
2 ∂x2
q
Z
2
∂
p
p 1 ∂
= ap
dx. (26)
log + bp 2 log
∂x
q
2 ∂x
q
Note that
∂
p
q ∂ p
log =
∂x
q
p ∂x q
and
p
q ∂2 p
∂2
log
=
−
∂x2
q
p ∂x2 q
p
∂
log
∂x
q
2
.
Plugging these relations into equation (26) and using integration by partsRand the Fokker-Planck equation (6) for q(x, t),
p
we find that ∂p
∂t log q dx is equal to
2 !
Z
∂ p 1 ∂2 p 1
p
∂
aq
dx
+ bq 2 − bp
log
∂x q
2 ∂x q
2
∂x
q
Z
1
∂(aq) 1 ∂ 2 (bq) p
+
dx − Jbt (pt k qt )
−
=
∂x
2 ∂x2
q
2
Z
∂q p
1
=
dx − Jbt (pt k qt ).
∂t q
2
Substituting this into equation (25), we conclude that
1
d
K(pt k qt ) = − Jbt (pt k qt ),
dt
2
as desired.
C. Proof of Theorem 7
Theorem 5 holds for any initial distribution. In particular,
if X0 is a point mass at x0 ∈ R, we have
1
d
H(Xt | X0 = x0 ) = Jbt (Xt | X0 = x0 )
dt
2
1 ∂2
∂
b(X
,
t)
X
=
x
+E
a(Xt , t) −
t
0
0 . (27)
∂x
2 ∂x2
For any initial distribution X0 , we may average equation (27)
over X0 to obtain
1
d
H(Xt | X0 ) = Jbt (Xt | X0 )
dt
2
∂
1 ∂2
+E
b(X
,
t)
.
a(Xt , t) −
t
∂x
2 ∂x2
Combining this with Theorem 5, we conclude that
d
d
I(X0 ; Xt ) = (H(Xt ) − H(Xt | X0 ))
dt
dt
1
= (Jbt (Xt ) − Jbt (Xt | X0 ))
2
1
= − Jbt (X0 ; Xt ),
2
as desired.
A PPENDIX D
D ETAILS FOR S ECTION III-D
For the Ornstein-Uhlenbeck process, the explicit solution to
the Fokker-Planck equation (8) is given by
r
α(x − e−αt x0 )2
α
p(x, t) =
.
exp −
π(1 − e−2αt )
1 − e−2αt
For the geometric Brownian motion process, the explicit
solution to the Fokker-Planck equation (8) is given by
!
2
(log x − log x0 − (µ − σ2 )t)2
1
1
p(x, t) = √
.
exp −
2σ 2 t
2πσ 2 t x
The mmse expressions in for the examples in Section III-D
are then obtained through direct computations.
A PPENDIX E
M ULTIVARIATE SETTING
In the multivariate setting, our channel definition still takes
the form of equation (5), where a(x, t) ∈ Rd is a drift vector,
σ(x, t) ∈ Rd×d is a covariance matrix, and the weight matrix
b(x, t) = σ(x, t)σ(x, t)⊤ is assumed to be uniformly positive
definite. The Fokker-Planck equation (6) now takes the form
∂p(x, t)
1
= −∇ · (a(x, t)p(x, t)) + ∇ · (∇ · (b(x, t)p(x, t))),
∂t
2
where ∇ is the gradient and ∇· is the divergence operator.
More explicitly, we have
d
X
∂p(x, t)
∂
(ai (x, t)p(x, t))
=−
∂t
∂x
i
i=1
d
+
d
1 X X ∂2
(bij (x, t)p(x, t)).
2 i=1 j=1 ∂xi ∂xj
The derivation for the Fokker-Planck equation in the multivariate setting is analogous to the derivation in the univariate
setting (cf. Mackey [15]).
As explained in Section III-E, the generalized Fisher information is defined using the Mahalanobis inner product
with respect to the weight matrix b, and the statistical Fisher
information is defined by averaging the pointwise Fisher
information with respect to the weight matrix b. The weighted
van Trees’ inequality in Theorem 3 becomes
h
i
1
1
E kT (Y ) − Xk2b(X)−1 ≥
=
,
Φb (Y |X) + Jb (X)
Jb (X|Y )
where kvk2b(x)−1 = v ⊤ b(x)−1 v = Tr(b(x)−1 vv ⊤ ) is the
Mahalanobis inner product with respect to the inverse weight
matrix b−1 . The proof proceeds in the same way as in the
univariate case (see also Gill and Levit [21, §4]).
The expression for the time derivative of entropy in Theorem 5 now becomes
d
1
1
H(Xt ) = Jbt (Xt )+E ∇· a(Xt , t) − ∇· (∇· b(Xt , t)) ,
dt
2
2
or more explicitly,
d
X
d
∂ai (Xt , t)
1
E
H(Xt ) = Jbt (Xt ) +
dt
2
∂xi
i=1
2
d
d
∂ bij (Xt , t)
1 XX
E
.
−
2 i=1 j=1
∂xi ∂xj
The statements of the other theorems remain unchanged and
the proofs remain valid.
| 10math.ST
|
arXiv:1801.02644v2 [math.AC] 31 Jan 2018
ON MONOMIAL IDEALS AND THEIR SOCLES
GEIR AGNARSSON AND NEIL EPSTEIN
Abstract. For a finite subset M ⊂ [x1 , . . . , xd ] of monomials, we describe how to constructively
obtain a monomial ideal I ⊆ R = K[x1 , . . . , xd ] such that the set of monomials in Soc(I) \ I is
precisely M , or such that M ⊆ R/I is a K-basis for the the socle of R/I. For a given M we obtain a
natural class of monomials I with this property. This is done by using solely the lattice structure of
the monoid [x1 , . . . , xd ]. We then present some duality results by using anti-isomorphisms between
upsets and downsets of (Zd , ). Finally, we define and analyze zero-dimensional monomial ideals
of R of type k, where type 1 are exactly the Artinian Gorenstein ideals, and describe the structure
of such ideals that correspond to order-generic antichains in Zd .
1. Introduction
Aside from the algebraic benefits of the socle of a module over a commutative ring, when studying
local rings, Cohen-Macaulay rings or Gorenstein rings [BH97, Eis95, Vil01] the socle of an ideal, in
particular that of a monomial ideal of the polynomial ring over a field, has a rich lattice structure
that many times can be studied solely by using combinatorial arguments. Consider the following:
any range of mountains on a given piece of land where each mountain has the same cone-type shape
is, of course, determined by the loci of the mountain tops. But is it determined by the mere valleys
these mountains form? The answer depends on what additional conditions we know the mountain
range has. This is one of the motivating questions in this article when we view a monomial ideal
as a mountain range and the valleys it forms as the elements in the socle of the ideal that are not
in the ideal.
For a field K, monomial ideals of the polynomial ring R = K[x1 , . . . , xd ] play a pivotal role in
the investigation of Gröbner bases of general ideals of R [BW93, AL94]. The underlying reason as
to why monomial ideals are so convenient, yet so important, in investigating reduction systems for
general ideals of R stems in part from the fact that for every monomial ideal I = (m1 , . . . , mk ),
Date: February 1, 2018.
2010 Mathematics Subject Classification. 13B25, 13A02, 06A07.
Key words and phrases. poset, upset, downset, polynomial ring, monomial ideal, socle, duality, Artinian ideal,
Gorenstein ideal, type k monomial ideal.
1
where no mi divides another mj , has a Gröbner basis M = {m1 , . . . , mk }; exactly the minimum set
of generators for I. For A ⊆ {1, . . . , k} let mA denote the least common multiple of the mi : i ∈ A.
With this convention one can define the Scarf complex ∆(I) of the monomial ideal I as the simplicial
complex consisting of all the subsets A ⊆ {1, . . . , k} with unique least common multiple mA , that
is
∆(I) = {A ⊆ {1, . . . , k} : mA 6= mB for all B 6= A}.
The Scarf complex discussed in [BPS98] and [MS05] was first introduced in [Sca86]. It is easy
to see that the facets Fd−1 (∆(I)) of the Scarf complex ∆(I) are in bijective correspondence with
the maximal monomials of R \ I (w.r.t. the divisibility partial order), which is exactly the set
of monomials of the socle Soc(I) that are not in I (see definition in the following section.) The
cardinality of this very set of monomials has many interesting combinatorial interpretations, two
of which we will briefly describe here below.
We say that an ideal I of R is co-generated by a set F of K-linear functionals R → K if I
is the largest ideal of R contained in all the kernels of the functionals in F. A celebrated result
by Macaulay from 1916 [Mac94] states that every ideal of R is finitely co-generated, so R is coNoetherian in this sense. More specifically, it turns out that any monomial ideal I of R has at
least |Fd−1 (∆(I))| co-generators and can always be co-generated by |Fd−1 (∆(I))| + 1 functionals
(see [Agn00]). Further, for a given ideal I (not necessarily monomial) of R and a fixed term
order, then I has a Gröbner basis where the head or leading terms of the basis elements form
a corresponding monomial ideal L(I) of R, and I can then be co-generated by one functional if
|Fn−1 (∆(L(I)))| < d and by |Fn−1 (∆(L(I)))| + 1 functionals otherwise [Agn00].
The cardinality |Fd−1 (∆(I))| is also linked to the number of edges in a simple graph on k
vertices in the following way. For given d, k ∈ N let cd (k) denote the maximum number of facets
|Fd−1 (∆(I))| among all monomial ideals I of R that are minimally generated by k monomials. In
general, the Scarf complex ∆(I) is always a sub-complex of the boundary complex of a simplicial
polytope P (I) on k vertices where one facet is missing. When I is Artinian (or zero-dimensional)
and generic (i.e. the powers of all xi in the generators for I are distinct) then ∆(I) actually equals
this boundary complex P (I) with one facet missing [MS05]. Using the Dehn-Sommerville equations
for the simplicial polytope P (I) one can show that determining cd (k) for d ≤ 5 is equivalent to
determining the maximum number of edges of the 1-skeleton of P (I). This turns out to be the same
problem of determining the maximum number of a simple graph on k vertices of order dimension
at most d [Agn97].
2
We therefore see that the number of monomials in Soc(I)\I, the socle of the monomial ideal I that
are not in I, has some interesting combinatorial interpretations in addition to the known algebraic
ones, in particular, the maximum number that the set Soc(I) \ I can have. Trivially we have
c2 (k) = k − 1 for any k ∈ N and less trivially we know that c3 (k) = 2k − 5 for k ≥ 3 [Agn97, MS05].
For each d ≥ 4 the exact value of cd (k) is still unknown, although the asymptotic behavior does
satisfy cd (k) = Θ(k⌊d/2⌋ ) as k tends to infinity and d is fixed [Agn97]. In any case, we see from
this that a monomial ideal minimally generated by k monomials, can “generate” monomials in its
socle of cardinality considerably larger order than that of k. In order for that to occur though, the
k monomials that generate the given ideal must be special and relate to each other in a singular
way. By the same token, the resulting monomials in Soc(I) \ I also relate in a special way if they
are generated by “few” generating monomials of I. Looking away from these extreme cases for a
moment, some natural questions about monomial ideals and their socles, in particular about those
monomials in the socle and not in the ideal, arise.
(i) Assume that for a given (apriori unknown) monomial ideal I the set Soc(I) \ I is given, is it
always possible to retrieve the ideal I from it, if not, what additional structure is needed?
In which cases is it unique?
(ii) Given any set of monomials that form an antichain w.r.t. divisibility order, can one always
find a monomial ideal I such that this set of monomials has the form Soc(I) \ I? Is the
ideal I unique? If not, can they be characterized in some way?
The purpose of this article is to address these questions, discuss uniqueness and non-uniqueness,
present up-down duality results, discuss generalizations to general ideals of R and address what is
similar with the monomial ideal case and what is not.
This article was in part inspired by the M.S. thesis of Anna-Rose Wolff [Wol16]. She analyzed the
survival complex Σ(R/I) of R/I where I is a monomial ideal that contains powers of all the variables
of R. The survival complex is a simplicial complex whose vertices are the monomials of R that are
not in I, where a simplex consists of a set of monomials whose product is not in I. She showed
[Wol16, Prop. 2.2.1] that the truly isolated points of that complex correspond to the monomial
basis of the socle of I, a set which in this document we call ∂oc(I) (see Definition 2.1.) As such,
it was natural for her to address the question we address here, completely solving the two-variable
case [Wol16, Algorithm 3.3.1 and Proposition 3.3.3] and making headway on the three-variable case
[Wol16, Algorithm 3.4.1].
The rest of this article is organized as follows:
3
In Section 2 we define some basic concepts about partially ordered sets that we will be using and
referring to throughout the article.
In Section 3 we show that if for a monomial ideal I the set Soc(I) \ I is given, then we can
retrieve a monomial ideal I back from a derived version of the set Soc(I) \ I.
In Section 4 we present and use an intuitively obvious duality between monomial ideals of R on
the one hand and “down-sets” of monomials on the other hand, that contain all elements dividing
a given element of the down-set (see definitions in Section 4).
In Section 5 we elaborate more explicitly the discussion presented in Section 3 and show that
if Soc(I) \ I is given, then there is a unique zero-dimensional ideal I yielding the given set of
monomials. As a corollary we obtain a well-known description of Artinian monomial ideals of R
that are Gorenstein. We then use what we have established to describe Artinian monomial ideals
that are type 2 Gorenstein (where type 1 Gorenstein is the usual Gorenstein notion (see definitions
in Section 5)).
In Section 6 we discuss further Artinian type k Gorenstein monomial ideals and describe their
order-generic case.
In Section 7 we discuss the socle of ideals of R = K[x1 , . . . , xd ] in more general settings when K
is an algebraically closed field.
Finally, in Section 8 we summarize our main results and pose some questions.
2. Partially ordered sets, basic definitions and notations
The set of integers will be denoted by Z, the set of natural numbers by N and the set of nonnegative numbers N ∪ {0} by N0 . For each n ∈ N we let [n] := {1, . . . , n}.
Throughout, R = K[x1 , . . . , xd ] will denote the polynomial ring over a field K. By the socle of
an ideal I ⊆ R w.r.t. the maximal ideal m = (x1 , . . . , xd ) of R, we will mean the ideal
Soc(I) = Socm (I) := (I : m) = {f ∈ R : xi f ∈ I for each i}.
Note that for a monomial ideal I ⊆ R, the set Soc(I) \ I contains monomials a such that (i) a 6∈ I
and (ii) xi a ∈ I for every i ∈ [d].
Definition 2.1. For a monomial ideal I and the maximal ideal m = (x1 , . . . , xd ) of R, let ∂oc(I)
denote the set of the monimials in Soc(I) \ I.
4
Remarks.
(i) By the above Definition 2.1, we have Soc(I) = I ⊕ SpanK (∂oc(I)) as a K-vector spaces.
(ii) The image Soc(I) = (0 : m/I) = Soc(R/I) in R/I has a basis as a vector space over K
consisting of the images of the monomials in ∂oc(I) (see [Vil01, AF74].) Hence, ∂oc(I)
consists exactly of the maximal monomials in R \ I w.r.t. the divisibility partial order and
|∂oc(I)| = dimK (Soc(R/I)).
Assume for a moment we have a general ideal I and a general maximal ideal m of R. If m does
not contain I, then m + I = R and hence for f ∈ (I : m) we have f ∈ f I + f m ⊆ I. Therefore
(I : m) = I and so Socm (I) is trivial. Therefore, for an ideal I and a maximal ideal m of R,
the socle Socm (I) is only of interest when I ⊆ m. Hence, unless otherwise stated, we will always
assume I ⊆ m.
Most of what we establish from now on about monomial ideals of R = K[x1 , . . . , xd ] will only
use the the monoid [x1 , . . . , xd ] as a partially ordered set (poset) with the partial order given by
divisibility. We present some basic definitions and notations for general posets that we will use
throughout the article.
For a poset (P, ≤) recall that C ⊆ P is a chain if C forms a linearly or totally ordered set within
(P, ≤). A subset N ⊆ P is an antichain or a clutter if no two elements in N are comparable in
(P, ≤).
For a poset (P, ≤) call a subset U ⊆ P an upset, or an up-filter, of P if x ≥ u ∈ U ⇒ x ∈ U .
Call a subset D ⊆ P a downset, or a down-filter, of P if if x ≤ d ∈ D ⇒ x ∈ D.
For an upset U of a poset P let Sd (U ) := max(P \ U ) be the maximal elements of P \ U . For a
downset D of a poset P let Su (D) := min(P \ D) be the minimal elements of P \ D.
For a subset A ⊆ P , let U (A) := {x ∈ P : x ≥ a for some a ∈ A} be the upset generated by A,
and D(A) := {x ∈ P : x ≤ a for some a ∈ A} the downset generated by A. If A = {a1 , . . . , an } is
finite, then we write U (a1 , . . . , an ) (resp. D(a1 , . . . , an )) for U (A) (resp. D(A).)
Recall that Zd has a natural partial order where for ã = (a1 , . . . , ad ) and b̃ = (b1 , . . . , bd )
we have ã b̃ ⇔ ai ≤ bi for each i ∈ [d]. This is a generalization of the partial order given
in [GP06, Ex. 19, p. 71], and we will henceforth be using “≤” for this partial order “”. The usual
componentwise addition +, which makes (Zd , +) an abelian group, respects the mention partial
order ≤.
5
For any fixed point c̃ ∈ Zd , the map τc̃ : Zd → Zd given by τc̃ (x̃) = c̃ + x̃ is an order preserving
translation, and hence an order automorphism, of the poset (Zd , ≤). The following is easy to obtain.
Observation 2.2. For any A ⊆ Zd and c̃ ∈ Zd we have
τc̃ (U (A)) = U (τc̃ (A)), τc̃ (D(A)) = D(τc̃ (A)).
Further, for any upset U ⊆ Zd and downset D ⊆ Zd we have
τc̃ (Sd (U )) = Sd (τc̃ (U )), τc̃ (Su (D)) = Su (τc̃ (D)).
Writing out what the above observation states in terms of c̃ we get
U (A + c̃) = U (A) + c̃,
D(A + c̃) = D(A) + c̃,
Sd (U + c̃) = Sd (U ) + c̃,
Su (D + c̃) = Su (D) + c̃.
An upset U ⊆ Zd is cofinite if here is a c̃ ∈ Zd such that (i) U ⊆ U (c̃) and (ii) U (c̃) \ U is finite.
Likewise a downset D ⊆ Zd is cofinite if there is a c̃ ∈ Zd such that (i) D ⊆ D(c̃) and (ii) D(c̃) \ D
is finite.
The map given by p̃ = (p1 , . . . , pd ) 7→ xp11 · · · xpdd = x̃p̃ is an isomorphism between the additive
monoid Nd0 and the multiplicative monoid [x1 , . . . , xd ]. Moreover, the map is also an order isomor-
phism from the poset (Nd0 , ≤), where ≤ is the partial order of Zd from above, to [x1 , . . . , xd ] as a
poset given by divisibility. Hence as the map is an order preserving monoid isomorphism, we will
for the most part in the next two sections deal with the additive monoid Nd0 and the abelian group
Zd instead of the multiplicative monoid [x1 , . . . , xd ] within the ring R. Via this isomorphism we
have a bijective correspondence between monomial ideals of R and upsets of the poset Nd0 .
3. Retrieving an upset from maximal elements not in the upset
We will in this section and the following section describe everything in terms of Nd0 or Zd instead
of the ring R = K[x1 , . . . , xd ] as we justified in the last paragraph in the previous section.
Let G = {p̃1 , . . . , p̃k } ⊆ Nd be an antichain. For the moment we assume that G is positive, i.e. all
the coordinates of each p̃i are strictly positive. (This is by no means a restrictive assumption as
will become clear as we go along.) If p̃i = (pi1 , . . . , pid ) for each i, then for each j ∈ [d] let
(1)
mj = min{pij : 1 ≤ i ≤ d} − 1, Mj = max{pij : 1 ≤ i ≤ d} + 1
6
and
B ∗ (G) =
{(M1 , m2 , m3 . . . , md ), (m1 , M2 , m3 , . . . , md ), (m1 , m2 , M3 , . . . , md ),
(2)
. . . , (m1 , m2 , m3 , . . . , Md )}.
Let G∗ = G∪B ∗ (G). That r̃ ∈ Sd (U (G)), means precisely that r̃ 6∈ U (G) and r̃+ ẽj ∈ U (G) for each
j, where ẽj is the usual basis vector for Rd with 1 in the j-coordinate and 0 everywhere else. Since
r̃ 6∈ U (G) and r̃ + ẽj ∈ U (G) for each given j, then for some p̃i ∈ G we have rj = pij −1 < Mj . Since
this holds for each j, then r̃ cannot be greater than any element of B ∗ (G) and so r̃ 6∈ U (B ∗ (G))
and hence r̃ 6∈ U (G∗ ). Since G ⊆ G∗ and so U (G) ⊆ U (G∗ ), we have r̃ + ẽj ∈ U (G∗ ) for each j,
and so r̃ ∈ Sd (U (G∗ )).
Also we have U (G) ⊆ U (G∗ ) ⊆ U (m̃), where m̃ = (m1 , . . . , md ) from (1), and U (m̃) \ U (G∗ ) ⊆
U (m̃) \ U (B ∗ (G)), which is finite. We summarize in the following.
Proposition 3.1. For any antichain G = {p̃1 , . . . , p̃k } ⊆ Nd we have
Sd (U (G)) ⊆ Sd (U (G∗ )) ⊆ U (m̃) \ U (G∗ ),
which is a finite set, and hence U (G∗ ) is cofinite.
Note that the Sd (U (G∗ )) corresponds to the socle of the monomial ideal whose generators correspond to G∗ . The main result in this section is the following.
Theorem 3.2. For a given antichain G = {p̃1 , . . . , p̃k } ⊆ Nd we have
G = Su (D(Sd (U (G∗ )))).
Before proving Theorem 3.2 we need the following.
Lemma 3.3. Let (P, ≤) be a poset. For an upset U ⊆ P and a downset D ⊆ P we have
D(Sd (U )) ∩ U = U (Su (D)) ∩ D = ∅.
Proof. If x̃ ∈ D(Su (U )), then x̃ ≤ ỹ ∈ Sd (U ). If now x̃ ∈ U as well, then by definition of an upset
we must have ỹ ∈ U , which is a contradiction since U ∩ Sd (U ) = ∅. In the same way we obtain
U (Su (D)) ∩ D = ∅.
⊔
⊓
7
Conventions.
(i) For a set A ⊆ Zd and i ∈ [d] let Axi =ai := {x̃ ∈ A : xi = ai }.
(ii) For a finite set G ⊆ Zd let U (G) = U (m̃) \ U (G∗ ), where m̃ is as in (1), which then is a
finite set.
(iii) For I = {i1 , . . . , ih } ⊆ [d] let πI = πi1 ,...,ih : Rd → Rh denote the natural projection onto
coordinates i1 , . . . , ih . In particular πi (x̃) = xi and πı̂ (x̃) = (x1 , . . . , xi−1 , xi+1 , . . . , xd ).
Proof of Theorem 3.2. Let p̃ = (p1 , . . . , pd ) ∈ G. Since U (G∗ ) is cofinite, then for each i ∈ [d]
the set U (G)xi =pi −1 is also a finite poset. Also, it is nonempty since p̃ − ẽi = (p1 , . . . , pi−1 , pi −
1, pi+1 , . . . , pd ) ∈ U (G)xi =pi −1 as G∗ is an antichain. Hence, there is a maximal element q̃ ∈
U (G)xi =pi −1 with p̃ − ẽi ≤ q̃. Since q̃ + ẽi ≥ p̃, then q̃ + ẽi ∈ U (p̃) ⊆ U (G) ⊆ U (G∗ ). Also,
since q̃ is maximal in U (G)xi =pi −1 , we have q̃ + ẽℓ ∈ U (G∗ )xi =pi −1 ⊆ U (G∗ ) for all ℓ 6= i, and so
q̃ ∈ Sd (U (G∗ )). So, for each i ∈ [d] we have p̃ − ẽi ∈ D(Sd (U (G∗ ))). Since p̃ ∈ G ⊆ U (G∗ ), then
by Lemma 3.3 we have p̃ 6∈ D(Sd (U (G∗ ))), and therefore p̃ ∈ Su (D(Sd (U (G∗ )))). This proves that
G ⊆ Su (D(Sd (U (G∗ )))).
For the other direction, we first verify that
Su (D(Sd (U (G∗ )))) ⊆ U (m̃ + 1̃),
(3)
where 1̃ = (1, . . . , 1) ∈ Zd . First note that by Proposition 3.1 we have Sd (U (G∗ )) ⊆ U (m̃)\U (G∗ ) ⊆
U (m̃).
If r̃ ∈ Su (D(Sd (U (G∗ )))), then r̃ − ẽi ∈ D(Sd (U (G∗ ))) for each i, and so r̃ − ẽi ≤ q̃ for some
q̃ ∈ Sd (U (G∗ )) ⊆ U (m̃) for each i. Consider a fixed i. Since r̃ 6≤ q̃ and r̃ − ẽi ≤ q̃, we must have
ri − 1 = qi . Since q̃ ∈ Sd (U (G∗ )) ⊆ U (m̃), we have qi ≥ mi and therefore ri = qi + 1 ≥ mi + 1.
Since this holds for each i, we have thus (3).
For r̃ ∈ Su (D(Sd (U (G∗ )))) we have for each i that r̃ − ẽi ∈ D(Sd (U (G∗ ))), and hence, by
Lemma 3.3, r̃ − ẽi 6∈ U (G∗ ) for each i. If now r̃ ∈ U (G∗ ), then r̃ is a minimal element of U (G∗ )
and hence r̃ ∈ G∗ . Hence
r̃ ∈ G∗ ∩ Su (D(Sd (U (G∗ )))) ⊆ G∗ ∩ U (m̃ + 1̃) = G.
Therefore it suffices to show that Su (D(Sd (U (G∗ )))) ⊆ U (G∗ ).
Assume r̃ ∈ Su (D(Sd (U (G∗ ))))\U (G∗ ). We then have r̃ ∈ U (m̃+ 1̃)\U (G∗ ) ⊆ U (m̃)\U (G∗ ). By
Proposition 3.1 U (G∗ ) is cofinite, and so there is a maximal element q̃ ∈ U (m̃) \ U (G∗ ) with r̃ ≤ q̃.
In this case we have, by the definition of Sd , that r̃ ≤ q̃ ∈ Sd (U (G∗ )) and hence r̃ ∈ D(Sd (U (G∗ )))
8
contradicting that r̃ ∈ Su (D(Sd (U (G∗ )))). This proves that Su (D(Sd (U (G∗ )))) ⊆ U (G∗ ), and
hence, by previous paragraph, Su (D(Sd (U (G∗ )))) ⊆ G.
⊔
⊓
Remark. The values m1 , . . . , md and M1 , . . . , Md from (1), used to define G∗ , do not play a major
role, except for merely being small and respectively large enough. In fact, if ã ≤ m̃ and b̃ ≥ M̃ and
(4)
B ∗ (ã, b̃) = {(b1 , a2 , a3 . . . , ad ), (a1 , b2 , a3 , . . . , ad ), (a1 , a2 , b3 , . . . , ad ), . . . , (a1 , a2 , a3 , . . . , bd )},
then we can define G∗ (ã, b̃) = G ∪ B ∗ (ã, b̃), and we then, as in Proposition 3.1, obtain that
Sd (U (G)) ⊆ Sd (U (G∗ (ã, b̃))) ⊆ U (ã) \ U (G∗ (ã, b̃)),
which is a finite set, and hence U (G∗ (ã, b̃)) is cofinite, and the proof of Theorem 3.2 works verbatim
for the following slight generalization.
Theorem 3.4. For a given antichain G = {p̃1 , . . . , p̃k } ⊆ Nd and for ã ≤ m̃ and b̃ ≥ M̃ where m̃
and M̃ are as in (1), we have
G = Su (D(Sd (U (G∗ (ã, b̃))))).
This will be used in the next Section 4.
Note that Theorem 3.2 states that one can retrieve the antichain, or the generating points, G
from Sd (U (G∗ )) alone. However, this does not mean that Sd (U (G∗ )) can be arbitrary; the set is
derived from the (apriori unknown) set G.
In the next section we use poset duality of Zd to obtain some related results from Theorem 3.2
where we start with an arbitrary antichain Q and show how it corresponds to the socle of a certain
monomial ideal.
4. Up-down duality
As in the derivation of Observation 2.2, for any fixed point c̃ ∈ Zd , the map ρc̃ : Zd → Zd given
by ρc̃ (x̃) = c̃ − x̃ is a reverse-order preserving rotation, and hence an anti-automorphism of the
poset (Zd , ≤). In particular, for N ⊆ Zd we have that N is an antichain iff ρc̃ (N ) is an antichain.
Clearly ρc̃ is its own inverse, and so we have the following.
Observation 4.1. For any c̃ ∈ Zd we have ρ2c̃ = IZd , the identify map of Zd . For any A ⊆ Zd and
c̃ ∈ Zd we have
ρc̃ (U (A)) = D(ρc̃ (A)), ρc̃ (D(A)) = U (ρc̃ (A)).
9
Further, for any upset U ⊆ Zd and downset D ⊆ Zd then ρc̃ (U ) is a downset, ρc̃ (D) is an upset and
ρc̃ (Sd (U )) = Su (ρc̃ (U )), ρc̃ (Su (D)) = Sd (ρc̃ (D)).
Let G = {p̃1 , . . . , p̃k } ⊆ Nd be an antichain, and G∗ = G ∪ B ∗ (G), where B ∗ (G) is as in (2). If
ρ = ρm̃+M̃ where m̃ and M̃ are as in (1), then for any coordinate i we have
min(πi (ρ(G))) = mi + Mi − (Mi − 1) = mi + 1, max(πi (ρ(G))) = mi + Mi − (mi + 1) = Mi − 1,
and so G, ρ(G) ⊆ U (m̃ + 1̃) ∩ D(M̃ − 1̃).
Suppose that Q ⊆ Nd is such that min(πi (Q)) = mi + 1 and max(πi (Q)) = Mi − 1 for each i.
Since ρ is its own inverse, there is a unique G ⊆ U (m̃ + 1̃) ∩ D(M̃ − 1̃) with Q = ρ(G) (namely
G = ρ(Q),) such that min(πi (G)) = mi + 1 and max(πi (G)) = Mi − 1 for each i, as well. Hence we
have
ρ(G∗ ) = ρ(G ∪ B ∗ (G))
= ρ(G) ∪ ρ(B ∗ (G))
= Q ∪ {(m1 , M2 , M3 . . . , Md ), (M1 , m2 , M3 , . . . , Md ), (M1 , M2 , m3 , . . . , Md ),
. . . , (M1 , M2 , M3 , . . . , md )}.
Hence, if we define
B∗ (Q) =
{(m1 , M2 , M3 . . . , Md ), (M1 , m2 , M3 , . . . , Md ), (M1 , M2 , m3 , . . . , Md ),
(5)
. . . , (M1 , M2 , M3 , . . . , md )},
where for each i min(πi (Q)) = mi + 1 and max(πi (Q)) = Mi − 1, and let Q∗ = Q ∪ B∗ (Q), then
ρ(G∗ ) = Q∗ and so by Theorem 3.2 and Observation 4.1 we get
(6)
Q = ρ(G) = ρ(Su (D(Sd (U (G∗ ))))) = Sd (U (Su (D(ρ(G∗ ))))) = Sd (U (Su (D(Q∗ )))).
Since Q ⊆ U (m̃ + 1̃) ∩ D(M̃ − 1̃) ⊆ Nd is a finite set, we have by (6) the following dual theorem of
Theorem 3.2.
Theorem 4.2. For an antichain Q = {p̃1 , . . . , p̃k } ⊆ Nd we have
Q = Sd (U (Su (D(Q∗ )))),
where Q∗ = Q ∪ B∗ (Q) and B∗ (Q) is as in (5).
10
By the above Theorem 4.2 we see that given any antichain Q ⊆ Nd , then the corresponding
monomials are exactly the monomials in ∂oc(I(Q)), where I(Q) is the monomial ideal generated
by the monomials that correspond to Su (D(Q∗ )) ⊆ Nd0 . This does give a positive answer to one of
our motivating questions in Section 1. Note, however, that this monomial ideal is not unique, as
stated in the dual theorem of Theorem 4.3 here below.
Let Q = {p̃1 , . . . , p̃k } ⊆ Nd be an antichain and ã ≤ m̃ and b̃ ≥ M̃ where m̃ and M̃ are as in (5).
Further, let ρ = ρã+b̃ as defined in the beginning of this section. If now G = ρ(Q), then we have,
as above, that ρ(G) = ρ2 (Q) = Q and for G∗ (ã, b̃) as in the Remark right before Theorem 3.4, that
ρ(G∗ (ã, b̃)) = ρ(G ∪ B ∗ (ã, b̃))
= ρ(G) ∪ ρ(B ∗ (ã, b̃))
= Q ∪ {(a1 , b2 , b3 . . . , bd ), (b1 , a2 , b3 , . . . , bd ), (b1 , b2 , a3 , . . . , bd ), . . . , (b1 , b2 , b3 , . . . , ad )}.
So, again, we can define
B∗ (ã, b̃) = {(a1 , b2 , b3 . . . , bd ), (b1 , a2 , b3 , . . . , bd ), (b1 , b2 , a3 , . . . , bd ), . . . , (b1 , b2 , b3 , . . . , ad )}.
(7)
If now
Q∗ (ã, b̃) := Q ∪ B∗ (ã, b̃),
then ρ(G∗ (ã, b̃)) = Q∗ (ã, b̃), and so by Theorem 3.4 and Observation 4.1 we get
Q = ρ(G) = ρ(Su (D(Sd (U (G∗ (ã, b̃)))))) = Sd (U (Su (D(ρ(G∗ (ã, b̃)))))) = Sd (U (Su (D(Q∗ (ã, b̃))))).
Since this holds for any ã ≤ m̃ and b̃ ≥ M̃ , we have a following dual theorem of Theorem 3.4.
Theorem 4.3. For an antichain Q = {p̃1 , . . . , p̃k } ⊆ Nd and for ã ≤ m̃ and b̃ ≥ M̃ where m̃ and
M̃ are as in (5), we have
Q = Sd (U (Su (D(Q∗ (ã, b̃))))).
where Q∗ (ã, b̃) = Q ∪ B∗ (ã, b̃) and B∗ (ã, b̃) is as in (7).
Remarks.
(i) Note that Theorem 4.3 is valid for any ã, b̃ ∈ Zd that satisfy ã ≤ m̃ and b̃ ≥ M̃ . This will
be used in the next section.
(ii) Note that if Q∗ (ã, b̃) is given (without the prior knowledge of B∗ (ã, b̃)), then we can retrieve
B∗ (ã, b̃) as in in (7) from Q∗ (ã, b̃), and hence the set Q, if ã ≤ m̃ and b̃ ≥ M̃ .
11
(iii) Last but not least, note that by Observations 2.2 and 4.1, all Theorems 3.2, 3.4, 4.2 and 4.3
hold for any antichain Q ⊆ Zd , and not merely those of Nd .
Example. Consider the case d = 3 and the set Q = {(2, 2, 3), (3, 3, 2)} ⊆ N3 . We use the above Theorem 4.2 to obtain an upset U (Sd (D(Q∗ ))), whose minimal points Su (D(Q∗ )) ⊆ N30 correspond to
the generators of a monomial ideal I where Soc(I) is spanned by the image of M = {x2 y 2 z 3 , x3 y 3 z 2 }
in R/I, i.e. the monomials corresponding to the set Q, as follows.
By (5) we have here that B∗ (Q) = {(4, 4, 1), (4, 1, 4), (1, 4, 4)} and hence
Q∗ = Q ∪ B∗ (Q) = {(2, 2, 3), (3, 3, 2), (4, 4, 1), (4, 1, 4), (1, 4, 4)},
and so Su (D(Q∗ )) = {(2, 2, 4), (2, 3, 3), (2, 4, 2), (3, 2, 3), (4, 2, 2)}. By Theorem 4.2 we now have
Sd (U (Su (D(Q∗ ))))
= Sd (U (Su (D((2, 2, 3), (3, 3, 2), (4, 4, 1), (4, 1, 4), (1, 4, 4)))))
= Sd (U ((2, 2, 4), (2, 3, 3), (2, 4, 2), (3, 2, 3), (4, 2, 2)))
= {(2, 2, 3), (3, 3, 2)}
= Q.
Therefore, the monomial ideal I1 = (x2 y 2 z 4 , x2 y 3 z 3 , x2 y 4 z 2 , x3 y 2 z 3 , x4 y 2 z 2 ) ⊆ K[x, y, z] has Soc(I1 )
spanned by the image of M = {x2 y 2 z 3 , x3 y 3 z 2 } in R/I1 as a k-vector space.
Since ã = (0, 0, 1) ≤ (1, 1, 1) = m̃ and b̃ = (5, 6, 7) ≥ (4, 4, 4) = M̃ in Z3 , we have for
Q∗ (ã, b̃) = Q∗ ((0, 0, 1), (5, 6, 7)) = {(2, 2, 3), (3, 3, 2), (5, 6, 1), (5, 0, 7), (0, 6, 7)},
by Theorem 4.3 that
Sd (U (Su (D(Q∗ ((0, 0, 1), (5, 6, 7))))))
= Sd (U (Su (D((2, 2, 3), (3, 3, 2), (5, 6, 1), (5, 0, 7), (0, 6, 7)))))
= Sd (U ((1, 1, 4), (1, 3, 3), (1, 4, 2), (3, 1, 3), (4, 1, 2)))
= {(2, 2, 3), (3, 3, 2)}
= Q.
as well. Hence, the monomial ideal I2 = (xyz 4 , x1 y 3 z 3 , xy 4 z 2 , x3 yz 3 , x4 yz 2 ) ⊆ K[x, y, z] also has
Soc(I2 ) spanned by the image of M = {x2 y 2 z 3 , x3 y 3 z 2 } in R/I2 as a k-vector space.
12
Remarks.
(i) As with many formulae, when it comes utilizing them to compute specific values, the compact forms and shortness is not always a guarantee for a fast evaluation. Conversely, a
seemingly ugly expression can many times be much better to use to obtain specific values
in a fast and an efficient manner. The computation of Su (D(Q∗ )), from an antichain Q
consisting of k points from Nd as in Theorem 4.2, can for each fixed k be done in polynomial
time in k alone. In fact, it can be done in O(kd )-time, although the exact or a tight upper
bound of its complexity is hard to come by.
(ii) We have so far assumed G, Q ⊆ Nd , only to guarantee that G∗ , Q∗ ⊆ Nd0 . However, general
G, Q ⊆ Nd0 will imply G∗ , Q∗ ⊆ {−1, 0, 1, 2, . . .}d ⊆ Zd which does not correspond to a set
−1
of monomials from [x1 , . . . , xd ] but rather monomials from [x1 , x−1
1 , . . . , xd , xd ].
5. Zero-dimensional monomial ideals
For a field K, the ring R = K[x1 , . . . , xd ], the maximal ideal m of R and a monomial ideal I of
R, the motivation for this section is the following question.
Question 5.1. Under what circumstances can we retrieve a unique monomial ideal I from the
monomials of Soc(I) = Socm (I) that are not in I?
We saw in the last example in the previous Section 4 that both the following ideals of R =
K[x, y, z]
I1 = (x2 y 2 z 4 , x2 y 3 z 3 , x2 y 4 z 2 , x3 y 2 z 3 , x4 y 2 z 2 ) and
I2 = (xyz 4 , x1 y 3 z 3 , xy 4 z 2 , x3 yz 3 , x4 yz 2 )
have Soc(I1 ) and resp. Soc(I2 ) spanned by the image of M = {x2 y 2 z 3 , x3 y 3 z 2 } in R/I1 and
resp. R/I2 as a k-vector space. So Question 5.1 has in general a negative answer. In fact, for any
monomial ideal I it is easy to construct a rich family of monomial ideals such that ∂oc(I ′ ) = ∂oc(I)
for any ideal I ′ in that family.
If, however, we assume dim(I) = 0 then, we will see shortly, Question 5.1 has a positive answer.
Zero-dimensional monomial ideals of R = K[x1 , . . . , xd ] constitute an interesting yet fairly general
class of monomial ideals for numerous reasons: the quotient ring R/I is a local ring with a unique
prime and maximal ideal, R/I is finite dimensional over k, and their variety consist of a single
point 0̃, to name a few.
13
For a zero-dimensional ideal I 6= m of R we note that the image of Socm (I) in R/I corresponds to
an antichain Q of Nd0 . As mentioned in the remark after Theorem 4.3, we then have for our Q ⊆ Nd0
that Q = Sd (U (Su (D(Q∗ (ã, b̃))))) for any suitable ã, b̃ ∈ Zd . In particular, since we have −1̃ =
(−1, . . . , −1) ≤ m̃ from Theorem 4.3, then for any b̃ ≥ M̃ we have Q = Sd (U (Su (D(Q∗ (−1̃, b̃))))).
We now briefly argue that the upset U (Su (D(Q∗ (−1̃, b̃)))) corresponds to a monomial ideal I of R
of dimension zero: first we note that by definition (7) we have
D(Q∗ (−1̃, b̃)) ⊇ D(B∗ (−1̃, b̃))
= D((−1, b2 , b3 . . . , bd ), (b1 , −1, b3 , . . . , bd ), . . . , (b1 , b2 , b3 , . . . , −1)
and hence Su (D(Q∗ (−1̃, b̃))) ⊆ U (0̃) = Nd0 , and so the generators for the upset U (Su (D(Q∗ (−1̃, b̃))))
correspond to monomials of R = K[x1 , . . . , xd ]. Secondly, since b̃ ≥ M̃ as in Theorem 4.3, then for
each i ∈ [d] we have bi ẽi = (0, . . . , 0, bi , 0, . . . , 0) 6∈ D(Q∗ (−1̃, b̃)) and −ẽi = (0, . . . , 0, −1, 0, . . . , 0) ∈
D(Q∗ (−1̃, b̃)), and so there is a smallest ci ∈ N0 , ci ≤ bi with ci ẽi 6∈ D(Q∗ (−1̃, b̃)). Since ci ẽi − ẽj ∈
D(Q∗ (−1̃, b̃)) for each j 6= i and ci ẽi − ẽi = (ci − 1)ẽi ∈ D(Q∗ (−1̃, b̃)), we have that ci ẽi ∈
Su (D(Q∗ (−1̃, b̃))) for each i, which means that xci i is an element of the monomial ideal of I that
corresponds to U (Su (D(Q∗ (−1̃, b̃)))), showing that I is indeed of dimension zero.
To see that the zero-dimensional ideal I is unique, it suffices to show that the upset from above
U = U (Su (D(Q∗ (−1̃, b̃)))) is the unique upset with Q = Sd (U ). Assume I1 and I2 are two zerodimensional monomial ideals with the same socle w.r.t. the maximal ideal m of R. As each Ii is
zero-dimensional, it corresponds to an upset U (G∗i (0̃, b̃i )) where G∗i (0̃, b̃i ) = Gi ∪ B ∗ (0̃, b̃i ) is an
antichain of Nd0 as in (4). By our assumption we have Sd (U (G∗1 (0̃, b̃1 ))) = Q = Sd (U (G∗2 (0̃, b̃2 )))
and hence by Theorem 3.4 we then get
G1 = Su (D(Sd (U (G∗1 (0̃, b̃1 ))))) = Su (D(Q)) = Su (D(Sd (U (G∗1 (0̃, b̃1 ))))) = G2 .
Also, directly from Sd (U (G∗1 (0̃, b̃1 ))) = Q = Sd (U (G∗2 (0̃, b̃2 ))), we obtain for each i ∈ [d] that
b1i = πi (b̃1 ) = max({πi (q̃) : q̃ ∈ Q}) = πi (b̃2 ) = b12 ,
and hence b̃1 = b̃2 . Therefore
G∗1 (0̃, b̃1 ) = G1 ∪ B ∗ (0̃, b̃1 ) = G2 ∪ B ∗ (0̃, b̃2 ) = G∗2 (0̃, b̃2 ),
and so U (G∗1 (0̃, b̃1 ) = U (G∗2 (0̃, b̃2 ). This means that the corresponding zero-dimensional monomial
ideals are equal I1 = I2 . This yields a positive answer to Question 5.1 for zero-dimensional monomial
14
ideals, as stated in the following proposition, in which the set S of monomials corresponds to our
antichain Q ⊆ Nd0 from above.
Proposition 5.2. For any non-empty set S of non-comparable monomials of R = K[x1 , . . . , xd ]
there is a unique zero-dimensional ideal I of R with ∂oc(I) = S.
For the remainder of this section we discuss in further detail the cases when the set S of monomials
in Proposition 5.2, that corresponds to an antichain Q ⊆ Nd0 , has cardinality 1 and 2, that is
|S| = |Q| ∈ {1, 2}.
The case when Q is a singleton. An interesting case of Proposition 5.2 this is when |Q| = 1,
so the antichain Q of Nd0 consists of just one point, say Q = {p̃}. As before, the unique zerodimensional monomial ideal corresponds to the upset U = U (Su (D(Q∗ (−1̃, b̃)))) where Q = Sd (U ).
If now p̃ = (p1 , . . . , pd ) ∈ Nd0 , we note that the downset D(Q∗ (−1̃, b̃)) is generated by an antichain
consisting of d + 1 elements
Q∗ (−1̃, b̃) = {p̃, (−1, b2 , b3 . . . , bd ), (b1 , −1, b3 , . . . , bd ), . . . , (b1 , b2 , b3 , . . . , −1)}.
Since each element r̃ ∈ Su (D(Q∗ (−1̃, b̃))) is, by definition of Su (D), uniquely determined by d
distinct elements of Q∗ (−1̃, b̃) (by the conditions (i) r̃ 6∈ Q∗ (−1̃, b̃) and (ii) r̃ − ẽi ∈ Q∗ (−1̃, b̃) for
= d + 1 elements
each i ∈ [d]), we have that Su (D(Q∗ (−1̃, b̃))) is a subset of the possible d+1
d
{0̃, (p1 + 1)ẽ1 , . . . , (pd + 1)ẽd }, and hence, as it is an antichain, we obtain
Su (D(Q∗ (−1̃, b̃))) = {(p1 + 1)ẽ1 , . . . , (pd + 1)ẽd }},
which means that the upset U = U (Su (D(Q∗ (−1̃, b̃)))) corresponds to a monomial ideal I of the
form I = (xp11 +1 , . . . , xdpd +1 ).
That |S| = 1 means that ∂oc(I) = S contains the unique generator of the ideal Socm (I) of R/I.
This means that R/I is a zero dimensional local Gorenstein ring [Eis95, Prop. 21.5]. Clearly, if
our monomial ideal I is generated by pure powers of the variables x1 , . . . , xd , then R/I is a zero
dimensional local Gorenstein ring with a unique monomial in ∂oc(I). Hence, we obtain as a corollary
the following description of Artinian Gorenstein rings that are defined by monomials [Bei93], see
also [Wol16].
Corollary 5.3. Let R = K[x1 , . . . , xd ] and I be a monomial ideal of R. Then R/I is a zero
dimensional local Gorenstein ring if, and only if, I = (xp11 +1 , . . . , xpdd +1 ) for some p̃ ∈ Nd0 , in which
case ∂oc(I) = {x̃p̃ = xp11 · · · xpdd }.
15
For a general Q ⊆ Zd we have, as mentioned here above, that every element of Su (D(Q)) is
uniquely determined by d distinct elements of Q. More specifically we have the following.
Claim 5.4. For an antichain G ⊆ Zd and r̃ ∈ Sd (U (G)) there are distinct p̃1 , . . . , p̃d ∈ G such that
(i) r̃ ∈ Sd (U (p̃1 , . . . , p̃d )) and (ii) r̃ = (p11 + 1, . . . , pdd + 1).
Dually, for an antichain Q ⊆ Zd and r̃ ∈ Su (D(Q)) there are distinct p̃1 , . . . , p̃d ∈ Q such that
(i) r̃ ∈ Sd (U (p̃1 , . . . , p̃d )) and (ii) r̃ = (p11 − 1, . . . , pdd − 1).
We can in similar fashion, as for Corollary 5.3, consider the case when |S| = k, that is the
corresponding antichain Q of Zd0 contains k points Q = {p̃1 , . . . , p̃k }, where each p̃i = (pi1 , . . . , pid ).
In this case we have Q∗ (−1̃, b̃) = Q ∪ B∗ (−1̃, b̃) where B∗ (−1̃, b̃) is as in (7), and hence bi ≥
max(pi1 + 1, . . . , pid + 1) for each index i, and so the downset D(Q∗ (−1̃, b̃)) is generated by an
antichain of d + k elements. For convenience we let
b̃(i) := (b1 , . . . , bi−1 , −1, bi+1 , . . . , bd ),
for each i ∈ [d], so B∗ (−1̃, b̃) = {b̃(1), . . . , b̃(d)}. The following lemma provides a useful tool.
d), . . . , b̃(i
d), . . . , b̃(d))
Lemma 5.5. For I = {i1 , . . . ik } ⊆ [d] and D = D(p̃1 , . . . , p̃k , b̃(1), . . . , b̃(i
1
k
we have
Su (D) =
nP
k
ℓ=1 riℓ ẽiℓ
∅
o
6 ∅
if Su (D(πI (p̃1 ), . . . , πI (p̃k ))) = {(ri1 , . . . , rik )} =
otherwise.
Proof. For r̃ ∈ Su (D) we have by definition that r̃ 6∈ D and r̃ − ẽℓ ∈ D for each ℓ. Since D ⊆ D(b̃)
we have then rk ≤ bk for each k 6= ℓ. As this holds for each ℓ we have r̃ ≤ b̃.
Further, for r̃ ∈ Su (D), there is, analogous to Claim 5.4, a permutation σ ∈ Sd such that
r̃ − ẽσ(ℓ) ≤ b̃(ℓ) for each index ℓ 6∈ I and r̃ − ẽσ(iℓ ) ≤ p̃ℓ for 1 ≤ ℓ ≤ k. Since r̃ 6∈ D we hence have
for each ℓ 6∈ I that
(8)
rσ(ℓ)
b
σ(ℓ) + 1
=
0
if σ(ℓ) 6= ℓ
if σ(ℓ) = ℓ
and rσ(ℓ) = pℓσ(ℓ) + 1 for each ℓ ∈ I. Since r̃ ≤ b̃ we must by (8) have that rσ(ℓ) = 0 for each ℓ 6∈ I
and therefore σ(ℓ) = ℓ for each ℓ 6∈ I. Hence we have rℓ = 0 for each ℓ 6∈ I. Consequently σ ∈ S(I)
is a permutation only on I and leaves every other element of [d] \ I fixed. In particular we have
that
r̃ =
k
X
riℓ ẽiℓ =
X
i∈I
ℓ=1
16
ri ẽi .
Since r̃ 6∈ D we have r̃ 6∈ D(p̃i1 , . . . , p̃ik ) and since r̃ − ẽσ(iℓ) ≤ p̃ℓ for each ℓ ∈ [k], or r̃ − ẽℓ ≤
p̃γ(ℓ) for each ℓ ∈ I, where γ : I → [k] is the map iℓ 7→ ℓ, we have (ri1 , . . . , rik ) = πI (r̃) ∈
Su (D(πI (p̃1 ), . . . , πI (p̃k ))), which by Claim 5.4 is uniquely determined.
⊔
⊓
From the above proof we note that if Su (D) 6= ∅ and r̃ ∈ Su (D), then rσ(iℓ ) = Pℓσ(iℓ ) + 1 for
1 ≤ ℓ ≤ k and so rσ(iℓ ) = Pγ(iℓ )σ(iℓ ) + 1 for 1 ≤ ℓ ≤ k, or rℓ = pβ(ℓ)ℓ + 1 for each ℓ ∈ I, where
β = γ ◦ σ −1 , and so
r̃ = (pβ(ii )i1 + 1)ẽi1 + · · · + (pβ(ik )ik + 1)ẽik
for the bijection β : I → [k]. From this we get the following.
d), . . . , b̃[
Observation 5.6. Let I = {i1 , . . . ik } ⊆ [d] and D = D(p̃1 , . . . , p̃k , b̃(1), . . . , b̃(i
(ik ), . . . , b̃(d)).
1
If Su (D) 6= ∅ and r̃ ∈ Su (D) is its unique element, then there is a permutation α ∈ Sk such that
πI (r̃) = (pα(1)i1 + 1, . . . , pα(k)ik + 1) ∈ Su (D(πI (p̃1 ), . . . , πI (p̃k ))).
The case when k = |Q| = 2. The second special case of k = 2, so Q = {p̃, q̃}, is, as we will
see here below, simple enough so that we can state some conclusion in a direct and uncluttered
manner. The following follows directly from the above Lemma 5.5 and Observation 5.6.
d
d
Corollary 5.7. For two distinct indices i, j ∈ [d] and D = D(p̃, q̃, b̃(1), . . . , b̃(i), . . . , b̃(j), . . . , b̃(d))
we have
{(pi + 1)ẽi + (qj + 1)ẽj }
Su (D) =
{(qi + 1)ẽi + (pj + 1)ẽj }
∅
if pi < qi and qj < pj
if qi < pi and pj < qj
otherwise.
For the antichain Q = {p̃, q̃} of Nd0 we have a pseudo-partition [d] = A ∪ B ∪ C (some parts could
be empty) where pi < qi for all i ∈ A, pi = qi for all i ∈ B, and pi > qi for all i ∈ C. We have here
that
Q∗ (−1̃, b̃) = Q ∪ B∗ (−1̃, b̃) = {p̃, q̃, b̃(1), . . . , b̃(d)}.
Since Su (D) forms an antichain, for any downset D ⊆ Zd , and each element of Su (Q∗ (−1̃, b̃) Is
uniquely determined by a d-subset of Q∗ (−1̃, b̃), then Su (Q∗ (−1̃, b̃) is among the at most d+2
=
d
S
d+2
elements of Q′ Su (D(Q′ )), where the union is taken over all Q′ ⊆ Q∗ (−1̃, b̃) with cardinality
2
d. Since [d] = A ∪ B ∪ C is a pseudo-partition, then by letting |A| = a, |B| = b and |C| = c we
have |A| + |B ∪ C| = a + b + c = d, and therefore, by Corollary 5.7, we then have the following.
17
Proposition 5.8. For an antichain Q = {p̃, q̃} of of Zd0 and the corresponding pseudo-partition
[d] = A ∪ B ∪ C we have
Su (D(Q∗ (−1̃, b̃))) =
{(pi + 1)ẽi + (qj + 1)ẽj : (i, j) ∈ A × C} ∪ {(pi + 1)ẽi : i ∈ B ∪ C} ∪ {(qi + 1)ẽi : i ∈ A},
and hence |Su (D(Q∗ (−1̃, b̃)))| = ac + d.
Note that
{(pi + 1)ẽi : i ∈ B ∪ C} ∪ {(qi + 1)ẽi : i ∈ A}
= {(pi + 1)ẽi : i ∈ B ∪ C} ∪ {(qi + 1)ẽi : i ∈ A ∪ B}
= {(pi + 1)ẽi : i ∈ C} ∪ {(qi + 1)ẽi : i ∈ A ∪ B}.
In the above display, the first and the last unions are pseudo-partitions and not symmetric, whereas
the middle union is symmetric but not disjoint.
Similar to Corollary 5.3 describing Artinian Gorenstein rings defined by monomial from [Bei93],
we can interpret the above Proposition 5.8 for Artinian rings defined by monomials that are almost
Gorenstein as defined in [HV06] as those monomial ideals I of R = K[x1 , . . . , xd ] with ∂oc(I) = S
from Proposition 5.2 being a set of “small” cardinality. More specifically, if I is a zero dimensional
monomial ideal of R with |S| = k ≥ 1, then I is said to be of type k, and so type 1 zero dimensional
ideals are exactly the Artinian Gorenstein monomial ideals [HV06]. The above Proposition 5.8
gives a complete description of the type 2 zero dimensional monomial ideals of R.
Corollary 5.9. Let R = K[x1 , . . . , xd ] and I be a monomial ideal of R. Then R/I is a zero
dimensional local type 2 ring if, and only if, for two incomparable p̃, q̃ ∈ Zd0 with corresponding
pseudo-partition [d] = A ∪ B ∪ C we have
q +1
I = (xipi +1 xj j
, xphh +1 , xqkk +1 : (i, j) ∈ A × C, h ∈ B ∪ C, k ∈ A),
in which case ∂oc(I) = S = {x̃p̃ , x̃q̃ }.
Remarks.
(i) Note that Corollary 5.9 characterizes completely those monomial ideals I of R = K[x1 , . . . , xd ]
that make R/I a local Artinian ring of type 2, in a purely combinatorial way, using only
the poset structure of Zd ,
18
(ii) Comparing with [HV06, Example 4.3] we see that the analysis of type 2 monomial ideal done
there is slightly different from ours from Corollary 5.9, in that there in [HV06] the authors
describe when exactly the condition J1 : J2 + J2 : J1 ⊇ m = (x1 , . . . , xd ) from [HV06,
Theorem 4.2] holds, where I = J1 ∩ J2 is the irredundant intersection of two irreducible
ideals J1 and J2 , each of which must be monomial ideals generated by powers of the variables
x1 , . . . , xd .
6. Almost Gorenstein monomial ideals with k ≥ 3
The general case for k = |Q| ≥ 3. We have so far analysed fully the structure of the antichain
Su (Q∗ (−1̃, b̃)) of Zd0 in the cases when k = |Q| ∈ {1, 2}, the antichain corresponding to the gener-
ators of a zero-dimensional monomial ideal with ∂oc(I) corresponding to Q. These two cases are
unique in that the form of the antichain Q is quite simple: an antichain Q = {p̃, q̃} in Zd with the
corresponding pseudo-partition [d] = A ∪ B ∪ C as indicated in Proposition 5.8 must have both
parts A and C non-empty and so ac ≥ 1. This means that the parts (here A and C) corresponding
to the two possible strict orderings of the coordinates, namely pi < qi or pi > qi , are each nonempty.
Needless to say, for an antichain Q ⊆ Zd where k = |Q| ≥ 3, this need not be case. The possibilities
are simply to great. There are plenty antichains (in fact, vastly most of them, as we shall see
momentarily) where not all strict orderings of the coordinates are present.
Analogous to the pseudo-partition [d] = A ∪ B ∪ C from Proposition 5.8 we have for general
k ∈ N a pseudo-partition
(9)
[d] =
[
Aω ,
ω∈W (k)
where W (k) is the collection of weak orderings of the set [k]: the collection of distinct orderings of
the elements where equality is allowed. In particular we clearly have for each k ≥ 2 that |W (k)| > k!,
the number of strict orderings of [k], and so the union in (9) is large. More specifically, for k = 1, 2
and 3 there are 1, 3 and 13 weak orderings in W (k) respectively, and, in general, a(k) = |W (k)| is
the k-th ordered Bell number or Fubini number given explicitly by
a(k) =
k
X
k
i=1
19
i
i!,
where
k
i
= S(k, i) is the Stirling number of the second kind [Sti] and where the exponential
generating function for the corresponding sequence (a(k))k≥0 = (1, 1, 3, 13, 75, 541, . . .) is given by
∞
X
a(k)
k=0
Asymptotically a(k) ≈
n!
2(log 2)n+1
k!
xk =
1
.
2 − ex
= 12 (1.4427 · · · )n+1 n!, where log is the natural logarithm and so
log 2 < 1, and so we see that the number k! of strict orderings of [k] constitutes only a small fraction
of the a(k) weak orderings of [k]. The ordered Bell numbers have been studied extensively [OBn,
A00].
Definition 6.1. An antichain Q = {p̃1 , . . . , p̃k } ⊆ Zd is called order-generic if in the corresponding
pseudo-partition (9) for each ω ∈ W (k) of strict orderings of [k] we have Aω 6= ∅.
Note that in order for an antichain Q ⊆ Zd to be order-generic we must have d ≥ k!. We will
see that the structure of the antichain Su (Q∗ (−1̃, b̃)) of Zd0 when Q is order-generic is nice enough
for some specific enumerations.
The following slight reformulation follows directly from Lemma 5.5 and Observation 5.6.
Corollary 6.2. For I = {i1 , . . . ik } ⊆ [d] containing distinct elements and
we have
d
d
D = D(p̃1 , . . . , p̃k , b̃(1), . . . , b̃(i1 ), . . . , b̃(ik ), . . . , b̃(d))
Su (D) = {(pα(1)i1 + 1)ẽi1 + · · · + (pα(k)ik + 1)ẽik },
if there is a unique permutation α ∈ Sk with pα(h)ih < pα(1)ih , . . . , p\
α(h)ih , . . . , pα(k)ih for each h ∈ [k]
and Su (D) = ∅ otherwise.
For k ∈ N we have that Su (Q∗ (−1̃, b̃), the set of Nd0 corresponding to the generators of our desired
= d+k
possible
monomial ideal, is the antichain formed by the maximal elements of all the d+k
d
k
singletons Su (D), where D ⊆ Q∗ (−1̃, b̃) contains exactly d elements, and hence there are k types of
points in Su (Q∗ (−1̃, b̃): for each ℓ ∈ [k] the ones that are formed by a d-set D that contains exactly
ℓ points from Q = {p̃1 , . . . , p̃k }. Since our antichain is order-generic each mentioned type occurs in
Su (Q∗ (−1̃, b̃). To effectively list of these points we make the following convention. For each a ∈ [k]
and a subset C ⊆ [k] we let
(10)
Ba (C) := {i ∈ [d] : pai < pbi for all b ∈ C \ {a}} .
20
For C ⊆ C ′ ⊆ [k] we clearly have Ba (C ′ ) ⊆ Ba (C). With this convention the sets of the mentioned
type ℓ ∈ [k], that contain those points formed by d-set D containing ℓ points from Q, are those
labeled by a C ⊆ [k] with |C| = ℓ, and are of the form
(
)
X
Y
[ Y
′
(11)
PC =
(ptit + 1)ẽit : (it )t∈C ∈
Bt (C) \
Bt (C ) .
t∈C
t∈C
C ′ ⊃C t∈C
In particular, for ℓ = k we have
P[k] = {(p1i1 + 1)ẽi1 + · · · + (pkik + 1)ẽik : (i1 , . . . , ik ) ∈ B1 ([k]) × · · · × Bk ([k])}.
(12)
and for ℓ = k − 1 we obtain the sets labeled by [â] := {1, . . . , â, . . . , k} for each a ∈ [k] as follows
(13)
P[â] = {(p1i1 + 1)ẽi1 + · · · + (pkik + 1)ẽik : (i1 , . . . , iba , . . . , ik )
\
∈ B1 ([â]) × · · · × B\
a ([â]) × · · · × Bk ([â]) \ B1 ([k]) × · · · × Ba ([k]) × · · · × Bk ([k])}}.
Note that when ℓ = 1, so our d-set contains a single element p̃a from Q, then we have by this
construction that
P{a} = {(pai + 1)ẽi : pai ≥ pbi for all b 6= a}.
Remarks.
(i) Note that the sets PC are constructed directly in terms of the sets Ba (C) from (10), each
which is union of sets Aω from (9).
(ii) The fact that Q is order-generic ensures that the sets Ba (C) are all non-empty, which in
return will imply that each PC from (11) does not contain comparable elements, and is
therefore an antichain, as we will see below.
We clearly have Su (D(Q∗ (−1̃, b̃))) ⊆
S
∅6=C⊆[k] PC ,
but we have not verified that this union of
these sets PC is actually an antichain. By the following two lemmas we will see that this is indeed
the case.
Lemma 6.3. Let Q = {p̃1 , . . . , p̃k } be an order-generic antichain in Zd . If r̃ = (pa1 i1 + 1)ẽi1 +
· · · + (paℓ iℓ + 1)ẽiℓ ∈ P{a1 ,...,aℓ } and s̃ = (pb1 i1 + 1)ẽi1 + · · · + (pbℓ iℓ + 1)ẽiℓ ∈ P{b1 ,...,bℓ } , then r̃ and s̃
are comparable in Zd if and only if r̃ = s̃.
Proof. Suppose r̃ < s̃ so r̃ 6= s̃. In this case there is a coordinate, say i1 for simplicity, with
pa1 i1 < pb1 i1 .
21
Suppose that b1 = ah ∈ {a1 , . . . , aℓ }. We first note that h 6= 1, since if h = 1, then pb1 i1 =
pah i1 = pa1 i1 . We then have pah ih = pb1 ih > pbh ih since s̃ ∈ P{b1 ,...,bℓ } , and so looking at coordinates
i1 and ih we see that r̃ and s̃ are in this case incomparable, which is a contradiction. Therefore we
have that b1 6∈ {a1 , . . . , aℓ }.
By definitions of Ba (C) in (10) we then have ij ∈ Baj ({b1 , a1 , . . . , aℓ }) for 1 ≤ j ≤ ℓ and
therefore (i1 , . . . , iℓ ) ∈ Ba1 (C ′ ) × · · · × Baℓ (C ′ ) for C ′ = {b1 , a1 , . . . , aℓ } ⊃ {a1 , . . . , aℓ } contradicting
⊔
⊓
r̃ =∈ P{a1 ,...,aℓ } . Hence, r̃ = s̃ must hold. This completes the proof.
Remark. We can say a bit more than in the above proof: since Q is order-generic there is an
iℓ+1 ∈ Bb1 ({b1 , a1 , . . . , aℓ }) and so
r̃ ′ := r̃ + (pb1 iℓ+1 + 1)ẽℓ+1 ∈ PC ′ = P{b1 ,a1 ,...,aℓ } ,
or is dominated by an element in PC ′′ where {b1 , a1 , . . . , aℓ } ⊆ C ′′ .
The claims of Lemma 6.3 actually hold in a slightly more general setting: suppose Q = {p̃1 , . . . , p̃k }
is an order-generic antichain in Zd . Suppose also that C ⊆ C ′ ⊆ [k], r̃ ∈ PC , s̃′ ∈ PC ′ and
r̃ ≤ s̃′ in Zd . In this case we can write C = {a1 , . . . , aℓ } and C ′ = {b1 , . . . , bh } where ℓ ≤ h. If
r̃ = (pa1 i1 + 1)ẽi1 + · · · + (paℓ iℓ + 1)ẽiℓ ∈ PC and s̃′ = (pb1 i1 + 1)ẽi1 + · · · + (pbh ih + 1)ẽih ∈ PC ′ ,
then, since r̃ ≤ s̃′ there is a projective image s̃ of s̃′ , which we can assume has the form s̃ =
(pb1 i1 + 1)ẽi1 + · · · + (pbℓ iℓ + 1)ẽiℓ , such that r̃ ≤ s̃. As in the proof of the previous Lemma 6.3,
if r̃ 6= s̃ then there is a coordinate, say i1 for simplicity, with pa1 i1 < pb1 i1 . Also, since r̃ < s̃ we
cannot have b1 ∈ {a1 , . . . , aℓ } and so C ′ = {b1 , a1 , . . . , aℓ } ⊂ {a1 , . . . , aℓ }. Since ij ∈ Baj (C ′ ) for
1 ≤ j ≤ ℓ we also get in this case a contradiction to the fact that r̃ ∈ PC . We summarize in the
following Lemma.
Lemma 6.4. Let Q = {p̃1 , . . . , p̃k } be an order-generic antichain in Zd and suppose C ⊆ C ′ ⊆ [k],
r̃ ∈ PC and s̃′ ∈ PC ′ . If r̃ ≤ s̃′ , then r̃ = s̃ where s̃ is a projective image of s̃′ , and this can only
occur when C ′ = C and s̃′ = s̃ = r̃.
By Lemmas 6.3 and 6.4 we have in particular that the union
We therefore have the following.
S
∅6=C⊆[k] PC
is an antichain in Zd .
Proposition 6.5. For an order-generic antichain Q = {p̃1 , . . . , p̃k } in Zd we have
[
(14)
Su (D(Q∗ (−1̃, b̃))) =
PC ,
∅6=C⊆[k]
22
where the sets PC are as in (11).
Remarks.
(i) Note that (14) does not hold when Q is not order-generic, since the union on the right is
not in general an antichain.
(ii) Although the union on the right in (14) is not, in general, disjoint, it does yield an
inclusion/exclusion-like formula for explicit enumeration for an order-generic Q. This will
be demonstrated here below for the case k = 3.
The case when k = 3. Here we assume that our order-generic antichain Q = {p̃1 , p̃2 , p̃3 } ⊆ Zd0
contains three points. This case, as we will see, is considerably more complicated than the case
k = 2.
For distinct i1 , i2 , i2 ∈ [d] and
d
d
d
D = D(p̃1 , p̃2 , p̃3 , b̃(1), . . . , b̃(i1 ), . . . , b̃(i2 ), . . . , b̃(i3 ), . . . , b̃(d))
we have by Corollary 6.2 that
Su (D) = {(pα(1)i1 + 1)ẽi1 + (pα(2)i2 + 1)ẽi2 + (pα(3)i3 + 1)ẽi3 },
if there is a unique permutation α ∈ S3 with pα(1)i1 < pα(2)i1 , pα(3)i1 ; pα(2)i2 < pα(1)i2 , pα(3)i2 ; and
pα(3)i3 < pα(1)i3 , pα(2)i3 and Su (D) = ∅ otherwise.
For {u, v, w} = {1, 2, 3} we will write the parts of [d] as defined in (9) in the following way:
Au,v,w = {i ∈ [d] : pui < pvi < pwi },
Auv,w = {i ∈ [d] : pui = pvi < pwi },
Au,vw = {i ∈ [d] : pui < pvi = pwi },
and A123 = {i ∈ [d] : p1i = p2i = p3i }. In this way our pseudo-partition of [d] from in (9) becomes:
[d] = A123 ∪ A12,3 ∪ A13,2 ∪ A23,1 ∪ A1,23 ∪ A2,13 ∪ A3,12
(15)
∪ A1,2,3 ∪ A1,3,2 ∪ A2,1,3 ∪ A2,3,1 ∪ A3,1,2 ∪ A3,2,1 .
As Su (Q∗ (−1̃, b̃) is the antichain formed by the maximal elements of all the
d+3
d
=
d+3
3
possible
singletons Su (D), where D ⊆ Q∗ (−1̃, b̃) contains exactly d elements, then there are three types of
points in Su (Q∗ (−1̃, b̃): those ones obtained from a d-set D that contains all the three points of
23
Q = {p̃1 , p̃2 , p̃3 }, those obtained from a D that contain exactly two points of Q, and finally those
obtained from a D that contain exactly one of p̃1 , p̃2 or p̃3 .
In order to facilitate our notation and presentation we define the following subsets of [d], where
{u, v, w} = {1, 2, 3}:
Bhu,vi := Auw,v ∪ Aw,u,v ,
B hui := Av,w,u ∪ Aw,v,u ∪ Avw,u ,
B hu∗i := Aw,uv ∪ Av,uw ,
that is Bhu,vi contains all the coordinates i ∈ [d] where pui < pvi and pui is not the sole minimum,
B hui contains all the coordinates i ∈ [d] where pui is the sole maximum among pui , pvi and pwi and
B hu∗i contains those coordinates i where pui and exactly one of the other two, pvi or pwi , are the
maximum ones.
By (12) the set of points in Su (Q∗ (−1̃, b̃) that are formed by the sets Su (D) where D ⊆ Q∗ (−1̃, b̃)
contain all the three points of Q = {p̃1 , p̃2 , p̃3 }, is then given by
(16) P[3] = {(p1i1 + 1)ẽi1 + (p2i2 + 1)ẽi2 + (p3i3 + 1)ẽi3 : (i1 , i2 , i3 ) ∈ B1 ([3]) × B2 ([3]) × B3 ([3])}.
When {u, v, w} = {1, 2, 3} then we have
Bu ({u, v}) × Bv ({u, v}) \ Bu ([3]) × Bv ([3]) = Bhu,vi × Bhv,ui ∪ Bhu,vi × Bv ([3]) ∪ Bu ([3]) × Bhu,vi
and hence the set Pŵ = Puv of points in Su (Q∗ (−1̃, b̃) formed by the sets Su (D) where D contains
exactly two points p̃u and p̃v are by (13) given by
(17) P{u,v} = {(pui + 1)ẽi + (pvj + 1)ẽj : (i, j) ∈ Bhu,vi × Bhv,ui ∪ Bhu,vi × Bv ([3]) ∪ Bu ([3]) × Bhu,vi }.
Lastly, the set of points in Su (Q∗ (−1̃, b̃) formed by the sets Su (D) where D contains exactly one
point p̃i ∈ Q is given by
P{u} = {(pui + 1)ẽi : i ∈ B hui ∪ B hu∗i ∪ A123 }.
(18)
Note that for {u, v, w} = {1, 2, 3} we have
∩
Bhu,vi × Bhv,ui ∪ Bhu,vi × Bv ([3]) ∪ Bu ([3]) × Bhu,vi
Bhu,wi × Bhw,ui ∪ Bhu,wi × Bv ([3] ∪ Bu ([3]) × Bhu,wi )
= Bhui × Avw,u
and hence by Lemma 6.3 we have the following.
24
Claim 6.6. Two points r̃ ∈ P{u,v} and s̃ ∈ P{u,w} are comparable in Zd0 if and only they are equal
in which case r̃ = s̃ = (pui + 1)ẽi + (pvj + 1)ẽj , where (i, j) ∈ Bhui × Avw,u .
This claim together with (17) makes it possible to enumerate the maximal elements in the union
P{u,v} ∪ P{u,w} in a direct manner.
By symmetry and the above Claim 6.6 we have that if two points in P{u,v} and P{v,w} respectively
are comparable, then the must be equal, say to (pui + 1)ẽi + (pvj + 1)ẽj where (i, j) ∈ Auw,v × Bhvi .
Since (Bhui × Avw,u ) ∩ (Auw,v × Bhvi ) = ∅ we have, in particular, that no three points in P{u,v} ,
P{u,w} and P{v,w} respectively, are pairwise comparable in Zd0 .
Lastly, consider the sets P{1} , P{2} and P{3} from (18). Since for any distinct u, v ∈ {1, 2, 3} we
have
(B hui ∪ B hu∗i ∪ A123 ) ∩ (B hvi ∪ B hv∗i ∪ A123 ) = Aw,uv ∪ A123 ,
then clearly two comparable elements in P{u} and P{v} respectively must be equal. Also note that
P{1} ∩ P{2} ∩ P{3} = {(pui + 1)ẽi : i ∈ A123 }.
Convention. For an upper case letter X ∈ {A, B} we denote the cardinality of X∗ , X ∗ by the
corresponding lower case boldface letter x∗ and x∗ respectively, so |A12,3 | = a12,3 , |B h2∗i | = bh2∗i
etc.
With this convention we can now list and also enumerate the elements in Su (Q∗ (−1̃, b̃) by the
inclusion/exclusion principle in the following.
Proposition 6.7. For an order-generic antichain Q = {p̃1 , p̃2 , p̃3 } of Zd0 and the corresponding
pseudo-partition of [d] as in (15), we have
Su (D(Q∗ (−1̃, b̃))) = P{1,2,3} ∪ P{1,2} ∪ P{1,3} ∪ P{2,3} ∪ P{1} ∪ P{2} ∪ P{3} ,
where P{1,2,3} , P{u,v} and each P{u} are as in (16), (17) and (18) respectively.
Further, the cardinality is given by
|Su (D(Q∗ (−1̃, b̃)))| = b1 ([3])b2 ([3])b3 ([3]) + bh1,2i bh2,1i + bh1,2i bh2i + bh1i bh2,1i
+ bh1,3i bh3,1i + bh1,3i bh3i + bh1i bh3,1i + bh2,3i bh3,2i + bh2,3i bh3i + bh2i bh3,2i
− (bh1i a23,1 + bh2i a13,2 + bh3i a12,3 ) + bh1i + bh2i + bh3i + a3,12 + a2,13 + a1,23 + a123 .
Remark. Unlike the presentation in Proposition 5.8 and Corollary 5.9 the presentation for the
antichain Su (D(Q∗ (−1̃, b̃))) as a union in Proposition 6.7 is symmetric, however it is not a disjoint
union, as that would be quite convoluted and confusing.
25
Corollary 6.8. For an order-generic antichain Q = {p̃1 , p̃2 , p̃3 } of Zd0 , the unique local zero di-
mensional type 3 monomial ideal I of R = K[x1 , . . . , xd ] with ∂oc(I) = {x̃p̃1 , x̃p̃2 , x̃p̃3 } is given
by
I = (x̃r̃ : r̃ ∈ Su (D(Q∗ (−1̃, b̃)))}
where Su (D(Q∗ (−1̃, b̃))) is as in Proposition 6.7.
Remarks.
(i) Corollary 6.8 characterizes completely those monomial ideals I of R = K[x1 , . . . , xd ] that
make R/I a local Artinian ring of type 3, where ∂oc(I) corresponds to an order-generic
antichain in Zd in a purely combinatorial way, using only the poset structure of Zd .
(ii) Note that since Su (D(Q∗ (−1̃, b̃))) is an antichain, then I as presented in Corollary 6.8 is
exactly its minimal generation, that is, {x̃r̃ : r̃ ∈ Su (D(Q∗ (−1̃, b̃)))} is the unique Gröbner
basis for I.
7. Socles of general ideals
We have so far only been interested in the combinatorial properties of monomial ideals of R =
K[x1 , . . . , xd ] that can described solely in terms of the monoid [x1 , . . . , xd ] viewed as a poset, where
the partial order is given by divisibility. As a consequence, the role of the field K has so far not
played a major role in our investigations on monomial ideals. We have at times further assumed
our monomial ideal I ⊆ R to be positive, that is I ⊆ (x1 · · · xd ), or each (generating) monomial x̃p̃
of I having p̃ ∈ Nd . Also, we so far only considered the the socle Soc(I) = Socm (I) of R w.r.t. the
specific maximal ideal m = (x1 , . . . , xd ) ⊆ R.
Just to quickly iterate that we are not gaining anything by allowing K-linear combination when
considering monomial ideals, we note that that for a monomial ideal I = (x̃p̃1 , . . . , x̃p̃k ) ⊆ R the
set of its generators M = {x̃p̃1 , . . . , x̃p̃k } is always a Gröbner basis for I w.r.t. any term order on
[x1 , . . . , xd ]. As a result, a fully reduced polynomial f ∈ R \ {0} is exactly a linear combination of
fully reduced monomials w.r.t. I, that is, those monomials that do not reduce to zero in R/I. In
particular we have the following.
Observation 7.1. If I ⊆ R = K[x1 , . . . , xd ] is a monomial ideal, then the socle of R/I w.r.t. the
maximal ideal m = (x1 , . . . , xd ) is exactly the set of all K-linear combinations of the monomials of
the socle, Soc(I) = SpanK ({x̃r̃ : x̃r̃ ∈ Soc(I)}).
26
In this section we show that even in these mentioned seemingly restrictive case, we have managed
to capture many of the interesting combinatorial properties of the socle of a general ideal w.r.t. a
maximal ideal.
√
First we recall some notations and facts from commutative ring theory. For an ideal I ⊆ R let
I = {f ∈ R : f n ∈ I for some n ∈ N} denote the radical of I. For a field K and R = K[x1 , . . . , xd ]
then V (I) = {ã ∈ K d : f (ã) = 0 for all f ∈ I} is the affine variety defined by I. For a subset
U ⊆ K d we let I(U ) = {f ∈ R : f (ã) = 0 for all ã ∈ U } denote the ideal of all polynomials in R that
vanish on U . For any field K and a point ã ∈ K d then clearly (x1 −a1 , . . . , xd −ad ) ⊆ R is a maximal
ideal. For the converse, we recall an important corollary of Hilbert’s Nullstellensatz [AM69, p. 85]
and [Alu09, p. 410] as the following [Alu09, Cor. 2.10, p. 406].
Theorem 7.2. If K is an algebraically closed field, then an ideal of K[x1 , . . . , xd ] is maximal if
and only if it has the form mã = (x1 − a1 , . . . , xd − ad ) for some ã ∈ K d .
Recall the following theorem from [AM69, Prop. 1.11] on prime avoidance:
Theorem 7.3. Let p1 , . . . , pk be prime ideals in a commutative ring. If I ⊆
then I ⊆ pi for some i.
Sk
i=1
pi is an ideal,
Since R = K[x1 , . . . , xd ] is Noetherian, every ideal I ⊆ R has a primary decomposition I =
Tk
√
qi be the associated
q
p
i=1 i which we can assume to be minimal (or reduced). For each i let i =
prime ideal. The set {p1 , . . . , pk } of associated primes of the ideal I is uniquely determined by I
and is always a finite set.
Suppose m ⊆ R is a maximal ideal containing the ideal I is such that Socm (I) is nontrivial, so
Socm (I) = (I : m) 6= I. Since
I 6= (I : m) =
\
(I : f ),
f ∈m
then for each f ∈ m we have (I : f ) 6= I as well, and so by [AM69, Prop. 4.7] we then have
m ⊆ {f ∈ R : (I : f ) 6= I} =
k
[
pi .
i=1
By Theorem 7.3 above we therefore have that m ⊆ pi for some i, and since m is maximal we have
m = pi . We summarize in the following.
Proposition 7.4. If R is a Noetherian ring, I ⊆ R an ideal and m ⊆ R a maximal ideal containing
I such that Socm (I) = (I : m) 6= I, then m is one of the finitely many associated prime ideals of I.
27
If now I ⊆ R = K[x1 , . . . , xd ] is a monomial ideal then every associated prime of I is a face
ideal, or a Stanley-Reisner ideal [MS05, p. 19], that is an ideal of the form (xi1 , . . . , xih ) (See [Vil01,
Prop. 5.1.3].) Hence, if m ⊆ R is a maximal ideal such that Socm (I) 6= I, then by Proposition 7.4
m must be a face ideal and so we have the following.
Proposition 7.5. If K is a field, I ⊆ R = K[x1 , . . . , xd ] a monomial ideal and m ⊆ R a maximal
ideal containing I such that Socm (I) = (I : m) 6= I, then m = (x1 , . . . , xd ).
If now K is algebraically closed, then by Theorem 7.2 we have a bijective correspondence between
points ã ∈ K d and maximal ideals mã = (x1 − a1 , . . . , xd − ad ) of R = K[x1 , . . . , xd ].
Observation 7.6. Let R = K[x1 , . . . , xd ] where K is algebraically closed and I ⊆ R a proper
ideal. A maximal ideal m of R contains I if and only if m = mã for some ã ∈ V (I).
For an ideal I ⊆ R = K[x1 , . . . , xd ] where K is algebraically closed and mã a maximal ideal
containing I let Socã (I) := Socmã (I) = (I : mã ) denote the socle of I w.r.t. the maximal ideal mã .
By Proposition 7.4 the following.
Corollary 7.7. If K is an algebraically closed field and I ⊆ R = K[x1 , . . . , xd ] is an ideal of R,
then for all but finitely many points ã ∈ V (I) the socle Socã (I) = I is trivial.
Also, directly by Proposition 7.5 do we have the following corollary.
Corollary 7.8. If K is an algebraically closed field, I ⊆ R = K[x1 , . . . , xd ] is a monomial ideal of
R and Socã (I) 6= I, then ã = 0̃.
As the combinatorial properties we want to investigate can be phrased in terms of the monoid
[x1 , . . . , xd ] of indeterminates as a poset, we obtain virtually the same combinatorial properties of
our monomial ideal I ⊆ R = K[x1 , . . . , xd ] regardless of what the field K is. Hence, we might as
well look at the extended ideal I = IK[x1 , . . . , xd ] where K is the algebraic closure of K. This is
so, since for any vector space (in particular ideals) W over K, then B is a basis for W as a vector
space over K if, and only if, B is a basis for W ⊗K K as a vector space over K. In particular
R := K[x1 , . . . , xd ] = K ⊗K K[x1 , . . . , xd ], and in general for any ideal I ⊆ R the corresponding
ideal I of R has the form I = I ⊗K K. Hence, as we have been almost exclusively focusing on the
bases of the socle of a monomial ideal consisting of monomials alone, we will not lose any of the
combinatorial poset structure of the bases by tensoring with K, that is, assuming K is algebraically
28
closed. In that case, as stated earlier, Theorem 7.2 yields a bijective correspondence between points
of K d and maximal ideals of R.
At this point two natural questions arise:
(1) What ideals I ⊆ R have their socles Socã (I) trivial for all ã ∈ V (I)?
(2) Knowing just the socle Socã (I) of an a priori unknown ideal I of R, can we retrieve the
ideal I?
In general, the second question has a negative answer. However, in numerous specific cases the
answer is positive, for example if we know I has dimension zero, or if we know the socle to be
trivial.
We consider the first question for I = (f ) a principal ideal of R = K[x1 , . . . , xd ]. Since R is
a UFD we can write f = r1b1 · · · rhbh where each ri is irreducible in R. In this case we have the
following minimal primary decomposition of I as
I=
h
\
(ribi ),
i=1
and the unique set of associated primes {p1 , . . . , ph } where pi =
q
(ribi ) = (ri ) for each i. For
d ≥ 2 then clearly none of these principal prime ideals pi are maximal ideals of R, and hence, by
Proposition 7.4, Socã (I) must be trivial.
Observation 7.9. If I ⊆ R = K[x1 , . . . , xd ] is a principal ideal where K is algebraically closed
and d ≥ 2, then Socã (I) = I is trivial for all ã ∈ V (I).
When I is generated by two or more elements from R, things are, needless to say, more involved,
and it seems we must apply the Euclidean algorithm to obtain Gröbner bases for the ideal, in order
to conclude something fruitful – we leave that as a question in the next section.
8. Summary and further questions
We briefly discuss some of the main results in this article and post some some relevant and
motivating questions.
Presenting everything in terms of Zd or Nd0 provided with the componentwise partial order, we
showed in Theorem 3.2 and its slight generalization Theorem 3.4 that a given antichain G of Zd
(corresponding to a monomial ideal of the polynomial ring in d variables) can always be retrieved
from the set Sd (U (G∗ )) alone, or more generally from Sd (U (G∗ (ã.b̃))) as stated in Theorem 3.4.
29
As a consequence, the monomial ideal I can always be retrieved from a specific derived version of
∂oc(I) and also from a more general derived version of it socle as stated in Theorem 3.4.
The main results in Section 4 are Theorem 4.2 and Theorem 4.3 which are duals of the previously
mentioned Theorems 3.2 and 3.4. This shows that any antichain Q in Nd is indeed the socle for
a monomial ideal whose minimal generators correspond to U (Su (D(Q∗ ))) or more generally to
U (Su (D(Q∗ (ã, b̃)))) in Theorem 3.4. As a consequence, for any set of incomparable monomials
in the monoid [x1 , . . . , xd ] w.r.t. divisibility, there are monomials ideals I with the given set of
monomials as ∂oc(I).
In Section 5 we focused on Artinian monomial ideals and obtained some structure theorems for
zero dimensional local type k rings R/I where I is a monomial ideal and k = 1, 2.
In Section 6 we obtained some results for zero dimensional local type k rings R/I when k ≥ 3
and where we assumeed ∂oc(I), or rather the power points of the corresponding monomials, to be
order-generic. This yielded a specific enumeration of the minimal generators of I in Proposition 6.7.
In the last Section 7 we stated some observations about socles of general ideals of Noetherian
rings and argued that a lot of the interesting combinatorics really is captured by monomial ideals.
There are still many questions worth considering related to what has been covered.
Question 8.1. Let K be algebraically closed, and I ⊆ R = K[x1 , . . . , xd ] an ideal such that that
Socã (I) = I is trivial for all ã ∈ V (I). Can we conclude something about the Gröbner basis for I
w.r.t. some term order?
Question 8.2. Can we generalize the results to general ideals, using Gröbner bases in terms of
some term orders? If not, perhaps one can obtain similar results for toric ideals, since there all
relations can be obtained and described within the monoid [x1 , . . . , xd ] ⊆ R.
Question 8.3. When translating the results in previous sections into results for polynomial rings
R = K[x1 , . . . , xd ], or general Noetherian rings for that matter, and their ideals, can anything
be phrased simpler by using the ring structure instead or in addition to the partial order on the
monomials induced by divisibility?
References
[A00]
https://oeis.org/A000670.
[AF74]
Frank W. Anderson and Kent R. Fuller, Rings and categories of modules, Graduate Texts in Mathematics,
vol. 13, Springer-Verlag, New York-Heidelberg, 1974.
30
[Agn97] Geir Agnarsson, The number of outside corners of monomial ideals, J. Pure Appl. Algebra 117/118 (1997),
3–21.
[Agn00]
, Co-generators for algebras over fields and commutative applications, Comm. Algebra 28 (2000),
no. 9, 4071–4087.
[AL94]
William W. Adams and Philippe Loustaunau, An introduction to Gröbner bases, Graduate Studies in Mathematics, vol. 3, Amer. Math. Soc., Providence, RI, 1994.
[Alu09] Paolo Aluffi, Algebra: chapter 0, Graduate Studies in Mathematics, vol. 104, Amer. Math. Soc., Providence,
RI, 2009.
[AM69] Michael Francis Atiyah and Ian G. Macdonald, Introduction to commutative algebra, Addison-Wesley, Reading, Mass.-London-Don Mills, Ont., 1969.
[Bei93]
Mark B. Beintema, A note on Artinian Gorenstein algebras defined by monomials, Rocky Mountain J. Math.
23 (1993), no. 1, 1–3.
[BH97]
Winfried Bruns and Jürgen Herzog, Cohen-Macaulay rings, revised ed., Cambridge Studies in Advanced
Mathematics, no. 39, Cambridge Univ. Press, Cambridge, 1997.
[BPS98] Dave Bayer, Irena Peeva, and Bernd Sturmfels, Monomial resolutions, Math. Res. Lett. 5 (1998), no. 1-2,
31–46.
[BW93] Thomas Becker and Volker Weispfenning, Gröbner bases, Graduate Texts in Mathematics, vol. 141, SpringerVerlag, New York, 1993, A computational approach to commutative algebra. In cooperation with Heinz
Kredel.
[Eis95]
David Eisenbud, Commutative algebra with a view toward algebraic geometry, Graduate Texts in Mathematics, vol. 150, Springer-Verlag, New York, 1995.
[GP06]
Edgar G. Goodaire and Michael M. Parmenter, Discrete mathematics with graph theory, 3rd ed., Pearson
Prentice Hall, 2006.
[HV06] Craig Huneke and Adela Vraciu, Rings that are almost Gorenstein, Pacific J. Math. 225 (2006), no. 1,
85–102.
[Mac94] Francis Sowerby Macaulay, The algebraic theory of modular systems, Cambridge Mathematical Library,
Cambridge Univ. Press, Cambridge, 1994, Revised reprint of the 1916 original. With an introduction by
Paul Roberts.
[MS05]
Ezra Miller and Bernd Sturmfels, Combinatorial commutative algebra, Graudate Texts in Mathematics, vol.
227, Springer-Verlag, New York, 2005.
[OBn]
https://en.wikipedia.org/wiki/Ordered_Bell_number.
[Sca86] Herbert E. Scarf, Neighborhood systems for production sets with indivisibilities, Econometrica 54 (1986),
no. 3, 507–534.
[Sti]
https://en.wikipedia.org/wiki/Stirling_numbers_of_the_second_kind.
[Vil01]
Rafael H. Villarreal, Monomial algebras, Monographs and Textbooks in Pure and Applied Mathematics, vol.
238, Marcel Dekker, Inc., New York, 2001.
[Wol16] Anna-Rose G. Wolff, The survival complex, arXiv:1602.08998 [math.AC], 2016.
31
Department of Mathematical Sciences, George Mason University, Fairfax, VA 22030
E-mail address: [email protected]
Department of Mathematical Sciences, George Mason University, Fairfax, VA 22030
E-mail address: [email protected]
32
| 0math.AC
|
Minor-free graphs have light spanners
arXiv:1711.00821v1 [cs.DS] 2 Nov 2017
Glencora Borradaile
Oregon State University
[email protected]
Hung Le
Oregon State University
[email protected]
Christian Wulff-Nilsen
University of Copenhagen
[email protected]
Abstract
We show that every H-minor-free graph has a light (1+)-spanner, resolving an open problem
of Grigni and Sissokho [13] and proving a conjecture of Grigni and Hung [12]. Our lightness
bound is
σH
1
log
O
3
p
where σH = |V (H)| log |V (H)| is the sparsity coefficient of H-minor-free graphs. That is,
it has a practical dependency on the size of the minor H. Our result also implies that the
polynomial time approximation scheme (PTAS) for the Travelling Salesperson Problem (TSP)
in H-minor-free graphs by Demaine, Hajiaghayi and Kawarabayashi [7] is an efficient PTAS
1
1
whose running time is 2OH ( 4 log ) nO(1) where OH ignores dependencies on the size of H. Our
techniques significantly deviate from existing lines of research on spanners for H-minor-free
graphs, but build upon the work of Chechik and Wulff-Nilsen for spanners of general graphs [6].
1
Introduction
Peleg and Schäffer [18] introduced t-spanners of graphs as a way to sparsify graphs while approximately preserving pairwise distances between vertices. A t-spanner of a graph G is a subgraph S
of G such that dS (x, y) ≤ t · dG (x, y) for all vertices x, y 1 . Two parameters of t-spanners that are of
interest are their sparsity and lightness. The sparsity of S is the ratio of the number of edges to the
number of vertices of S. The lightness of S is the ratio of the total weight of the edges of S to the
weight of an MST of G; generally, we assume that MST(G) ⊆ S (and so MST(S) = MST(G)).
Here, we are concerned with the lightness of (1 + )-spanners, where < 1, and so we refer to
(1 + )-spanners simply as spanners.
We say that a spanner is light if the lightness does not depend on the number of vertices in the
graph. Grigni and Sissokho [13] showed that H-minor-free graphs have spanners of lightness
O 1 σH log n .
(1)
p
where σH = |V (H)| log |V (H)| is the sparsity coefficient
of H-minor-free graphs; namely that
p
an H-minor-free graph of n vertices has O(|V (H)| log |V (H)|n) edges2 [17]. Later Grigni and
1
2
We use standard graph terminology, which can be found in Appendix A.
This bound is tight [21].
1
Hung [12], in showing that graphs of bounded pathwidth have light spanners, conjectured that
H-minor-free graphs also have light spanners; that is, that the dependence on n can be removed
from the lightness above. In this paper, we resolve this conjecture positively, proving:
Theorem 1. Every H-minor-free graph G has a (1 + )-spanner of lightness
σH
1
O
log
.
3
(2)
Our algorithm consists of a reduction phase and a greedy phase. In the reduction phase, we
adopt a technique of Chechik and Wulff-Nilsen [6]: edges of the graph are subdivided and their
weights are rounded and scaled to guarantee that every MST-edge has unit weight and we include
all very low weight edges in the spanner (Appendix C). In the greedy phase, we use the standard
greedy algorithm for constructing a spanner to select edges from edges of the graph not included
in the reduction phase (Appendix B).
As a result of the reduction phase, our spanner is not the ubiquitous greedy spanner. However, since Filtser and Solomon have shown that greedy spanners are (nearly) optimal in their
lightness [10], our result implies that the greedy spanner for H-minor-free graphs is also light.
1.1
Implication: Approximating TSP
Light spanners have been used to give PTASes, and in some cases efficient PTASes, for the traveling
salesperson problem (TSP) on various classes of graphs. A PTAS, or polynomial-time approximation scheme, is an algorithm which, for a fixed error parameter , finds a solution whose value is
within 1 ± of optimal in polynomial time. A PTAS is efficient if its running time is f ()nO(1)
where f () is a function of . Rao and Smith [19] used light spanners of Euclidean graphs to give
an EPTAS for Euclidean TSP. Arora, Grigni, Karger, Klein and Woloszyn [2] used light spanners
of planar graphs, given by Althöfer, Das, Dobkin, Joseph and Soares [1], to design a PTAS for TSP
1
1
in planar graphs with running time nO( 2 ) . Klein [15] improved upon this running time to 2O( 2 ) n
by modifying the PTAS framework, using the same light spanner. Borradaile, Demaine and Tazari
generalized Klein’s EPTAS to bounded genus graphs [4].
In fact, it was in pursuit of a PTAS for TSP in H-minor-free graphs that Grigni and Sissokho
discovered the logarithmic bound on lightness (Equation (1)); however, the logarithmic bound
implies only a quasi-polynomial time approximation scheme (QPTAS) for TSP [13]. Demaine,
Hajiaghayi and Kawarabayashi [7] used Grigni and Sissokho’s spanner to give a PTAS for TSP
1
in H-minor-free graphs with running time nO(poly( )) ; that is, not an efficient PTAS. However,
Demaine, Hajiaghayi and Kawarabayashi’s PTAS is efficient if the spanner used is light. Thus, the
main result of this paper implies an efficient PTAS for TSP in H-minor-free graphs.
1.2
Techniques
In proving the lightness of spanners in planar graphs [1] and bounded genus graphs [11], the
embedding of the graph was heavily used. Thus, it is natural to expect that showing minorfree graphs have light spanners would rely on the decomposition theorem of minor-free graphs by
Robertson and Seymour [20], which shows that graphs excluding a fixed minor can be decomposed
into the clique-sum of graphs nearly embedded on surfaces of fixed genus. Borradaile and Le [5]
use this decomposition theorem to show that if graphs of bounded treewidth have light spanners,
2
then H-minor-free graphs also have light spanners. As graphs of bounded treewidth are generally
regarded as easy instances of H-minor-free graphs, it may be possible to give a simpler proof of
lightness of spanners for H-minor-free graphs using this implication.
However, relying on the Robertson and Seymour decomposition theorem generally results in
constants which are galactic in the size of the the minor [16, 14]. In this work, we take a different
approach which avoids this problem. Our method is inspired from the recent work of Chechik and
Wulff-Nilsen [6] on spanners for general graphs which uses an iterative super-clustering technique [3,
8]. Using the same technique in combination with amortized analysis, we show that H-minor-free
graphs not only have light spanners, but also that the dependency of the lightness on and |V (H)|
is practical (Equation (2)).
At a high level, our proof shares several ideas with the work of Chechik and Wulff-Nilsen [6]
who prove that (general) graphs have (2k − 1) · (1 + )-spanners with lightness O (n1/k ), removing
a factor of k/ log k from the previous best-known bound and matching Erdős’s girth conjecture [9]
up to a 1 + factor. Our work differs from Chechik and Wulff-Nilsen in two major aspects. First,
Chechik and Wulff-Nilsen reduce their problem down to a single hard case where the edges of the
graph have weight at most g k for some constant g. In our problem, we must partition the edges
according to their weight along a logarithmic scale and deal with each class of edges separately.
Second, we must employ the fact that H-minor-free graphs (and their minors) are sparse in order
to get a lightness bound that does not depend on n.
1.3
Future directions
Since we avoid relying on Robertson and Seymour’s decomposition theorem and derive bounds using
only the sparsity of graphs excluding a fixed minor, it is possible this technique could be extended to
related spanner-like constructions that are used in the design of PTASes for connectivity problems.
Except for TSP, many connectivity problems [4] have PTASes for bounded genus graphs but are not
known to have PTASes for H-minor-free graphs – for example, subset TSP and Steiner tree. The
PTASes for these problems rely on having a light subgraph that approximates the optimal solution
within 1 + (and hence is spanner-like). The construction of these subgraphs, though, rely heavily
on the embedding of the graph on a surface and since the Robertson and Seymour decomposition
gives only a weak notion of embedding for H-minor-free graphs, pushing these PTASes beyond
surface embedded-graphs does not seem likely. The work of this paper may be regarded as a first
step toward designing spanner-like graphs for problems such as subset TSP and Steiner tree that
do not rely on the embedding.
2
Bounding the lightness of a (1 + )-spanner
As we already indicated, we start with a reduction that allows us to assume that the edges of
the MST of the graph each have unit weight. (For details, see Appendix C.) For simplicity of
presentation, we will also assume that the spanner is a greedy (1 + s · ))-spanner for a sufficiently
large constant s; this does not change the asymptotics of our lightness bound.
Herein, we let S be the edges of a greedy (1 + s · )-spanner of graph G with an MST having
edges all of unit weight. We simply refer to S as the spanner. The greedy spanner considers the
edges in non-decreasing order of weights and adds an edge xy if (1 + s · )w(xy) is at most the
x-to-y distance in the current spanner (see Appendix B for a review).
3
We partition the edges of S according to their weight as it will be simpler to bound the weight
of subsets of S. Let J0 be the edges of S of weight in the range [1, 1 ); note that MST ⊆ J0 and,
since G has O(σH n) edges and w(MST) = n − 1,
σ
H
w(J0 ) = O(σH n/) = O
w(MST)
(3)
j
j+1
Let Πij be the edges of S of weight in the range [ 2i , 2 i ) for every i ∈ Z+ and j ∈ {0, 1, . . . , dlog 1 e}.
Let Jj = ∪i Πij . We will prove that
Lemma 2. There exists a set of spanner edges B such that w(B) = O( 12 w(MST)) and for every
j ∈ {0, . . . , dlog 1 e},
σ
H
w(MST).
w (MST ∪(Jj \ B)) = O
3
Combined with Equation (3), Lemma 2 gives us
dlog 1 e
w(S) = w(B) +
X
w(Jj \ B) = O
j=0
σH
1
log
3
w(MST)
which, combined with the reduction to unit-weight MST-edges, proves Theorem 1 (noting that the
stretch condition of S is satisfied since S is a greedy spanner of G).
In the remainder, we prove Lemma 2 for a fixed j ≥ 0. Let Ei = Πij for this fixed j and some
j+1
i ∈ Z+ . Let `i = 2 i ; then, the weight of the edges in Ei are in the range [`i /2, `i ). Let E0 = MST.
We refer to the indices 0, 1, 2, . . . of the edge partition as levels.
2.1
Proof overview
To prove Lemma 2,we use an amortized analysis, initially assigning each edge of E0 = MST a
credit of c = O σH3 . For each level, we partition the vertices of the spanner into clusters where
each cluster is defined by a subgraph of the graph formed by the edges in levels 0 through i. (Note
that not every edge of level 0 through i may belong to a cluster; some edges may go between
clusters.) Level i − 1 clusters are a refinements of level i clusters. We prove (by induction over the
levels), that the clusters for each level satisfy the following diameter-credit invariants:
DC1 A cluster in level i of diameter k has at least c · max{k, `2i } credits.
DC2 A cluster in level i has diameter at most g`i for some constant g > 2 (specified later).
We achieve the diameter-credit invariants for the base case (level 0) as follows. Although a
simpler proof could be given, the following method we use will be revisited in later, more complex,
constructions. Recall that E0 = MST and that, in a greedy spanner, the shortest path between
endpoints of any edge is the edge itself. If the diameter of E0 is < `0 /2 = O(1), edges in the
spanner have length at most `0 /2. Thus, it is trivial to bound the weight of all the spanner edges
across all levels using the sparsity of H-minor-free graphs. Assuming a higher diameter, let T be
a maximal collection of vertex-disjoint subtrees of E0 , each having diameter d`0 /2e (chosen, for
example, greedily). Delete T from E0 . What is left is a set of trees T 0 , each of diameter < `0 /2.
4
For each tree T ∈ T , let CT be the union of T with any neighboring trees in T 0 (connected to T
by a single edge of E0 ). By construction, CT has diameter at most 3`0 /2 + 1 ≤ 2`0 (giving DC2).
CT is assigned the credits of all the edges in the cluster each of which have credit c (giving DC1).
We build the clusters for level i from the clusters of level i − 1 in a series of four phases
(Section 3). We call the clusters of level i − 1 -clusters, since the diameter of clusters in level i − 1
are an -fraction of the diameters of clusters in level i. A cluster in level i is induced by a group of
-clusters.
We try to group the -clusters so that the diameter of the group is smaller than the sum of the
diameters of the -clusters in the group (Phases 1 to 3). This diameter reduction will give us an
excess of credit beyond what is needed to maintain DC1 which allows us to pay for the edges of Ei .
We will use the sparsity of H-minor free graphs to argue that each -cluster needs to pay for, on
average, a constant number of edges of Ei . In Phase 4, we further grow existing clusters via MST
edges and unpaid edges of Ei .
Showing that the clusters for level i satisfy invariant DC2 will be seen directly from the construction. However, satisfying invariant DC1 is trickier. Consider a path D witnessing the diameter
of a level-i cluster B. Let D be the graph obtained from D by contracting -clusters; we call D the
cluster-diameter path. The edges of D are a subset of MST ∪Ei . If D does not contain an edge of
Ei , the credits from the -clusters and MST edges of D are sufficient for satisfying invariant DC1
for B. However, since edges of Ei are not initialized with any credit, when D contains an edge
of Ei , we must use credits of the -clusters of B outside D to satisfy DC1 as well as pay for Ei .
Finally, we need to pay for edges of Ei that go between clusters. We do so in two ways. First,
some edges of Ei will be paid for by this level by using credit leftover after satisfying DC1. Second,
the remaining edges will be paid for at the end of the entire process (over all levels); we show that
there are few such edges over all levels (the edges B of Lemma 2).
In our proof below, the fixed constant g required in DC2 is roughly 100 and is sufficiently
smaller than g1 . For simplicity of presentation, we make no attempt to optimize g. We note that a
(1 + )-spanner is also a (1 + 2)-spanner for any constant and the asymptotic dependency of the
lightness on remains unchanged. That is, requiring that is sufficiently small is not a limitation
on the range of the parameter .
3
Achieving diameter-credit invariants
In this section, we construct clusters for level i that satisfy DC2 using the induction hypothesis
that -clusters (clusters of level i − 1) satisfy the diameter-credit invariants (DC1 and DC2). Since
`i−1 = `i , we let ` = `i , and drop the subscript in the remainder. For DC2, we need to group
-clusters into clusters of diameter Θ(`). Let C be the collection of -clusters and C be the set
of clusters that we construct for level i. Initially, C = ∅. We define a cluster graph K(C , Ei )
whose vertices are the -clusters and edges are the edges of Ei . K(C , Ei ) can be obtained from the
subgraph of G formed by the edges of the -clusters and Ei by contracting each -cluster to a single
vertex. Recall each -cluster is a subgraph of the graph formed by the edges in levels 0 through
i − 1.
Observation 3. K(C , Ei ) is a simple graph.
Proof. Since g` ≤ 2` when is sufficiently smaller than g1 , there are no self-loops in K(C , Ei ).
Suppose that there are parallel edges x1 y1 and x2 y2 where x1 , x2 ∈ X ∈ C and y1 , y2 ∈ Y ∈ C .
5
Let w(x2 y2 ) ≤ w(x1 y1 ), w.l.o.g.. Then, the path P consisting of the shortest x1 -to-x2 path in X,
edge x2 y2 and the shortest y1 -to-y2 path in Y has length at most w(x2 y2 ) + 2g` by DC2. Since
w(x2 y2 ) ≤ w(x1 y1 ) and w(x1 y1 ) ≥ `/2, P has length at most (1 + 4g)w(x1 y1 ). Therefore, if our
spanner is a greedy (1 + 4g)-spanner, x1 y1 would not be added to the spanner.
We call an -cluster X high-degree if its degree in the cluster graph is at least 20
, and low-degree
otherwise. For each -cluster X, we use C(X) to denote the cluster in C that contains X. To
both maintaining diameter-credit invariants and buying edges of Ei , we use credits of -clusters in
C(X) and MST edges connecting -clusters in C(X). We save credits of a subset S(X) of -clusters
of C(X) and MST-edges connecting -clusters in S(X) for maintaining invariant DC1. We then
reserve credits of another subset R(X) of -clusters to pay for edges of of Ei incident to -clusters
in S(X) ∪ R(X). We let other -clusters in C(X) \ (S(X) ∪ R(X)) release their credits to pay for
their incident edges of Ei ; we call such -clusters releasing -clusters. We designate an -cluster
in C(X) to be its center and let the center collect the credits of -clusters in R(X). The credits
collected by the center are used to pay for edges of Ei incident to non-releasing -clusters.
3.1
Phase 1: High-degree -clusters
In this phase, we group high-degree -clusters. The goal is to ensure that any edge of Ei not incident
to a low-degree -cluster has both endpoints in the new clusters formed (possibly in distinct clusters).
Then we can use sparsity of the subgraph of K(C , Ei ) induced by the -clusters that were clustered
to argue that the clusters can pay for all such edges; this is possible since this subgraph is a minor
of G. The remaining edges that have not been paid for are all incident to low-degree -clusters
which we deal with in later phases.
With all -clusters initially unmarked, we apply Step 1 until it no longer applies and then apply
Step 2 to all remaining high-degree -clusters at once and breaking ties arbitrarily:
Step 1 If there is a high-degree -cluster X such that all of its neighbor -clusters in K are unmarked,
we group X, edges in Ei incident to X and its neighboring -cluster into a new cluster C(X).
We then mark all -clusters in C(X). We call X the center -cluster of C(X).
Step 2 After Step 1, any unmarked high-degree -cluster, say Y , must have at least one marked
neighboring -cluster, say Z. We add Y and the edge of Ei between Y and Z to C(Z) and
mark Y .
In the following, the upper bound is used to guarantee DC2 and the lower bound will be used
to guarantee DC1.
Claim 4. The diameter of each cluster added in Phase 1 is at least ` and at most (4 + 5g)`.
Proof. Since the clusters formed are trees each containing at least two edges of Ei and since each
edge of Ei has weight at least `/2, the resulting clusters have diameter at least `.
b in
Consider an -cluster X that is the center of a cluster C in Step 1 that is augmented to C
b
b
Step 2 (where, possibly C = C). The upper bound on the diameter of C comes from observing
b are connected via at most 5 -clusters and via at most 4 edges of Ei
that any two vertices in C
(each -cluster that is clustered in Step 2 is the neighbor of a marked -cluster from Step 1). Since
b is
-clusters have diameter at most g` and edges of Ei have weight at most `, the diameter of C
at most (4 + 5g)`.
6
C(X)
N(X)
X
S(X)
R(X)
Releasing
Step 1
Step 2
Phase 4a
Phase 1 Phase 1 ediam < 4ℓ
Phase 4b
Phase 4b
Step 1
Step 2
Figure 1: A cluster C(X) formed in Phase 1 is enclosed in the dotted blue curve. The set S(X)
consists of five gray -clusters inside the dotted blue closed curve and S(X) consists of green-shaded
-clusters. Remaining hollow -clusters are releasing. Cluster C(X) will be augmented further in
Phase 4 and augmenting -clusters are outside the dotted blue curve. Solid blue edges are in Ei
and thin black edges are in MST. The diameter path D is highlighted by the dashed red curve and
-clusters in D are gray-shaded.
Let C(X) be a cluster in Phase 1 with the center X. Let N (X) be the set of X’s neighbors
in the cluster graph K(C , Ei ). By construction, C(X) is a tree of -clusters. Thus, at most five
-clusters in C(X) would be in the cluster-diameter path D while at most three of them are in
N (X) ∪ {X}. We use the credit of X and of two -clusters in N (X) for maintaining DC1. Let
this set of three -clusters be S(X). Since X is high-degree and < 1, N (X) \ S(X) has at least
20
18
18
− 2 ≥ -clusters. Let R(X) be any subset of -clusters in N (X) \ S(X). The center X
collects the credits of -clusters in R(X). We let other -clusters in C \ (R(X) ∪ S(X)) release
their own credits; we call such -clusters releasing -clusters. By diameter-credit invariants for level
i − 1, each -cluster has at least c`
2 credits. Thus, we have:
Observation 5. The center X of C(X) collects at least 9c` credits.
Let A1 be the set of edges of Ei that have both endpoints in marked -clusters.
Claim 6. If c = Ω σH , we can buy edges of A1 using c` credits deposited in the centers and credit
of releasing -clusters.
Proof. Since the subgraph of K induced by marked -clusters and edges of A1 is H-minor-free, each
marked -cluster, on average, is incident to at most O(σH ) edges of A1 . Thus, each -cluster must
be responsible for buying Ω(σH ) edges of A1 .
Consider a cluster C(X). The total credits of each releasing -clusters is at least c`
2 , which is
Ω(σH )` when c = Ω( σH ). For non-releasing -clusters, we use c` credits from their center X to pay
for incident edges of A1 . Recall that non-releasing -clusters are in R(X) ∪ S(X) and:
|R(X) ∪ S(X)| ≤ 5 +
7
18
(4)
X
Phase 2
Phase 4a
Phase 4b
ediam~ℓ
ediam < 4ℓ
Step 1
Figure 2: A cluster C(X) formed in Phase 2 is enclosed in the dotted blue curve. C(X) will be
augmented further in Phase 4 and augmenting -clusters are outside the dotted blue curve. Edges
connecting -clusters are MST edges.
Thus, non-releasing -cluster are responsible for paying at most O( σH ) edges of A1 and c` credits
suffice if c = Ω( σH ).
By Claim 6, each center -cluster has at least 8c` credits remaining after paying for A1 . We
note that clusters in Phase 1 could be augmented further in Phase 4. We will use these remaining
credits at the centers to pay for edges of Ei in Phase 4.
3.2
Phase 2: Low-degree, branching -clusters
Let F be a maximal forest whose nodes are the -clusters that remain unmarked after Phase 1 and
whose edges are MST edges between pairs of such -clusters.
Let diam(P) be the diameter of a path P in F, which is the diameter of the subgraph of G
formed by edges inside -clusters and MST edges connecting -clusters of P. We define the effective
diameter ediam(P) to be the sum of the diameters of the -clusters in P. Since the edges of F
have unit weight (since they are MST edges), the true diameter of a path in F is bounded by
the effective diameter of P plus the number of MST edges in the path. Since each -cluster has
diameter at least 1 (by construction of the base case), we have:
Observation 7. diam(P) ≤ 2ediam(P).
We define the effective diameter of a tree (in F) to be the maximum effective diameter over all
paths of the tree. Let T be a tree in F that is not a path and such that ediam(T ) ≥ 2`. Let X
be a branching vertex of T , i.e., a vertex of T of degree is at least 3, and let C(X) be a minimal
subtree of T that contains X and X’s neighbors and such that ediam(C(X)) ≥ 2`. We add C(X)
to C and delete C(X) from T ; this process is repeated until no such tree exists in F. We refer to
X as the center -cluster of C(X).
Claim 8. The diameter of each cluster added in Phase 2 is at most (4 + 2g)`.
Proof. Since C(X) is minimal, its effective diameter is at most 2` + g`. The claim follows from
Observation 7.
8
Let X be a set of -clusters. We define a subset of X as follows:
(
X
if |X | ≤ 2g/
bX c2g/ =
any subset of 2g/ of X otherwise
By definition, we have:
2g
(5)
Let S(X) = bC(X) ∩ Dc2g/ where D is the diameter path of C(X). We save credits of -clusters
in S(X) for maintaining DC1 and we use credits of -clusters in C \S(X) to buy edges of Ei incident
to -clusters in C(X). Since X is branching, at least one neighbor -cluster of X, say Y , is not
in S(X). Let R(X) = {Y }. The center collects credits of clusters in R(X); other -clusters in
C(X) \ {S(X) ∪ R(X)} release their credits.
Let A2 be the set of unpaid edges of Ei incident to -clusters grouped in Phase 2.
|bX c2g/ | ≤
Claim 9. If c = Ω( g3 ), we can buy edges of A2 using
the credit from releasing -clusters.
c`
6
credits from the center -clusters and half
Proof. Consider a cluster C(X) formed in Phase 2. Recall -clusters in Phase 2 are low-degree.
Thus, each -cluster in C(X) is incident to at most 20
edges of A2 . We need to argue that each
20`
`
-cluster has at least = Ω( ) credits to pay for edges of A2 . By invariant DC1 for level i − 1,
1
1
half credits of releasing -clusters are at least c`
4 , which is Ω( )` when c = Ω( 2 ).
Since |R(X)| = 1, the center X collects at least c`
2 credits by invariant DC1 for level i − 1.
Recall non-releasing -clusters are all in S(X). Thus, by Equation 5, the total number of edges of
A2 incident to -clusters in S(X) ∪ R(X) is at most:
g
2g
20
+1
=O 2
g`
Since c = Ω( g3 ), c`
6 credits of the center X is at least Ω( 2 ) which suffices to buy all edges of
A2 incident to -clusters in S(X) ∪ R(X).
We use remaining half the credit of releasing -clusters to achieve invariant DC1. More details
will be given later when we show diameter-credit invariants of C(X).
3.3
Phase 3: Grouping -clusters in high-diameter paths
In this phase, we consider components of F that are paths with high effective diameter. To that
end, we partition the components of F into HD-components (equiv. HD-paths), those with (high)
effective diameter at least 4` (which are all paths) and LD-components, those with (low) effective
diameter less that 4` (which may be paths or trees).
Phase 3a: Edges of Ei within an HD-path
Consider an HD-path P that has an edge e ∈ Ei with endpoints in -clusters X and Y of P
such that the two disjoint affices ending at X and Y both have effective diameter at least 2`. We
choose e such that there is no other edge with the same property on the X-to-Y subpath of P (By
Observation 3, there is no edge of Ei parallel to e). Let PX,Y be the X-to-Y subpath of P. By the
9
X
Y
PX
PX,Y
X
PX
PY
Y
QX
(a)
QY
PY
(b)
Figure 3: (a) A cluster of C in Case 1 of Phase 3a and (b) a cluster of C in Case 2 of Phase 3a.
Thin edges are edges of MST, solid blue edges are edges of Ei and vertices are -clusters. Edges
and vertices inside the dashed red curves are grouped into a new cluster.
stretch guarantee of the spanner, diam(PX,Y ) ≥ (1 + s)w(e). Let PX and PY be minimal subpaths
of the disjoint affices of P that end at X and Y , respectively, such that the effective diameters of
PX and PY are at least 2`. PX and PY exist by the way we choose e.
Case 1: ediam(PX,Y ) ≤ 2` We construct a new cluster consisting of (the -clusters and MST
edges of) PX,Y , PX , PY and edge e (see Figure 3(a)). We refer to, w.l.o.g, X as the center -cluster
of the new cluster.
Claim 10. The diameter of each cluster added in Case 1 of Phase 3a is at least
(12 + 4g)`.
`
2
and at most
Proof. Since the new cluster contains edge e of Ei and, in spanner S, the shortest path between
endpoints of any edge is the edge itself, we get the lower bound of the claim. The effective diameters
of PX and PY are each at most (2 + g)` since they are minimal. By Observation 7, we get that
the diameter is at most:
2(ediam(PX ) + ediam(PY ) + ediam(PX,Y )) ≤ 4(2 + g)` + 4` = (12 + 4g)`
Claim 11. Let x, y be any two vertices of G in a cluster C(X) added in Case 1 of Phase 3a. Let
Px,y be the shortest x-to-y path in C(X) as a subgraph of G. Let Px,y be obtained from Px,y by
contracting -clusters into a single vertex. Then, Px,y is a simple path.
Proof. By construction, the only cycle of -clusters in C(X) is PX,Y ∪ {e} (see Figure 3(a)). Therefore, if Px,y is not simple, e ∈ Px,y and Px,y must enter and leave PX,Y at some -cluster Z. In
this case, D could be short-cut through Z, reducing the weight of the path by at least w(e) ≥ `/2
and increasing its weight by at most diam(Z) ≤ g`. This contradicts the shortness of Px,y for
sufficiently smaller than g1 (g < 12 ).
Since PX,Y ∪ {e} is the only cycle of -clusters, by Claim 11, -clusters in D ∩ C(X) form a
simple subpath of D where D is the diameter path of C(X). We have:
Observation 12. PX,Y 6⊆ D.
Proof. For otherwise, D could be shortcut through e at a cost of
≤ diam(X) + diam(Y ) + w(e) − (diam(PX,Y ) − diam(X) − diam(Y ))
{z
} |
|
{z
}
cost of shortcut
≤ +w(e) + 4g` − (1 + s)w(e)
≤ 4g` − s`/2
lower bound on diameter
(by the stretch condition for e)
(since w(e) ≥ `/2)
10
This change in cost is negative for s ≥ 8g + 1.
Let S(X) = bD ∩ C(X)c2g/ and R(X) = bC(X) \ Dc2g/ . The center X collects the credits of
-clusters in R(X) and MST edges outside D connecting -clusters of C(X). We let other -clusters
in C(X) \ (R(X) ∪ S(X)) release their credits.
Claim 13. If D does not contain e, then X has at least
cw(e) + c`
2 credits.
c`
2
credits. Otherwise, X has at least
Proof. If C(X) \ D contains at least 2g
-clusters, then |R(X)| =
level i − 1, the total credit of -clusters in R(X) is at least:
2g c`
c`
·
= gc` ≥ c` +
2
2
≥ cw(e) +
2g
.
Thus, by invariant DC1 for
(for g ≥ 2 and < 1)
c`
2
(since w(e) ≤ `)
Thus, we can assume that C(X) \ D contains less than 2g
-clusters. In this case, R(X) = C(X) \ D.
Since PX,Y 6⊆ D by Observation 12, D does not contain, w.l.o.g., PX . Thus, R(X) contains at least
one -cluster and the claim holds for the case that e 6∈ D.
Suppose that D contains e and an internal -clusters of PX,Y , then w.l.o.g., D does not contain
PX \ X. PX \ X has credit 2c` − gc`. Since ` ≥ w(e) and ` − g` ≥ `
2 when is sufficiently small
2
( ≤ 1+2g
), the claim holds.
If D contains e but no internal -clusters of PX,Y , then
diam(PX,Y \ {X, Y })
≥ diam(PX,Y ) − diam(X) − diam(Y )
≥ (1 + s)w(e) − diam(X) − diam(Y )
(by the stretch condition)
(6)
≥ we + s`/2 − 2g` (by bounds on w(e) and DC2)
≥ we + `/2
(for s ≥ 8g + 1, as previously required)
The credit of the MST edges and -clusters of PX,Y \ {X, Y } is at least:
c · (MST(PX,Y \ {X, Y }) + ediam(PX,Y \ {X, Y }))
≥ c · diam(PX,Y \ {X, Y })
≥ c(we + `/2)
Let A3 be the set of unpaid edges of Ei incident to -clusters of clusters in Case 1 of Phase 3a.
Claim 14. If c = Ω( g3 ), we can buy edges of A3 using
releasing -clusters.
c`
6
credits from each center and credits of
Proof. Consider a cluster C(X) in Phase 3. Similar to Claim 9, releasing -clusters can pay for
their incident edges in A3 when c = Ω( 12 ). By construction, non-releasing clusters of C(X) are in
2g
S(X) ∪ R(X). Since |R(X)| ≤ 2g
and |S(X)| ≤ by Equation (5) and since clusters now we are
considering have low degree, there are at most
g
4g 20
·
=O 2
g
edges of A3 incident to non-releasing -clusters. Thus, if c = Ω( g3 ), c`
6 ≥ Ω( 2 )`. That implies
credits of X suffice to pay for all edges of A3 incident to non-releasing -clusters.
11
c`
4
Case 2: ediam(PX,Y ) > 2` Refer to Figure 3(b). Let QX and QY be minimal affices of PX,Y
such that each has effective diameter at least `. We construct a new cluster consisting of (the
-clusters and MST edges of) PX , PY , QX and QY and edge e. We refer to X as the center of the
new cluster.
We apply Case 1 to all edges of Ei satisfying the condition of Case 1 until no such edges exist.
We then apply Case 2 to all remaining edges of Ei satisfying the conditions of Case 2. After each
new cluster is created (by Case 1 or 2), we delete the -clusters in the new cluster from P, reassign
the resulting components of P to the sets of HD- and LD-components. At the end, any edge of Ei
with both endpoints in the same HD-path have both endpoints in two disjoint affixes of effective
diameter less than 2`.
We bound the diameter and credit of the centers of clusters in Case 2 of Phase 3a in Phase 3b.
Phase 3b: Edges of Ei between HD-paths
Let e be an edge of Ei that connects -cluster X of HD-path P to -cluster Y of different HD-path
Q such that none affix of effective diameter less than 2` of P contains X and none affix of effective
diameter less than 2` of Q contains Y . Such edge e is said to have both endpoints far from endpoint
-clusters of P and Q.
Let PX and QX be minimal edge-disjoint subpaths of P that end at X and each having effective
diameter at least 2`. (PX and QX exist by the way we choose edge e.) Similarly, define PY and
QY . We construct a new cluster consisting of (the -clusters and MST edges of) PX , PY , QX , QY
and edge e (see Figure 4). We refer to X as the center of the new cluster. We then delete the
-clusters in the new cluster from P and Q, reassign the resulting components of P and Q to the
sets of HD- and LD-components. We continue to create such new clusters until there are no edges
of Ei connecting HD-paths with far endpoints.
We now bound the diameter and credits of the center of a cluster, say C(X), that is formed
in Case 2 of Phase 3a or in Phase 3b. By construction in both cases, C(X) consists of two paths
PX ∪ QX and PY ∩ QY connected by an edge e.
Claim 15. The diameter of each cluster in Case 2 of Phase 3a and Phase 3b is at least
most (9 + 4g)`.
`
2
and at
Proof. The lower bound follows from the same argument as in the proof of Claim 10. Since the
effective diameters of QX and QY are smaller than the effective diameters of PX and PY , the
diameter of the new cluster is bounded by the sum of the diameters of PX and PY and w(e). The
upper bound follows from the upper bounds on these diameters as given in the proof of Claim 10.
We show how to pay for unpaid edges of Ei incident to -clusters in Case 2 of Phase 3a and
Phase 3b. W.l.o.g, we refer to X as the center -cluster of C(X). Let S(X) = bD ∩ C(X)c2g/ and
R(X) = bC(X) \ Dc2g/ where D is the cluster-diameter path of C(X). We save credits of -clusters
in S(X) for maintaining invariant DC1. The center X collects credits of -clusters in R(X). We
let other -clusters in C(X) \ (S(X) ∪ R(X)) to release their credits.
Claim 16. The center of a cluster in Case 2 of Phase 3a or in Phase 3b has at least 2c(1 − g)`
credits.
2g
Proof. If |C(X) \ D| ≥ 2g
, R(X) has -clusters which have at least gc` total credits by invariant
DC1 for level i − 1. Since gc` > 2c(1 − g)` when g > 2, the claim holds. Thus, we assume that
12
X
QX
PX
Y
PY
QY
Figure 4: A cluster of C in Phase 4a. Thin edges are edges of MST, solid blue edges are edges of
Ei and vertices are -clusters. Edges and vertices inside the dashed red curves are grouped into a
new cluster.
|C(X) \ D| < 2g
which implies R(X) = C(X) \ D. By construction, S(X) contains -clusters of
at most two of four paths PX , PY , QX , QY . Since each path has effective diameter at least `, the
-clusters of each path in R(X) have total diameter at least ` − g`. By invariant DC1 for level
i − 1, each path in R(X) has at least c(1 − g)` credits that implies the claim.
Let A4 be the set of unpaid edges of Ei incident to -clusters of clusters in Case 2 of Phase 3a
and clusters in Phase 3b.
Claim 17. If c = Ω( g2 ), we can buy edges of A4 using c(1 − 3g)` credits of the centers of clusters
in Case 2 of Phase 3a and Phase 3b and credits of releasing -clusters.
Proof. Similar to the proof of Claim 9, releasing -clusters of C(X) can buy their incident edges in A4
when c = Ω( 12 ). By construction, non-releasing -clusters are in S(X) ∪ R(X). Since |R(X)| ≤ 2g
g
and |S(X)| ≤ 2g
,
there
are
at
most
O(
)
edges
of
A
incident
to
non-releasing
-clusters.
When
4
2
g
g
c`
c`
1
is sufficiently small ( < 6g ), c(1−3g)` > 2 . Thus, if c = Ω( 2 ), 2 = Ω( 2 )` and hence, c(1−3g)`
credits suffice to pay for all edges of A4 incident to non-releasing -clusters of C(X).
3.4
Phase 4: Remaining HD-paths and LD-components
We assume that C 6= ∅ after Phase 3. The case when C = ∅ will be handled at the end of this
section.
Phase 4a: LD-components
Consider a LD-component T , that has effective diameter less than 4`. By construction, T must
have an MST edge to a cluster, say C(X), in C formed in a previous phase. We include T and an
MST edge connecting T and C(X) to C(X). Let A5 be the set of unpaid edges of Ei that incident
to -clusters merged into new clusters in this phase. We use credit of the center X and -clusters
in this phase to pay for A5 . More details will be given in Phase 4b.
Phase 4b: Remaining HD-paths
Let P be a HD-path. By construction, there is at least one MST edge connecting P to an existing
cluster in C. Let e be one of them. Greedily break P into subpaths such that each subpath has
effective diameter at least 2` and at most 4`. We call a subpath of P a long subpath if it contains
13
at least 2g
+ 1 -clusters and short subpath otherwise. We process subpaths of P in two steps. In
Step 1, we process affixes of P, long subpaths of P and the subpath of P containing an endpoint
-cluster of e. In Step 2, we process remaining subpaths of P.
Step 1 If a subpath P 0 of P contain an -cluster that is incident to e, we merge P 0 to the cluster
in C that contains another endpoint -cluster of e. We call P 0 the augmenting subpath of P. We
form a new cluster from each long subpath of P and each affix of P. It could be that one of two
affixes of P is augmenting. We repeatedly apply Step 1 for all HD-paths. The remaining cluster
paths which are short subpaths of HD-paths would be handled in Step 2. We then pay for every
unpaid edges of Ei incident to -clusters in this step. We call a cluster a long cluster if it is a long
subpath of P and a short cluster if it is a short subpath of P.
Let A6 be the set of unpaid edges of Ei incident to -clusters of long clusters. We show below
that each long cluster can both maintain diameter-credit invariant and pay for its incident edges
in A6 using credits of its -clusters.
Let A7 be the set of unpaid edges of Ei incident to remaining -clusters involved in this step;
those belong to augmenting subpaths and short affices of HD-paths. We can pay for edges of
A7 incident to -clusters in augmenting subpaths using the similar argument in previous phases.
However, we must be careful when paying for other edges of A7 that are incident to -clusters in
short affices of P. Since short affices of P spend all credits of their children -clusters to maintain
invariant DC1, we need to use credits of -clusters in P 0 to pay for edges of A7 incident to short
affices of P.
Step 2 Let P 0 be a short subpath of P. If edges of Ei incident to -clusters of P 0 are all paid, we
let P 0 become a new cluster. Suppose that -clusters in P 0 are incident to at least one unpaid edge
of Ei , say e. We have:
Observation 18. Edge e must be incident to an -cluster merged in Phase 1.
Proof. Recall that edges of Ei incident to -clusters of clusters initially formed in previous phases
except Phase 1 are in A2 ∪ . . . ∪ A7 ; thus, they are all paid. By construction, edges of Ei between
two -clusters in the same cluster initially formed in Phase 1 are in A1 ∪A5 ∪A7 which are also paid.
Since P 0 is not an affix of P, there is no unpaid edge between two -clusters of P 0 since otherwise
P 0 would become a new cluster in Phase 3a; that implies the observation.
We merge -clusters, MST edges of P 0 and e to the cluster in C that contains another endpoint
of e. This completes the clustering process. Let A8 be the set of remaining unpaid edges of Ei
incident to -clusters involved in Step 2.
We now analyze clusters of C which are formed or modified in Phase 4.
Claim 19. Let B be a short cluster. Then, diam(B) ≤ 8` and credits of -clusters and MST edges
connecting -clusters in B suffice to maintain invariant DC1 for B.
Proof. Since ediam(B) ≤ 4`, by Observation 7, diam(B) ≤ 8`. The total credit of -clusters and
MST edges in B is at least:
c(| MST(B)| + ediam(B)) ≥ c · diam(B))
Since ediam(B)) ≥ 2`, B has at least 2c` credits. Thus, B has at least c · max(diam(B), `/2)
credits.
14
We show how to pay for edges of Ei in A6 and maintain diameter-credit invariants of long
clusters. We use cr(X ) to denote the total credit of -clusters of a set of -clusters X .
Claim 20. Let B be a long cluster. If c = Ω( g3 ) and g ≥ 8, we can maintain diameter-credit
invariants of B and pay for edges in A6 incident to -clusters in B using credits of -clusters in B.
Proof. By construction, B has effective diameter at most 4`. By Observation 7, B has diameter at
most 8`. Thus, B satisfies invariant DC2 if g ≥ 8. Since B is a long cluster, it has at least 2g
+1
-clusters
in
B
and
X
be
an
-cluster
in
B
\
S.
Let
R
=
{X}.
We
-clusters. Let S be a set of 2g
save credits of S for maintaining invariant DC1 of B and use credits of R to pay for edges of A6
incident to -clusters in S ∪ R. Since |S ∪ R| = 2g
+ 1 and -clusters in S ∪ R are low-degree, there
g
are at most O( 2 ) edges of A6 incident to -clusters in S ∪ R. By invariant DC1 for level i − 1, R
g
g
has at least c`
2 credits which is sufficient to pay for O( 2 ) edges of A6 when c = Ω( 3 ). We let other
-clusters in B \ (S ∪ R) pay for their incident edges of A6 using their credits. This is sufficient
c`
when c = Ω( 12 ) since each -cluster is incident to at most 20
edges and has at least 2 credits.
We use credits of S to maintain invariant DC1. Since |S| = 2g
and each -clusters has at
least c`
credits,
cr(S)
≥
g`.
Since
diam(B)
≤
g`
by
DC2,
cr(S)
≥
cdiam(B). Thus, cr(S) ≥
2
cmax(diam(B),
˙
`/2); invariant DC1 is satisfied.
Let C(X) be a cluster in C before Phase 4. Let C 0 (X), C 00 (X) and C 000 (X) be the corresponding
clusters that are augmented from C(X) in Phase 4a, Step 1 of Phase 4b and Step 2 of Phase 4b,
respectively. It could be that any two of three clusters are the same.
By construction in Phase 4a, LD-components are attached to C(X) via MST edges. Recall each
LD-component has effective diameter at most 4` and hence, diameter at most 8` by Observation 7.
Thus, diam(C 0 (X)) − diam(C(X)) ≤ 16` + 2. By construction in Step 1 of Phase 4b, subpaths
of effective diameter at most 4b` are attached to C 0 (X) via MST edges. Thus, diam(C 00 (X)) −
diam(C 0 (X)) ≤ 16` + 2. We have:
Claim 21. diam(C 00 (X)) − diam(C(X)) ≤ 32` + 4.
By construction in Step 2 of Phase 4b, subpaths of HD-paths are attached to C 00 (X) via edges
of Ei . Since attached subpaths have effective diameter at most 4`, by Observation 7, we have:
Claim 22. diam(C 000 (X)) − diam(C 00 (X)) ≤ 18`
We are now ready to show invariant DC2 for C 000 (X).
Claim 23. diam(C 000 (X)) ≤ g` when g ≥ 70.
Proof. From Claim 4, Claim 8, Claim 10 and Claim 15, C(X) has diameter at most:
max((4 + 5g)`, (4 + 2g)`, (12 + 4g)`, (9 + 4g)`) = (12 + 4g)`
which is at most 16` when is sufficiently small ( < 1/g). By Claim 21 and Claim 22, C 000 (X) has
diameter at most:
16` + 32` + 4 + 18` = 66` + 4 ≤ 70`
since ` ≥ 1.
15
Recall we show how to pay for edges in A1 , A2 , A3 , A4 , A6 before. It remains to show how to
pay for edges in A5 ∪ A7 ∪ A8 . We first consider edges in A5 ∪ A7 . Recall S(X) = b(D ∩ C(X))c2g/
where D is the cluster-diameter path. We call -clusters in C 00 (X)\C(X) augmenting -clusters. Let
S 00 (X) = b(D ∩ (C 00 (X) \ C(X))c2g/ be the set of augmenting -clusters that are in the diameter
path D. We save credits of -clusters in S 00 (X) for maintaining DC1 and let other augmenting
-clusters release their credits.
Claim 24. If c = Ω( g2 ), we can buy edges in A5 ∪ A7 using
credits of releasing augmenting -clusters.
c`
3
credits of the cluster centers and
00
Proof. We use c`
6 credits of X to pay for edges of A5 ∪ A7 incident to -clusters in S (X). Recall
20
each -cluster is incident to at most edges of Ei since it is low-degree. Thus, -clusters in S 00 (X)
g
are incident to most O( g2 ) edges of A5 ∪ A7 . Hence, c`
6 credits suffice when c = Ω( 2 ). We let
releasing augmenting -clusters of LD-components to pay for their incident edges of A5 . This is
sufficient when c = Ω( 12 ). Thus, all edges of A5 are paid. We now turn to edges of A7 .
Let P1 , P2 , P3 be three segments of a HD-path P in Step 1 where P1 , P2 are affixes of P and
P3 is the augmenting subpath of P. It could be that P1 = P3 or P2 = P3 . Since edges of Ei
incident to long clusters are paid in Claim 20, -clusters of Pi , 1 ≤ i ≤ 2, are incident to unpaid
edges of Ei only when Pi is a short cluster and thus, incident to at most O( g2 ) edges of A7 . Note
that in Claim 19, we use all credits of -clusters and MST edges of Pi to maintain diameter-credit
invariants and we need to pay for edges of A7 incident to Pi . We consider two cases:
1. If P3 ∩ S 00 (X) = ∅, then -clusters in P3 are releasing. Recall P3 has effective diameter at
least 2`. We let each -cluster in P3 pay for its incident edges of A7 using half of its credits,
which is at least c`
4 by invariant DC1 for level i − 1. This amount of credits is enough when
1
c = Ω( 2 ). The total remaining credit from -clusters of P3 is at least c`, that is sufficient to
pay for O( g2 ) edges of A7 incident to -clusters of P1 ∪ P2 when c = Ω( g2 ).
2. If P3 ∩S 00 (X) 6= ∅, we use c`/6 credits of the center X of C 00 (X) to pay for edges of A7 incident
to -clusters in X = (P1 ∩ S 00 (X)) ∪ P2 ∪ P3 . Recall S 00 (X) has at most 2g
-clusters, X has
120g
6g
at most -clusters. Thus, -clusters in X are incident to at most 2 edges in A7 . Since
there are at most two augmenting subpaths that contain -clusters of the cluster diameter
path D, X only need to pay for at most 240g
= O( g2 ) edges. Thus, c`
6 credits are sufficient if
2
g
00
c = Ω( 2 ). Other -clusters of P3 \ S (X) are releasing and we can use their released credits
to pay for their incident edges of A7 .
We now show how to pay for edges of A8 which consists of edges of Ei incident to -clusters in
Step 2 of Phase 4b. By Observation 18, C(X) is formed in Phase 1. Let S 000 (X) be augmenting
-clusters in D of C 000 (X) that are not in S(X) ∪ S 00 (X). We save credit of S 000 (X) for maintaining
DC1 and let other augmenting -clusters release their credits.
Claim 25. If c = Ω( g2 ), we can pay for edges of A8 incident to -clusters in C 000 (X) using credits
of releasing -clusters and c` credits of the center X.
Proof. Since augmenting -clusters are low-degree, each augmenting -cluster is incident to at most
20
1
c`
edges of A8 . When c = Ω( 2 ), 2 credits of each releasing -cluster suffice to buy their incident
edges of A8 .
16
By construction, the augmenting subpath P 0 in Step 1 of Phase 4b is a short path. Since the
cluster-diameter path D contains -clusters of at most two short subpaths of HD-paths, |S 000 (X)| ≤
g
4g
. Thus, there are at most O( 2 ) edges of A8 incident to non-releasing -clusters. Hence, c` credits
of X suffice to pay for such edges when c = Ω( g2 ).
It remains to maintain invariant DC1 for clusters in C. We have:
Claim 26. If any of the sets S(X), S 00 (X) and S 000 (X) has at least
satisfies invariant DC1.
Proof. Suppose, w.l.o.g, say S(X) has at least
credits of -clusters in S(X) is at least:
2g
2g
-clusters, then C 000 (X)
-clusters. Then, by DC1 for level i − 1, the total
2g c`
·
= gc`
2
which is at least c · max(diam(C 000 (X)), `/2) since diam(C 000 (X)) ≤ g` by Claim 23 and g > 1.
Claim 27. If c = Ω( g3 ), we can maintain invariant DC1 of C 000 (X) using credits of -clusters and
MST edges in D and the credits of the cluster center X.
Proof. By Claim 26, credits of all -clusters and MST edges of D are saved for maintaining DC1.
We prove the claim by case analysis.
Case 1: C(X) is formed in Phase 1. Recall D contains at most six edges of Ei where four edges
of Ei are in C(X) and two more edges of Ei are by the augmentation in Step 2 of Phase 4b. We
use 6c` credits from X and credits of -clusters and MST edges in D. The total credit is:
6c` + c(| MST(D)| + ediam(D)) ≥ c · diam(D) = c · diam(C 000 (X)))
Since C 000 (X) contains an edge in Ei , diam(C 000 (X)) ≥ `/2. Thus, c · diam(C 000 (X))) ≥ c`/2.
To complete the proof, we need to argue that X has non-negative credits after paying for edges
of Ei and maintaining invariant DC1 of C 000 (X). Recall X initially has 9c` credits by Observation 5
and loses:
• c` credits in Claim 6.
•
c`
3
credits in Claim 24.
• c` credits to pay for the edges of A8 incident to non-releasing augmenting -clusters in Step
2 of Phase 4b.
• 6c` credits for maintaining DC1 of C 000 (X).
Thus, X still has:
9c` − 8c` −
c`
= c(1 − )`
3
3
which is non-negative since < 1.
Case 2: C(X) is formed in Phase 2. Recall the center X collects at least `
2 from a neighbor
Y of X (R(X) = {Y }). We observe that credits in X is taken totally by at most `
2 in Claim 9
17
and Claim 24. Thus, the center still has non-negative credits after buying incident edges Ei when
c = Ω( g3 ).
Since credits of -clusters and MST edges in D are reserved and D does not contain any edge
of Ei , the total reserved credit is:
c(| MST(D)| + ediam(D)) ≥ c · diam(D) = c · diam(C 00 (X)))
(7)
It remains to argue that C 00 (X) has at least c`
2 credits. Note that we do not have lower bound on
the diameter of C 00 (X) as in other cases. Let X be the set of releasing -clusters of C(X) and cr(X )
be the total credits of -clusters in X . Since ediam(C(X)) ≥ 2`, we have:
cr(X ) + cr(S(X)) ≥ 2c`
(8)
Recall half credit of X is taken in Claim 9. We use the remaining half to guarantee that the
credit of C 00 (X) is at least c`/2.
Case 3: C(X) is formed in Case 1 of Phase 3a. Recall (in Claim 13) the center X collects
at least `
2 credits if D does not contain e (we are using notation in Case 1 Phase 3a) and at
least cw(e) + `
2 credits if D contains e. We observe that credits in X is taken totally by at most
`
2 in Claim 14 and Claim 21. By construction, D contains at most one edge of Ei which is e
(in this case X has at least c · w(e) + `
2 credits). Thus, the remaining credits of X and credits
from reserved -clusters and MST edges in D are sufficient for maintaining invariant DC1. Since
diam(C 00 (X)) ≥ `/2 by Claim 10, c · diam(C 00 (X)) ≥ c`/2.
Case 4: C(X) is formed in Case 2 of Phase 3a or in Phase 3b. Recall in Claim 16, we argue
that the center of cluster X collects at least 2c(1 − g)` credits. By construction, D can contain
at most one edge of Ei , which connects two cluster paths in Case 2 of Phase 3a or Phase 3b. We
observe that credits in X is taken totally by at most c(1 − 3g − /3)` in Claim 17 and Claim 24.
Thus, X has at least:
c(2 − 2g)` − c(1 − 3g − /3)` > c`
remaining credits. That implies the remaining credits of X and credits from reserved -clusters
and MST edges in D are sufficient for maintaining invariant DC1. Since diam(C 00 (X)) ≥ `/2 by
Claim 15, c · diam(C 00 (X)) ≥ c`/2.
Proof of Lemma 2. Recall in the beginning of Phase 4, we assume that C 6= ∅ after Phase 3 and in
this case, we already paid for every edges of Ei with:
c = max(
Θ(g) Θ(σH )
σH
,
) = O( 3 )
3
and sufficiently small.
We only need to consider the case when C = ∅ after Phase 3. We have:
Observation 28. The case when C = ∅ after Phase 3 only happens when: (i) there is a single
cluster-path P that contains all -clusters, (ii) every edge of Ei is incident to an -cluster in an
affix of P of effective diameter at most 2` and (iii) -clusters of P are low-degree in K(C , Ei ).
18
We greedily break P into subpath of -clusters of effective diameter at least 2` and at most 4`
as in Phase 4b and form a new cluster from each subpath. Recall a long cluster is formed from
0
0
a subpath containing at least 2g
+ 1 -clusters. Let P be a subpath of P. If P is long, we can
0
both buy edges of Ei incident to -clusters of P and maintain two diameter-credit invariants as
in Claim 20. If P 0 is short, we use credits of and MST edges of P 0 to maintain DC1. Recall P 0
has effective diameter at least `, thus, has at least c` credits by DC1 for level i − 1. That implies
c · diam(P 0 ) ≥ c max( 2` , diam(P 0 )).
We put remaining unpaid edges of Ei to the holding bag B. Recall unpaid edges of Ei must be
incident to -clusters of short clusters, which are affixes of P. By Observation 28, B holds at most
O( g2 ) = O( 12 ) edges of Ei . Thus, the total weight of edges of B in all levels is at most:
X
X
1
1
O 2
i , where `max = maxe∈S {w(e)}
`i ≤ O 2 `max
i
i
X
1
≤ O 2 w(MST)
i
(9)
i
1
1
1
≤ O 2 w(MST)
= O 2 w(MST)
1−
Acknowledgments:
References
[1] I. Althöfer, G. Das, D. Dobkin, D. Joseph, and J. Soares. On sparse spanners of weighted
graphs. Discrete Computational Geometry, 9(1):81–100, 1993.
[2] S. Arora, M. Grigni, D. R. Karger, P. N. Klein, and A. Woloszyn. A polynomial-time approximation scheme for weighted planar graph TSP. In Proceedings of the 9th Annual ACM-SIAM
Symposium on Discrete Algorithms, SODA ’98, pages 33–41, 1998.
[3] B. Awerbuch, M. Luby, A. V. Goldberg, and S. A. Plotkin. Network decomposition and locality
in distributed computation. In Proceedings of the 30th Annual Symposium on Foundations of
Computer Science, FOCS ’89, pages 364–369, 1989.
[4] G. Borradaile, E. D. Demaine, and S. Tazari. Polynomial-time approximation schemes for
subset-connectivity problems in bounded-genus graphs. Algorithmica, 68(2):287–311, 2014.
[5] G. Borradaile and H. Le. Light spanners for bounded treewidth graphs imply light spanners
for h-minor-free graphs. CoRR, abs/1703.10633, 2017.
[6] S. Chechik and C. Wulff-Nilsen. Near-optimal light spanners. In Proceedings of the 27th Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA’16, pages 883–892, 2016.
[7] E. D. Demaine, M. Hajiaghayi, and K. Kawarabayashi. Contraction decomposition in H-minorfree graphs and algorithmic applications. In Proceedings of the 43rd Annual ACM Symposium
on Theory of Computing, STOC ’ 11, pages 441–450, 2011.
19
[8] M. Elkin and D. Peleg. (1 + , β)-spanner constructions for general graphs. SIAM Journal on
Computing, 33(3):608–631, 2004.
[9] P. Erdős. Extremal problems in graph theory. Theory of Graphs and Its Applications (Proc.
Sympos. Smolenice), pages 29–36, 1964.
[10] A. Filtser and S. Solomon. The greedy spanner is existentially optimal. In Proceedings of the
2016 ACM Symposium on Principles of Distributed Computing, PODC ’16, pages 9–17, 2016.
[11] M. Grigni. Approximate TSP in graphs with forbidden minors. In Proceedings of the 27th
International Colloquium on Automata, Languages and Programming, ICALP ’00, pages 869–
877, 2000.
[12] M. Grigni and H. Hung. Light spanners in bounded pathwidth graphs. In Proceedings of the
37th International Conference on Mathematical Foundations of Computer Science, MFCS’12,
pages 467–477, 2012.
[13] M. Grigni and P. Sissokho. Light spanners and approximate TSP in weighted graphs with
forbidden minors. In Proceedings of the 13th Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA ’02, pages 852–857, 2002.
[14] D. S. Johnson. The NP-completeness column: An ongoing guide (column 19). Journal of
Algorithms, 8(3):438–448, 1987.
[15] P. N. Klein. A linear-time approximation scheme for planar weighted TSP. In Proceedings
of the 46th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’05, pages
647–657, 2005.
[16] R. J. Lipton and K. W. Regan. People, Problems, and Proofs: Essays from Gödel’s Lost
Letter: 2010, chapter David Johnson: Galactic Algorithms, pages 109–112. Springer Berlin
Heidelberg, 2013.
[17] W. Mader. Homomorphiesätze für graphen. Mathematische Annalen, 178(2):154–168, 1968.
[18] D. Peleg and A. A. Schäffer. Graph spanners. Journal of Graph Theory, 13(1):99–116, 1989.
[19] S. B. Rao and W. D. Smith. Approximating geometrical graphs via “spanners” and “banyans”.
In Proceedings of the 30th Annual ACM Symposium on Theory of Computing, STOC ’98, pages
540–550, 1998.
[20] N. Robertson and P. D. Seymour. Graph minors. XVI. Excluding a non-planar graph. Journal
of Combinatoral Theory Series B, 89(1):43–76, 2003.
[21] A. Thomason. The extremal function for complete minors. Journal of Combinatorial Theory
Series B, 82(2):318–338, 2001.
20
A
Notation and definitions
Let G(V (G), E(G)) be a connected and undirected graph with a positive edge weight function
w : E(G) → <+ \ {0}. We denote |V (G)| and |E(G)| by n and m, respectively. Let MST(G) be
a minimum spanning tree of G; when the graph is clear from the context, we simply write MST.
A walk of length p is a sequence of alternating vertices and edges {v0 , e0 , v1 , e1 , . . . , ep−1 , vp } such
that ei = vi vi+1 for every i such that 1 ≤ 0 ≤ p − 1. A path is a simple walk where every vertex
appears exactly once in the walk. For two vertices x, y of G, we use dG (x, y) to denote the shortest
distance between x and y.
P
Let S be a subgraph of G. We define w(S) = e∈E(S) w(e).
Let X ⊆ V (G) be a set of vertices. We use G[X] to denote the subgraph of G induced by X.
Let Y ⊆ E(G) be a subset of edges of G. We denote the graph with vertex set V (G) and edge set
Y by G[Y ].
We call a graph K a minor of G if K can be obtained from G from a sequences of edge
contraction, edge deletion and vertex deletion operations. A graph G is H-minor-free if it excludes
a fixed graph H as a minor. If G excludes a fixed graph H as a minor, it also excludes the complete
h-vertex graph Kh as a minor where h = |V (H)|.
Observation 29. If a graph G excludes Kh as a minor for h ≥ 3, then any graph obtained from
G by subdividing an edge of G also excludes Kh as a minor.
Proof. We can assume that G is connected. If h = 3, G is acyclic and the observation follows easily.
Let K be the graph obtained from G by subdividing an arbitrary edge, say e, of G. Let v be the
subdividing vertex. Suppose that K contains Kh as a minor. Then there are h vertex-disjoint trees
{T1 , T2 , . . . , Th } that are subgraphs of K such that each Ti corresponds to a vertex of the minor
Kh and there is an edge connecting every two trees. We say {T1 , . . . , Th } witnesses the minor Kh
in K. If v 6∈ V (T1 ∪ . . . ∪ Th ), then {T1 , . . . , Th } witnesses Kh in G, contradicts that G excludes Kh
as a minor. Thus, we can assume, w.l.o.g, v ∈ T1 . Since h ≥ 4 and v has degree 2, T1 \ {v} 6= ∅.
By contracting v to any of its neighbors in T1 , we get a set of h trees witnessing the minor Kh in
G, contradicting that G is Kh minor-free.
B
Greedy spanners
A subgraph S of G is a (1 + )-spanner of G if V (S) = V (G) and dS (x, y) ≤ (1 + )dG (x, y) for all
x, y ∈ V (G). The following greedy algorithm by Althöfer et al. [1] finds a (1 + )-spanner of G:
GreedySpanner(G(V, E), )
S ← (V, ∅).
Sort edges of E in non-decreasing order of weights.
For each edge xy ∈ E in sorted order
if (1 + )w(xy) < dS (x, y)
E(S) ← E(S) ∪ {e}
return S
Observe that as algorithm GreedySpanner is a relaxation of Kruskal’s algorithm, MST(G) =
MST(S). Since we only consider (1 + )-spanners in this work, we simply call an (1 + )-spanners
w(S)
. We call S light if its
a spanner. We define the lightness of a spanner S to be the ratio w(MST(G))
lightness is independent of the number of vertices or edges of G.
21
C
Reduction to unit-weight MST edges
We adapt the reduction technique of Chechik and Wulff-Nilsen [6] to analyze the increase in lightness
due to this simplification for H-minor-free graphs. Let G be the input graph and let w : E(G) → <+
be the edge weight function for G. Let w̄ = w(MST)
be the average weight of the MST edges. We
n−1
do the following:
1. Round up the weight of each edge of E(G) to an integral multiple of w̄.
2. Subdivide each MST edge so that each resulting edge has weight exactly w̄. Let G0 be the
resulting graph.
3. Scale down the weight of every edge by w̄. Let w0 be the resulting edge weights of G0 . G0 is
minor-free by Observation 29.
4. Find a (1 + )-spanner S 0 of G0 .
5. Let S be a graph on V (G) with edge set equal to the union of E(S 0 ) ∩ E(G), the edges of
MST(G), and every edge e in G of weight w(e) ≤ w̄ .
Lemma 30. If S 0 is a (1 + )-spanner of G0 with lightness f (), then S is a (1 + O())-spanner of
G with lightness 2f () + O(σH /).
Proof. We adapt the proof of Chechik and Wulff-Nilsen [6].
We first bound w(S). For an edge e in E(S 0 ) ∩ E(G), w(e) ≤ w̄ · w0 (e) since weights are rounded
up before scaling down. Since G is H-minor-free, G has O(σH n) edges and so has O(σH n) edges
of weight at most w̄ . Thus, the weight of the edges returned in Step 5 is:
w(S) ≤ w̄ · w0 (S 0 ) + w(MST(G)) +
w̄
· O(σH n)
Since S 0 has lightness f () and since w(MST(G)) = (n − 1)w̄, we get
w(S) ≤ w̄ · f ()w0 (MST(G0 )) + O(σH /) · w(MST(G))
The MST of G0 is comprised of the subdivided edges (Step 2) of the MST of G. Since the weight of
each edge of MST(G) is rounded up to an integral multiple of w̄, at most (n − 1)w̄ = w(MST(G))
is added to the weight of the MST of G. Therefore w0 (MST(G0 )) ≤ 2w(MST(G))/w̄, giving
w(S) ≤ 2f () · w(MST(G)) + O(σH /) · w(MST(G))
This proves the bound on the lightness of S.
We next show that S is a (1 + O())-spanner of G. It is sufficient to show that for any edge
e∈
/ E(S) there is a path in S of weight at most (1 + O())w(e). Since S contains all edges of weight
at most w̄ , we may assume that w(e) > w̄ . Let Se0 be a path in S 0 between e’s endpoints of length
at most (1 + )w0 (e). Let Se be the path in S that naturally corresponds to the path Se0 . As above,
we have w(Se ) ≤ w̄ · w0 (Se0 ). Therefore w(Se ) ≤ (1 + )w̄ · w0 (e). Since edge weights are rounded
up by at most w̄, w̄ · w0 (e) ≤ w(e) + w̄ which in turn is ≤ w(e) + w(e) since w(e) > w̄ . We get
w(Se ) ≤ (1 + )2 w(e) = (1 + O())w(e).
By Lemma 30, we may assume that all edges of MST(G) have weight 1. We find the (1 + )spanner S of G by using the greedy algorithm. Thus, the stretch condition of S is satisfied.
22
| 8cs.DS
|
arXiv:1404.1022v1 [math.SG] 3 Apr 2014
ON COMPOSITIONS WITH x2 /(1 − x)
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
Abstract. In the past, empirical evidence has been presented that Hilbert series of symplectic quotients of unitary representations obey a certain universal
system of infinitely many constraints. Formal series with this property have
been called symplectic. Here we show that a formal power series is symplectic
if and only if it is a formal composite with the formal power series x2 /(1 − x).
Hence the set of symplectic power series forms a subalgebra of the algebra of
formal power series. The subalgebra property is translated into an identity for
the coefficients of the even Euler polynomials, which can be interpreted as a
cubic identity for the Bernoulli numbers. Furthermore we show that a rational
power series is symplectic if and only if it is invariant under the idempotent
Möbius transformation x 7→ x/(x − 1). It follows that the Hilbert series of a
graded Cohen-Macaulay algebra A is symplectic if and only if A is Gorenstein
with its a-invariant and its Krull dimension adding up to zero. It is shown
that this is the case for algebras of regular functions on symplectic quotients
of unitary representations of tori.
Contents
1. Introduction
Acknowledgements
2. Proof of Proposition 1.4
3. Proof of Theorem 1.6
4. Proof of Theorem 1.3
5. Applications to even Euler polynomials and Bernoulli Numbers
6. Sample Calculations
6.1. Quotients by finite unitary group representations
6.2. Symplectic quotients by S1
References
1
4
4
6
7
9
11
12
13
14
1. Introduction
Let G → U(V ) be a unitary representation of a compact Lie group G on a
hermitian vector space (V, h , i). Here, V is viewed as a symplectic manifold or
real variety. The R-algebra of smooth functions on V is denoted C ∞ (V ), and its
subalgebra of real regular functions is denoted R[V ]. Note that R[V ] is actually a
Poisson subalgebra of C ∞ (V ). The symplectic form on V is given by the imaginary
2010 Mathematics Subject Classification. Primary 05A15; Secondary 11B68, 13A50, 53D20.
The first and second author were supported by the grant GA CR P201/12/G028. The third
author was supported by a Rhodes College Faculty Development Grant as well as the E.C. Ellett
Professorship in Mathematics.
1
2
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
part of the scalar product h , i, and the G-action on V is Hamiltonian with moment
map
√
−1
∗
hv, ξ.vi.
J : V → g , Jξ (v) := (J(v), ξ) :=
2
Here, ξ.v := d/dtt=0 (exp(−tξ).v) denotes the infinitesimal action of ξ ∈ g on v ∈ V
and ( , ) stands for the dual pairing between the dual space g∗ and the Lie algebra
g of G.
Let us denote by Z := J −1 (0) the zero fibre of the moment map. If G is finite,
then J = 0 by convention and Z = V . Since J is G-equivariant, we can consider
the space M0 := Z/G of G-orbits in Z, the so-called symplectic quotient. It is a
stratified symplectic space and can be viewed in a natural way as a semialgebraic set
(for more information the reader may consult [10]). In order to define the smooth
structure on M0 , one introduces the vanishing ideal IZ of Z inside C ∞ (V ). Then the
algebra of smooth functions on M0 is given by C ∞ (M0 ) := C ∞ (V )G /(IZ ∩C ∞ (V )G ).
Note that C ∞ (M0 ) carries a canonical Poisson bracket. The N-graded R-algebra of
regular functions R[M0 ] := R[V ]G /(IZ ∩R[V ]G ) is a Poisson subalgebra of C ∞ (M0 ).
In this paper, we are concerned with the Hilbert series of the N-graded algebra
R[M0 ]. This is the generating function counting the dimensions dimR (R[M0 ]i ) of
the spaces of regular functions of degree i ∈ N:
X
HilbR[M0 ] (t) :=
dimR (R[M0 ]i ) ti ∈ N[[t]] ⊂ C[[t]].
i≥0
The Poisson brackets will play no role in the considerations to follow.
The main motivation for our investigation is Conjecture 1.2 below, that has been
formulated in [8]. We recall the following definition from [8].
P
Definition 1.1. For a formal power series ϕ(x) = i≥0 γi xi ∈ C[[x]] and m ≥ 1
we introduce the linear constraint
m−1
X
m−1
γm+k = 0.
(Sm )
(−1)k
k
k=0
We say that ϕ(x) is symplectic if condition (Sm ) holds for each m ≥ 1. A meromorphic function ψ(t) in the variable t is said to be symplectic at a ∈ C of pole order
d ∈ Z if the formal power series xd ψ(a − x) ∈ C[[x]] is symplectic. Here we assume
that the order of the pole of ψ(t) at t = a is ≤ d.
P
The reader is invited to check that in a symplectic power series ϕ(x) = i≥0 γi xi ∈
C[[x]] the odd coefficients γ1 , γ3 , γ5 , . . . are uniquely determined by the even ones
γ0 , γ2 , γ4 , . . . . Moreover, for each choice of the even coefficients
γ0 , γ2 , γ4 , . . . there
P
is a uniquely determined symplectic power series ϕ(x) = i≥0 γi xi .
The curious sign convention (we expand in powers of (a − x) instead of (x − a))
appears to be more natural, because in this way our typical examples render nonnegative coefficients. When we say a meromorphic function is symplectic at x = a
of order d = 0, we mean that it is analytic at x = a and symplectic as a series
expanded in (a − x). Note that we only use this sign convention for a formal power
series in the context of Lemma 3.1.
Conjecture 1.2 ([8]). Let G → U (V ) be a unitary representation of a compact
Lie group G and let R[M0 ] be the graded R-algebra of regular functions on the
ON COMPOSITIONS WITH x2 /(1 − x)
3
corresponding symplectic quotient M0 . Then HilbR[M0 ] (t) is symplectic at t = 1 of
order d = dimR (M0 ).
There is an analogue of this conjecture for cotangent lifted representations of
reductive complex Lie groups. Certainly, over the complex numbers, there exist
also symplectic quotients that arise from non-cotangent lifted representations whose
Hilbert series are symplectic. For instance, the invariant ring of any unimodular
representation of a finite group has a symplectic Hilbert series; for more details, see
Section 6. To name a specific example, for n ≥ 2 the action of the binary dihedral
group Dn ⊂ SL2 (C) on C2 cannot be cotangent lifted as there are no quadratic
invariants.
Our aim is to give a simple proof of the following statement.
Theorem 1.3. Conjecture 1.2 holds if G is a torus.
The crucial insight that helps us to achieve our goal is the following reformulation
of what it means for a generating function to be symplectic.
Proposition 1.4. A formal power series ϕ(x) is symplectic if and only if it is a
formal composite with x2 /(1 − x), i.e., if there exists a formal power series ρ(y) ∈
C[[y]] such that ϕ(x) = ρ(x2 /(1 − x)).
As a corollary, we obtain the following.
Corollary 1.5. The space of symplectic power series forms a subalgebra of C[[x]].
A meromorphic function ψ(t) is symplectic of order d at a ∈ C if and only if there
exists a formal power series ρ(y) ∈ C[[y]] such that the Laurent expansion of ψ(t) at
t = a is
(a − t)2
1
.
ρ
(a − t)d
1−a+t
If ψ1 (t) is symplectic at a ∈ C of order d1 and ψ2 (t) is symplectic at a ∈ C of order
d2 , then the product ψ1 (t)ψ2 (t) is symplectic at a ∈ C of order d1 + d2 .
It is tempting to think of x2 /(1 − x) as some sort of fundamental (rational or
formal) invariant of a group action. In fact, the requisite transformation is provided
by the order two Möbius transformation x 7→ x/(x − 1).
Theorem 1.6. A formal power series ϕ(x) is symplectic if and only if it is invariant under the substitution x 7→ x/(x − 1). If ϕ(x) is rational, then the following
statements are equivalent:
(1) ϕ(x) is symplectic,
(2) there exists a rational function ρ(y) such that ϕ(x) = ρ(x2 /(1 − x)),
(3) ϕ(x) = ϕ(x/(x − 1)).
Corollary 1.7. A rational function ψ(t) is symplectic of order d at t = a if and
only if
2
a − 2a + (1 − a)t
ψ
(1.1)
= (a − 1 − t)d ψ(t).
a−1−t
This type of functional equation one encounters in the theory of Gorenstein
algebras (cf. [2, Section 4.4]). Namely, by a theorem of Richard P. Stanley [12],
an N-graded Cohen-Macaulay
P algebra R = ⊕i≥0 Ri is Gorenstein if and only if its
Hilbert series HilbR (t) = i≥0 dim(Ri ) ti fulfills
(1.2)
HilbR (t−1 ) = (−1)d t−a(R) HilbR (t),
4
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
where d = dim(R) and a(R) is the so-called a-invariant. By comparison with (1.1)
for a = 1 we finally obtain the following result.
Corollary 1.8. The Hilbert series HilbR (t) of a graded Cohen-Macaulay algebra
R is symplectic of order d = dim R if and only if R is Gorenstein with a-invariant
a(R) = −d.
Remark 1.9. In particular, this implies that if the graded ring R = ⊕i≥0 Ri is
Gorenstein of Krull dimension d and in the Laurent expansion
X
γi
HilbR (t) =
(1.3)
,
(1 − t)d−i
i≥0
the coefficient γ1 = 0, then HilbR (t) is symplectic of order d. Here we make use of
the fact [15, Equation (3.32)] that −2γ1 /γ0 = a(R) + d.
Let us give an outline of the paper. In Section 2 we prove Proposition 1.4 and,
as a side remark, discuss relations to the sequence of Genocchi numbers. In Section
3 we use Proposition 1.4 to show Theorem 1.6. The latter is used in Section 4 to
give a proof of our main result, Theorem 1.3, that is based on Molien’s formula
and the fact that a moment map of a faithful torus representation forms a regular
sequence in the ring of invariants [7, 5]. In Section 5 we deduce from Corollary
1.5 an identity for the coefficients of the even Euler polynomials. In Section 6 we
illustrate our results by discussing specific examples.
Acknowledgements. We would like to thank Srikanth Iyengar for suggesting that
condition (Sm ) might be fulfilled termwise in the Molien formula for a finite unitary
group. The third author would like to thank Eric Gottlieb for helpful conversations.
2. Proof of Proposition 1.4
As, for each i ≥ 1, the alternating sum over the ith row of the Pascal triangle is
zero, the power series x2 /(1 − x) is symplectic. Based on this observation we are
able to find more examples.
Lemma 2.1. For each n ≥ 0 the formal power series (x2 /(1 − x))n is symplectic.
Proof. First let us observe that for a formal power series ϕ(x) we have
d2m−1
m−1
(1
−
x)
ϕ(x)
= 0.
(Sm ) ⇐⇒
d2m−1 x x=0
n
Let us introduce the shorthand notation fn,m (x) := (1 − x)m−1 x2 /(1 − x) . The
rational function fn,m (x) is regular at x = 0 and vanishes there to the order 2n.
So if m ≤ n, then
(2m−1)
fn,m
(0) = 0.
On the other hand, if m > n, then fn,m (x) is a polynomial of degree n + m − 1 <
2m − 1 and hence the (2m − 1)-fold derivative of fn,m vanishes identically.
It will be convenient to introduce some terminology.
Definition 2.2. By a symplectic basis we mean a sequence (ϕn (x))n∈N of symplectic power series ϕn (x) ∈ C[[x]] such that each ϕn (x) ∈ m2n and its class in
m2n /m2n+1 is nonzero. Here m denotes the maximal ideal x C[[x]] of the complete
local ring C[[x]].
ON COMPOSITIONS WITH x2 /(1 − x)
5
Lemma 2.3. Let (ϕn (x))n∈N be a symplectic basis. Then for each symplectic power
series ϕ(x) there exists a unique sequence (an )n∈N of numbers such that for each
k≥0
(2.1)
ϕ(x) −
It follows that ϕ(x) =
of C[[x]].
P
i≥0
k
X
i=0
ai ϕi (x) ∈ m2k+2 .
ai ϕi (x), where the sum converges in the m-adic topology
Proof.
We start with a preparatory observation. Suppose that k ≥ 0 and f (x) =
P
i
2k+1
, i.e., α0 = α1 = · · · = α2k = 0. Then (Sk+1 )
i≥0 αi x is symplectic and in m
implies that α2k+1 = 0 as well, that is f (x) ∈ m2k+2 .
Assume now for induction that
ϕ(x) −
k−1
X
i=0
ai ϕi (x) ∈ m2k .
Pk
It follows that there is a unique number ak such that ϕ(x)− i=0 ai ϕi (x) ∈ m2k+1 .
Since the latter series is symplectic, the above argument tells us that it is in fact
in m2k+2 .
As a consequence, with the choice of the symplectic basis
2 n
x
(2.2)
1−x
n∈N
we can P
write each symplectic series ϕ(x) as a formal composite ρ(x2 /(1 − x)), where
ρ(y) = i≥0 ai y i ∈ C[[y]]. This proves Proposition 1.4.
Remark 2.4. There are of course plenty of other symplectic bases. In fact, any
symplectic power series ϕ1 (x) that is in m2 and whose class in m2 /m3 does not vanish
generates a symplectic basis ((ϕ1 (x))n )n∈N . A choice different from x2 /(1 − x) is
provided by the sequence of Genocchi numbers. The sequence of Genocchi numbers
(Gn )n∈N (cf. entry A036968 in the online encyclopedia [11]) is defined by the
exponential generating function
2z
ez + 1
=
X
n≥0
Gn
zn
n!
z2
z4
3z 6 17z 8 155z 10 2073z 12
+
−
+
−
+
− . . . ∈ C[[z]].
2!
4!
6!
8!
10!
12!
P
Setting Gen(x) := i≥0 Gn+1 xn , it follows from [6] (see also Section 5) that
= z−
(2.3)
ϕ1 (x) := x2 Gen(−x)
is symplectic, and hence generates a symplectic basis as described above. As Gn =
O(n!/π n ), Gen(x) as well as ϕ1 (x) cannot be rational. Also note that the only even
monomial occurring in the expansion of ϕ1 (x) := x2 Gen(−x) is x2 .
6
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
3. Proof of Theorem 1.6
First let us prove that a formal power series ϕ(x) =
and only if
(3.1)
P
i≥0
γi xi is symplectic if
ϕ(x) = ϕ(x/(x − 1)).
The implication =⇒ is a consequence of Proposition 1.4. Conversely, let us assume
that ϕ(x) fulfills Equation (3.1). Using the identity
n X
n + i − 1 n+i
x
=
x ,
(−1)n
x−1
n−1
i≥0
for n ≥ 1, we see that Equation (3.1) is tantamount to
m
X
n m−1
(3.2)
γn
(−1)
γm =
n−1
i=1
for all m ≥ 1. Without loss of generality we can assume that γ2n = 0 for all
n ≥ 0. This can be achieved by subtracting a suitable symplectic power series.
With this assumption it follows recursively from (3.2) that γn = 0 for all n ≥ 0.
Since ϕ(x) = 0 is symplectic, this shows implication ⇐=.
This establishes the first claim of Theorem 1.6. The rest of the statement will
follow from the following.
Lemma 3.1. Let ϕ(x) = P (x)/Q(x) be a rational symplectic function. Then there
exists a rational function ρ(y) such that ϕ(x) = ρ(x2 /(1 − x)).
Proof. Assume that ϕ is nonzero, and then we may express
ϕ(x) = Cxk (x − 1)ℓ (x − 2)m
(3.3)
×
r
Y
i=1
(x − λi )ni
where C ∈ C , each λi ∈ C r {0, 1, 2}, r is a nonnegative integer, and k, ℓ, m, and
ni for i = 1, . . . , r are integers. Let q = deg(Q(x)) − deg(P (x)), and then we have
(3.4)
q = −k − ℓ − m −
r
X
ni .
i=1
By a simple computation,
ni
r
Y
x
λi
(3.5) ϕ
(1 − λi )ni x −
.
= C(−1)m xk (x − 1)q (x − 2)m
x−1
λi − 1
i=1
We have that ϕ(x) is symplectic by hypothesis so that by the substitution theorem
[1, Theorem 9.25], Equation (3.1) holds for ϕ(x). Hence, a comparison of Equations
(3.3) and (3.5) yields
(3.6)
m
(−1)
r
Y
i=1
(3.7)
(3.8)
r
Y
x−
i=1
(1 − λi )ni = 1,
λi
λi − 1
q = ℓ,
and
ni Y
r
(x − λi )ni .
=
i=1
ON COMPOSITIONS WITH x2 /(1 − x)
7
Qr
From Equation (3.8), we have that for each factor x − λi in i=1 (x − λi )ni , a factor
x − λi /(λi − 1) must also appear. Hence we may rewrite
n′j
r′
r
Y
Y
µj
n′j
ni
x−
(x − µj )
(x − λi ) =
(3.9)
µj − 1
j=1
i=1
for a nonnegative integer r′ , nonnegative integers n′j and µj ∈ C r {0, 1, 2} for
j = 1, . . . , r′ . Combining Equations (3.6) and (3.9) and observing that for each j,
(1 − µj )(1 − µj /(µj − 1)) = 1, we obtain
n′j
r′
Y
µj
n′j
m
1−
(1 − µj )
1 = (−1)
= (−1)m
µ
−
1
j
j=1
so that m = 2m′ for some m′ ∈ Z. Similarly, Equations (3.4) and (3.7) can now be
used to express
r′
X
2n′j
k = −2ℓ − 2m′ −
j=1
′
′
so that k = 2k for some k ∈ Z, and then we have
′
′
′
ℓ = −k − m −
r
X
n′j .
j=1
Substituting (3.9) into (3.3) and applying the above observations yields
n′j
r′
P ′
Y
′
µj
−k′ −m′ − rj=1 n′j
2k′
2m′
(x − µj )nj x −
ϕ(x) = Cx (x − 1)
(x − 2)
µj − 1
j=1
!n′j
k′
m′ Y
r′
µ2j
x2
x2
x2
−
−
−4
−
,
=C −
1−x
1−x
1 − x µj − 1
j=1
a rational function of x2 /(1 − x), completing the proof.
4. Proof of Theorem 1.3
In this section, we let G = Tℓ = (S1 )ℓ , let V be a unitary representation of G
with dimC V = n, and let M0 denote the corresponding symplectic quotient. We
choose a (complex) basis for V with respect to which the G-action is diagonal, and
then the action of G is described by a weight matrix A ∈ Zℓ×n . Specifically, we
let z := (z1 , . . . , zℓ ) ∈ G with each zi ∈ S1 and introduce the notation z aj :=
a
a
a
z1 1,j z2 2,j · · · zℓ ℓ,j for each j = 1, . . . , n. Then the action of z on V as a unitary
transformation is given with respect to this basis by
z 7→ diag(z a1 , . . . , z an ).
Concatenating our basis for V with its complex conjugate to produce a real basis
for V , the action of z on V as real linear transformations is given by
z 7→ diag(z a1 , . . . , z an , z −a1 , . . . , z −an ).
Let J : V → g∗ denote the homogeneous quadratic moment map, let Z :=
J (0), and let M0 := Z/G denote the symplectic quotient; see Section 1. As G is
abelian, the components of J are elements of R[V ]G . We may assume without loss
−1
8
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
of generality that 0 is in the convex hull of the columns of A in Rℓ and the rank of
A is ℓ; see [7, Section 2] or [5, Section 3].
Using Molien’s formula, see [3, Section 4.6.1], the Hilbert series of the invariant
ring R[V ]G is given by
Z
dz dz · · · dzℓ
1
Qn
Qn 1 2
.
HilbR[V ]G (t) =
ℓ
(2πi)
( j=1 zj ) j=1 (1 − tz aj )(1 − tz −aj )
z∈Tℓ
Then by [8, Proposition 2.1], the Hilbert series of the real regular functions on the
symplectic quotient M0 is given by
Z
1
(1 − t2 )ℓ dz1 dz2 · · · dzℓ
Qn
Qn
.
HilbR[M0 ] (t) =
ℓ
(2πi)
( j=1 zj ) j=1 (1 − tz aj )(1 − tz −aj )
z∈Tℓ
Define the function
(1 − t2 )ℓ
Qn
h(z, t) = Qn
,
( j=1 zj ) j=1 (1 − tz aj )(1 − tz −aj )
and then we have
(1 − t−2 )ℓ
Qn
h(z, t−1 ) = Qn
( j=1 zj ) j=1 (1 − t−1 z aj )(1 − t−1 z −aj )
t2(n−ℓ) (t2 − 1)ℓ
Qn
= Qn
( j=1 zj ) j=1 (1 − tz aj )(1 − tz −aj )
= (−1)ℓ t2(n−ℓ) h(z, t).
Fix t ∈ C with |t| < 1, and then
HilbR[M0 ] (t−1 ) =
1
(2πi)ℓ
Z
h(z, t−1 )dz1 dz2 · · · dzℓ
Z
(−1)ℓ h(z, t)dz1 dz2 · · · dzℓ .
z∈Tℓ
2(n−ℓ)
=
t
(2πi)ℓ
z∈Tℓ
Choose an i and fix arbitrary values zk ∈ S1 for k 6= i. Dividing the numerator and
a
denominator by zi i,j for each ai,j < 0 to express h(z, t) in terms of positive powers
of zi , and using the fact that each row of A contains at least one nonzero entry, it
is easy to see that
Reszi =∞ h(z, t) = − Reszi =0
1
h(z1 , . . . , 1/zi , . . . , zn , t) = 0.
zi2
A computation demonstrates that the transformation t 7→ t−1 induces a bijection
between the poles in zi inside the unit disk with those outside the unit disk. Then
considering each S1 -factor of Tℓ as a negatively-oriented curve about the point at
infinity, introducing a factor of (−1)ℓ , we have
HilbR[M0 ] (t−1 ) = t2(n−ℓ) HilbR[M0 ] (t).
Then Theorem 1.3 follows from Corollary 1.7 and the fact that R[M0 ] has dimension
d = 2(n − ℓ).
ON COMPOSITIONS WITH x2 /(1 − x)
9
5. Applications to even Euler polynomials and Bernoulli Numbers
Our aim is to derive from the fact that the space of symplectic power series forms
a subalgebra (cf. Corollary 1.5) a certain combinatorial identity, Equation (5.3),
that might be of independent interest. We recall that the Euler polynomials En (x),
n = 0, 1, 2, . . . , are defined by the expansion
X
2ext
tn
(5.1)
=
En (x) .
t
e +1
n!
n≥0
We introduce numbers
n
i
via
x(x2n − E2n (x)) =:
X n
i
i
x2i ,
observing that the even Euler polynomials, apart from their leading
monomials,
n
contain
only
monomials
that
are
odd
powers
of
x.
Note
that
are
integers
and
i
n
i = 0 for i ≤ 0 or i > n. The coefficients of the even indexed Euler polynomials
are listed in the
online encyclopedia [11] as entry A060083. The first six lines in
the triangle of ni are:
1
3
155
−2073
−17
3410
−1
28
−255
2
−5
126
−1683
3
−14
396
4
−30
5
−55
6
,
where the line and diagonal indexing starts with n = 1 (read from top to bottom)
and
i = 1 (read from left to right). We warn the reader that the recursions for the
n
i have no resemblance to those for the binomial coefficients.
P
i
Theorem 5.1. Let ϕ(x) =
i≥0 γi x be a formal power series. Then ϕ(x) is
symplectic if and only if for each n ≥ 0,
X n
γ2i .
(5.2)
γ2n+1 =
i
i
In particular, for each choice of γ0 , γ2 , γ4 , γ6 , . . . there is a uniquely defined symplectic power series ϕ(x) determined by the above rule.
Corollary 5.2. For all integers n, k, ℓ we have
X X
n r−1 i−r
n−k
n−ℓ
n
(5.3)
+
.
=
+
i
k
ℓ
ℓ
k
k+ℓ
r
i
We would like to mention that the Euler coefficients are related to the Genocchi
numbers Gn , respectively the Bernoulli numbers Bn , using the formula
G2(n−i+1)
n
4n−i+1 − 1
2n
2n
=−
=
B2(n−i+1)
.
i
2(n − i + 1) 2i − 1
n−i+1
2i − 1
This means that Equation (5.3) can be interpreted as a cubic relation for the
Bernoulli numbers.
10
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
P
P
Proof of Corollary 5.2. Let i≥0 γi xi and j≥0 δj xj be symplectic power series.
P
From Corollary 1.5 we know that their Cauchy product m≥0 ϑm xm is symplectic
P
with, for each m ∈ N, ϑm = i+j=m γi δj . The left hand side of Equation (5.3)
arises from expressing
X
ϑ2n+1 =
γr δ s
r+s=2n+1
in terms of even γ’s and δ’s using Equation (5.2). Similarly, the right hand side of
Equation (5.3) arises from expressing
n
n
X
X n
X
n
γr δ s
ϑ2i =
ϑ2n+1 =
i
i
i=1 r+s=2i
i=1
in terms of even γ’s and δ’s. In the argument, we also use the fact that the even
γ’s and δ’s can be chosen freely.
Proof of Theorem 5.1. The argument is inspired by [6, Section 7]. There the situation is studied when two sequences (cn )n∈N and (dn )n∈N are related by
n
X
n
(5.4)
ci
dn =
i
i=0
for all n. Then [6, Theorem 7.4] states that for all nonnegative integers m and n,
n
m
X
X
m
n+j n
(5.5)
dm+j .
(−1)
cn+i =
j
i
j=0
i=0
By inspection of the generating function (5.1), we derive the recursion
n
X
n
Ei (x) = 2xn .
En (x) +
i
i=0
The idea is to put ci := Ei (x) and dn := 2xn − En (x) and observe that condition
(5.4) holds. As the special case n = m ≥ 0 of (5.5), we find
n
n
X
X
n
n+i n
(−1)
(5.6)
2xn+i − En+i (x) .
En+i (x) =
i
i
i=0
i=0
This can be rewritten as
n
X
n
xn+i − En+i (x) =
i
i=0
n+i even
n
X
n n+i
x ,
i
i=0
n+i odd
which is equivalent to
X n+i
n
n
X
X
n
n n+i
2
(−1)i
(5.7)
x
= 0.
(−1)i
x2j−1 +
i
i
j
j
i=0
i=0
n+i even
n+i odd
Let now (γnP
)n∈N be a number sequence such that (5.2) holds. It will be enough
to show that i≥0 γi xi is symplectic. We interpret x as an umbral variable (cf.
for example [6]) and define the functional Γ : C[x] → C by Γ(xn ) = γn+1 for n
ON COMPOSITIONS WITH x2 /(1 − x)
11
odd. For even n, we put Γ(xn ) = 0 (this choice will not affect the considerations
to follow). Applying Γ to Equation (5.7), we end up with
X n+i
n
n
X
X
n
n
2
γ2j +
γn+i+1 = 0,
(−1)i
(−1)i
j
i
i
i=0
j
i=0
n+i even
|
{z
} n+i odd
=γn+i+1
showing that
γn+i+1 = 0, i.e., condition (Sn+1 ) of Definition 1.1.
Pn
i n
i=0 (−1) i
To complete this section, we use the above observations to indicate an alternate
symplectic basis than those considered in Section 2.
Lemma 5.3. For all integers k ≥ 0, the formal power series
∞
∞
X
X
1
i 2i+1
(2k−1)
(−1)i−1 Ei−1 (0) xi = −x2k −
(5.8) ψk (x) :=
x
(2k − 1)! i=0
k
i=k
is symplectic.
Proof. Using Equation (5.6), we write
n
n
n
X
1X
1X n
n
n+i n
(−1)
2xn+i − En+i (x)
En+i (x) +
En+i (x) =
i
2
i
2
i
i=0
i=0
i=0
n
n
X
1X
n
n
n+i
=
(−x)n+i
En+i (x) +
1 − (−1)
i
2 i=0
i
i=0
n
n
X
X
n
n n+i
x ,
En+i (x) − xn+i +
=
i
i
|
{z
}
i=0
i=0
n+i odd
n+i even
=: (∗)
where (∗) contains only even powers of x. Thus in i ni En+i (x), all coefficients
of odd degree vanish, meaning that for all k ≥ 1,
n
X
1
n
(2k−1)
En+i (0) = 0.
(2k − 1)! i=0 i
P
It follows that
∞
P
i=0
(2k−1)
(−1)i−1 En−1
(0) xi fulfills (Sn+1 ).
As a consequence of Lemma 5.3, we have that (ψn (x))n∈N forms a symplectic
basis in the sense of Definition 2.2. Note that ψ1 is essentially the generating
function of the Genocchi sequence (2.3), namely we have ψ1 (x) = −ϕ1 (x). The
idea of the above proof can be used to argue that for each λ ∈ C the power series
X
(−1)i−1 (Ei−1 (λ) − Ei−1 (−λ)) xi ∈ C[[x]]
i≥0
is symplectic.
6. Sample Calculations
In this section, we survey a few special cases of Conjecture 1.2 and Theorem 1.3
that can be verified by direct computations. We first consider the case of a unitary
representation of a finite group. In this case, as a consequence of Corollary 1.7,
Conjecture 1.2 is a special case of Watanabe’s Theorem [16, 17], see in particular
[13, Theorem 7.1].
12
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
6.1. Quotients by finite unitary group representations. Let G be a finite
group and G → U(V ) a unitary representation. For g ∈ G, we let gV : V → V
denote the corresponding linear transformation. Let W := V × V , and then G
acts on W via gW : (u, v) 7→ (gV u, (gV−1 )t v). We identify R[V ] with the subring
C[W ]− of C[W ] given by those elements fixed by complex conjugation, and then
by Molien’s formula [9], see also [3, 14], the Hilbert series of real regular invariants
is given by
(6.1)
HilbR[V ]G |R (t) =
1 X
1
−1 .
|G|
det(id
−gW
t)
W
g∈G
Fix g ∈ G and choose a basis for V with respect to which gV is diagonal, say
gV = diag(λ1 , . . . , λn ) where |λi | = 1 for each i. Choosing the conjugate basis for V
and concatenating to form a basis for W , we have gW = diag(λ1 , . . . , λn , λ1 , . . . , λn ).
It then follows that
(6.2)
1
−1
det(idW −gW
t)
=
n
Y
1
.
(1
−
λ
t)(1
− λ−1
i
i t)
i=1
In this case, each term of the sum in Equation (6.1) is symplectic of order 2n at
t = 1. Specifically, for λ ∈ C, define
fλ (t) :=
1
.
(1 − λt)(1 − λ−1 t)
Then by the above observations, we have that each term in Molien’s formula is
given by a product of fλi (t). We claim the following.
Lemma 6.1. For nonzero λ ∈ C, the function fλ (t) is symplectic at t = 1 of order
2.
Proof. If λ = 1, then f1 (t) = (1 − t)−2 and the result is trivial, so assume not.
Define
λy
,
ρλ (y) =
1 + 2λ + λ2 + λy
and then by a simple computation,
1
(1 − t)2
= fλ (t).
ρλ
(1 − t)2
t
The result then follows from Corollary 1.5.
We remark that Lemma 6.1 can also be seen using the expansion
fλ (t) =
∞
X
−λ(λk+1 − (−1)k+1 )
(1 − t)k
(λ2 − 1)(λ − 1)k+1
k=−2
and verifying (Sm ) directly, or by checking that fλ (t) satisfies (1.1) for a = 1 and
d = 2.
By Lemma 6.1 and Corollary 1.5, it follows that the expression in Equation (6.2)
is symplectic of order 2n at t = 1.
ON COMPOSITIONS WITH x2 /(1 − x)
13
6.2. Symplectic quotients by S1 . We observe that Corollary 1.5 (in the case
a = 1) is a sufficient tool to verify Conjecture 1.2 for symplectic S1 -quotients caseby-case. When the action is generic, i.e. no two weights have the same absolute
value, an algorithm for computing the Hilbert series is described in [8, Section 4]
and has been implemented on Mathematica [18]. To check that a concrete Hilbert
series is symplectic of order d = dim(M0 ), we use the substitution
t 7→
p
1
y + 2 ± y(y + 4)
2
as a heuristic to find a rational function ρ(y) as in Corollary 1.5.
As an example, if M0 is a reduced space associated to the weight vector (±1, ±2, ±3)
where not all weights have the same sign, then
HilbR[M0 ] (t) =
t10 + t8 + 3t7 + 4t6 + 4t5 + 4t4 + 3t3 + t2 + 1
.
(1 − t2 )(1 − t3 )(1 − t4 )(1 − t5 )
One easily checks that
ρ(y) =
y 5 + 10y 4 + 36y 3 + 59y 2 + 50y + 22
(y + 2)(y + 3)(y + 4)(y 2 + 5y + 5)
satisfies the condition of Corollary 1.5.
Similarly, let S1 act on Cn with weight vector (±1, . . . , ±1) and let M0 denote
the symplectic quotient, which we note has dimension 2n − 2. We assume that
n ≥ 2 and not all weights have the same sign, for otherwise M0 is a point and the
result is trivial. By [8, Section 5.3], the Hilbert series R[M0 ] is given by
n−1
X n − 12
1
HilbR[M0 ] (t) = (1 − t ) 2 F1 (n, n, 1, t ) =
t2k
(1 − t2 )2n−2
k
2
2
k=0
where 2 F1 denotes the hypergeometric function; see [4]. Theorem 1.3 can be verified
directly using Corollary 1.5 for this case as follows.
Define
n−1
X n − 12k
1
ρ(y) =
y n−k−1 .
(y + 4)n−1
k
k
k=0
Let (a)b denote the Pochhammer symbol, i.e. (a)b := a(a + 1) · · · (a + b − 1) for
b > 0 and (a)0 = 1, and note that (1/2)k = (2k)!/(4k k!). Applying the definition
of 2 F1 , we compute
k
n−1
X n − 12k
1
1
(1 − t)2
t
=
ρ
(1 − t)2n−2
t
(1 + t)2n−2
(1 − t)2
k
k
k=0
n−1
X (1 − n)k (1/2)k −4t k
1
=
(1 + t)2n−2
(1)k k!
(1 − t)2
k=0
1
2
=
2 F1 (1 − n, 1/2, 1, −4t/(1 − t) ).
(1 + t)2n−2
14
HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
We then apply [4, page 113, Equation (36)] and continue
=
n−1
X (1 − n)k (1 − n)k
1
1
2
F
(1
−
n,
1
−
n,
1,
t
)
=
t2k
2
1
2
2n−2
2
2n−2
(1 − t )
(1 − t )
(1)k k!
k=0
n−1
X n − 12
1
=
t2k
(1 − t2 )2n−2
k
k=0
= HilbR[M0 ] (t).
References
1. Tom M. Apostol, Mathematical analysis, second ed., Addison-Wesley Publishing Co., Reading,
Mass.-London-Don Mills, Ont., 1974.
2. Winfried Bruns and Jürgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced
Mathematics, vol. 39, Cambridge University Press, Cambridge, 1993.
3. Harm Derksen and Gregor Kemper, Computational invariant theory, Invariant Theory and
Algebraic Transformation Groups, I, Springer-Verlag, Berlin, 2002, Encyclopaedia of Mathematical Sciences, 130.
4. Arthur Erdélyi, Wilhelm Magnus, Fritz Oberhettinger, and Francesco G. Tricomi, Higher
transcendental functions. Vols. I, II, McGraw-Hill Book Company, Inc., New York-TorontoLondon, 1953, Based, in part, on notes left by Harry Bateman.
5. Carla Farsi, Hans-Christian Herbig, and Christopher Seaton, On orbifold criteria for symplectic toric quotients, SIGMA Symmetry Integrability Geom. Methods Appl. 9 (2013), Paper
032, 18.
6. Ira M. Gessel, Applications of the classical umbral calculus, Algebra Universalis 49 (2003),
no. 4, 397–434, Dedicated to the memory of Gian-Carlo Rota.
7. Hans-Christian Herbig, Srikanth B. Iyengar, and Markus J. Pflaum, On the existence of star
products on quotient spaces of linear Hamiltonian torus actions, Lett. Math. Phys. 89 (2009),
no. 2, 101–113.
8. Hans-Christian Herbig and Christopher Seaton, The Hilbert series of a linear symplectic circle
quotient, to appear in Experimental Mathematics (2013), arXiv:1302.2662 [math.SG].
9. Th. Molien, Über die Invarianten der linearen Substitutionsgruppen, Sitzungsber. der Königl.
Preuss. Akad. d. Wiss. zweiter Halbband (1897), 1152–1156.
10. Reyer Sjamaar and Eugene Lerman, Stratified symplectic spaces and reduction, Ann. of Math.
(2) 134 (1991), no. 2, 375–422.
11. N. J. A. Sloane, Online Encyclopaedia of Integer Sequences, https://oeis.org (cited February
2014).
12. Richard P. Stanley, Hilbert functions of graded algebras, Advances in Math. 28 (1978), no. 1,
57–83.
, Invariants of finite groups and their applications to combinatorics, Bull. Amer. Math.
13.
Soc. (N.S.) 1 (1979), no. 3, 475–511.
14. Bernd Sturmfels, Algorithms in invariant theory, Texts and Monographs in Symbolic Computation, Springer-Verlag, Vienna, 1993.
15. È. B. Vinberg and V. L. Popov, Invariant theory, Algebraic geometry. IV, Encyclopaedia of
Mathematical Sciences, vol. 55, Springer-Verlag, Berlin, 1994, Linear algebraic groups. Invariant theory, A translation of ıt Algebraic geometry. 4 (Russian), Akad. Nauk SSSR Vsesoyuz.
Inst. Nauchn. i Tekhn. Inform., Moscow, 1989 [ MR1100483 (91k:14001)], Translation edited
by A. N. Parshin and I. R. Shafarevich, pp. vi+284.
16. Keiichi Watanabe, Certain invariant subrings are Gorenstein. I, Osaka J. Math. 11 (1974),
1–8.
, Certain invariant subrings are Gorenstein. II, Osaka J. Math. 11 (1974), 379–388.
17.
18. Wolfram
Research,
Mathematica
edition:
Version
7.0,
(2008),
http://www.wolfram.com/mathematica/.
ON COMPOSITIONS WITH x2 /(1 − x)
15
Charles University in Prague, Faculty of Mathematics and Physics, Sokolovská 83,
186 75 Praha 8, Czech Republic
E-mail address: [email protected]
University of Duisburg-Essen, Department of Mathematics, Campus Essen, 45117 Essen, Germany
E-mail address: [email protected]
Department of Mathematics and Computer Science, Rhodes College, 2000 N. Parkway, Memphis, TN 38112
E-mail address: [email protected]
| 0math.AC
|
Electronic Letters on Computer Vision and Image Analysis 16(2):17-20, 2017
Learning audio and image representations with
bio-inspired trainable feature extractors
Nicola Strisciuglio∗
∗
Johann Bernoulli Institute for Mathematics and Computer Science, University of Groningen, Netherlands
arXiv:1801.00688v1 [cs.CV] 2 Jan 2018
This work was carried out at the University of Groningen (Netherlands) and at the University of Salerno (Italy)
Received 24th Jul 2017; accepted 24th Nov 2017
Abstract
Recent advancements in pattern recognition and signal processing concern the automatic learning of
data representations from labeled training samples. Typical approaches are based on deep learning and
convolutional neural networks, which require large amount of labeled training samples. In this work, we
propose novel feature extractors that can be used to learn the representation of single prototype samples in
an automatic configuration process. We employ the proposed feature extractors in applications of audio and
image processing, and show their effectiveness on benchmark data sets.
1 Introduction
Since when very young, we can quickly learn new concepts, and distinguish between different kinds of object
or sound. If we see a single object or hear a particular sound, we are then able to recognize such sample or even
different versions of it in other scenarios. For example, if one sees an iron chair and associates the object to the
general concept of “chairs”, he will be able to detect and recognize also wooden or wicker chairs. Similarly,
when we hear the sound of a particular event, such as a scream, we are then able to recognize other kinds of
scream that occur in different environments. We continuously learn representations of the real world, which we
then use in order to understand new and changing environments.
In the field of pattern recognition, traditional methods are typically based on representations of the real world
that require a careful design of a suitable feature set (i.e. a data representation), which involves considerable
domain knowledge and effort by experts. Recently, approaches for automated learning of representations from
training data were introduced. Representation learning aims at avoiding engineering of hand-crafted features
and providing automatically learned features suitable for the recognition tasks. Nowadays, widely popular
approaches for representation learning are based on deep learning techniques and convolutional neural networks
(CNN). These techniques are very powerful, but are computationally expensive and require large amount of
labeled training data to learn effective models for the applications at hand.
In this paper we report the main achievements included in the doctoral thesis titled ‘Bio-inspired algorithms
for pattern recognition in audio and image processing’, in which we proposed novel approaches for representation learning for audio and image signals [5].
Correspondence to: Nicola Strisciuglio <[email protected]>
Recommended for acceptance by Anjan Dutta and Carles Sánchez
http://dx.doi.org/10.5565/rev/elcvia.1128
ELCVIA ISSN:1577-5097
Published by Computer Vision Center / Universitat Autònoma de Barcelona, Barcelona, Spain
18
Electronic Letters on Computer Vision and Image Analysis 16(2):17-20, 2017
feature
learning
input
data
pre-processing
feature
extraction
feature
selection
model
training
classification/rejection
decision
Figure 1: Overview of a pattern recognition system. The input data are pre-processed and then features are
computed to extract important properties from such data. The features to be computed can be determined by
an engineering process or can be learned from the data (representation learning). Feature selection procedures
can be employed to determine a sub set of discriminant features that are then used to train a classifier, which
determines a model of the training data. Such model is then used in the operating phase of the system, while a
classifier takes decisions on the input data.
2 Motivation and contribution
Motivated by the fact that we can learn effective representations of a new category of objects or sounds from
a single example and successively generalize to a wide range of real-world samples, we studied the possibility
of learning data representations from small amounts of training samples. We investigated the design of feature
extractors that can be automatically trained by showing single prototype samples, and employed them into
pattern recognition systems to solve practical problems.
We proposed representation learning techniques for audio and image processing based on novel trainable
feature extractors. The design and implementation of the proposed feature extractors are inspired by some
functions of the human auditory and visual systems. The structure of the proposed feature extractors is learned
from training samples in an automatic configuration step, rather than fixed a-priori in the implementation [5].
We employed the newly designed methodologies into systems for audio event detection and classification in
noisy environments and for delineation of blood vessels in retinal fundus images. The contributions of this
work can be listed as: a) novel bio-inspired trainable feature extractors for representation learning of audio
and image signals, respectively called COPE and B-COSFIRE; b) a system for audio event detection based on
COPE feature extractors; c) the release of two data sets of audio events of interest mixed to various background
sounds and with different signal to noise ratios (SNR); d) a method for delineation of elongated and curvilinear
patterns in images based on B-COSFIRE filters; e) feature selection mechanisms based on information theory
and machine learning approaches.
3 Methods
We introduced a novel approach for representation learning, based on trainable feature extractors. We extended
the traditional scheme of pattern recognition systems with feature learning algorithms (dashed box at the top of
Figure 1), which construct a suitable representation of training data by automatic configuring a set of feature
extractors.
Electronic Letters on Computer Vision and Image Analysis 16(2):17-20, 2017
19
We proposed trainable COPE (Combination of Peaks of Energy) feature extractors for sound analysis, that
can be trained to detect any sound pattern of interest. In an automatic configuration process performed on a
single prototype sound pattern, the structure of a COPE feature extractor is learned by modeling the constellation of peak points in a time-frequency representation of the input sound [10]. In the application phase, a
COPE feature has high value when computed on the same sound used for configuration, but also to similar or
corrupted versions of it due to noise or distortion. This accounts for generalization capabilities and robustness
of detection of the patterns of interest. The response of a COPE feature extractor is computed as the combination of the weighted score of its constituent constellation of energy peaks. For further detail we refer the
reader to [10]. For the design of COPE feature extractors, we were inspired by some functions of the cochlea
membrane and the inner hair cells in the auditory system, which convert the sound pressure waves into neural
stimuli on the auditory nerve. We employed COPE feature extractors together with a multi-class Support Vector
Machine (SVM) classifier to perform audio event detection and classification, also in cases where sounds have
null or negative SNR.
We proposed B-COSFIRE (that stands for Bar-selective Combination of Shifted Filter Responses) filters for
detection of elongated and curvilinear patterns in images and apply them to the delineation of blood vessels in
retinal images [1, 8]. The B-COSFIRE filters are trainable, that is their structure is automatically configured
from prototype elongated patterns. The design of the B-COSFIRE filters is inspired by the functions of some
neurons, called simple cells, in area V1 of the visual system, which fire when presented with line or contour
stimuli. A B-COSFIRE filter achieves orientation selectivity by computing the weighted geometric mean of the
output of a pool of Difference-of-Gaussians (DoG) filters, whose supports are aligned in a collinear manner.
Rotation invariance is efficiently obtained by appropriate shiftings of the DoG filter responses. For further
detail we refer the reader to [1].
After configuring a large bank of B-COSFIRE filters selective for vessels (i.e. lines) and vessel-endings (i.e.
line-endings) of various thickness (i.e. scale), we proposed to use several approaches based on information
theory and machine learning to select an optimal subset of B-COSFIRE filters for the vessel delineation task [7,
9]. We indicate this procedure with the dashed box named ‘feature learning’ in Figure 1. We consider the
selected filters as feature extractors to construct a pixel-wise feature vector which we used in combination with
a SVM classifier to classify the pixels in the testing image as vessel and non-vessel pixels.
4 Experiments and Results
We released two data sets for benchmark of audio event detection and classification methods, namely the
MIVIA audio events [3] and the MIVIA road events [2] data sets. We reported baseline results (recognition
rate of 86.7% and 82% on the two data sets) by using a real-time method for event detection that is based on
an adaptation of the bag of features classification scheme to noisy audio streams [3, 2]. The results that we
achieved by using COPE feature extractors show a considerable improvement with respect to the ones of the
bag of features approach. We obtained a recognition rate of 95.38% on the MIVIA audio event and of 94%
(with standard deviation on cross-validation experiments equal to 4.32) on the MIVIA road event data sets.
We performed t-Student tests and observed a statistically significant improvement of the recognition rate with
respect to baseline performance on both data sets.
We evaluated the performance of the proposed B-COSFIRE filters on four data sets of retinal fundus images
for benchmarking of blood vessel segmentation algorithms, namely the DRIVE, STARE, CHASE DB1 and
HRF data sets. The results that we achieved (DRIVE: Se = 0.7655, Sp = 0.9704; STARE: Se = 0.7716,
Sp = 0.9701; CHASE DB1: Se = 0.7585, Sp = 0.9587; HRF: Se = 0.7511, Sp = 0.9745) are higher than the ones
reported by many state-of-the-art methods based on filtering approaches. The filter selection procedure based
on supervised learning that we proposed in [9] contributes to a statistically significant increase of performance
results, which are higher than or comparable to the ones of other methods based on machine learning techniques.
We extended the application range of the B-COSFIRE filters to aerial images for the delineation of roads and
rivers, natural and textured images [4], and to pavement and road surface images for the detection of cracks and
20
Electronic Letters on Computer Vision and Image Analysis 16(2):17-20, 2017
damages [6]. The results that we achieved are better than or comparable to the ones achieved by existing methods, which are usually designed to solve specific problems. The proposed B-COSFIRE filters demonstrated
to be effective in various applications and with different types of images (retinal fundus photography, aerial
photography, laser scans) for delineation of elongated and curvilinear patterns.
We studied the computational requirements of the proposed algorithms in order to evaluate their applicability
in real-world applications and the fulfillment of real-time constraints given by the considered problems. The
MATLAB implementation of the proposed algorithms are publicly released for research purposes* .
5 Conclusions
In this work, we proposed novel trainable feature extractors and employed them in applications of sound and
image processing. The trainable character of the proposed feature extractors is in that their structure is learned
directly from training data in an automatic configuration process, rather the fied in the implementation. This
provides flexibility and adaptability of the proposed methods to different applications. The experimental results
that we achieved, compared to the ones of other existing approaches, demonstrated the effectiveness of the
proposed methods in various applications.
This work contributes to the development of techniques for representation learning in audio and image
processing, suitable for domains where there is no availability of large amount of labeled training data.
References
[1] Azzopardi, G., Strisciuglio, N., Vento, M., Petkov, N.: Trainable COSFIRE filters for vessel delineation
with application to retinal images. Medical Image Analysis 19(1), 46 – 57 (2015)
[2] Foggia, P., Petkov, N., Saggese, A., Strisciuglio, N., Vento, M.: Audio surveillance of roads: A system
for detecting anomalous sounds. IEEE Trans. Intell. Transp. Syst. 17(1), 279–288 (2016)
[3] Foggia, P., Petkov, N., Saggese, A., Strisciuglio, N., Vento, M.: Reliable detection of audio events in
highly noisy environments. Pattern Recogn. Lett. 65, 22 – 28 (2015)
[4] Strisciuglio, N., Petkov, N.: Delineation of line patterns in images using b-cosfire filters. In: 2017 International Conference and Workshop on Bioinspired Intelligence (IWOBI). pp. 1–6 (July 2017)
[5] Strisciuglio, N.: Bio-inspired algorithms for pattern recognition in audio and image processing. University
of Groningen (2016), http://www.cs.rug.nl/~nick/strisciuglio phd.pdf
[6] Strisciuglio, N., Azzopardi, G., Petkov, N.: Detection of curved lines with b-cosfire filters: A case study
on crack delineation. In: CAIP 2017, pp. 108–120 (2017)
[7] Strisciuglio, N., Azzopardi, G., Vento, M., Petkov, N.: Multiscale blood vessel delineation using BCOSFIRE filters. In: CAIP, LNCS, vol. 9257, pp. 300–312 (2015)
[8] Strisciuglio, N., Azzopardi, G., Vento, M., Petkov, N.: Unsupervised delineation of the vessel tree in
retinal fundus images. In: VIPIMAGE, pp. 149–155 (2015)
[9] Strisciuglio, N., Azzopardi, G., Vento, M., Petkov, N.: Supervised vessel delineation in retinal fundus
images with the automatic selection of B-COSFIRE filters. Mach. Vis. Appl. pp. 1–13 (2016)
[10] Strisciuglio, N., Vento, M., Petkov, N.: Bio-inspired filters for audio analysis. In: BrainComp 2015,
Revised Selected Papers. pp. 101–115 (2016)
* The
code is available from the GitLab repositories at http://gitlab.com/nicstrisc
| 1cs.CV
|
Verified Parallel String Matching in Haskell
Niki Vazou1,2 and Jeff Polakow2
arXiv:1610.07118v1 [cs.PL] 23 Oct 2016
1
UC San Diego
2
Awake Networks
Abstract. In this paper, we prove correctness of parallelizing a string
matcher using Haskell as a theorem prover. We use refinement types
to specify correctness properties, Haskell terms to express proofs and
Liquid Haskell to check correctness of proofs. First, we specify and prove
that a class of monoid morphisms can be parallelized via parallel monoid
concatenation. Then, we encode string matching as a morphism to get a
provably correct parallel transformation. Our 1839LoC prototype proof
shows that Liquid Haskell can be used as a fully expressive theorem
prover on realistic Haskell implementations.
1
Introduction
In this paper, we prove correctness of parallelization of a naı̈ve string matcher
using Haskell as a theorem prover. We use refinement types to specify correctness properties, Haskell terms to express proofs and Liquid Haskell to check
correctness of proofs.
Optimization of sequential functions via parallelization is a well studied technique [14,4]. Paper and pencil proofs of program have been developed to support
the correctness of the transformation [6]. However, these paper and pencil proofs
show correctness of the parallelization algorithm and do not reason about the
actual implementation that may end up being buggy.
Dependent Type Systems (like Coq [3] and Adga [19] ) enable program equivalence proofs for the actual implementation of the functions to be parallelized.
For example, SyDPaCC [18] is a Coq extension that given a naı̈ve Coq implementation of a function, returns an Ocaml parallelized version with a proof of
program equivalence. The limitation of this approach is that the initial function should be implemented in the specific dependent type framework and thus
cannot use features and libraries from one’s favorite programming language.
Refinement Types [8,12,24] on the other hand, enable verification of existing general purpose languages (including ML [34,2,22], C [7,23], Haskell [31],
Racket [15] and Scala [26]). Traditionally, refinement types are limited to “shallow” specifications, that is, they are used to specify and verify properties that
only talk about behaviors of program functions but not functions themselves.
This restriction critically limits the expressiveness of the specifications but allows for automatic SMT [1] based verification. Yet, program equivalence proofs
were out of the expressive power of refinement types.
Recently, we extended refinement types with Refinement Reflection [32], a
technique that reflects each function’s implementation into the function’s type
and is implemented in Liquid Haskell [31]. We claimed that Refinement Reflection can turn any programming language into a proof assistant. In this paper
we check our claim and use Liquid Haskell to prove program equivalence. Specifically, we define in Haskell a sequential string matching function, toSM, and
its parallelization, toSMPar, using existing Haskell libraries; then, we prove in
Haskell that these two functions are equivalent, and we check our proofs using
Liquid Haskell.
Theorems as Refinement Types Refinement Types refine types with properties
drawn from decidable logics. For example, the type {v:Int | 0 < v} describes
all integer values v that are greater than 0. We refine the unit type to express
theorems, define unit value terms to express proofs, and use Liquid Haskell to
check that the proofs prove the theorems. For example, Liquid Haskell accepts
the type assignment () :: {v:()| 1+1=2}, as the underlying SMT can always
prove the equality 1+1=2. We write {1+1=2} to simplify the type {v:()| 1+1=2}
from the irrelevant binder v:().
Program Properties as Types The theorems we express can refer to program
functions. As an example, the type of assoc expresses that ♦ is associative.
assoc :: x : m → y : m → z : m → { x ♦ ( y ♦ z ) = ( x ♦ y ) ♦ z }
In § 2 we explain how to write Haskell proof terms to prove theorems like assoc
by proving that list append (++) is associative. Moreover, we prove that the
empty list [] is the identity element of list append, and conclude that the list
(with [] and (++), i.e., the triple ([a], [], (++))) is provably a monoid.
Corectness of Parallelization In § 3, we define the type Morphism n m f that
specifies that f is a morphism between two monoids (n, η, ) and (m, , ♦), i.e.,
f :: n → m where f η = and f (x
y) = f x ♦ f y.
A morphism f on a “chunkable” input type can be parallelized by:
1. chunking up the input in j chunks (chunk j),
2. applying the morphism in parallel to all chunks (pmap f), and
3. recombining the mapped chunks via ♦, also in parallel (pmconcat i).
We specify correctness of the above transformation as a refinement type.
parallelismEquivalence
:: f :( n → m ) → Morphism n m f → x : n → i : Pos → j : Pos
→ { f x = pmconcat i ( pmap f ( chunk j x ) ) }
§ 3 describes the parallelization transformation in details, while Correctness of
Parallelization Theorem 4 proves correctness by providing a term that satisfies
the above type.
Case Study: Parallelization of String Matching We use the above theorem to
parallelize string matching. We define a string matching function toSM ::
RString → toSM target from a refined string to a string matcher. A refined
string (§ 4.1) is a wrapper around the efficient string manipulation library
ByteString that moreover assumes various string properties, including the monoid
laws. A string matcher SM target (§ 4.2) is a data type that contains a refined
string and a list of all the indices where the type level symbol target appears in
the input. We prove that SM target is a monoid and toSM is a morphism, thus
by the aforementioned Correctness of Parallelization Theorem 4 we can correctly
parallelize string matching.
To sum up, we present the first realistic proof that uses Haskell as a theorem
prover: correctness of parallelization on string matching. Our contributions are
summarized as follows
– We explain how theorems and proofs are encoded and checked in Liquid
Haskell by formalizing monoids and proving that lists are monoids (§ 2).
– We formalize morphisms between monoids and specify and prove correctness
of parallelization of morphisms (§ 3).
– We show how libraries can be imported as trusted components by wrapping
ByteStrings as refined strings which satisfy the monoid laws (§ 4.1).
– As an application, we prove that a string matcher is a morphism between the
monoids of refined strings and string matchers, thus we get provably correct
parallelization of string matching (§ 4).
– Based on our 1839LoC proof we evaluate the approach of using Haskell as a
theorem prover (§ 5).
2
Proofs as Haskell Functions
Refinement Reflection [32] is a technique that lets you write Haskell functions
that prove theorems about other Haskell functions and have your proofs machinechecked by Liquid Haskell [31]. As an introduction to Refinement Reflection, in
this section, we prove that lists are monoids by
– specifying monoid laws as refinement types,
– proving the laws by writing the implementation of the law specifications, and
– verifying the proofs using Liquid Haskell.
2.1
Reflection of data types into logic.
To start with, we define a List data structure and teach Liquid Haskell basic
properties about List, namely, how to check that proofs on lists are total and
how to encode functions on List into the logic.
The data list definition L is the standard recursive definition.
data L [ length ] a = N | C a ( L a )
With the length annotation in the definition Liquid Haskell will use the length
function to check termination of functions recursive on Lists. We define length
as the standard Haskell function that returns natural numbers. We lift length
into logic as a measure [31], that is, a unary function whose (1) domain is the
data type, and (2) body is a single case-expression over the datatype.
type Nat = { v : Int | 0 ≤ v }
measure length
length
:: L a → Nat
length N
= 0
length ( C x xs ) = 1 + length xs
Finally, we teach Liquid Haskell how to encode functions on Lists into logic.
The flag "--exact-data-cons" automatically derives measures which (1) test
if a value has a given data constructor, and (2) extract the corresponding field’s
value. For example, Liquid Haskell will automatically derive the following List
manipulation measures from the List definition.
isN :: L a → Bool
isC :: L a → Bool
-- Haskell 0 s null
-- Haskell 0 s not . null
selC1 :: L a → a
selC2 :: L a → L a
-- Haskell 0 s head
-- Haskell 0 s tail
Next, we describe how Liquid Haskell uses the above measures to automatically
reflect Haskell functions on Lists into logic.
2.2
Reflection of Haskell functions into logic.
Next, we define and reflect into logic the two monoid operators on Lists. Namely,
the identity element (which is the empty list) and an associative operator (♦)
(which is list append).
reflect
:: L a
= N
reflect (♦)
(♦) :: L a → L a → L a
N
♦ ys = ys
( C x xs ) ♦ ys = C x ( xs ♦ ys )
The reflect annotations lift the Haskell functions into logic in three steps.
First, check that the Haskell functions indeed terminate by checking that the
length of the input list is decreasing, as specified in the data list definition. Second, in the logic, they define the respective uninterpreted functions and (♦).
Finally, the Haskell functions and the logical uninterpreted functions are related
by strengthening the result type of the Haskell function with the definition of the
function’s implementation. For example, with the above reflect annotations,
Liquid Haskell will automatically derive the following strengthened types for the
relevant functions.
:: { v : L a | v = ∧ v = N }
(♦) :: xs : L a → ys : L a
→ { v : L a | v = xs ♦ ys
∧ v = if isN xs then ys
else C ( selC1 xs ) ( selC2 xs ♦ ys )
}
2.3
Specification and Verification of Monoid Laws
Now we are ready to specify the monoid laws as refinement types and provide
their respective proofs as terms of those type. Liquid Haskell will verify that our
proofs are valid. Note that this is exactly what one would do in any standard
logical framework, like LF [13].
The type Proof is defined as an alias of the unit type (()) in the library
ProofCombinators that comes with Liquid Haskell. Figure 1 summarizes the
definitions we use from ProofCombinators. We express theorems as refinement
types by refining the Proof type with appropriate refinements. For example, the
following theorem states the is always equal to itself.
trivial :: { = }
Where { = } is a simplification for the Proof type {v:Proof | = }, since
the binder v is irrelevant, and trivial is defined in ProofCombinators to be
unit. Liquid Haskell will typecheck the above code using an SMT solver to check
congruence on .
Definition 1 (Monoid). The triple (m, , ♦) is a monoid (with identity element and associative operator ♦), if the following functions are defined.
idLeftm :: x : m → { ♦ x = x }
idRightm :: x : m → { x ♦ = x }
assocm :: x : m → y : m → z : m → { x ♦ ( y ♦ z ) = ( x ♦ y ) ♦ z }
Using the above definition, we prove that our list type L is a monoid by defining
Haskell proof terms that satisfy the above monoid laws.
Left Identity is expressed as a refinement type signature that takes as input a
list x:L a and returns a Proof type refined with the property ♦ x = x
idLeft :: x : L a → { ♦ x = x }
idLeft x = empty ♦ x ==. N ♦ x ==. x *** QED
type Proof = ()
data QED
= QED
trivial :: Proof
trivial = ()
(==.) :: x : a → y :{ a | x = y } → { v : a | v = x }
x ==. _ = x
(***) :: a → QED → Proof
_ *** _ = ()
(∴) :: ( Proof → a ) → Proof → a
f ∴ y = f y
Fig. 1. Operators and Types defined in ProofCombinators
We prove left identity using combinators from ProofCombinators as defined in
Figure 1. We start from the left hand side empty ♦ x, which is equal to N ♦ x
by calling empty thus unfolding the equality empty = N into the logic. Next,
the call N ♦ x unfolds into the logic the definition of (♦) on N and x, which is
equal to x, concluding our proof. Finally, we use the operators p *** QED which
basically casts p into a proof term. In short, the proof of left identity, proceeds
by unfolding the definitions of and (♦) on the empty list.
Right identity is proved by structural induction. We encode inductive proofs
by case splitting on the base and inductive case, and enforcing the inductive
hypothesis via a recursive call.
idRight :: x : L a → { x ♦ = x }
idRight N = N ♦ empty ==. N *** QED
idRight ( C x xs )
=
( C x xs ) ♦ empty
==. C x ( xs ♦ empty )
==. C x xs ∴ idRight xs
*** QED
The recursive call idRight xs is provided as a third optional argument in the
(==.) operator to justify the equality xs ♦ empty = xs, while the operator
(∴) is merely a function application with the appropriate precedence. Note that
LiquiHaskell, via termination and totality checking, is verifying that all the proof
terms are well formed because (1) the inductive hypothesis is only applying to
smaller terms, and (2) all cases are covered.
Associativity is proved in a very similar manner, using structural induction.
assoc
assoc
=
==.
==.
***
:: x : L a → y : L a → z : L a
→ { x ♦ (y ♦ z) = (x ♦ y) ♦ z}
N y z
N ♦ (y ♦ z)
y ♦ z
(N ♦ y) ♦ z
QED
assoc ( C x xs ) y z
= ( C x xs ) ♦ ( y ♦ z )
==. C x ( xs ♦ ( y ♦ z ) )
==. C x (( xs ♦ y ) ♦ z ) ∴ associativity xs y z
==. ( C x ( xs ♦ y ) ) ♦ z
==. (( C x xs ) ♦ y ) ♦ z
*** QED
As with the left identity, the proof proceeds by (1) function unfolding (or rewriting in paper and pencil proof terms), (2) case splitting (or case analysis), and
(3) recursion (or induction).
Since our list implementation satisfies the three monoid laws we can conclude
that L a is a monoid.
Theorem 1. (L a, , ♦) is a monoid.
Proof. L a is a monoid, as the implementation of idLeft, idRight, and assoc
satisfy the specifications of idLeftm , idRightm , and assocm , with m = L a.
t
u
3
Verified Parallelization of Monoid Morphisms
A monoid morphism is a function between two monoids which preserves the
monoidal structure; i.e., a function on the underlying sets which preserves identity and associativity. We formally specify this definition using a refinement type
Morphism.
Definition 2 (Monoid Morphism). A function f :: n → m is a morphism
between the monoids (m, , ♦) and (n, η, ) if Morphism n m f has an inhabitant.
type Morphism n m F =
x:n → y:n → {F η = ∧ F (x
y) = F x ♦ F y}
A monoid morphism can be parallelized when its domain can be cut into
chunks and put back together again, a property we refer to as chunkable and
expand upon in § 3.1. A chunkable monoid morphism is then parallelized by:
1. chunking up the input,
2. applying the morphism in parallel to all chunks, and
3. recombining the chunks, also in parallel, back to a single value.
In the rest of this section we implement and verify to be correct the above
transformation.
3.1
Chunkable Monoids
Definition 3 (Chunkable Monoids). A monoid (m, , ♦) is chunkable if the
following four functions are defined on m.
lengthm :: m → Nat
dropm :: i : Nat → x : MGEq m i → MEq m ( lengthm x - i )
takem :: i : Nat → x : MGEq m i → MEq m i
takeDropProp m :: i : Nat → x : m →
{ x = takem i x ♦ dropm i x }
Where the type aliases MLeq m I (and MEq m I) constrain the monoid m to
have lengthm greater than (resp. equal) to I.
type MGEq m I = { x : m | I ≤ lengthm x }
type MEq m I = { x : m | I = lengthm x }
Note that the “important” methods of chunkable monoids are the take and
drop, while the length method is required to give pre- and post-condition on
the other operations. Finally, takeDropProp provides a proof that for each i and
monoid x, appending take i x to drop i x will reconstruct x.
Using takem and dropm we define for each chunkable monoid (m, , ♦) a
function chunkm i x that splits x in chunks of size i.
chunkm :: i : Pos → x : m → { v : L m | chunkResm i x v }
chunkm i x
| lengthm x ≤ i = C x N
| otherwise
= takem i x 8 C 8 chunkm i ( dropm i x )
chunkResm i x v
| lengthm x ≤ i = lengthm v == 1
| i == 1
= lengthm v == lengthm xs
| otherwise
= lengthm v < lengthm xs
The function chunkm provably terminates as dropm i x will return a monoid
smaller than x, by the Definition of dropm . The definitions of both takem and
dropm are also used from Liquid Haskell to verify the lengthm constraints in
the result of chunkm .
3.2
Parallel Map
We define a parallelized map function pmap using Haskell’s library parallel.
Concretely, we use the function Control.Parallel.Strategies.withStrategy
that computes its argument in parallel given a parallel strategy.
pmap :: ( a → b ) → L a → L b
pmap f xs = withStrategy parStrategy ( map f xs )
The strategy parStrategy does not affect verification. In our codebase we choose
the traversable strategy.
parStrategy :: Strategy ( L a )
parStrategy = parTraversable rseq
Parallelism in the Logic. The function withStrategy is an imported Haskell
library function, whose implementation is not available during verification. To
use it in our verified code, we make the assumption that it always returns its
second argument.
assume withStrategy :: Strategy a → x : a → { v : a | v = x }
Moreover, we need to reflect the function pmap and represent its implementation
in the logic. Thus, we also need to represent the function withStrategy in the
logic. LiquidHaskell represents withStrategy in the logic as a logical function
that merely returns its second argument, withStrategy _ x = x, and does not
reason about parallelism.
3.3
Monoidal Concatenation
The function chunkm lets us turn a monoidal value into several pieces. In the
other direction, for any monoid m, there is a standard way of turning L m back
into a single m 1
mconcat :: L m → m
mconcat N
=
mconcat ( C x xs ) = x ♦ mconcat xs
For any chunkable monoid n, monoid morphism f :: n → m, and natural number i > 0 we can write a chunked version of f as
mconcat . pmap f . chunk n i :: n → m .
Before parallelizing mconcat, we will prove that the previous function is equivalent to f.
Theorem 2 (Morphism Distribution). Let (m, , ♦) be a monoid and (n,
η, ) be a chunkable monoid. Then, for every morphism f :: n → m, every
positive number i and input x, f x = mconcat (pmap f (chunkn i x)) holds.
1
mconcat is usually defined as foldr mappend mempty
morphismDistribution
:: f :( n → m ) → Morphism n m f → x : n → i : Pos
→ { f x = mconcat ( pmap f ( chunk n i x ) ) }
Proof. We prove the theorem by providing an implementation of morphismDistribution
that satisfies its type. The proof proceeds by induction on the length of the input.
m o r p h i s m D i s t r i b u t i o n f thm x i
| length n x ≤ i
=
mconcat ( pmap f ( chunk n i x ) )
==. mconcat ( map f ( chunk n i x ) )
==. mconcat ( map f ( C x N ) )
==. mconcat ( f x 8 C 8 map f N )
==. f is ♦ mconcat N
==. f is ♦
==. f is ∴ idRightm ( f is )
*** QED
m o r p h i s m D i s t r i b u t i o n f thm x i
=
mconcat ( pmap f ( chunk n i x ) )
==. mconcat ( map f ( chunk n i x ) )
==. mconcat ( map f ( C takeX ) ( chunk n i dropX ) ) )
==. mconcat ( f takeX 8 C 8 map f ( chunk n n dropX ) )
==. f takeX ♦ f dropX
∴ m o r p h i s m D i s t r i b u t i o n f thm dropX i
==. f ( takeX
dropX )
∴ thm takeX dropX
==. f x
∴ takeDropProp n i x
*** QED
where
dropX = drop n i x
takeX = take n i x
In the base case we use rewriting and right identity on the monoid f x. In the
inductive case, we use the inductive hypothesis on the input dropX = dropn i
x, that is provably smaller than x as 1 < i. Then, the fact that f is a monoid
morphism, as encoded by our assumption argument thm takeX dropX we get
basic distribution of f, that is f takeX ♦ f dropX = f (takeX
dropX). Finally, we merge takeX
dropX to x using the property takeDropPropn of the
chunkable monoid n.
t
u
3.4
Parallel Monoidal Concatenation
We now parallelize the monoid concatenation by defining a pmconat i x function that chunks the input list of monoids and concatenates each chunk in parallel.
We use the chunk function of § 3.1 instantiated to L m to define a parallelized
version of monoid concatenation pmconcat.
pmconcat :: Int → L m → m
pmconcat i x | i ≤ 1 || length x ≤ i
= mconcat x
pmconcat i x
= pmconcat i ( pmap mconcat ( chunk i x ) )
The function pmconcat i x calls mconcat x in the base case, otherwise it (1)
chunks the list x in lists of size i, (2) runs in parallel mconcat to each chunk, (3)
recursively runs itself with the resulting list. Termination of pmconcat holds, as
the length of chunk i x is smaller than the length of x, when 1 < i.
Next, we prove equivalence of parallelized monoid concatenation.
Theorem 3 (Correctness of Parallelization). Let (m, , ♦) be a monoid.
Then, the parallel and sequential concatenations are equivalent.
p m co n c at E q ui v a le n c e
:: i : Int → x : L m → { pmconcat i x = mconcat x }
Proof. We prove the theorem by providing a Haskell implementation of pmconcatEquivalence
that satisfies its type. The details of the proof can be found in [33], here we provide the sketch of the proof.
First, we prove that mconcat distributes over list splitting
mconcatSplit
:: i : Nat → xs :{ L m | i ≤ length xs }
→ { mconcat xs = mconcat ( take i xs )
♦ mconcat ( drop i xs ) }
The proofs proceeds by structural induction, using monoid left identity in the
base case and monoid associativity associavity and unfolding of take and drop
methods in the inductive step.
We generalize the above lemma to prove that mconcat distributes over list
chunking.
mconcatChunk
:: i : Pos → xs : L m
→ { mconcat xs = mconcat ( map mconcat ( chunk i xs ) ) }
The proofs proceeds by structural induction, using monoid left identity in the
base case and lemma mconcatSplit in the inductive step.
Lemma mconcatChunk is sufficient to prove pmconcatEquivalence by structural induction, using monoid left identity in the base case.
t
u
3.5
Parallel Monoid Morphism
We can now replace the mconcat in our chunked monoid morphism in § 3.3 with
pmconcat from § 3.4 to provide an implementation that uses parallelism to both
map the monoid morphism and concatenate the results.
Theorem 4 (Correctness of Parallelization). Let (m, , ♦) be a monoid and
(n, η, ) be a chunkable monoid. Then, for every morphism f :: n → m, every
positive numbers i and j, and input x, f x = pmconcat i (pmap f (chunkn
j x)) holds.
parallelismEquivalence
:: f :( n → m ) → Morphism n m f → x : n → i : Pos → j : Pos
→ { f x = pmconcat i ( pmap f ( chunk n j x ) ) }
Proof. We prove the theorem by providing an implementation of parallelismEquivalence
that satisfies its type.
p a r a l l e l i s m E q u i v a l e n c e f thm x i j
=
pmconcat i ( pmap f ( chunk n j x ) )
==. mconcat ( pmap f ( chunk n j x ) )
∴ p m co n c at E q ui v a le n c e i ( pmap f ( chunk n j x ) )
==. f x
∴ m o r p h i s m D i s t r i b u t i o n f thm x j
*** QED
The proof follows merely by application of the two previous Theorems 2 and 3.
t
u
4
Case Study: Correctness of Parallel String Matching
§ 3 showed that any monoid morphism whose domain is chunkable can be parallelized. We now make use of that result to parallelize string matching. We start
by observing that strings are a chunkable monoid. We then turn string matching
for a given target into a monoid morphism from a string to a suitable monoid,
SM target, defined in § 4.2. Finally, in § 4.4, we parallelize string matching by
a simple use of the parallel morphism function of § 3.5.
4.1
Refined Strings are Chunkable Monoids
We define a new type RString, which is a chunkable monoid, to be the domain of our string matching function. Our type simply wraps Haskell’s existing
ByteString.
data RString = RS BS . ByteString
Similarly, we wrap the existing ByteString functions we will need to show
RString is a chunkable monoid.
η = RS ( BS . empty )
( RS x )
( RS y ) = S ( x
8
BS . append 8 y )
lenStr
( RS x ) = BS . length x
takeStr i ( RS x ) = RS ( BS . take i x )
dropStr i ( RS x ) = RS ( BS . take i x )
Although it is possible to explicitly prove that ByteString implements a chunkable monoid [30], it is time consuming and orthogonal to our purpose. Instead, we
just assume the chunkable monoid properties of RString– thus demonstrating
that refinement reflection is capable of doing gradual verification.
For instance, we define a logical uninterpreted function and relate it to the
Haskell function via an assumed (unchecked) type.
assume ( )
:: x : RString → y : RString → { v : RString | v = x
Then, we use the uninterpreted function
like associativity.
y}
in the logic to assume monoid laws,
assume assocStr :: x : RString → y : RString → z : RString
→ { x
(y
z) = (x
y)
z }
assocStr _ _
= trivial
Haskell applications of
are interpreted in the logic via the logical
that
satisfies associativity via theorem assocStr.
Similarly for the chunkable methods, we define the uninterpreted functions
takeStr, dropStr and lenStr in the logic, and use them to strengthen the result
types of the respective functions. With the above function definitions (in both
Haskell and logic) and assumed type specifications, Liquid Haskell will check (or
rather assume) that the specifications of chunkable monoid, as defined in the
Definitions 1 and 3, are satisfied. We conclude with the assumption (rather that
theorem) that RString is a chunkable monoid.
Assumption 5 (RString is a Chunkable Monoid) (RString, η, ) combined with the methods lenStr, takeStr, dropStr and takeDropPropStr is a
chunkable monoid.
4.2
String Matching Monoid
String matching is determining all the indices in a source string where a given
target string begins; for example, for source string ababab and target aba the
results of string matching would be [0, 2].
We now define a suitable monoid, SM target, for the codomain of a string
matching function, where target is the string being looked for. Additionally, we
will define a function toSM :: RString → SM target which does the string
matching and is indeed a monoid morphism from RString to SM target for a
given target.
String Matching Monoid We define the data type SM target to contain a
refined string field input and a list of all the indices in input where the target
appears.
data SM ( target :: Symbol ) where
SM :: input : RString
Fig. 2. Mappend indices of String Matcher
→ indices :[ GoodIndex input target ]
→ SM target
We use the string type literal 2 to parameterize the monoid over the target being
matched. This encoding allows the type checker to statically ensure that only
searches for the same target can be merged together. The input field is a refined
string, and the indices field is a list of good indices. For simplicity we present lists
as Haskell’s built-in lists, but our implementation uses the reflected list type, L,
defined in § 2.
A GoodIndex input target is a refined type alias for a natural number i for
which target appears at position i of input. As an example, the good indices
of "abcab" on "ababcabcab" are [2,5].
type GoodIndex Input Target
= { i : Nat | isGoodIndex Input ( fromString Target ) i }
isGoodIndex :: RString →
isGoodIndex input target
= ( subString i ( lenStr
∧ ( i + lenStr target ≤
RString → Int → Bool
i
target ) input == target )
lenStr input )
subString :: Int → Int → RString → RString
subString o l = takeStr l . dropStr o
Monoid Methods for String Matching Next, we define the mappend and
identity elements for string matching.
The identity element of SM t, for each target t, is defined to contain the
identity RString (η) and the identity List ([]).
:: ∀ ( t :: Symbol ) . SM t
= SM η []
The Haskell definition of ♦, the monoid operation for SM t, is as follows.
(♦) :: ∀ ( t :: Symbol ) . KnownSymbol t ⇒ SM t → SM t → SM t
( SM x xis ) ♦ ( SM y yis )
2
Symbol is a kind and target is effectively a singleton type.
= SM ( x
where
tg
=
xis 0 =
xyis =
yis 0 =
y ) ( xis 0 ++ xyis ++ yis 0 )
fromString ( symbolVal ( Proxy :: Proxy t ) )
map ( castGo odInd exLeft tg x y ) xis
makeNewIndices x y tg
map ( shiftStringRight tg x y ) yis
Note again that capturing target as a type parameter is critical, otherwise there
is no way for the Haskell’s type system to specify that both arguments of (♦)
are string matchers on the same target.
The action of (♦) on the two input fields is straightforward; however, the
action on the two indices is complicated by the need to shift indices and the
possibility of new matches arising from the concatenation of the two input
fields. Figure 2 illustrates the three pieces of the new indices field which we
now explain in more detail.
1. Casting Good Indices If xis is a list of good indices for the string x and the
target tg, then xis is also a list of good indices for the string x
y and the
target tg, for each y. To prove this property we need to invoke the property
subStrAppendRight on Refined Strings that establishes substring preservation
on string right appending.
assume
::
→
→
su bStrAp pendRi ght
sl : RString → sr : RString → j : Int
i :{ Int | i + j ≤ lenStr sl }
{ subString sl i j = subString ( sl
sr ) i j }
The specification of subStrAppendRight ensures that for each string sl and sr
and each integer i and j whose sum is within sl, the substring from i with
length j is identical in sl and in (sl
sr). The function castGoodIndexLeft
applies the above property to an index i to cast it from a good index on sl to
a good index on (sl
sr)
cast GoodIn dexLef t
:: tg : RString → sl : RString → sr : RString
→ i : GoodIndex sl tg
→ { v : GoodIndex ( sl
sr ) target | v = i }
cast GoodIn dexLef t tg sl sr i
= cast ( subStr Append Right sl sr ( lenStr tg ) i ) i
Where cast p x returns x, after enforcing the properties of p in the logic
cast :: b → x : a → { v : a | v = x }
cast _ x = x
Moreover, in the logic, each expression cast p x is reflected as x, thus allowing
random (i.e., non-reflected) Haskell expressions to appear in p.
2. Creation of new indices The concatenation of two input strings sl and sr
may create new good indices. For instance, concatenation of "ababcab" with
"cab" leads to a new occurence of "abcab" at index 5 which does not occur
in either of the two input strings. These new good indices can appear only at
the last lenStr tg positions of the left input sl. makeNewIndices sl sr tg
detects all such good new indices.
makeNewIndices
:: sl : RString → sr : RString → tg : RString
→ [ GoodIndex { sl
sr } tg ]
makeNewIndices sl sr tg
| lenStr tg < 2 = []
| otherwise
= makeIndices ( sl
sr ) tg lo hi
where
lo = maxInt ( lenStr sl - ( lenStr tg - 1) ) 0
hi = lenStr sl - 1
If the length of the tg is less than 2, then no new good indices are created.
Otherwise, the call on makeIndices returns all the good indices of the input sl
sr for target tg in the range from maxInt (lenStr sl-(lenStr tg-1)) 0
to lenStr sl-1.
Generally, makeIndices s tg lo hi returns the good indices of the input
string s for target tg in the range from lo to hi.
makeIndices
:: s : RString → tg : RString → lo : Nat
→ hi : Int → [ GoodIndex s tg ]
makeIndices s tg lo hi
| hi < lo
= []
| isGoodIndex s tg lo = lo : rest
| otherwise
= rest
where
rest = makeIndices s tg ( lo + 1) hi
It is important to note that makeNewIndices does not scan all the input,
instead only searching at most lenStr tg positions for new good indices. Thus,
the time complexity to create the new indices is linear on the size of the target
but independent of the size of the input.
3. Shift Good Indices If yis is a list of good indices on the string y with target
tg, then we need to shift each element of yis right lenStr x units to get a list
of good indices for the string x
y.
To prove this property we need to invoke the property subStrAppendLeft
on Refined Strings that establishes substring shifting on string left appending.
assume subStrAppendLeft
:: sl : RString → sr : RString
→ j : Int → i : Int
→ { subStr sr i j = subStr ( sl
sr ) ( lenStr sl + i ) j }
The specification of subStrAppendLeft ensures that for each string sl and sr
and each integers i and j, the substring from i with length j on sr is equal to
the substring from lenStr sl + i with length j on (sl
sr). The function
shiftStringRight both shifts the input index i by lenStr sl and applies the
subStrAppendLeft property to it, casting i from a good index on sr to a good
index on (sl
sr)
Thus, shiftStringRight both appropriately shifts the index and casts the
shifted index using the above theorem:
shiftStringRight
:: tg : RString → sl : RString → sr : RString
→ i : GoodIndex sr tg
→ { v :( GoodIndex ( sl
sr ) tg ) | v = i + lenStr sl }
shiftStringRight tg sl sr i
= subStrAppendLeft sl sr ( lenStr tg ) i
8
cast 8 i + lenStr sl
String Matching is a Monoid Next we prove that the monoid methods
and (♦) satisfy the monoid laws.
Theorem 6 (SM is a Monoid). (SM t, , ♦) is a monoid.
Proof. According to the Monoid Definition 1, we prove that string matching is a
monoid, by providing safe implementations for the monoid law functions. First,
we prove left identity.
idLeft :: x : SM t → { ♦ x = xs }
idLeft ( SM i is )
=
( :: SM t ) ♦ ( SM i is )
==. ( SM η []) ♦ ( SM i is )
==. SM ( η
i ) ( is1 ++ isNew ++ is2 )
∴ idLeftStr i
==. SM i ([] ++ [] ++ is )
∴ ( mapShiftZero tg i is ∧ newIsNullRight i tg )
==. SM i is
∴ idLeftList is
*** QED
where
tg
= fromString ( symbolVal ( Proxy :: Proxy t ) )
is1
= map ( ca st Go od In dex Ri gh t tg i η ) []
isNew = makeNewIndices η i tg
is2
= ( map ( shiftStringRight tg η i ) is )
The proof proceeds by rewriting, using left identity of the monoid strings and
lists, and two more lemmata.
– Identity of shifting by an empty string.
mapShiftZero :: tg : RString → i : RString
→ is :[ GoodIndex i target ]
→ { map ( shiftStringRight tg η i ) is = is }
The lemma is proven by induction on is and the assumption that empty
strings have length 0.
– No new indices are created.
newIsNullLeft :: s : RString → t : RString
→ { makeNewIndices η s t = [] }
The proof relies on the fact that makeIndices is called on the empty range
from 0 to -1 and returns [].
Next, we prove right identity.
idRight :: x : SM t → { x ♦ = x }
idRight ( SM i is )
=
( SM i is ) ♦ ( :: SM t )
==. ( SM i is ) ♦ ( SM η [])
==. SM ( i
η ) ( is1 ++ isNew ++ is2 )
∴ idRightStr i
==. SM i ( is ++ N ++ N )
∴ ( mapCastId tg i η is ∧ newIsNullLeft i tg )
==. SM i is
∴ idRightList is
*** QED
where
tg
= fromString ( symbolVal ( Proxy :: Proxy t ) )
is1
= map ( ca st Go od Ind ex Ri gh t tg i η ) is
isNew = makeNewIndices i stringEmp tg
is2
= map ( shiftStringRight tg i η ) []
The proof proceeds by rewriting, using right identity on strings and lists and
two more lemmata.
– Identity of casting is proven
mapCastId :: tg : RString → x : RString → y : RString
→ is :[ GoodIndex x tg ] →
→ { map ( ca st Go od In de xR ig ht tg x y ) is = is }
We prove identity of casts by induction on is and identity of casting on a
single index.
– No new indices are created.
newIsNullLeft :: s : RString → t : RString
→ { makeNewIndices s η t = [] }
The proof proceeds by case splitting on the relative length of s and t. At
each case we prove by induction that all the potential new indices would be
out of bounds and thus no new good indices would be created.
- Finally we prove associativity. For space, we only provide a proof sketch.
The whole proof is available online [33]. Our goal is to show equality of the left
and right associative string matchers.
assoc :: x : SM t → y : SM t → z : SM t
→ { x ♦ (y ♦ z) = (x ♦ y) ♦ z}
To prove equality of the two string matchers we show that the input and indices
fields are respectively equal. Equality of the input fields follows by associativity
of RStrings. Equality of the index list proceeds in three steps.
1. Using list associativity and distribution of index shifting, we group the indices in the five lists shown in Figure 3: the indices of the input x, the new
indices from mappending x to y, the indices of the input y, the new indices
from mappending x to y, and the indices of the input z.
2. The representation of each group depends on the order of appending. For example, if zis1 (resp. zis2) is the group zis when right (resp. left) mappend
happened first, then we have
zis1 = map ( shiftStringRight tg xi ( yi
zi ) )
( map ( shiftStringRight tg yi zi ) zis )
zis2 = map ( shiftStringRight tg ( xi
yi ) zi ) zis
That is, in right first, the indices of z are first shifted by the length of yi
and then by the length of xi, while in the left first case, the indices of z
are shifted by the length of xi
yi. In this second step of the proof we
prove, using lemmata, the equivalence of the different group representations.
The most interesting lemma we use is called assocNewIndices and proves
equivalence of all the three middle groups together by case analysis on the
relative lengths of the target tg and the middle string yi.
3. After proving equivalence of representations, we again use list associativity
and distribution of casts to wrap the index groups back in string matchers.
Fig. 3. Associativity of String Matching
We now sketch the three proof steps, while the whole proof is available online [33].
assoc x@ ( SM xi xis ) y@ ( SM yi yis ) z@ ( SM zi zis )
-- Step 1: unwrapping the indices
=
x ♦ (y ♦ z)
==. ( SM xi xis ) ♦ (( SM yi yis ) ♦ ( SM zi zis ) )
...
-- via list associativity and distribution of shifts
==. SM i ( xis1 ++ (( xyis1 ++ yis1 ++ yzis1 ) ++ zis1 ) )
-- Step 2: Equivalence of representations
==. SM i ( xis2 ++ (( xyis1 ++ yis1 ++ yzis1 ) ++ zis1 ) )
∴ castConcat tg xi yi zi xis
==. SM i ( xis2 ++ (( xyis1 ++ yis1 ++ yzis1 ) ++ zis2 ) )
∴ mapLenFusion tg xi yi zi zis
==. SM i ( xis2 ++ (( xyis2 ++ yis2 ++ yzis2 ) ++ zis2 ) )
∴ assocNewIndices y tg xi yi zi yis
-- Step 3: Wrapping the indices
...
-- via list associativity and distribution of casts
==. ( SM xi xis ♦ SM yi yis ) ♦ SM zi zis
=
(x ♦ y) ♦ z
*** QED
where
i
= xi
( yi
zi )
yzis1 = map ( shiftStringRight tg xi ( yi
yzis2 = makeNewIndices ( xi
yi ) zi tg
yzis = makeNewIndices yi zi tg
...
zi ) ) yzis
t
u
4.3
String Matching Monoid Morphism
Next, we define the function toSM :: RString → SM target which does the
actual string matching computation for a set target 3
toSM :: ∀ ( target :: Symbol ) . ( KnownSymbol target )
⇒ RString → SM target
toSM input = SM input ( makeSMIndices input tg ) where
tg = fromString ( symbolVal ( Proxy :: Proxy target ) )
makeSMIndices
:: x : RString → tg : RString → [ GoodIndex x tg ]
makeSMIndices x tg
= makeIndices x tg 0 ( lenStr tg - 1)
3
toSM assumes the target is clear from the calling context; it is also possible to write
a wrapper function taking an explicit target which gets existentially reflected into
the type.
The input field of the result is the input string; the indices field is computed by
calling makeIndices within the range of the input, that is from 0 to lenStr
input - 1.
We now prove that toSM is a monoid morphism.
Theorem 7 (toSM is a Morphism). toSM :: RString → SM t is a morphism between the monoids (RString, η, ) and (SM t, , ♦).
Proof. Based on definition 2, proving toSM is a morphism requires constructing
a valid inhabitant of the type
Morphism RString ( SM t ) toSM
= x : RString → y : RString
→ { toSM η = ∧ toSM ( x
y ) = toSM x ♦ toSM y }
We define the function distributestoSM :: Morphism RString (SM t) toSM
to be the required valid inhabitant.
The core of the proof starts from exploring the string matcher toSM x ♦
toSM y. This string matcher contains three sets of indices as illustrated in
Figure 2: (1) xis from the input x, (2) xyis from appending the two strings,
and (3) yis from the input y. We prove that appending these three groups of
indices together gives exactly the good indices of x
y, which are also the value
of the indices field in the result of toSM (x
y).
distributestoSM x y
=
( toSM x :: SM target ) ♦ ( toSM y :: SM target )
==. ( SM x is1 ) ♦ ( SM y is2 )
==. SM i ( xis ++ xyis ++ yis )
==. SM i ( makeIndices i tg 0 hi1 ++ yis )
∴ ( mapCastId tg x y is1 ∧ mergeNewIndices tg x y )
==. SM i ( makeIndices i tg 0
hi1
++ makeIndices i tg ( hi1 +1) hi )
∴ shi ftIndi cesRig ht 0 hi2 x y tg
==. SM i is
∴ mergeIndices i tg 0 hi1 hi
==. toSM ( x
y)
*** QED
where
xis = map ( ca st Go od In dex Ri gh t tg x y ) is1
xyis = makeNewIndices x y tg
yis = map ( shiftStringRight
tg x y ) is2
tg
= fromString ( symbolVal ( Proxy :: Proxy target ) )
is1 = makeSMIndices x tg
is2 = makeSMIndices y tg
is
= makeSMIndices i tg
i
= x
y
hi1 = lenStr x - 1
hi2
hi
= lenStr y - 1
= lenStr i - 1
The most interesting lemma we use is mergeIndices x tg lo mid hi that
states that for the input x and the target tg if we append the indices in the
range from to to mid with the indices in the range from mid+1 to hi, we get
exactly the indices in the range from lo to hi. This property is formalized in
the type of the lemma.
mergeIndices
:: x : RString → tg : RString
→ lo : Nat → mid :{ Int | lo ≤ mid } → hi :{ Int | mid ≤ hi }
→ {
makeIndices x tg lo hi
= makeIndices x tg lo mid
++ makeIndices x tg ( mid +1) hi }
The proof proceeds by induction on mid and using three more lemmata:
– mergeNewIndices states that appending the indices xis and xyis is equivalent to the good indices of x
y from 0 to lenStr x - 1. The proof
case splits on the relative sizes of tg and x and is using mergeIndices on
mid = lenStr x1 - lenStr tg in the case where tg is smaller than x.
– mapCastId states that casting a list of indices returns the same list.
– shiftIndicesRight states that shifting right i units the indices from lo
to hi is equivalent to computing the indices from i + lo to i + hi on the
string x
y, with lenStr x = i.
t
u
4.4
Parallel String Matching
We conclude this section with the definition of a parallelized version of string
matching. We put all the theorems together to prove that the sequential and
parallel versions always give the same result.
We define toSMPar as a parallel version of toSM using machinery of section 3.
toSMPar :: ∀ ( target :: Symbol ) . ( KnownSymbol target )
⇒ Int → Int → RString → SM target
toSMPar i j = pmconcat i . pmap toSM . chunkStr j
First, chunkStr splits the input into j chunks. Then, pmap applies toSM at each
chunk in parallel. Finally, pmconat concatenates the mapped chunks in parallel
using ♦, the monoidal operation for SM target.
Correctness. We prove correctness of toSMPar directly from Theorem 4.
Theorem 8 (Correctness of Parallel String Matching). For each parameter i and j, and input x, toSMPar i j x is always equal to toSM x.
correctness :: i : Int → j : Int → x : RString
→ { toSM x = toSMPar i j x }
Proof. The proof follows by direct application of Theorem 4 on the chunkable
monoid (RString, η, ) (by Assumption 5) and the monoid (SM t, , ♦) (by
Theorem 6).
correctness i j x
=
toSMPar i j x
==. pmconcat i ( pmap toSM ( chunkStr j x ) )
==. toSM is
∴ p a r a l l e l i s m E q u i v a l e n c e toSM distributestoSM x i j
*** QED
Note that application of the theorem parallelismEquivalence requires a proof
that its first argument toSM is a morphism. By Theorem 2, the required proof is
provided as the function distributestoSM.
t
u
5
Evaluation: Strengths & Limitations
Verification of Parallel String Matching is the first realistic proof that uses (Liquid) Haskell to prove properties about program functions. In this section we use
the String Matching proof to quantitatively and qualitatively evaluate theorem
proving in Haskell.
Quantitative Evaluation. The Correctness of Parallel String Matching proof can
be found online [33]. Verification time, that is the time Liquid Haskell needs to
check the proof, is 75 sec on a dual-core Intel Core i5-4278U processor. The proof
consists of 1839 lines of code. Out of those
–
–
–
–
226 are Haskell “runtime” code,
112 are liquid comments on the “runtime” Haskell code,
1307 are Haskell proof terms, that is functions with Proof result type, and
194 are liquid comments to specify theorems.
Counting both liquid comments and Haskell proof terms as verification code, we
conclude that the proof requires 7x the lines of “runtime” code. This ratio is
high and takes us to 2006 Coq, when Leroy [17] verified the initial CompCert C
compiler with the ratio of verification to compiler lines being 6x.
Strengths. Though currently verbose, deep verification using Liquid Haskell has
many benefits. First and foremost, the target code is written in the general purpose Haskell and thus can use advanced Haskell features, including type literals, deriving instances, inline annotations and optimized library functions like
ByteString. Even diverging functions can coexist with the target code, as long
as they are not reflected into logic [31].
Moreover, SMTs are used to automate the proofs over key theories like linear
arithmetic and equality. As an example, associativity of (+) is assumed throughout the proofs while shifting indices. Our proof could be further automated by
mapping refined strings to SMT strings and using the automated SMT string
theory. We did not follow this approach because we want to show that our
techinique can be used to prove any (and not only domain specific) program
properties.
Finally, we get further automation via Liquid Type Inference [22]. Properties
about program functions, expressed as type specifications with unit result, often
depend on program invariants, expressed as vanilla refinement types, and vice
versa. For example, we need the invariant that all indices of a string matcher are
good indices to prove associativity of (♦). Even though Liquid Haskell cannot
currently synthesize proof terms, it performs really well at inferring and propagating program invariants (like good indices) via the abstract interpretation
framework of Liquid Types.
Limitations. There are severe limitations that should be addressed to make
theorem proving in Haskell a pleasant and usable technique. As mentioned earlier
the proofs are verbose. There are a few cases where the proofs require domain
specific knowledge. For example, to prove associativity of string matching x
♦ (y ♦ z) = (x ♦ y) ♦ z we need a theorem that performs case analysis
on the relative length of the input field of y and the target string. Unlike this
case split though, most proofs do not require domain specific knowledge and
merely proceed by term rewriting and structural inductuction that should be
automated via Coq-like [3] tactics or/and Dafny-like [16] heuristics. For example,
synquid [21] could be used to automatically synthesize proof terms.
Currently, we suffer from two engineering limitations. First, all reflected function should exist in the same module, as reflection needs access to the function
implementation that is unknown for imported functions. This is the reason why
we need to use a user defined, instead of Haskell’s built-in, list. In our implementation we used CPP as a current workaround of the one module restriction.
Second, class methods cannot be currently reflected. Our current workaround
is to define Haskell functions instead of class instances. For example (append,
nil) and (concatStr, emptyStr) define the monoid methods of List and Refined
String respectively.
Overall, we believe that the strengths outweigh the limitations which will be
addressed in the near future, rendering Haskell a powerful theorem prover.
6
Related Work
SMT-Based Verification SMT solvers have been extensively used to automate
reasoning on verification languages like Dafny [16], Fstar [28] and Why3 [10].
These languages are designed for verification, thus have limited support for
commonly used language features like parallelism and optimized libraries that
we use in our verified implementation. Refinement Types [8,12,24] on the other
hand, target existing general purpose languages, such as ML [34,2,22], C [7,23],
Haskell [31], Racket [15] and Scala [26]. However, before Refinement Reflection [32] was introduced, they only allowed “shallow” program specifications,
that is, properties that only talk about behaviors of program functions but not
functions themselves.
Dependent Types Unlike Refinement Types, dependent type systems, like Coq [3],
Adga [19] and Isabelle/HOL [20] allow for “deep” specifications which talk about
program functions, such as the program equivalence reasoning we presented.
Compared to (Liquid) Haskell, these systems allow for tactics and heuristics that
automate proof term generation but lack SMT automations and general-purpose
language features, like non-termination, exceptions and IO. Zombie [5,27] and
Fstar [28] allow dependent types to coexist with divergent and effectful programs, but still lack the optimized libraries, like ByteSting, which come with a
general purpose language with long history, like Haskell.
Parallel Code Verification Dependent type theorem provers have been used before to verify parallel code. BSP-Why [11] is an extension to Why2 that is using
both Coq and SMTs to discharge user specified verification conditions. Daum [9]
used Isabelle to formalize the semantics of a type-safe subset of C, by extending
Schirmer’s [25] formalization of sequential imperative languages. Finally, Swierstra [29] formalized mutable arrays in Agda and used them to reason about
distributed maps and sums.
One work closely related to ours is SyDPaCC [18], a Coq library that automatically parallelizes list homomorphisms by extracting parallel Ocaml versions of user provided Coq functions. Unlike SyDPaCC, we are not automatically generating the parallel function version, because of engineering limitations
(§ 5). Once these are addressed, code extraction can be naturally implemented
by turning Theorem 4 into a Haskell type class with a default parallelization
method. SyDPaCC used maximum prefix sum as a case study, whose morphism
verification is much simpler than our string matching case study. Finally, our implementation consists of 2K lines of Liquid Haskell, which we consider verbose
and aim to use tactics to simplify. On the contrary, the SyDPaCC implementation requires three different languages: 2K lines of Coq with tactics, 600 lines of
Ocaml and 120 lines of C, and is considered “very concise”.
7
Conclusion
We made the first non-trivial use of (Liquid) Haskell as a proof assistant. We
proved the parallelization of chunkable monoid morphisms to be correct and
applied our parallelization technique to string matching, resulting in a formally
verified parallel string matcher. Our proof uses refinement types to specify equivalence theorems, Haskell terms to express proofs, and Liquid Haskell to check
that the terms prove the theorems. Based on our 1839LoC sophisticated proof we
conclude that Haskell can be successfully used as a theorem prover to prove arbitrary theorems about real Haskell code using SMT solvers to automate proofs
over key theories like linear arithmetic and equality. However, Coq-like tactics
or Dafny-like heurestics are required to ease the user from manual proof term
generation.
References
1. C. Barrett, A. Stump, and C. Tinelli. The SMT-LIB Standard: Version 2.0. 2010.
2. J. Bengtson, K. Bhargavan, C. Fournet, A.D. Gordon, and S. Maffeis. Refinement
types for secure implementations. In CSF, 2008.
3. Y. Bertot and P. Castéran. Coq’Art: The Calculus of Inductive Constructions.
Springer Verlag, 2004.
4. Guy E. Blelloch. Synthesis of Parallel Algorithms. Morgan Kaufmann Pub, 1993.
5. C. Casinghino, V. Sjöberg, and S. Weirich. Combining proofs and programs in a
dependently typed language. In POPL, 2014.
6. Murray Cole. Parallel programming, list homomorphisms and the maximum segment sum problem. 1993.
7. Jeremy Condit, Matthew Harren, Zachary R. Anderson, David Gay, and George C.
Necula. Dependent types for low-level programming. In ESOP, 2007.
8. R. L. Constable and S. F. Smith. Partial objects in constructive type theory. In
LICS, 1987.
9. M Daum. Reasoning on Data-Parallel Programs in Isabelle/Hol. In C/C++ Verification Workshop, 2007.
10. Jean-Christophe Filliâtre and Andrei Paskevich. Why3 – Where Programs Meet
Provers. In ESOP, 2013.
11. J Fortin and F. Gava. BSP-Why: A tool for deductive verification of BSP algorithms with subgroup synchronisation. In Int J Parallel Prog, 2015.
12. T. Freeman and F. Pfenning. Refinement types for ML. In PLDI, 1991.
13. Robert Harper, Furio Honsell, and Gordon Plotkin. A framework for defining
logics. J. ACM, 1993.
14. Joseph JáJá. Introduction to Parallel Algorithms. 1992.
15. Andrew M. Kent, David Kempe, and Sam Tobin-Hochstadt. Occurrence typing
modulo theories. In PLDI, 2016.
16. K. Rustan M. Leino. Dafny: An automatic program verifier for functional correctness. LPAR, 2010.
17. Xavier Leroy. Formal certification of a compiler back-end, or: programming a
compiler with a proof assistant. In POPL 06, 2006.
18. Frédéric Loulergue, Wadoud Bousdira, and Julien Tesson. Calculating Parallel
Programs in Coq using List Homomorphisms. In International Journal of Parallel
Programming, 2016.
19. U. Norell. Towards a practical programming language based on dependent type
theory. PhD thesis, Chalmers, 2007.
20. L. C. Paulson. Isabelle A Generic Theorem prover. Lecture Notes in Computer
Science, 1994.
21. Nadia Polikarpova, Ivan Kuraj, and Armando Solar-Lezama. Program synthesis
from polymorphic refinement types. In PLDI, 2016.
22. P. Rondon, M. Kawaguchi, and R. Jhala. Liquid types. In PLDI, 2008.
23. P. Rondon, M. Kawaguchi, and R. Jhala. Low-level liquid types. In POPL, 2010.
24. J. Rushby, S. Owre, and N. Shankar. Subtypes for specifications: Predicate subtyping in pvs. IEEE TSE, 1998.
25. N Schirmer. Verification of Sequential Imperative Programs in Isabelle/HOL. PhD
thesis, TU Munich, 2006.
26. Georg Stefan Schmid and Viktor Kuncak. SMT-based Checking of PredicateQualified Types for Scala. In Scala, 2016.
27. Vilhelm Sjöberg and Stephanie Weirich. Programming up to congruence. POPL,
2015.
28. Nikhil Swamy, Cătălin Hriţcu, Chantal Keller, Aseem Rastogi, Antoine DelignatLavaud, Simon Forest, Karthikeyan Bhargavan, Cédric Fournet, Pierre-Yves Strub,
Markulf Kohlweiss, Jean-Karim Zinzindohoue, and Santiago Zanella-Béguelin. Dependent types and multi-monadic effects in F*. In POPL, 2016.
29. Wouter Swierstra. More dependent types for distributed arrays. 2010.
30. N. Vazou, E. L. Seidel, and R. Jhala. Liquidhaskell: Experience with refinement
types in the real world. In Haskell Symposium, 2014.
31. N. Vazou, E. L. Seidel, R. Jhala, D. Vytiniotis, and S. Peyton-Jones. Refinement
Types for Haskell. In ICFP, 2014.
32. Niki Vazou and Ranjit Jhala. Refinement Reflection. arXiv:1610.04641, 2016.
33. Niki Vazou and Jeff Polakow. Code for verified string indexing. 2016. https:
//github.com/nikivazou/verified_string_matching.
34. H. Xi and F. Pfenning. Eliminating array bound checking through dependent
types. In PLDI, 1998.
| 6cs.PL
|
Workshop track - ICLR 2018
S TABLE D ISTRIBUTION A LIGNMENT
U SING THE D UAL OF THE A DVERSARIAL D ISTANCE
Ben Usman
Boston University
[email protected]
Kate Saenko
Boston University
[email protected]
Brian Kulis
Boston University
[email protected]
arXiv:1707.04046v4 [cs.LG] 30 Jan 2018
A BSTRACT
Methods that align distributions by minimizing an adversarial distance between
them have recently achieved impressive results. However, these approaches are
difficult to optimize with gradient descent and they often do not converge well
without careful hyperparameter tuning and proper initialization. We investigate
whether turning the adversarial min-max problem into an optimization problem by
replacing the maximization part with its dual improves the quality of the resulting
alignment and explore its connections to Maximum Mean Discrepancy. Our
empirical results suggest that using the dual formulation for the restricted family of
linear discriminators results in a more stable convergence to a desirable solution
when compared with the performance of a primal min-max GAN-like objective
and an MMD objective under the same restrictions. We test our hypothesis on the
problem of aligning two synthetic point clouds on a plane and on a real-image
domain adaptation problem on digits. In both cases, the dual formulation yields an
iterative procedure that gives more stable and monotonic improvement over time.
1
I NTRODUCTION
Adversarial methods have recently become a popular choice for learning distributions of highdimensional data. The key idea is to learn a parametric representation of a distribution by aligning it
with the empirical distribution of interest according to a distance given by a discriminative model. At
the same time, the discriminative model is trained to differentiate between true and artificially obtained
samples. Generative Adversarial Networks (GANs) that use neural networks to both discriminate
samples and parameterize a learned distribution, have achieved particularly impressive results in many
applications such as generative modeling of images (Goodfellow et al., 2014; Denton et al., 2015;
Radford et al., 2015), image super-resolution (Ledig et al., 2016), and image-to-image translation
(Isola et al., 2016). Adversarial matching of empirical distributions has also shown promise for
aligning train and test data distributions n scenarios involving domain shift (Ganin & Lempitsky,
2015; Tzeng et al., 2015; 2017).
However, GANs and related models have proved to be extremely difficult to optimize. It has been
widely reported that training GANs is a tricky process that often diverges and requires very careful
parameter initialization and tuning. Arjovsky and Bottou (Arjovsky & Bottou, 2017) have recently
identified several theoretical problems with loss functions used by GANs, and have analyzed how
they contribute to instability and saturation during training.
In this paper, we focus on one of the major barriers to stable optimization of adversarial methods,
namely their min-max nature. Adversarial methods seek to match the generated and real distributions
by minimizing some notion of statistical distance between the two, which is often defined as a maximal
difference between values of certain test (witness) functions that could differentiate these distributions.
More specifically in the case of GANs, the distance is usually considered to be equal to the likelihood
of the best neural network classifier that discriminates between the distributions, assigning "real
or generated?" labels to the input points. This way, in order to align these distributions one has to
minimize this maximum likelihood w.r.t. the parameters of the learned or aligned distribution.
Unfortunately, solving min-max problems using gradient descent is inherently very difficult. Below
we use a simple example to demonstrate that different flavors of gradient descent are very unstable
when it comes to solving problems of this kind.
1
Workshop track - ICLR 2018
To address this issue, we explore the possibility of replacing the maximization part of the adversarial
alignment problem with a dual minimization problem for linear and kernelized linear discriminators.
The resulting dual problem turns out to be much easier to solve via gradient descent. Moreover,
we make connections between our formulation and existing objectives such as the Maximum Mean
Discrepancy (MMD) (Gretton, 2012). We show that it is strongly related to the iteratively reweighted
empirical estimator of MMD.
We first evaluate how well our dual method can handle a point alignment problem on a lowdimensional synthetic dataset. Then, we compare its performance with the analogous primal method
on a real-image domain adaptation problem using the Street View House Numbers (SVHN) and
MNIST domain adaptation dataset pair. Here the goal is to align the feature distributions produced
by the network on the two datasets so that a classifier trained to label digits on SVHN does not
loose accuracy on MNIST due to the domain shift. In both cases, we show that our proposed dual
formulation of the adversarial distance often shows improvement over time, whereas using the primal
formulation results in drifting objective values and often does not converge to a solution.
Our contributions can be summarized as follows:
• we explore the dual formulation of the adversarial alignment objective for linear and
kernelized linear discriminators and how they relate to the Maximum Mean Discrepancy;
• we demonstrate experimentally on both synthetic and real datasets that the resulting objective
leads to more stable convergence and better alignment quality;
• we apply this idea to a domain adaptation scenario and show that the stability of reaching
high target classification accuracy is also positively impacted by the dual formulation.
2
R ELATED W ORK
There has been a long line of work on unsupervised generative machine learning, a review of which
is beyond the scope of this work. Recently, adversarial methods for learning generative neural
networks have achieved popularity due to their ability to effectively model high dimensional data,
e.g., generating very realistic looking images. These include the original Generative Adversarial
Networks (GANs) (Goodfellow et al., 2014) and follow up work proposing improved formulations
such as Wasserstein GANs (Arjovsky et al., 2017) and Conditional GANs (Mirza & Osindero, 2014).
Related ideas have been proposed for unsupervised domain adaptation. Neural domain adaptation
methods seek to improve the performance of a classifier network on a target distribution that is
different from the original training distribution by introducing an additional objective that minimizes
the difference between representations learned for source and target data. Some models align feature
representations across domains by minimizing the distance between first or second order feature
space statistics (Tzeng et al., 2014; Long & Wang, 2015; Sun & Saenko, 2016). When adversarial
objectives are used for domain adaptation, a domain classifier is trained to distinguish between
the generated source and target representations, both using the standard minimax objective (Ganin
& Lempitsky, 2015), as well as alternative losses (Tzeng et al., 2015; 2017). Instead of aligning
distributions in feature space, several models perform generative alignment in pixel-space, e.g., using
Coupled GANs (Liu & Tuzel, 2016) or conditional GANs (Bousmalis et al., 2017). These models
adapt by “hallucinating” what the raw training images might look like in the target test domain. A
related line of work uses adversarial training to translate images in one domain to the “style” of a
different domain to create artistic effects (Mathieu et al., 2016). Several recent works perform the
alignment in an unsupervised way without aligned image pairs (Zhu et al., 2017; Onsiderations et al.,
2017). Other objectives for distribution matching that have been proposed in the literature, including
Maximum Mean Discrepancy (Gretton, 2012), f-discrepancy (Nowozin et al., 2016) and others, have
also been used for generative modeling (Dziugaite et al., 2015; Li et al., 2015). A single step of our
iterative reweighting procedure is similar to instance reweighting methods that were theoretically and
empirically shown to improve accuracy in the presence of domain shift. For example, Huang et al.
(2007) used sample reweighting that minimized empirical MMD between populations to plug it as
instance-weights in weighted classification loss, whereas Gong et al. (2013) did that to choose points
for a series of independent auxiliary tasks, so no iterative reweighting was performed in both cases.
In general, most statistical distances used for distribution alignment fall into one of two categories:
they are either f-divergences (e.g. GAN objective, KL-divergence), or integral probability metrics
(IPMs) that are differences in expected values of a test function at samples from different distributions
maximized over some function family (e.g. Maximum Mean Discrepancy, Wasserstein distance). In
2
Workshop track - ICLR 2018
this work we specifically consider the logistic adversarial objective (f-divergence), show that it is
useful to optimize its dual, and present a relation between this adversarial dual objective and MMD,
another statistical distance with a test function from reproducing kernel Hilbert space (RKHS).
3
A M OTIVATING E XAMPLE
We start with a well-known (Goodfellow, 2016) motivating
example of a simple min-max problem to show that, even
in this basic case, gradient descent might fail dramatically.
Let us consider the simplest min-max problem with a
unique solution: finding a saddle point of a hyperbolic
surface. Given the function f (x, y) = xy, our problem is
to solve minx maxy f (x, y), which has a unique solution
at (0, 0). Suppose that we want to apply gradient descent
to solve this problem. The intuitive analog of the gradient
vector that we might consider using in the update rule is
defined by the vector field g(x, y) = (x, −y). However,
at any given point the vector g(x, y) will be tangent to
a closed circular trajectory, thus following this trajectory
would never lead to the true solution (0, 0).
Figure 1: Gradient descent fails to solve
the saddle point problem minx maxy xy.
One can observe the trajectory produced by the update
Red line presents a trajectory of the grarule xt+1 = xt + αg(x, y) applied to the problem above
dient descent if vector field g(x, y) is
in Fig. 1. Neither block coordinate descent nor various
used at each iteration. Blue lines are exlearning rate schedules can significantly improve the peramples of vectors from this vector field.
formance of the gradient descent on this problem.
We want to mention that while there is a huge body of work on using alternative descent schemes for
convex-concave saddle point optimization, including, but not limited to different variants of mirror
descent, such as Nesterov’s dual averaging (Nesterov, 2009) and mirror prox (Nemirovski, 2004),
authors are not aware of any successful attempts to use it in the context of adversarial distribution
alignment. Some techniques developed for solving continuous games such as fictitious play were
successfully adopted by Salimans et al. (2016).
4
A PPROACH
We first propose a new formulation of the adversarial objective for distribution alignment problems.
Then we apply this approach to the domain adaptation scenario in Section 5.
Suppose that we are given a finite set of points A sampled from the distribution p, and a finite set of
points B sampled from the distribution q, and our goal is to match q with p by aligning B with A.
More specifically, we aim to learn a matching function Mθ (B) that maps B to be as close as possible
to A by minimizing some empirical estimate of a statistical distance d(·, ·) between them where θ are
parameters of the matching function: θ∗ = argminθ d(A, Mθ (B)).
Let us denote Bθ0 = Mθ (B) or just B 0 in contexts where dependence on θ is not important. The
regular adversarial approach obtains the distance function by finding the best classifier Dw (x) with
parameters w that discriminates points x ∈ A from points x ∈ B 0 and considers the distance between
A and B 0 to be equal to the likelihood of this classifier. A higher likelihood of separating A from B 0
means that A is far from B 0 . This can be any form of hypothesis in general, and is often chosen to be
a linear classifier (Tzeng et al., 2014) or a multi-layer neural network (Goodfellow et al., 2014). In
this work, we use the class of linear classifiers, specifically, logistic regression in its primal and dual
formulations. The solution can also be kernelized to obtain nonlinear discriminators.
We will define the distance between distributions to be equal to the maximum likelihood of the
logistic classifier parametrized by w:
d(A, B 0 ) = max
w
X
log(σ(wT xi )) +
X
xj ∈B 0
xi ∈A
3
log(1 − σ(wT xj )) −
λ T
w w
2
Workshop track - ICLR 2018
We can equivalently re-write this expression as:
Cθ = {(xi , yi ) : xi ∈ A ∪ Bθ0 , yi = 1 if xi ∈ A else –1}
X
λ
min d(A, Bθ0 ) = min max
log(σ(yi (wT xi + b))) − wT w
θ
θ
w,b
2
(1)
xi ,yi ∈Cθ
The duality derivation (Jaakkola & Haussler, 1999; Minka, 2003) follows from the fact that the
log-sigmoid has a sharp upper-bound
log(σ(u)) ≤ αT u + H(α) , αi ∈ [0, 1]
H(α) = αT log α + (1 − α)T log(1 − α)
thus we can upper-bound the distance as
d(A, Bθ0 ) = min max
0≤α≤1 w,b
X
αi yi (wT xi + b) + H(α) −
xi ,yi ∈Cθ
λ T
w w
2
where the dual variable αi corresponds to the weight of the point. Higher weight means that the point
is contributing more to the decision hyperplane. Optimal value of alpha attains this upper bound.
P
The w that maximizes the inner expression can be computed in a closed form, w∗ = λ1 ( j xj yj αj ).
P
Optimally of bias requires i αi yi = 0. By substituting w∗ we obtain a minimization problem:
d(A, Bθ0 ) = min
0≤αi ≤1
=
min
0≤αi ≤1/λ
1 X
1 T
αi αj (yi xi )T (yj xj ) + H(α) = min
α Qα + H(α) =
0≤αi ≤1 2λ
2λ ij
1 T
1 T
T
α QAA αA + αB
QBB αB − αA
QAB αB + H(αA ) + H(αB )
2 A
2
s.t. ||αA ||1 = ||αB ||1
(2)
(3)
The Eq. (3) is obtained by splitting the summation into blocks that include samples only from A,
only from B, and from both A and B. For example, matrix QAB consists of pairwise similarities
between points from A and B, and is equal to the dot product between corresponding data points
in the linear case. The factor of two in front of the cross term comes from the fact that off-diagonal
blocks in the quadratic form are equal. The constraint on alpha sums comes from splitting optimality
conditions on the bias into two term. We will denote the resulting objective as dD (α, A, B).
The above expression gives us a tight upper bound on the likelihood of the discriminator. Thus,
by minimizing this upper bound, we can minimize the likelihood itself, as in the original loss, and
therefore minimize the distance between the distributions:
θ∗ , α∗ = argmin dD (α, A, Mθ (B))
(4)
θ,α∈A
Note that the overall problem has changed from an unconstrained saddle point problem to a smooth
constrained minimization problem, which ultimately converges when gradient descent has a properly
chosen learning rate, whereas the descent iterations for the saddle point problem are not guaranteed
to converge at all.
The resulting smooth optimization problem consists of minimization over α to improve classification
scores and over θ to move points towards the decision boundary. Next section provides more intuition
behind the resulting iterative procedure.
4.1
R ELATIONSHIP TO MMD
In this section, we show that our dual formulation of the adversarial objective has an interesting relationship to another popular alignment objective. The integral probability metric between distributions
p and q with a given function family H is defined as
d(p, q) = sup Ex∼p f (x) − Ex∼q f (x) .
f ∈H
4
Workshop track - ICLR 2018
It was shown to have a closed form solution and a corresponding closed form empirical estimator
if H is a unit ball in reproducing kernel Hilbert space with the reproducing kernel k(x, y) and is
commonly referred to as Maximum Mean Discrepancy (Gretton, 2012):
1
1
d(p, q) = Ep×p k(x, x0 ) + Eq×q k(y, y 0 ) − Ep×q k(x, y)
2
2
X
1 X
1
1 X
0
k(x, x ) +
k(y, y 0 ) −
k(x, y).
d(A, B) =
2|A|
2|B|
|A||B|
i,j∈A
i,j∈B
A×B
From the definition, it is essentially the distance between means of vectors from p and q embedded
into the corresponding RKHS. The resulting empirical estimator combines average inner and outer
similarities between samples from the two distributions and goes to zero as the number of samples
increases if p = q.
Note that if sample weights in Eq. (3) are constant and equal across all samples, so αi = c, then
the dual distance introduced above becomes exactly an empirical estimate of the MMD plus the
constant from the entropic regularizer. Thus, the adversarial logistic distance introduced in Eq. (3)
can be viewed as an iteratively reweighted empirical estimator of the MMD distance. Intuitively,
what this means is that the optimization procedure consists of two alternating minimization steps:
(1) find the best sample weights assignment by changing α so that the regularized weighted MMD is
minimized, and then (2) use a fixed α to minimize the resulting weighted MMD distance by changing
the matching function Mθ . This makes the resulting procedure similar to the Iteratively Reweighted
Least Squares Algorithm (Green, 1984) for logistic regression. An interesting observation here is that
it turns out that high weights in this iterative procedure are given to the most mutually close subsets
of A and B 0 , where closeness is measured in terms of Maximum Mean Discrepancy. These happen to
be exactly support vectors of the corresponding optimal domain classifier. Therefore, the procedure
described above essentially brings sets of the support vectors of the optimal domain classifier from
different domains closer together.
We note that the computational complexity of a single gradient step of the proposed method grows
quadratically with the size of the dataset because of the kernelization step. However, our batched
GPU implementation of the method performed on par with MMD and outperformed primal methods, probably because inference in modern neural networks requires so many dot products that a
batch size × batch size multiplication is negligibly cheap compared to the rest of the network with
modern highly parallel computing architectures.
5
A PPLICATION TO D ISTRIBUTION A LIGNMENT IN D OMAIN A DAPTATION
We now show how the above formulation can be applied to distribution alignment for the specific
problem of unsupervised domain adaptation. In this scenario, we train our classifier in a supervised
fashion on some domain A and have to update it to perform well on a different domain B without
using any labeled samples from the latter. Common examples include adapting to a camera with
different image quality or to different weather conditions.
More rigorously, we assume that there exist two distinct distributions on X × Y: a source distribution
PS (X, Y ) and a target distribution PT (X, Y ). We assume that we observe a finite number of labeled
samples from the source distribution DS ⊂ [X × Y]n ∼ PS (X, Y ) and a finite number of unlabeled
samples from the target distribution DT ⊂ [X ]m ∼ PT (X). Our goal then is to find a labeling
function f : X → Y from a hypothesis space F that minimizes target risk RT , even though we only
have labels for samples from source.
RT (f ) = E(x,y)∼PT L(f (x), y) ≤ RS (f ) + d(PT , PS ) + C(V, n)
Ben-David et al. (2007) showed that, under mild restrictions on probability distributions, the target
risk is upper-bounded by the sum of three terms: (1) the source risk, (2) the complexity term involving
the dataset size and the VC-dimensionality of F, and (3) the discrepancy between source and target
distributions. Thus, in order to make the target risk closer to the source risk, we need to minimize the
discrepancy between distributions. They define discrepancy as a supremum of differences in measures
across all events in a given σ-algebra: d(p, q) = supA∈Σ |p(A) − q(A)|. Estimation of the indicated
expression is hard in practice, therefore it is usually replaced with more computationally feasible
statistical distances. The total variation between two distributions and the Kolmogorov-Smirnov test
5
Workshop track - ICLR 2018
Discriminator Accuracy
||Σ̂A (t) − Σ̂B ||2
Iteration
Iteration
Iteration
Dual Kernel
Dual
Primal
Objective
Figure 2: Convergence analysis for the synthetic point alignment problem. Top row: The procedure
of adversarial alignment in the primal discriminator weight space for points from A and B lying in a
two-dimensional plane does not converge to a single solution and oscillates over time. Middle row:
In contrast, the proposed dual minimization adversarial objective steadily converges to an optimum.
Bottom row: We observe similar behavior for the kernelized version. Even though the accuracy
of the learned discriminators (column two) drifts over time in all cases, the distance between the
covariance matrices (column three) and means (not shown in the figure) decreased in all cases;
however, in both cases marked as "dual" iterative procedure converged to a single stable solution that
indeed corresponds to the desired point cloud alignment. See Figure 3 for an illustration. Note that
absolute values of objectives are not comparable, and therefore not shown on plots above.
are closely related to the discrepancy definition above and are also often considered to be too strong
to be useful, especially in higher dimensions.
Our approach can be applied directly to this scenario if the discrepancy is replaced with an adversarial
objective that uses a logistic regression domain classifier. In Section 6, we consider an instance of
this problem where the main task is classification and the hypothesis space corresponds to multi-layer
neural networks. We compare the standard min-max formulation of the adversarial objective in
Eq. (1) with our min-min formulation in Eq. (3), and report the accuracy of the resulting classifier on
the target domain.
6
E XPERIMENTS
Synthetic Distribution Matching
We first test the performance of our proposed approach on a synthetic point cloud matching problem.
The data consists of two clouds of points on a two-dimensional plane and the goal is to match points
from one cloud with points from the other. There are no restrictions on the transformation of the
target point cloud, so Mθ includes all possible transformations, and is therefore parameterized by the
point coordinates themselves, so the coordinates themselves were updated on each gradient step. We
minimized the logistic adversarial distance in primal space by solving the corresponding min-max
problem in Eq (1) and compared this to maximizing the proposed negative adversarial distance given
by the dual of the logistic classifier in Eq (3) and the corresponding kernelized logistic classifier with
a Gaussian kernel.
As expected, the optimization of distances given by the dual versions of domain classifiers (linear
and kernelized) worked considerably better than the same distance given by a linear classifier in the
primal form. More specifically, the results in the primal case were very sensitive to the choice of
learning rate. In general, the resulting decent iterations for the saddle point problem did not converge
to a single solution, whereas both dual versions successfully converged to solutions that matched the
two clouds of points both visually and in terms of means and covariances, as presented in Fig. 2.
6
Dual Kernel
Dual
Primal
Workshop track - ICLR 2018
t0
...
tT
Figure 3: (Best viewed in color) When trained on a point cloud matching task, the primal approach
leads to an unstable solution that makes the decision boundary spin around data points when they
are almost aligned, whereas both the linear and kernel dual approaches lead to stable solutions that
gradually assign 0.5 probability of belonging to either A or B to all points, which is exactly the
desired behaviour. Yellow and blue points are the original point clouds, red points correspond to
positions of yellow points after transformation Mθ .
We suggest one more intuitive explanation of why the dual procedure might work better, in addition
to the fact that optimization problems are just inherently easier than saddle point problems. The
descision boundary of the classifier in the dual space is defined implicitly through a weighted average
of observed data points, so when these data points move, the decision boundary moves with them. If
points move too rapidly and the discriminator explicitly parametrizes the decision boundary, weights
of the discriminator may change drastically to keep up with moved points, leading to the overall
instability of the training procedure. In support of this hypothesis, we observed interesting patterns
in the behavior of the linear primal discriminator: when point clouds become sufficiently aligned,
the decision boundary starts "spinning" around these clouds, slightly pushing them in corresponding
directions. In contrast, both dual classifiers end up gradually converging to solutions that assigned
each point with 0.5 probability of corresponding to either of two domains.
Real-Image Unsupervised Domain Adaptation
We also evaluated the performance of the proposed dual objective on a visual domain adaptation task.
We performed a series of experiments on an SVHN-MNIST digit classification dataset pair in an
unsupervised domain adaptation setup: the task is to use a classifier trained on SVHN and unlabeled
samples from MNIST to improve test accuracy on the latter.
Following Tzeng et al. (2017) we used standard LeNet as a base model and outputs of the last
layer before the softmax as feature representations. We trained the source network to perform
well on source dataset and the discriminator to distinguish features computed by the source and
target networks. After training the source network on the source domain, we initialized the target
network with source weights and optimized it to make the distributions of source and target feature
representations less distinguishable from the discriminator perspective. More technical details are
given in the Supplementary Section 10.2.
We tested several primal objectives based on Adversarial Discriminative Domain Adaptation (ADDA)
(Tzeng et al., 2017), Improved Wasserstein GAN-based objective with a unit-norm gradient regularizer
(Salimans et al., 2016), and MMD (Long & Wang, 2015), and compared them to our Dual objective.
To eliminate the influence of a particular discriminator and examine the stability of the objective
structure, we restricted the discriminator hypothesis space H to linear classifiers, because primal
objectives (ADDA, Improved WGAN) cannot be kernelized and MMD does not support multilayer
discriminators. This restriction limits the power of the resulting discriminator, thus leading to scores
lower than reported state of the art (usually with carefully chosen hyperparameters), but we are more
interested in trends in the behaviour of these objectives rather than in absolute reached values.
7
Workshop track - ICLR 2018
For each model, we varied learning rates and regularization parameters and ran each experiment
for 50 epochs to examine the behavior of these models in the long run. More details on choice of
hyperparameters is given in Section 10.1). In unsupervised domain adaptation, we do not have access
to target labels and thus cannot perform validation of stopping criteria. In fact, if labeled target
data were available then it could be used for fine-tuning the source model, rather than just doing
unsupervised learning. Therefore we evaluate the behavior of the models over multiple training
epochs to see which would be more stable in the face of uncertain stopping criteria in practical
domain adaptation scenarios.
Figure 4 shows the digit classification accuracies obtained by the four models on the target MNIST
dataset. The top row presents the distribution of accuracies at different epochs and the bottom row
shows the evolution of individual runs. From these results we see that on average descent iterations
with our Dual objective converged to satisfactory solutions under a considerably higher number
of learning rate and hyperparameter combinations compared to other methods. Our model often
stayed at peak performance, whereas all other methods most often slowly deviated from it. The
amount of instability demonstrates how important it is to choose exactly the right hyperparameters and
stopping criteria for these models. In contrast, our Dual objective (third column) clearly performs well
under the majority of the learning rates. WGAN often performs better than MMD and ADDA, but
experiences significant oscillations. We tried using different validation heuristics, such as considering
only runs that resulted in a significant drop in distance, but this did not significantly change these
trends (Section 10.3).
We conclude that our Dual method leads to more stable optimization without the need for choosing
an optimal stopping criterion and learning rates by cross-validation on test data.
7
D ISCUSSION AND F UTURE W ORK
While exact dual exists only in the logistic discriminator case, when one can use the duality and solve
the inner problem in closed form, we want to stress that our paper presents a more general framework
for alignment that can be extended to other classes of functions. More specifically, one can rewrite
the quadratic form in kernel logistic regression (3) as a Frobenius inner product of a kernel matrix Q
with a symmetric rank 1 alignment matrix S (outer product of alpha with itself).
d(A, B) = min
0≤αi ≤1
1
hααT , QiF + H(α) = min hSα , QiF + H(Sα )
Sα
2λ
The kernel matrix specifies distances between points and S chooses pairs that minimize the total
distance. This way the problem reduces to “maximizing the maximum agreement between the
alignment and similarity matrices” that in turn might be seen as replacing “adversity” in the original
problem with “cooperation” in the dual maximization problem. In our paper, S is a rank 1 matrix, but
we could choose different alignment matrix parameterizations and a corresponding regularizer that
would correspond to having a neural network discriminator in the adversarial problem or a Wasserstein
distance in Earth Mover’s Distance form. The resulting problem is not dual to minimization of any
existing adversarial distances, but exploits same underlying principle of “iteratively-reweighted
alignment matrix fitting” discussed in this paper.
We assert that in order to understand the basic properties of the resulting formulation, an in-depth discussion of the well-studied logistic case is no less important than the discussion involving complicated
deep models, which deserves a paper of its own. This paper proposes a more stable “cooperative”
problem reformulation rather than a new adversarial objective as many recent papers do.
8
C ONCLUSION
We presented an adversarial objective that does not lead to a min-max problem. We proposed using
the dual of the discriminator objective to improve the stability of distribution alignment, showed its
connection to MMD, and presented quantitative results for alignment of toy datasets and unsupervised
domain adaptation results on real-image classification datasets. Our results suggest that the proposed
dual optimization objective is indeed better suited for learning with gradient descent than the saddle
point objective that naturally arises from the original primal formulation of adversarial alignment.
Further attempts to use duality to reformulate other notions of statistical distances in adversarial
settings as computationally feasible minimization problems may be promising.
8
Workshop track - ICLR 2018
ADDA
WGAN
6
0.0
0.2
0.4
Dual
MMD
10
0.6
0.8
1.0
0.0
0.2
0.4
40
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Target accuracies
Target accuracies
0.7
0.6
0.5
0.4
0
20
40
0
20
40
0
20
40
0
20
40
Epochs
Figure 4: (Best viewed in color) Top row: Distribution of target test accuracies at different epochs
with different objectives during SVHN-MNIST domain adaptation. The red dashed line represents
source accuracy, therefore, a larger accuracy distribution mass to the right of (above) the red line
is better. These results suggest that our Dual objective leads to very minimal divergence from
the optimal solution under the majority of learning rates and hyperparameter combinations. The
other methods have lower solution stability, in descending order: Improved WGAN, MMD, ADDA.
Bottom row: Evolution of target test accuracy over epochs. Our Dual objective (third column) clearly
performs well under the majority of the learning rates. WGAN often performs better than MMD and
ADDA, but experiences significant oscillations. Different validation heuristics, such as considering
only runs that resulted in a significant drop in distance, did not significantly change these trends
(Section 10.3). The proportion of runs that outperformed the source baseline after 40 epochs were:
52.3% for Dual, 21.5% for WGAN, 17.1% for MMD and 6.9% for ADDA.
9
ACKNOWLEDGMENTS
This work was supported by NSF Award #1723379 and NSF CAREER Award #1559558.
R EFERENCES
Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks.
2017.
Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein GAN. CoRR, 2017.
S. Ben-David, J. Blitzer, K. Crammer, and F. Pereira. Analysis of representations for domain adaptation. NIPS,
2007.
Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised
pixel-level domain adaptation with generative adversarial networks. 2017.
Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid
of adversarial networks. In NIPS, 2015.
Gintare Karolina Dziugaite, Daniel M Roy, and Zoubin Ghahramani. Training generative neural networks via
Maximum Mean Discrepancy optimization. Uai, 2015.
Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015.
9
Workshop track - ICLR 2018
Boqing Gong, Kristen Grauman, and Fei Sha. Connecting the dots with landmarks: Discriminatively learning
domain-invariant features for unsupervised domain adaptation. In ICML, 2013.
Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160, 2016.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron
Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS. 2014.
Peter J Green. Iteratively reweighted least squares for maximum likelihood estimation, and some robust and
resistant alternatives. Journal of the Royal Statistical Society. Series B (Methodological), pp. 149–192, 1984.
Arthur Gretton. A Kernel Two-Sample Test. Journal of Machine Learning Research, 13:723–773, 2012.
Jiayuan Huang, Arthur Gretton, Karsten M Borgwardt, Bernhard Schölkopf, and Alex J Smola. Correcting
sample selection bias by unlabeled data. In Advances in neural information processing systems, pp. 601–608,
2007.
Phillip Isola, Jun Zhu, Tinghui Zhou, and Alexei Efros. Image-to-image translation with conditional adversarial
networks. 2016.
Tommi S Jaakkola and David Haussler. Probabilistic kernel regression models. In AISTATS, 1999.
Christian Ledig, Lucas Theis, Ferenc Huszar, and Jose Caballero. Photo-realistic single image super-resolution
using a generative adversarial network. 2016.
Yujia Li, Kevin Swersky, and Rich Zemel. Generative Moment Matching Networks. ICML, 2015.
Ming-Yu Liu and Oncel Tuzel. Coupled Generative Adversarial Networks. NIPS, 2016.
Mingsheng Long and Jianmin Wang. Learning transferable features with deep adaptation nets. In ICML, 2015.
Michael Mathieu, Junbo Zhao, Pablo Sprechmann, Aditya Ramesh, and Yann Lecun. Disentangling factors of
variation in deep representations using adversarial training. (NIPS), 2016.
Thomas P Minka. A comparison of numerical optimizers for logistic regression. 2003.
Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. CoRR, abs/1411.1784, 2014.
Arkadi Nemirovski. Prox-method with rate of convergence o (1/t) for variational inequalities with lipschitz
continuous monotone operators and smooth convex-concave saddle point problems. SIAM, 2004.
Yurii Nesterov. Primal-dual subgradient methods for convex problems. Mathematical programming, 2009.
Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. f-GAN: Training Generative Neural Samplers using
Variational Divergence Minimization. arXiv, pp. 17, 2016.
C Onsiderations, D Iagnostic, C Hallenges, P Ostbiopsy, and C Linical M Anagement. Unsupervised CrossDomain Image Generation. pp. 1–14, 2017.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional
generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved
techniques for training gans. CoRR, abs/1606.03498, 2016.
Baochen Sun and Kate Saenko. Deep CORAL: Correlation alignment for deep domain adaptation. ECCV, 2016.
Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing
for domain invariance. CoRR, abs/1412.3474, 2014.
Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and
tasks. In ICCV, 2015.
Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In
CVPR, 2017.
Jun Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycleconsistent adversarial networks. arXiv, 2017.
10
Workshop track - ICLR 2018
10
10.1
S UPPLEMENTARY
H YPERPARAMETER SPACE
The space of hyperparameters (mostly, learning rates for different parts of the network) we explored
in those runs was built as follows:
1) on SVHN-MNIST we found reasonably working vectors of hyper-parameters for each
method essentially by cross-validation on test
2) then we combined all these vectors into a set H and then inflated H by adding many
combinations of parts from these vectors and other similar and in-between vectors; for
example, if for two models two optimal sets of parameters included (1e-6, 1e-3) for the first
model and (1-4, 1e-1) for the second one, we would inflate H by adding all nine combinations
of ([1e-6, 1e-5 1e-4], [1e-3, 1e-2, 1e-1]) and varying model-specific parameters such as
regularization parameters
3) then we conducted experiments with combinations of pairs [vector from H, model]
10.2
I MPLEMENTATION DETAILS
During domain adaptation experiments, discriminators of all models that have a trainable discriminator
(ADDA, WGAN, Dual) were trained for five epochs prior to the beginning of actual adaptation,
which improved performance for all models. Kernel density estimates on Figure 4 were obtained
with a fixed 0.03 bandwidth and a linear boundary correction method. All algorithms took greyscale
images
P resizedPto 28 × 28 as input. Restrictions on α were imposed by adding an error term
+λ1 | αA − αB | and positivity by −λ2 | min(0, α)|. Hyperparameters for the Dual method can
be set so that restrictions above hold by the end of adaptation. To implement the dual algorithm in
mini-batch fashion we extracted slices of α∗ that correspond to points processed in this batch on each
iteration. Search space was built as explained above, learning rates and weight decay for updating
target network and discriminator were as follows:
• for target network update we used Adam with lr in [1e-6, 1e-5, 1e-4] and beta in [0.5, 0.9]
• for discriminator updates we used SGD with lr in [1e-6, 1e-5, 1e-7] or Adam with learning
rates in [2e-4, 1e-6, 1e-5] and beta in [0.5, 0.9]
• discriminator weight decay parameter was chosen from [2.5e-05, 0]
resulting into 90 possible combinations + model specific parameters.
10.3
VALIDATION HEURISTICS
Considering runs such that the following holds for different values of λ did not change results
significantly.
d(P, Q(0) ) − d(P, Q(T ) )
> λ.
maxt d(P, Q(t) ) − mint d(P, Q(t) )
11
Workshop track - ICLR 2018
ADDA
WGAN
6
0.0
0.2
0.4
Dual
MMD
10
0.6
0.8
1.0
0.0
0.2
0.4
12
40
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
| 1cs.CV
|
arXiv:1802.00682v1 [cs.AI] 2 Feb 2018
How do Humans Understand Explanations from
Machine Learning Systems?
An Evaluation of the Human-Interpretability of
Explanation
Menaka Narayanan*1 , Emily Chen*1 , Jeffrey He*1 , Been Kim2 , Sam Gershman1
and Finale Doshi-Velez1
1
Harvard University
2
Google Brain
February 5, 2018
Abstract
Recent years have seen a boom in interest in machine learning systems that can provide a
human-understandable rationale for their predictions or decisions. However, exactly what kinds
of explanation are truly human-interpretable remains poorly understood. This work advances our
understanding of what makes explanations interpretable in the specific context of verification.
Suppose we have a machine learning system that predicts X, and we provide rationale for
this prediction X. Given an input, an explanation, and an output, is the output consistent with
the input and the supposed rationale? Via a series of user-studies, we identify what kinds of
increases in complexity have the greatest effect on the time it takes for humans to verify the
rationale, and which seem relatively insensitive.
1
Introduction
Interpretable machine learning systems provide not only decisions or predictions but also explanation for their outputs. Explanations can help increase trust and safety by identifying when the
recommendation is reasonable when it is not. While interpretability has a long history in AI [Michie,
1988], the relatively recent widespread adoption of machine learning systems in real, complex
environments has lead to an increased attention to interpretable machine learning systems, with
applications including understanding notifications on mobile devices [Mehrotra et al., 2017, Wang
et al., 2016], calculating stroke risk [Letham et al., 2015], and designing materials [Raccuglia
et al., 2016]. Techniques for ascertaining the provenance of a prediction are also popular within the
1
machine learning community as ways for us to simply understand our increasingly complex models
[Lei et al., 2016, Selvaraju et al., 2016, Adler et al., 2016].
The increased interest in interpretability has resulted in many forms of explanation being
proposed, ranging from classical approaches such as decision trees [Breiman et al., 1984] to input
gradients or other forms of (possibly smoothed) sensitivity analysis [Selvaraju et al., 2016, Ribeiro
et al., 2016, Lei et al., 2016], generalized additive models [Caruana et al., 2015], procedures [Singh
et al., 2016], falling rule lists [Wang and Rudin, 2015], exemplars [Kim et al., 2014, Frey and Dueck,
2007] and decision sets [Lakkaraju et al., 2016]—to name a few. In all of these cases, there is a
face-validity to the proposed form of explanation: if the explanation was not human-interpretable,
clearly it would not have passed peer review.
That said, these works provide little guidance about when different kinds of explanation might
be appropriate, and within a class of explanations—such as decision-trees or decision-sets—what
are the limitations of human reasoning. For example, it is hard to imagine that a human would
find a 5000-node decision tree as interpretable as 5-node decision tree, for any reasonable notion
of interpretable. The reason the explanation is desired is also often left implicit. In Doshi-Velez
and Kim [2017], we point to a growing need for the interpretable machine learning community to
engage with the human factors and cognitive science of interpretability: we can spend enormous
efforts optimizing all kinds of models and regularizers, but that effort is only worthwhile if those
models and regularizers actually solve the original human-centered task of providing explanation.
Carefully controlled human-subject experiments provide an evidence-based approach to identify
what kinds of regularizers we should be using.
In this work, we make modest but concrete strides toward this large goal of quantifying what
makes explanation human-interpretable. We shall assume that there exists some explanation system
that generates the explanation—for example, there exist a variety of approaches that use perturbations around a point of interest to produce a local decision boundary [Ribeiro et al., 2016, Singh
et al., 2016]. Our question is: What kinds of explanation can humans most easily utilize? Below, we
describe the kind of task we shall explore, as well as the form of the explanation.
Choice of Task The question of what kinds of explanation a human can utilize implies the
presence of a downstream task. The task may be intrinsic—relying on just the explanation alone—
or extrinsic—relying on the explanation and other facts about the environment.1 Intrinsic tasks
include problems such as verification—given an input, output, and explanation, can the human user
verify that the output is consistent with the input and provided explanation?—and counterfactual
reasoning—given an input, output, and explanation, can the human subject identify what input
changes would result in a different output? In contrast, extrinsic tasks include goals such as safety—
given the input, output, explanation, and observations of the world, does the explanation help the
human user identify when the agent is going to make a mistake?—and trust—given the input,
output, explanation, and observations of the world, does the explanation increase the human user’s
confidence in the agent? Evaluation on extrinsic tasks, while ultimately what we care about, require
careful experimental design that ensures all subjects have similar knowledge and assumptions about
1
As noted in Herman [2017], explanation can also be used to persuade rather than inform; we exclude that use-case
here.
2
the environment. One must also tease apart conflations between human perception of knowledge
with actual knowledge: for example, it may be possible to manipulate trust arbitrarily separately
from prediction accuracy. Thus, evaluating extrinsic tasks is more challenging than intrinsic ones.
In this work, we will focus on the simplest intrinsic setting: verification. Given a specific input,
explanation, and output, can we quickly determine whether the output is consistent with the input
and explanation? Such a setting may arise in consumer recommendation scenarios—for example,
a salesperson may wish to ensure that a specific product recommendation is consistent with a
consumer’s preferences. Starting simple also provides an opportunity to explore aspects relevant to
the experimental design.
Choice of Explanation Form As mentioned above, there have been many forms of explanation
proposed, ranging from decision trees to gradients of neural networks. In this work, we consider
explanations that are in the form of decision sets. Decision sets are a particular form of procedure
consisting of a collection of cases, each mapping some function of the inputs to a collection of
outputs. An example of a decision set is given below
Figure 1: Example of a decision set explanation.
where each line contains a clause in disjunctive normal form (an or-of-ands) of the inputs, which, if
true, provides a way to verify the output (also in disjunctive normal form). As argued in Lakkaraju
et al. [2016], decision sets are relatively easy for humans to parse given a specific instance, because
they can scan for the rule that applies and choose the accompanying output. Decision sets (also
known as rule sets) also enjoy a long history of optimization techniques, ranging from Frank and
Witten [1998], Cohen [1995], Clark and Boswell [1991] to Lakkaraju et al. [2016].
However, we can also already see that there are many factors that could potentially make the
decision set more or less challenging to follow: in addition to the number of lines, there are ways in
which terms interact in the disjunctions and conjunctions, and, more subtly, aspects such as how
often terms appear and whether terms represent intermediate concepts. Which of these factors are
most relevant when it comes to a human’s ability to process and utilize an explanation, and to what
extent? The answer to this question has important implications for the design of explanation systems
in interpretable machine learning, especially if we find that our explanation-processing ability is
relatively robust to variation in some factors but not others.
Contributions The core contribution of this work is to provide an empirical grounding for what
kinds of explanations humans can utilize. We find that while almost all increases in the complexity
of an explanation result in longer response times, some types of complexity—such as the number of
lines, or the number of new concepts introduced—have a much bigger effect than others—such as
variable repetition. We also find some unintuitive patterns, such as how participants seem to prefer
3
an explanation that requires a single, more complex line to one that spans multiple simpler lines
(each defining a new concept for the next line). While more work is clearly needed in this area,
we take initial steps in identifying what kinds of factors are most important to optimize for when
providing explanation to humans.
2
Related Work
Interpretable machine learning methods aim to optimize models for both succinct explanation and
predictive performance. Common types of explanation include regressions with simple, humansimulatable functions [Caruana et al., 2015, Kim et al., 2015a, Rüping, 2006, Bucilu et al., 2006,
Ustun and Rudin, 2016, Doshi-Velez et al., 2015, Kim et al., 2015b, Krakovna and Doshi-Velez,
2016, Hughes et al., 2016, Jung et al., 2017], various kinds of logic-based methods [Wang and Rudin,
2015, Lakkaraju et al., 2016, Singh et al., 2016, Liu and Tsang, 2016, Safavian and Landgrebe, 1991,
Wang et al., 2017], techniques for extracting local explanations from black-box models [Ribeiro
et al., 2016, Lei et al., 2016, Adler et al., 2016, Selvaraju et al., 2016, Smilkov et al., 2017, Shrikumar
et al., 2016, Kindermans et al., 2017, Ross et al., 2017], and visualization [Wattenberg et al., 2016].
There exist a range of technical approaches to derive each form of explanation, whether it be learning
sparse models [Mehmood et al., 2012, Chandrashekar and Sahin, 2014], monotone functions [Canini
et al., 2016], or efficient logic-based models [Rivest, 1987]. Related to our work, there also exists
a history of identify human-relevant concepts from data, including disentangled representations
[Chen et al., 2016] and predicate invention in inductive logic programming [Muggleton et al., 2015].
While the algorithms are sophisticated, the measures of interpretability are often not—it is common
for researchers to simply appeal to the face-validity of the results that they find (i.e., “this result
makes sense to the human reader”) [Caruana et al., 2015, Lei et al., 2016, Ribeiro et al., 2016].
In parallel, the literature on explanation in psychology also offers several general insights into
the design of interpretable AI systems. For example, humans prefer explanations that are both simple
and highly probable [Lombrozo, 2007]. Human explanations typically appeal to causal structure
[Lombrozo, 2006] and counterfactuals [Keil, 2006]. Miller [1956] famously argued that humans can
hold about seven items simultaneously in working memory, suggesting that human-interpretable
explanations should obey some kind of capacity limit (importantly, these items can correspond to
complex cognitive chunks—for example, ‘CIAFBINSA’ is easier to remember when it is chunked
as ‘CIA’, ‘FBI’, ‘NSA.’). Orthogonally, Kahneman [2011] notes that humans have different modes
of thinking, and larger explanations might push humans into a more careful, rational thinking
mode. Machine learning researchers can convert these concepts into notions such as sparsity or
simulatability, but the work to determine answers to questions such as “how sparse?” or “how long?”
requires empirical evaluation.
Existing studies evaluting the human-interpretability of explanation often fall into the A-B
test framework, in which a proposed model is being compared to some competitor, generally on
an intrinsic task. For example, Kim et al. [2014] showed that human subjects’ performance on a
classification task was better when using examples as representation than when using non-examplebased representation. Lakkaraju et al. [2016] performed a user study in which they found subjects
are faster and more accurate at describing local decision boundaries based on decision sets rather
4
than rule lists. Subramanian et al. [1992] found that users prefer decision trees to tables in games,
whereas Huysmans et al. [2011] found users prefer, and are more accurate, with decision tables
rather than other classifiers in a credit scoring domain. Hayete and Bienkowska [2004] found a
preference for non-oblique splits in decision trees (see Freitas [2014] for more detailed survey).
These works provide quantitative evaluations of the human-interpretability of explanation, but rarely
identify what properties are most essential for what contexts—which is critical for generalization.
Specific application areas have also evaluated the desired properties of an explanation within
the context of the application. For example, Tintarev and Masthoff [2015] provides a survey in
the context of recommendation systems, noting differences between the kind of explanations that
manipulate trust [Cosley et al., 2003] and the kind that increase the odds of a good decision [Bilgic
and Mooney, 2005]. In many cases, these studies are looking at whether the explanation has an effect,
sometimes also considering a few different kinds of explanation (actions of similar customers, etc.).
Horsky et al. [2012] describe how presenting the right clinical data alongside a decision support
recommendation can help with adoption and trust. Bussone et al. [2015] found that overly detailed
explanations from clinical decision support systems enhance trust but also create over-reliance;
short or absent explanations prevent over-reliance but decrease trust. These studies span a variety of
extrinsic tasks, and again given the specificity of each explanation type, identifying generalizable
properties is challenging.
Closer to the objectives of the proposed work, Kulesza et al. [2013] performed a qualitative study
in which they varied the soundness (nothing but the truth) and the completeness (the whole truth) of
an explanation in a recommendation system setting. They found completeness was important for
participants to build accurate mental models of the system. Allahyari and Lavesson [2011], Elomaa
[2017] also find that larger models can sometimes be more interpretable. Schmid et al. [2016] find
that human-recognizable intermediate predicates in inductive knowledge programs can sometimes
improve simulation time. Poursabzi-Sangdeh et al. [2017] manipulate the size and transparency
of an explanation and find that longer explanations and black-box models are harder to simulate
accurately (even given many instances) on a real-world application predicting housing prices. Our
work fits into this category of empirical study of explanation evaluation; we perform controlled
studies on a pair of synthetic application to assess the effect of a large set of explanation parameters.
3
Methods
Our main research question is to determine what properties of decision sets are most relevant for
human users to be able to utilize the explanations for verification. In order to carefully control
various properties of the explanation and the context, in the following we shall present human
subjects with explanations that could have been machine-generated, but were in fact generated by
us. Before describing our experiment, we emphasize that while our explanations are not actually
machine-generated, our findings provide suggestions to designers of interpretable machine learning
systems about what parameters affect the usability of an explanation, and which should be optimized
when producing explanations.
5
3.1
Factors Varied
Even within decision sets, there are a large number of ways in which the explanations could
be varied. Following initial pilot studies (see Appendix), we chose to focus on the three main
kinds of variation (described below). We also tested on two different domains—a faux recipe
recommendation domain and a faux clinical decision support domain—to see if the context would
result in different explanation processing while other factors were held constant.
Explanation Variation We explored the following sources of explanation variation:
• V1: Explanation Size. We varied the size of the explanation across two dimensions: the total
number of lines in the decision set, and the maximum number of terms within the output
clause. The first corresponds to increasing the vertical size of the explanation—the number of
cases—while the second corresponds to increasing the horizontal size of the explanation—the
complexity of each case. We focused on output clauses because they were harder to parse:
input clauses could be quickly scanned for setting-related terms, but output clauses had to
be read through and processed completely to verify an explanation. We hypothesized that
increasing the size of the explanation across either dimension would increase the time required
to perform the verification task.
• V2: Creating New Types of Cognitive Chunks. In Figure 2, the first line of the decision
set introduces a new cognitive chunk: if the alien is ‘checking the news’ or ‘coughing,’ that
corresponds to a new concept ‘windy.’2 On one hand, creating new cognitive chunks can make
an explanation more succinct. On the other hand, the human must now process an additional
idea. We varied two aspects related to new cognitive chunks. First, we simply adjusted the
number of new cognitive chunks present in the explanation. All of the cognitive chunks were
necessary for verification, to ensure that the participant had to traverse all of concepts instead
of skimming for the relevant one.
Second, we tested whether it was more helpful to introduce a new cognitive chunk or leave
it implicit: for example, instead of introducing a concept ‘windy’ for ‘checking the news or
coughing,’ (explicit) we could simply include ‘checking the news or coughing’ wherever
windy appeared (implicit). We hypothesized that adding cognitive chunks would increase the
time required to process an explanation, because the user would have to consider more lines
in the decision list to come to a conclusion. However, we hypothesized that it would still be
more time-efficient to introduce the new chunk rather than having long clauses that implicitly
contained the meaning of the chunk.
• V3: Repeated Terms in an Explanation. Another factor that might affect humans’ ability
to process an explanation is how often terms are used. For example, if input conditions in the
decision list have little overlap, then it may be faster to find the appropriate one because there
are fewer relevant cases to consider. We hypothesized that if an input condition appeared in
several lines of the explanation, this would increase the time it took to search for the correct
2
One can imagine this being akin to giving names to nodes in a complex model, such as a neural network.
6
rule in the explanation. (Repeated terms was also a factor used by Lakkaraju et al. [2016] to
measure interpretability.)
Domain Variation Below we describe the two contexts, or domains, that we used in our experiments: recipe recommendations (familiar) and clinical decision support (unfamiliar). The domains
were designed to feel very different but such that the verification tasks could be made to exactly
parallel each other, allowing us to investigate the effect of the domain context in situations when
the form of the explanation was exactly the same. We hypothesized that these trends would be
consistent in both the recipe and clinical domains.
Alien Recipes. In the first domain, study participants were told that the machine learning system
had studied a group of aliens and determined each of their individual food preferences in various
settings (e.g., snowing, weekend). Each task involved presenting participants with the setting
(the input), the systems’s description of the current alien’s preferences (the explanation), and a
set of recommended ingredients (the output). The user was then asked whether the ingredients
recommendation was a good one. This scenario represents a setting in which customers may wish
to know why a certain product or products were recommended to them. Aliens were introduced
in order to avoid the subject’s prior knowledge or preferences about settings and food affecting
their responses; each task involved a different alien so that each explanation could be unique. All
non-literals (e.g. what ingredients were spices) were defined in a dictionary so that all participants
would have the same cognitive chunks.
Alien Medicine In the second domain, study participants were told that the machine learning
system had studied a group of aliens and determined personalized treatment strategies for various
symptoms (e.g. sore throat). Each task involved presenting participants with the symptoms (the
input), the system’s description of the alien’s personalized treatment strategy (the explanation), and
a set of recommended drugs (the output). The user was then asked whether the drugs recommended
were appropriate. This scenario closely matches a clinical decision support setting in which a system
might suggest a treatment given a patient’s symptoms, and the clinician may wish to know why the
system chose a particular treatment.
As before, aliens were chosen to avoid the subject’s prior medical knowledge from affecting
their responses; each task involved personalized medicine so that each explanation could be unique.
We chose drug names that corresponded with the first letter of the illness (e.g. antibiotic medications
were Aerove, Adenon and Athoxin) so as to replicate the level of ease and familiarity of food names.
Again, all drug names and categories were provided in a dictionary so that participants would have
the same cognitive chunks.
In our experiments, we maintained an exact correspondence between inputs (setting vs. symptoms), outputs (foods vs. drugs), categories (food categories vs. drug categories), and the forms
of the explanation. These parallels allowed us to test whether changing the domain from a relatively familiar, low-risk product recommendation setting to a relatively unfamiliar, higher-risk
decision-making setting affected how subjects utilized the explanations for verification.
7
3.2
Experimental Design and Interface
The three kinds of variation and two domains resulted in six total experiments. In the recipe domain,
we held the list of ingredients, food categories, and possible input conditions constant. Similarly, in
the clinical domain, we held the list of symptoms, medicine categories, and possible input conditions
constant. The levels were as follows:
• Length of the explanation (V1). We manipulated the length of the explanation (varying
between 2, 6, and 10 lines) and the length of the output clause (varying between 2 and 5
terms). Each combination was tested twice within a experiment, for a total of 12 questions.
• Introducing new concepts (V2): We manipulated the number of cognitive chunks introduced
(varying from 1 to 5), and whether they were embedded into the explanation or abstracted out
into new cognitive chunks. Each combination was tested once within a experiment, for a total
of 10 questions.
• Repeated terms (V3): We manipulated the number of times the input conditions appeared
in the explanation (varying from 1 to 5) and held the number of lines and length of clauses
constant. Each combination was tested twice within a experiment, for a total of 10 questions.
The outputs were consistent with the explanation and the input 50% of the time, so subjects could
not simply learn that the outputs were always (in)consistent.
Participants were recruited via Amazon Mechanical Turk. Before starting the experiment, they
were given a tutorial on the verification task and the interface. Then they were given a set of
three practice questions. Following the practice questions, they started the core questions for each
experiment. They were told that their primary goal was accuracy, and their secondary goal was speed.
While the questions were the same for all participants, the order of the questions was randomized
for every participant. Each participant only participated in one of the experiments. For example, one
participant might have completed a 12-question experiment on the effect of varying explanation
length in the recipe domain, while another would have completed 10-question experiment on the
effect of repeated terms in the clinical domain. Experiments were kept short to avoid subjects getting
tired.
Metrics We recorded three metrics: response time, accuracy, and subjective satisfaction. Response
time was measured as the number of seconds from when the task was displayed until the subject
hit the submit button on the interface. Accuracy was measured if the subject correctly identified
whether the output was consistent with the input and the explanation (a radio button). After each
submitting their answer for each question, the participant was also asked to subjectively rate the
quality of the explanation on a scale from one to ten.
8
(a) Recipe Domain
(b) Clinical Domain
Figure 2: Screenshots of our interface for a task. In the Recipe Domain, the supposed machine
learning system has recommended the ingredients in the lower left box, based on its observations of
the alien (center box). The top box shows the system’s explanation. In this case, the recommended
ingredients are consistent with the explanation and the inputs: The input conditions are weekend
and windy (implied by coughing), and the recommendation of fruit and grains follows from the
last line of the explanation. In the Clinical Domain, the supposed machine learning system has
recommended the medication in the lower box, based on its observations of the alien’s symptoms
9
(center box). The top box shows the system’s reasoning.
The interface has exactly the same form as
the recipe domain.
Table 1: Participant Demographics. There were no patients over 69 years old. 4.2% of participants
reported “other” for their education level. The rates of participants from Australia, Europe, Latin
America, and South America were all less than 0.5%. (All participants were included in the analyses,
but we do not list specific proportions for them for brevity.)
Feature
Age
Education
Gender
Region
Category : Proportion
18-34 : 59.0%
35-50 : 35.1%
High School : 28.5% Bachelor’s : 52.4%
Male : 58.8%
Female : 41.2%
US/Canada : 87.1%
Asia : 11.4%
51-69 : 5.9%
Beyond Bachelor’s: 14.9%
Experimental Interface Figure 2 shows our interfaces for the Recipe and Clinical domains. The
observations section (middle) refers to the inputs into the algorithm. The recommendation section
refers to the output of the algorithm. The preferences section (top) contains the explanation—the
reasoning that the supposed machine learning system used to suggest the output (i.e., recommendation) given the input, presented as a procedure in the form of a decision set. Finally, the ingredients
section in the Recipe domain (and the disease medications section in the Clinical domain) contained
a dictionary of cognitive chunks relevant to the experiment (for example, the fact that bagels, rice,
and pasta are all grains). Including this list explicitly allowed us to control for the fact that some
human subjects may be more familiar with various concepts than others.
The choice of location for these elements was chosen based on pilot studies—while an ordering
of input, explanation, output might make more sense for an AI expert, we found that presenting the
information in the format of Figure 2 seemed to be easier for subjects to follow in our preliminary
explorations. We also found that presenting the decision set as a decision set seemed easier to follow
than converting it into paragraphs. Finally, we colored the input conditions in blue and outputs in
orange within the explanation. We found that this highlighting system made it easier for participants
to parse the explanations for input conditions.
4
Results
We recruited 100 subjects for each our six experiments, for a total of 600 subjects all together.
Table 1 summarizes the demographics of our subjects across the experiments. Most participants
were from the US or Canada (with the remainder being almost exclusively from Asia) and were less
than 50 years old. A majority had a Bachelor’s degree. There were somewhat more male participants
than female. We note that US and Canadian participants with moderate to high education dominate
this survey, and results may be different for people from different cultures and backgrounds.
All participants completed the full task (each survey was only 10-12 questions). In the analysis
below, however, we exclude participants who did not get all three initial practice questions or all two
of the additional practice questions correct. While this may have the effect of artificially increasing
the accuracy rates overall—we are only including participants who could already perform the task
to a reasonable extent—this criterion helped filter the substantial proportion of participants who
10
Table 2: Number of participants who met our inclusion criteria for each experiment.
Explanation Size
New Cognitive Chunks
Variable Repetition
Recipe Domain
N=88
N=77
N=70
Clinical Domain
N=73
N=73
N=71
(a) Recipe V1 Accuracy
(b) Recipe V2 Accuracy
(c) Recipe V3 Accuracy
(d) Clinical V1 Accuracy
(e) Clinical V2 Accuracy
(f) Clinical V3 Accuracy
Figure 3: Accuracy across the six experiments. Vertical lines indicate standard errors.
were simply breezing through the experiment to get their payment. We also excluded one participant
in the clinical version of the cognitive chunks experiment who did get sufficient practice questions
correct but then took more than ten minutes to answer a single question. Table 2 describes the total
number of participants that remained in each experiment out of the original 100 participants.
Figures 3 and 4 present the accuracy and response time across all six experiments, respectively.
Response time is shown for subjects who correctly answered the questions. Figure 5 shows the
trends in the participants’ subjective evaluation—whether they thought the explanation was easy to
follow or not. We evaluated the statistical significance of the trends in these figures using a linear
regression for the continuous outputs (response time, subjective score) and a logistic regression
for binary outputs (accuracy). For each outcome, one regression was performed for each of the
experiments V1, V2, and V3. If an experiment had more than one independent variable—e.g.
number of lines and terms in output—we performed one regression with both variables. Regressions
were performed with the statsmodels library [Seabold and Perktold, 2010] and included an intercept
term. Table 3 summarizes these results.
11
Table 3: Significance tests for each factor. Linear and logistic regression weights were computed for
continuous and binary outputs respectively (the subjective evaluations were treated as continuous).
A single regression was computed for each of V1, V2, and V3. Highlighted p-values are significant
at α = 0.05 with a Bonferroni multiple comparisons correction across all tests of all experiments.
Accuracy
Recipe
Factor
weight p-value
Explanation Length (V1)
-0.0116 0.00367
Number of Output Terms (V1)
-0.0161 0.0629
Number of Cognitive Chunks (V2)
0.0221 0.0377
Implicit Cognitive Chunks (V2)
0.0147 0.625
Number of Variable Repetitions (V3) -0.017
0.104
Response Time
Recipe
Factor
weight p-value
Explanation Length (V1)
3.77
2.24E-34
Number of Output Terms (V1)
1.34
0.0399
Number of Cognitive Chunks (V2)
8.44
7.01E-18
Implicit Cognitive Chunks (V2)
-15.3
2.74E-08
Number of Variable Repetitions (V3) 2.4
0.000659
Subjective Evaluation
Recipe
Factor
weight p-value
Explanation Length (V1)
-0.165
5.86E-16
Number of Output Terms (V1)
-0.187
2.12E-05
Number of Cognitive Chunks (V2)
-0.208
1.93E-05
Implicit Cognitive Chunks (V2)
0.297
0.0303
Number of Variable Repetitions (V3) -0.179
5.71E-05
12
Clinical
weight
p-value
-0.0171 0.000127
0.00685 0.48
0.0427
0.00044
0.0251
0.464
-0.0225 0.0506
Clinical
weight
p-value
3.3
5.73E-22
1.68
0.0198
4.6
1.71E-05
-11.8
0.000149
2.13
0.0208
Clinical
weight
p-value
-0.186
1.28E-19
-0.0335 0.444
-0.0208 0.703
0.365
0.018
-0.149
0.000771
(a) Recipe V1 Time
(b) Recipe V2 Time
(c) Recipe V3 Time
(d) Clinical V1 Time
(e) Clinical V2 Time
(f) Clinical V3 Time
Figure 4: Response times across the six experiments. Responses were normalized by subtracting
out subject mean to create centered values, and only response times for those subjects who got the
question right are included. Vertical lines indicate standard errors.
In general, greater complexity results in higher response times and lower satisfaction. Increasing the number of lines, the number of terms within a line, adding new concepts, and repeating
variables all increase the complexity of an explanation in various ways. In figure 4, we see that all
of these changes increase response time. The effect of adding lines to scan results in the biggest
increases in response time, while the effect of increasing the number of variable repetitions is more
subtle. Making new concepts explicit also consistently results in increased response time. This effect
may have partly been because adding a new concept explicitly adds a line, while adding it implicitly
increases the number of terms in a line—and from V1, we see that the effect of the number of
lines is larger than the effect of the number of terms. However, overall this effect seems larger than
just adding lines (note the scales of the plots). Subjective scores appear to correlate inversely with
complexity and response time.
Perhaps most unexpected was that participants both took longer and seemed less happy when
new cognitive chunks were made explicit rather than implicitly embedded in a line—we might
have expected that even if the explanation took longer to process, it would have been in some
senses easier to follow through. It would be interesting to unpack this effect in future experiments,
especially if participant frustration came from there now being multiple relevant lines, rather than
just one. Future experiments could also highlight terms from the input in the explanation, to make it
easier for participants to find the lines of potential relevance.
13
(a) Recipe V1 Evaluation
(b) Recipe V2 Evaluation
(c) Recipe V3 Evaluation
(d) Clinical V1 Evaluation
(e) Clinical V2 Evaluation
(f) Clinical V3 Evaluation
Figure 5: Subjective evaluation of explanations in three experiments. Participants were asked to
rate each explanation from 1 to 10. Responses were normalized by subtracting out subject mean to
create centered values.
Different explanation variations had little effect on accuracy. While there were strong, consistent effects from explanation variation on response time and satisfaction, these variations seemed
to have much less effect on accuracy. There existed general decreasing trends in accuracy with
respect to explanation length and the number of variable repetitions, and potentially some increasing
trends with the number of new concepts introduced. However, few were statistically significant.
This serves us as an interesting controlled comparison. In other words, we can now observe effects
of different factors, holding the accuracy constant. The lack of effect may be because subjects were
told to perform the task as quickly as they could without making mistakes, and our filtering also
removed subjects prone to making many errors in the practice questions. Thus, the effect of a task
being harder would be to increase response time, rather than decrease accuracy.
The differences in direction of the trends—some increases in complexity perhaps increasing
accuracy, others decreasing it—are consistent with findings that sometimes more complex tasks
force humans into slower but more careful thinking, while in other cases increased complexity
can lead to higher errors [Kahneman, 2011]. Is it the case that the factors that resulted in the
largest increases in response time (new concepts) also force the most concentration? While these
experiments cannot differentiate these effects, future work to understand these phenomena may help
us identify what kinds of increased complexities in explanation are innocuous or even useful and
which push the limits of our processing abilities.
14
Trends are consistent across recipe and clinical domains In all experiments, the general trends
in the metrics are consistent across both the recipe and clinical domains. Sometimes an effect is
much weaker or unclear, but never is an effect clearly reversed. We believe this bodes well for there
being a set of general principles for guiding explanation design, just as there exist design principles
for interfaces and human-computer interaction. However, one small pattern can be noted in Figure
5, which shows lower satisfaction for the clinical domain than the recipe domain. This could be due
to the fact that people felt more agitated about performing poorly in the medical domain than the
clinical domain.
5
Discussion and Conclusion
Identifying how different factors affect a human’s ability to utilize explanation is an essential piece
for creating interpretable machine learning systems—we need to know what to optimize. What
factors have the largest effect, and what kind of effect? What factors have relatively little effect?
Such knowledge can help us expand to faithfulness of the explanation to what it is describing with
minimal sacrifices in human ability to process the explanation.
In this work, we investigated how the ability of humans to perform a simple task—verifying
whether an output is consistent with an explanation and input—varies as a function of explanation
size, new types of cognitive chunks and repeated terms in the explanation. We tested across two
domains, carefully controlling for everything but the domain.
We summarized some intuitive and some counter intuitive discoveries—as any increase in
explanation complexity increases response time and decreases subjective satisfaction with the
explanation—some patterns were not so obvious. We had not expected that embedding a new
concept would have been faster to process and more appealing than creating a new definition. We
also found that new concepts and the number of lines increase response time more than variable
repetition or longer lines. It would be interesting to verify the magnitude of these sensitivities
on other tasks, such as forward simulation or counterfactual reasoning, to start building a more
complete picture of what we should be optimizing our explanations for.
More broadly, there are many interesting directions regarding what kinds of explanation are best
in what contexts. Are there universals that make for interpretable procedures, whether they be cast as
decision sets, decision trees, or more general pseudocode; whether the task is verification, forward
simulation, or counterfactual reasoning? Do these universals also carry over to regression settings?
Or does each scenario have its own set of requirements? When the dimensionality of an input gets
very large, do trade-offs for defining intermediate new concepts change? A better understanding of
these questions is critical to design systems that can provide rationale to human users.
Acknowledgements The authors acknowledge PAIR at Google and the Harvard Berkman Klein
Center for their support.
15
References
Philip Adler, Casey Falk, Sorelle A Friedler, Gabriel Rybeck, Carlos Scheidegger, Brandon Smith,
and Suresh Venkatasubramanian. Auditing black-box models for indirect influence. In Data
Mining (ICDM), 2016 IEEE 16th International Conference on, pages 1–10. IEEE, 2016.
Hiva Allahyari and Niklas Lavesson. User-oriented assessment of classification model understandability. In 11th scandinavian conference on Artificial intelligence. IOS Press, 2011.
Mustafa Bilgic and Raymond J Mooney. Explaining recommendations: Satisfaction vs. promotion.
In Beyond Personalization Workshop, IUI, volume 5, page 153, 2005.
Leo Breiman, Jerome Friedman, Charles J Stone, and Richard A Olshen. Classification and
regression trees. CRC press, 1984.
Cristian Bucilu, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings
of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining,
pages 535–541. ACM, 2006.
Adrian Bussone, Simone Stumpf, and Dympna O’Sullivan. The role of explanations on trust
and reliance in clinical decision support systems. In Healthcare Informatics (ICHI), 2015
International Conference on, pages 160–169. IEEE, 2015.
Kevin Canini, Andy Cotter, MR Gupta, M Milani Fard, and Jan Pfeifer. Fast and flexible monotonic
functions with ensembles of lattices. Advances in Neural Information Processing Systems (NIPS),
2016.
Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intelligible
models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings
of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
pages 1721–1730. ACM, 2015.
Girish Chandrashekar and Ferat Sahin. A survey on feature selection methods. Computers &
Electrical Engineering, 40(1):16–28, 2014.
Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:
Interpretable representation learning by information maximizing generative adversarial nets,
2016.
Peter Clark and Robin Boswell. Rule induction with cn2: Some recent improvements. In European
Working Session on Learning, pages 151–163. Springer, 1991.
William W Cohen. Fast effective rule induction. In Proceedings of the twelfth international
conference on machine learning, pages 115–123, 1995.
16
Dan Cosley, Shyong K Lam, Istvan Albert, Joseph A Konstan, and John Riedl. Is seeing believing?:
how recommender system interfaces affect users’ opinions. In Proceedings of the SIGCHI
conference on Human factors in computing systems, pages 585–592. ACM, 2003.
Finale Doshi-Velez and Been Kim. A roadmap for a rigorous science of interpretability. arXiv
preprint arXiv:1702.08608, 2017.
Finale Doshi-Velez, Byron Wallace, and Ryan Adams. Graph-sparse lda: a topic model with
structured sparsity. Association for the Advancement of Artificial Intelligence, 2015.
Tapio Elomaa. In defense of c4. 5: Notes on learning one-level decision trees. ML-94, 254:62, 2017.
Eibe Frank and Ian H Witten. Generating accurate rule sets without global optimization. 1998.
Alex A Freitas. Comprehensible classification models: a position paper. ACM SIGKDD explorations
newsletter, 15(1):1–10, 2014.
Brendan J Frey and Delbert Dueck. Clustering by passing messages between data points. science,
315(5814):972–976, 2007.
Boris Hayete and Jadwiga R Bienkowska. Gotrees: Predicting go associations from proteins.
Biocomputing 2005, page 127, 2004.
Bernease Herman. The promise and peril of human evaluation for model interpretability. arXiv
preprint arXiv:1711.07414, 2017.
Jan Horsky, Gordon D Schiff, Douglas Johnston, Lauren Mercincavage, Douglas Bell, and Blackford
Middleton. Interface design principles for usable decision support: a targeted review of best
practices for clinical prescribing interventions. Journal of biomedical informatics, 45(6):1202–
1216, 2012.
Michael C Hughes, Huseyin Melih Elibol, Thomas McCoy, Roy Perlis, and Finale Doshi-Velez.
Supervised topic models for clinical interpretability. arXiv preprint arXiv:1612.01678, 2016.
J. Huysmans, K. Dejaeger, C. Mues, J. Vanthienen, and B. Baesens. An empirical evaluation of the
comprehensibility of decision table, tree and rule based predictive models. DSS, 2011.
Jongbin Jung, Connor Concannon, Ravi Shroff, Sharad Goel, and Daniel G Goldstein. Simple rules
for complex decisions. 2017.
Daniel Kahneman. Thinking, fast and slow. Macmillan, 2011.
Frank Keil. Explanation and understanding. Annu. Rev. Psychol., 2006.
B. Kim, C. Rudin, and J.A. Shah. The Bayesian Case Model: A generative approach for case-based
reasoning and prototype classification. In NIPS, 2014.
17
Been Kim, Elena Glassman, Brittney Johnson, and Julie Shah. iBCM: Interactive bayesian case
model empowering humans via intuitive interaction. 2015a.
Been Kim, Julie Shah, and Finale Doshi-Velez. Mind the gap: A generative approach to interpretable
feature selection and extraction. In Advances in Neural Information Processing Systems, 2015b.
Pieter-Jan Kindermans, Kristof T Schütt, Maximilian Alber, Klaus-Robert Müller, and Sven Dähne.
Patternnet and patternlrp–improving the interpretability of neural networks. arXiv preprint
arXiv:1705.05598, 2017.
Viktoriya Krakovna and Finale Doshi-Velez. Increasing the interpretability of recurrent neural
networks using hidden markov models. arXiv preprint arXiv:1606.05320, 2016.
Todd Kulesza, Simone Stumpf, Margaret Burnett, Sherry Yang, Irwin Kwan, and Weng-Keen Wong.
Too much, too little, or just right? ways explanations impact end users’ mental models. In Visual
Languages and Human-Centric Computing (VL/HCC), 2013 IEEE Symposium on, pages 3–10.
IEEE, 2013.
Himabindu Lakkaraju, Stephen H Bach, and Jure Leskovec. Interpretable decision sets: A joint
framework for description and prediction. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pages 1675–1684. ACM, 2016.
Tao Lei, Regina Barzilay, and Tommi Jaakkola. Rationalizing neural predictions. arXiv preprint
arXiv:1606.04155, 2016.
Benjamin Letham, Cynthia Rudin, Tyler H McCormick, David Madigan, et al. Interpretable
classifiers using rules and bayesian analysis: Building a better stroke prediction model. The
Annals of Applied Statistics, 9(3):1350–1371, 2015.
Weiwei Liu and Ivor W Tsang. Sparse perceptron decision tree for millions of dimensions. In AAAI,
pages 1881–1887, 2016.
Tania Lombrozo. The structure and function of explanations. Trends in cognitive sciences, 10(10):
464–470, 2006.
Tania Lombrozo. Simplicity and probability in causal explanation. Cognitive psychology, 55(3):
232–257, 2007.
Tahir Mehmood, Kristian Hovde Liland, Lars Snipen, and Solve Sæbø. A review of variable
selection methods in partial least squares regression. Chemometrics and Intelligent Laboratory
Systems, 118:62–69, 2012.
Abhinav Mehrotra, Robert Hendley, and Mirco Musolesi. Interpretable machine learning for
mobile notification management: An overview of prefminer. GetMobile: Mobile Computing and
Communications, 21(2):35–38, 2017.
18
Donald Michie. Machine learning in the next five years. In Proceedings of the 3rd European
Conference on European Working Session on Learning, pages 107–122. Pitman Publishing, 1988.
George A Miller. The magical number seven, plus or minus two: some limits on our capacity for
processing information. Psychological review, 63(2):81, 1956.
Stephen H Muggleton, Dianhuan Lin, and Alireza Tamaddoni-Nezhad. Meta-interpretive learning
of higher-order dyadic datalog: Predicate invention revisited. Machine Learning, 100(1):49–73,
2015.
Forough Poursabzi-Sangdeh, Daniel G. Goldstein, Jake M. Hofman, Jennifer Wortman Vaughan,
and Hanna Wallach. Manipulating and measuring model interpretability. In NIPS Workshop on
Transparent and Interpretable Machine Learning in Safety Critical Environments, 2017.
Paul Raccuglia, Katherine C Elbert, Philip DF Adler, Casey Falk, Malia B Wenny, Aurelio Mollo,
Matthias Zeller, Sorelle A Friedler, Joshua Schrier, and Alexander J Norquist. Machine-learningassisted materials discovery using failed experiments. Nature, 533(7601):73–76, 2016.
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the
predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, pages 1135–1144. ACM, 2016.
Ronald L Rivest. Learning decision lists. Machine learning, 2(3):229–246, 1987.
Andrew Ross, Michael C Hughes, and Finale Doshi-Velez. Right for the right reasons: Training
differentiable models by constraining their explanations. In International Joint Conference on
Artificial Intelligence, 2017.
Stefan Rüping. Thesis: Learning interpretable models. PhD thesis, Universitat Dortmund, 2006.
S Rasoul Safavian and David Landgrebe. A survey of decision tree classifier methodology. IEEE
transactions on systems, man, and cybernetics, 21(3):660–674, 1991.
Ute Schmid, Christina Zeller, Tarek Besold, Alireza Tamaddoni-Nezhad, and Stephen Muggleton.
How does predicate invention affect human comprehensibility? In International Conference on
Inductive Logic Programming, pages 52–67. Springer, 2016.
Skipper Seabold and Josef Perktold. Statsmodels: Econometric and statistical modeling with python.
In Proceedings of the 9th Python in Science Conference, volume 57, page 61, 2010.
Ramprasaath R Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh,
and Dhruv Batra. Grad-cam: Why did you say that? visual explanations from deep networks via
gradient-based localization. arXiv preprint arXiv:1610.02391, 2016.
Avanti Shrikumar, Peyton Greenside, Anna Shcherbina, and Anshul Kundaje. Not just a black box:
Interpretable deep learning by propagating activation differences. ICML, 2016.
19
Sameer Singh, Marco Tulio Ribeiro, and Carlos Guestrin. Programs as black-box explanations.
arXiv preprint arXiv:1611.07579, 2016.
Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad:
removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017.
Girish H Subramanian, John Nosek, Sankaran P Raghunathan, and Santosh S Kanitkar. A comparison of the decision table and tree. Communications of the ACM, 35(1):89–94, 1992.
Nava Tintarev and Judith Masthoff. Explaining recommendations: Design and evaluation. In
Recommender Systems Handbook, pages 353–382. Springer, 2015.
Berk Ustun and Cynthia Rudin. Supersparse linear integer models for optimized medical scoring
systems. Machine Learning, 102(3):349–391, 2016.
Fulton Wang and Cynthia Rudin. Falling rule lists. In Artificial Intelligence and Statistics, pages
1013–1022, 2015.
Tong Wang, Cynthia Rudin, Finale Doshi, Yimin Liu, Erica Klampfl, and Perry MacNeille. Bayesian
ors of ands for interpretable classification with application to context aware recommender systems.
In ICDM, 2016.
Tong Wang, Cynthia Rudin, Finale Doshi-Velez, Yimin Liu, Erica Klampfl, and Perry MacNeille.
Bayesian rule sets for interpretable classification. In International Conference on Data Mining,
2017.
Martin Wattenberg, Fernanda Viégas, and Moritz Hardt. Attacking discrimination with smarter
machine learning. Google Research, 17, 2016.
Description of Pilot Studies
We conducted several pilot studies in the design of these experiments. Our pilot studies showed
that asking subjects to respond quickly or within a time limit resulted in much lower accuracies;
subjects would prefer to answer as time was running out rather than risk not answering the question.
That said, there are clearly avenues of adjusting the way in which subjects are coached to place
them in fast or careful thinking modes, to better identify which explanations are best in each case.
The experiment interface design also played an important role. We experimented with different
placements of various blocks, the coloring of the text, whether the explanation was presented as
rules or as narrative paragraphs, and also, within rules, whether the input was placed before or after
the conclusion (that is, ‘if A: B” vs. “B if A”). All these affected response time and accuracy, and
we picked the configuration that had the highest accuracy and user satisfaction.
Finally, in these initial trials, we also varied more factors: number of lines, input conjunctions,
input disjunctions, output conjunctions, output disjunctions and global variables. Upon running
20
preliminary regressions, we found that there was no significant difference in effect between disjunctions and conjunctions, though the number of lines, global variables, and general length of output
clause—regardless of whether that length came from conjunctions or disjunctions—did have an
effect on the response time. Thus, we chose to run our experiments based on these factors.
21
| 2cs.AI
|
Neural Question Answering at BioASQ 5B
Georg Wiese1,2 , Dirk Weissenborn2 and Mariana Neves1
1
Hasso Plattner Institute, August Bebel Strasse 88, Potsdam 14482 Germany
2
Language Technology Lab, DFKI, Alt-Moabit 91c, Berlin, Germany
[email protected],
[email protected], [email protected]
arXiv:1706.08568v1 [cs.CL] 26 Jun 2017
Abstract
This paper describes our submission to the
2017 BioASQ challenge. We participated
in Task B, Phase B which is concerned
with biomedical question answering (QA).
We focus on factoid and list question, using an extractive QA model, that is, we
restrict our system to output substrings of
the provided text snippets. At the core
of our system, we use FastQA, a state-ofthe-art neural QA system. We extended
it with biomedical word embeddings and
changed its answer layer to be able to
answer list questions in addition to factoid questions. We pre-trained the model
on a large-scale open-domain QA dataset,
SQuAD, and then fine-tuned the parameters on the BioASQ training set. With our
approach, we achieve state-of-the-art results on factoid questions and competitive
results on list questions.
1
Introduction
BioASQ is a semantic indexing, question answering (QA) and information extraction challenge
(Tsatsaronis et al., 2015). We participated in
Task B of the challenge which is concerned with
biomedical QA. More specifically, our system participated in Task B, Phase B: Given a question
and gold-standard snippets (i.e., pieces of text that
contain the answer(s) to the question), the system
is asked to return a list of answer candidates.
The fifth BioASQ challenge is taking place at
the time of writing. Five batches of 100 questions
each were released every two weeks. Participating
systems have 24 hours to submit their results. At
the time of writing, all batches had been released.
The questions are categorized into different
question types: factoid, list, summary and yes/no.
Our work concentrates on answering factoid and
list questions. For factoid questions, the system’s
responses are interpreted as a ranked list of answer candidates. They are evaluated using meanreciprocal rank (MRR). For list questions, the system’s responses are interpreted as a set of answers
to the list question. Precision and recall are computed by comparing the given answers to the goldstandard answers. F1 score, i.e., the harmonic
mean of precision and recall, is used as the official evaluation measure 1 .
Most existing biomedical QA systems employ
a traditional QA pipeline, similar in structure to
the baseline system by Weissenborn et al. (2013).
They consist of several discrete steps, e.g., namedentity recognition, question classification, and
candidate answer scoring. These systems require a
large amount of resources and feature engineering
that is specific to the biomedical domain. For example, OAQA (Zi et al., 2016), which has been
very successful in last year’s challenge, uses a
biomedical parser, entity tagger and a thesaurus to
retrieve synonyms.
Our system, on the other hand, is based on a
neural network QA architecture that is trained endto-end on the target task. We build upon FastQA
(Weissenborn et al., 2017), an extractive factoid
QA system which achieves state-of-the-art results
on QA benchmarks that provide large amounts of
training data. For example, SQuAD (Rajpurkar
et al., 2016) provides a dataset of ≈ 100, 000
questions on Wikipedia articles. Our approach
is to train FastQA (with some extensions) on the
SQuAD dataset and then fine-tune the model parameters on the BioASQ training set.
Note that by using an extractive QA network as
our central component, we restrict our system’s
1
The details of the evaluation can be found at
http://participants-area.bioasq.org/
Tasks/b/eval_meas/
Start Probabilities pstart
sigmoid
softmax
EndScores
Scoresee(s)(s)
End
End
Scores yend
Start Scores ystart
Extractive QA System
Biomedical Embeddings
...
...
GloVe Embeddings
Character Embeddings
Question Type Features
Context Embeddings
Question Embeddings
Figure 1: Neural architecture of our system. Question and context (i.e., the snippets) are mapped directly to start and end probabilities for each context token. We use FastQA (Weissenborn et al.,
2017) with modified input vectors and an output
layer that supports list answers in addition to factoid answers.
responses to substrings in the provided snippets.
This also implies that the network will not be able
to answer yes/no questions. We do, however, generalize the FastQA output layer in order to be able
to answer list questions in addition to factoid questions.
2
2014) which have been trained on a large collection of web documents.
End Probabilitiesp(e|s)
p(e|s)
End
End Probabilities
Probabilities pend
• Character embedding: This embedding is
computed by a 1-dimensional convolutional
neural network from the characters of the
words, as introduced by Seo et al. (2016).
• Biomedical Word2Vec embeddings: We
use the biomedical word embeddings provided by Pavlopoulos et al. (2014). These
are 200-dimensional Word2Vec embeddings
(Mikolov et al., 2013) which were trained on
≈ 10 million PubMed abstracts.
To the embedding vectors, we concatenate a
one-hot encoding of the question type (list or factoid). Note that these features are identical for all
tokens.
Following our embedding layer, we invoke
FastQA in order to compute start and end scores
for all context tokens. Because end scores are conditioned on the chosen start, there are O(n2 ) end
scores where n is the number of context tokens.
We denote the start index by i ∈ [1, n], the end
i
index by j ∈ [i, n], the start scores by ystart
, and
i,j
end scores by yend .
In our output layer, the start, end, and span probabilities are computed as:
Model
Our system is a neural network which takes as input a question and a context (i.e., the snippets) and
outputs start and end pointers to tokens in the context. At its core, we use FastQA (Weissenborn
et al., 2017), a state-of-the-art neural QA system.
In the following, we describe our changes to the
architecture and how the network is trained.
2.1
Network architecture
In the input layer, the context and question tokens are mapped to high-dimensional word vectors. Our word vectors consists of three components, which are concatenated to form a single
vector:
• GloVe embedding: We use 300-dimensional
GloVe embeddings 2 (Pennington et al.,
2
We use the 840B embeddings available here: https:
//nlp.stanford.edu/projects/glove/
i
pistart = σ(ystart
)
(1)
i,·
pi,·
end = sof tmax(yend )
(2)
i,j
i
pi,j
span = pstart · pend
(3)
where σ denotes the sigmoid function. By computing the start probability via the sigmoid rather
than softmax function (as used in FastQA), we enable the model to output multiple spans as likely
answer spans. This generalizes the factoid QA network to list questions.
2.2
Training & decoding
Loss We define our loss as the cross-entropy of
the correct start and end indices. In the case of
multiple occurrences of the same answer, we only
minimize the span of the lowest loss.
Factoid MRR
Ensemble
Batch
Single
1
2
3
4
5
Average
52.0% (2/10)
38.3% (3/15)
43.1% (1/16)
30.0% (3/20)
39.2% (3/17)
40.5%
Single
57.1% (1/10)
42.6% (2/15)
42.1% (2/16)
36.1% (1/20)
35.1% (4/17)
42.6%
List F1
Ensemble
33.6% (1/11)
29.0% (8/15)
41.5% (2/17)
24.2% (5/20)
36.1% (4/20)
32.9%
33.5%(2/11)
26.2%(9/15)
49.5%(1/17)
29.3%(4/20)
39.1%(2/20)
35.1%
Table 1: Preliminary results for factoid and list questions for all five batches and for our single and
ensemble systems. We report MRR and F1 scores for factoid and list questions, respectively. In parentheses, we report the rank of the respective systems relative to all other systems in the challenge. The
last row averages the performance numbers of the respective system and question type across the five
batches.
Optimization We train the network in two steps:
First, the network is trained on SQuAD, following
the procedure by Weissenborn et al. (2017) (pretraining phase). Second, we fine-tune the network
parameters on BioASQ (fine-tuning phase). For
both phases, we use the Adam optimizer (Kingma
and Ba, 2014) with an exponentially decaying
learning rate. We start with learning rates of
10−3 and 10−4 for the pre-training and fine-tuning
phases, respectively.
Ensemble In order to further tweak the performance of our systems, we built a model ensemble.
For this, we trained five single models using 5-fold
cross-validation on the entire training set. These
models are combined by averaging their start and
end scores before computing the span probabilities (Equations 1-3). As a result, we submit two
systems to the challenge: The best single model
(according to its development set) and the model
ensemble.
BioASQ dataset preparation During finetuning, we extract answer spans from the BioASQ
training data by looking for occurrences of the
gold standard answer in the provided snippets.
Note that this approach is not perfect as it can produce false positives (e.g., the answer is mentioned
in a sentence which does not answer the question)
and false negatives (e.g., a sentence answers the
question, but the exact string used is not in the synonym list).
Because BioASQ usually contains multiple
snippets for a given question, we process all snippets independently and then aggregate the answer
spans, sorting globally according to their probability pi,j
span .
Implementation We implemented our system
using TensorFlow (Abadi et al., 2016). It was
trained on an NVidia GForce Titan X GPU.
Decoding During the inference phase, we retrieve the top 20 answers span via beam search
with beam size 20. From this sorted list of answer strings, we remove all duplicate strings. For
factoid questions, we output the top five answer
strings as our ranked list of answer candidates. For
list questions, we use a probability cutoff threshold
t, such that {(i, j)|pi,j
span ≥ t} is the set of answers.
We set t to be the threshold for which the list F1
score on the development set is optimized.
3
Results & discussion
We report the results for all five test batches of
BioASQ 5 (Task 5b, Phase B) in Table 1. Note
that the performance numbers are not final, as the
provided synonyms in the gold-standard answers
will be updated as a manual step, in order to reflect
valid responses by the participating systems. This
has not been done by the time of writing3 . Note
also that – in contrast to previous BioASQ challenges – systems are no longer allowed to provide
an own list of synonyms in this year’s challenge.
In general, the single and ensemble system are
performing very similar relative to the rest of field:
Their ranks are almost always right next to each
other. Between the two, the ensemble model performed slightly better on average.
On factoid questions, our system has been very
successful, winning three out of five batches. On
3
The final results will be published at http:
//participants-area.bioasq.org/results/
5b/phaseB/
list questions, however, the relative performance
varies significantly. We expect our system to perform better on factoid questions than list questions, because our pre-training dataset (SQuAD)
does not contain any list questions.
Starting with batch 3, we also submitted responses to yes/no questions by always answering
yes. Because of a very skewed class distribution
in the BioASQ dataset, this is a strong baseline.
Because this is done merely to have baseline performance for this question type and because of the
naivety of the method, we do not list or discuss the
results here.
Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and
Hannaneh Hajishirzi. 2016. Bidirectional attention
flow for machine comprehension. arXiv preprint
arXiv:1611.01603 .
4
Dirk Weissenborn, Georg Wiese, and Laura Seiffe.
2017. Making neural qa as simple as possible but
not simpler. arXiv preprint arXiv:1703.04816 .
Conclusion
In this paper, we summarized the system design
of our BioASQ 5B submission for factoid and list
questions. We use a neural architecture which is
trained end-to-end on the QA task. This approach
has not been applied to BioASQ questions in previous challenges. Our results show that our approach achieves state-of-the art results on factoid
questions and competitive results on list questions.
References
Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene
Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado,
Andy Davis, Jeffrey Dean, Matthieu Devin, et al.
2016. Tensorflow: Large-scale machine learning on
heterogeneous distributed systems. arXiv preprint
arXiv:1603.04467 .
Diederik Kingma and Jimmy Ba. 2014. Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980 .
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing
systems. pages 3111–3119.
Ioannis Pavlopoulos, Aris Kosmopoulos, and
Ion Androutsopoulos. 2014.
Continuous
space word vectors obtained by applying
word2vec to abstracts of biomedical articles
http://bioasq.lip6.fr/info/BioASQword2vec/.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for
word representation. In Empirical Methods in Natural Language Processing (EMNLP). pages 1532–
1543. http://www.aclweb.org/anthology/D14-1162.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and
Percy Liang. 2016. Squad: 100,000+ questions
for machine comprehension of text. arXiv preprint
arXiv:1606.05250 .
George Tsatsaronis, Georgios Balikas, Prodromos
Malakasiotis, Ioannis Partalas, Matthias Zschunke,
Michael R Alvers, Dirk Weissenborn, Anastasia
Krithara, Sergios Petridis, Dimitris Polychronopoulos, et al. 2015. An overview of the bioasq largescale biomedical semantic indexing and question answering competition. BMC bioinformatics 16(1):1.
Dirk Weissenborn, George Tsatsaronis, and Michael
Schroeder. 2013. Answering factoid questions in the
biomedical domain. BioASQ@ CLEF 1094.
Yang Zi, Zhou Yue, and Eric Nyberg. 2016. Learning
to answer biomedical questions: Oaqa at bioasq 4b.
ACL 2016 page 23.
| 9cs.NE
|
Technische Universität Berlin
arXiv:1611.02072v1 [cs.SY] 7 Nov 2016
Institut für Mathematik
Data-driven Structured Realization
Philipp Schulze
Christopher Beattie
Benjamin Unger
Serkan Gugercin
Preprint 23-2016
Preprint-Reihe des Instituts für Mathematik
Technische Universität Berlin
http://www.math.tu-berlin.de/preprints
Preprint 23-2016
November 2016
Data-driven Structured Realization
P. Schulze∗1 , B. Unger†1 , C. Beattie‡2 , and S. Gugercin§2
1
2
Institut für Mathematik, TU Berlin, Germany, {pschulze,unger}@math.tu-berlin.de.
Department of Mathematics, Virginia Tech, Blacksburg, VA, {beattie,gugercin}@vt.edu.
November 8, 2016
Abstract
We present a framework for constructing structured P
realizations of linear dye K hk (s)A
ek )−1 B
e where
namical systems having transfer functions of the form C(
k=1
h1 , h2 , ..., hK are prescribed functions that specify the surmised structure of the model.
Our construction is data-driven in the sense that an interpolant is derived entirely from
measurements of a transfer function. Our approach extends the Loewner realization
framework to more general system structure that includes second-order (and higher)
systems as well as systems with internal delays. Numerical examples demonstrate the
advantages of this approach.
Keywords: structured realization, data-driven model reduction, interpolation, delay
system, second-order system
AMS(MOS) subject classification: 93B15, 30E05, 93C05
1
Introduction
The simulation of complex physical, chemical, or biological processes is a standard task
in science, engineering, and industry. The dynamics of such processes are commonly
modeled as dynamical systems, which then can be analyzed (often through simulation) for
optimization and control. The demand for higher fidelity models produces as a common
consequence ever more complex and larger dynamical systems, whose simulation may
require computational resources that become unmanageably large. This computational
cost is often directly related to the state space dimension of the underlying dynamical
system, thus creating a need for low-dimensional approximations of large-scale models.
Model order reduction (MOR) techniques using rational interpolation methods, such as
the iterative rational Krylov algorithm (IRKA) [20], or Gramian-based methods, such
as balanced truncation [30, 31], have become popular tools to create such reduced-order
models (ROMs); for an overview of these approaches, see the survey papers [2, 5, 6] and the
monograph [1]. There are many features that distinguish model reduction methods from
∗
The author was supported by the DFG Collaborative Research Center 1029 Substantial efficiency
increase in gas turbines through direct use of coupled unsteady combustion and flow dynamics, project A02.
†
The author was supported by the DFG Collaborative Research Center 910 Control of self-organizing
nonlinear systems: Theoretical methods and concepts of application, project A2.
‡
The work of this author was supported in part by the Einstein Foundation Berlin.
§
The work of this author was supported in part by the Alexander von Humboldt Foundation.
1
one another; we focus on the dichotomy between projection-based methods and data-driven
methods. Projection-based methods presuppose access to an explicit state space realization
of the original dynamical system and then identify low-dimensional, high-value subspaces
of the state space, using projections to constrain dynamics to these subspaces. Data-driven
model reduction methods as exemplified by vector fitting [15, 21] or the Loewner realization
framework of [27] are nonintrusive in the sense that the access to internal dynamics that an
explicit state space realization provides is not necessary. Such methods are able to produce
system realizations (generally low-dimensional) directly from measurements of the transfer
function. The greater flexibility that data-driven methods provide is balanced against the
general inability of data-driven methods to preserve structural features that may be present
in the original model, a capability that often is available to projection-based methods.
In some practical settings, the original (or approximate) model may be available only
implicitly either through response observation or simulation, leading one to data-driven
approaches as the only feasible strategy for MOR. Despite the inaccessibility of a description
of detailed internal dynamics, there may yet be significant ancillary information or at least
a good basic understanding of how the system should behave, allowing one to surmise
general structural features of the underlying dynamical system. For example, vibration
effects are naturally associated with subsystems that have second-order structure; internal
transport or signal propagation will naturally be associated with state delays.
Example 1.1. Consider acoustic transmission in a fluid-filled duct of length L that has an
acoustic driver positioned at one end. Suppose we are interested in the acoustic pressure
y(t) = p(ξ0 , t) at a fixed point ξ0 ∈ (0, L) in the duct, which we view as the output of an
abstract system that is driven by the input fluid velocity u(t), determined by an acoustic
driver positioned at ξ = 0. We assume that the output pressure depends linearly on the
input velocity in a way that is invariant to translation in time, and so the output could be
anticipated to involve some superposition of internal states that are lagged in time according
to propagation delays related to the distance traveled by the signal. Assuming a uniform
sound speed c throughout the duct, we allow for a direct propagation delay τ1 = ξ0 /c
between the input and output location and a second propagation delay τ2 = (2L − ξ0 )/c,
associated with a reflected signal. A semi-empirical model for the state evolution of a
system that has these basic features could have the form
A1 x(t) + A2 x(t − τ1 ) + A3 x(t − τ2 ) = bu(t),
with an output port map given by y(t) = cT x(t). The matrices A1 , A2 , and A3 , the
port maps associated with the vectors, b and c, as well as their dimensions are unknown.
We arrive at a (semi-empirical) transfer function for this system having the hypothesized
structure
−1
H(s) = cT A1 + A2 e−τ1 s + A3 e−τ2 s
b.
Based on observed or computed system response data, we wish to identify A1 , A2 , A3 , b
and c.
We present here a general data-driven model reduction framework that is capable of
preserving system structure present in an original model when it is known, and possibly
inducing hypothesized structure in other cases. We lay out our basic problem framework
in the following Section 2 and then show how to exploit this in developing structured
realizations in Section 3. Computational examples are offered in Section 4.
2
2
Problem Setting
Although the term system structure can have wide ranging meanings, for our purposes
we will understand the term to refer to equivalence classes of systems having realizations
associated with a linearly independent function family {h1 , h2 , . . . , hK } that appear as
H(s) = C
K
X
!−1
hk (s)Ak
B,
(2.1)
k=1
where C ∈ Rp×N , Ak ∈ RN ×N for k = 1, . . . , K, B ∈ RN ×m . We assume in all that follows
that the functions involved, hk : C → C, are meromorphic. For any given
PK function family,
we will refer to associated matrix-valued functions having the form k=1 hk (s)Ak as an
affine structure. By standard abuse of notation, we use H(s) to denote either the system
itself or the transfer function of the system evaluated at the point s ∈ C. The two systems
e
e
H(s) and H(s)
are called structurally equivalent if H(s), H(s)
∈ Cp×m for s ∈ C and if
they each have the form
H(s) = C
K
X
!−1
hk (s)Ak
B
and
e
e
H(s)
=C
k=1
K
X
!−1
ek
h̃k (s)A
e
B,
k=1
with span{h1 , h2 , . . . , hK } ≡ span{h̃1 , h̃2 , . . . , h̃K }. In particular, we allow different state
e ∈ Rp×n , A
ek ∈ Rn×n , and B
e ∈ Rn×m the integers N and n
space dimensions, i. e., for C
need not be the same. Given an original (full order) system associated with H(s), we aim
e
to construct a structurally equivalent system H(s),
with state space dimension n N ,
and we wish to accomplish this allowing only evaluations of H(s).
The general structure of (2.1) encompasses a variety of system formulations. Indeed, we
may observe immediately that linear time invariant dynamical systems often are presented
as standard first-order realizations
A1 ẋ(t) + A2 x(t) = Bu(t),
y(t) = Cx(t),
(2.2)
where A1 , A2 ∈ RN ×N , B ∈ RN ×m , and C ∈ Rp×N . The state is denoted by x(t) ∈ RN ,
while u(t) ∈ Rm and y(t) ∈ Rp are, respectively, the inputs and outputs of the system.
If the input u : [0, T ] → Rm is exponentially bounded, i. e., ku(t)k ∈ O(eγt ) for some
γ ∈ R, then x : [0, T ] → Rn and y : [0, T ] → Rp are also exponentially bounded. The
Laplace transform may be applied to (2.2) and rearranged to ŷ(s) = H(s)û(s) with
H(s) = C(sA1 − A2 )−1 B, which has the form of (2.1) with h1 (s) = s and h2 (s) ≡ −1.
In many practical applications, the underlying dynamical system comes in a form quite
different from (2.2) that will reflect the nature of the system, and we may wish to preserve
this structure in the realization. For example, a general RLC network may be modeled as
integro differential-algebraic equation [18], given by
Z t
A1 ẋ(t) + A2 x(t) + A3
x(τ )dτ = Bu(t),
y(t) = B T x(t).
(2.3)
0
The transfer function associated with (2.3) is given by
H(s) = B T
1
sA1 + A2 + A3
s
3
−1
B
Table 2.1: Different system structures (all with output mapping y(t) = Cx(t))
state space description
second-order
state delay
neutral delay
viscoelastic
A1 ẍ(t) + A2 ẋ(t) + A3 x(t) = Bu(t)
A1 ẋ(t) + A2 x(t) + A3 x(t − τ ) = Bu(t)
A1 ẋ(t) + A2 x + A3 ẋ(t − τ ) = Bu(t)
Rt
A1 ẍ(t) + 0 h(t−τ )A2 ẋ(τ )dτ + A3 x(t) = Bu(t)
transfer function
−1
B
C s2 A1 + sA2 + A3
−1
−τ s
B
C sA1 + A2 + e
A3
−1
B
C sA1 + A2 + se−τ s A3
−1
B
C s2 A1 + sĥ(s)A2 + A3
and we expect better approximation properties of the ROM by preserving this form. Further
examples for system structures are listed in Table 2.1.
Preserving structure often allows one to derive reduced models with smaller state space
dimension n, while maintaining comparable or at times even better accuracy than what
unstructured reduced models produce, see Section 5 in [7]. Additionally, since the internal
structure of models often reflects core phenomenological properties, structured models may
behave in ways that remain qualitatively consistent with the phenomena that are being
modeled – possibly more so than unstructured models having higher objective fidelity.
Structure-preserving model reduction has for the most part developed in a projection-based
context that presupposes access to internal dynamics. Projection-based techniques are
often able to retain special structural features in the reduced models that may reflect
underlying physical properties of the systems under study [7, 11, 12, 18, 25, 28, 33]. Datadriven techniques for system identification and model reduction do not generally have
this capacity, however, the recent contributions of [17, 32] provide a notable exception for
time-delay systems. In the present work, we build on the results of [32] and extend its
domain of applicability to a wide range of structured dynamical systems. These ideas
originate in the Loewner realization framework of [27].
The Loewner realization framework is an effective and broadly applicable approach for
constructing rational approximants directly from interpolation data; it has been extended
to parametric systems [3, 23] and to realization independent methods for optimal H2
approximation [8]. However, the Loewner framework is only capable of producing rational
approximants and, so in particular, it cannot capture the transcendental character of
transfer functions for dynamical systems containing distributed parameter subsystems that
model convection or transport (cf. [14]).
We begin by describing in detail the type of system response data that we will assume
henceforth to be available. Suppose we have 2n points in the complex plane, which may
be interpreted as complex driving frequencies, {µ1 , . . . , µn } and {σ1 , . . . , σn }. We assume
for the time being that these two sets, {µi }ni=1 and {σi }ni=1 , are each made up of n distinct
points, although we allow the two point sets to have nontrivial intersection (so that it
could happen that µi = σj for some index pairs (i, j)). In addition to these complex
frequencies, we have the so-called left tangential direction vectors {`1 , . . . , `n } and the right
tangential direction vectors {r 1 , . . . , r n } where `i ∈ Rp and r i ∈ Rm for i = 1, . . . , n. In
the single-input/single-output (SISO) case, these tangential directions are assigned the
value one, i. e., `i = r i = 1. Unlike projection-based model reduction, which requires access
to the state space quantities, data-driven interpolatory model reduction only assumes
access to the action of the transfer function evaluated at the driving frequencies along the
4
tangential directions, i. e.,
`Ti H(µi ) = f Ti
and
H(σi )r i = g i
for i = 1, . . . , n.
(2.4)
If the direction vectors `i and `j are linearly independent, one can allow µi to coincide
with µj , and similarly for σi ’s. However, for simplicity the only coincidence of interpolation
points that we admit will be between left and right interpolation points, i. e., µi = σj . If
this is the case for an index pair (i, j), then bitangential derivative data is assumed to
be available. Since we assume that each of the two sets {µi }ni=1 and {σi }ni=1 consists of
n distinct points, if µi = σj for an index pair (i, j), without loss of generality, we assume
i = j. Then, the corresponding bitangential derivative data is defined as
`Ti H 0 (µi )r i = θi .
Following [2, 27], we summarize the interpolation data as
left interpolation data: {(µi , `i , f i ) | µi ∈ C, `i ∈ Cp , f i ∈ Cm , i = 1, . . . , n},
right interpolation data: {(σi , r i , g i ) | σi ∈ C, r i ∈ Cm , g i ∈ Cp , i = 1, . . . , n},
(2.5)
bitangential derivative data: {(i, θi ) | i ∈ {1, . . . , n} for which µi = σi , θi ∈ C},
with the understanding that the last category may be empty if {µi }ni=1 ∩ {σi }ni=1 = ∅.
Note that in the case µi = σi , the compatibility of the conditions (2.4) requires that
f Ti r i = `Ti g i .
For ease of presentation, we introduce the matrices
M := diag(µ1 , . . . , µn ) ∈ Cn×n ,
L := `1 . . . `n ∈ Cp×n ,
F := f 1 . . . f n ∈ Cm×n ,
S := diag(σ1 , . . . , σn ) ∈ Cn×n ,
R := r 1 . . . r n ∈ Cm×n ,
G := g 1 . . . g n ∈ Cp×n .
e A
ek , and B
e (for k = 1, . . . , K) using the measureOur goal is to construct matrices C,
e
e PK hk (s)A
ek )−1 B
e satisfies the
ments (2.5), such that the transfer function H(s) = C(
k=1
interpolation conditions
e i ) = `Ti H(µi ) = f Ti
`Ti H(µ
e i )r i = H(σi )r i = g i
and H(σ
for i = 1, . . . , n.
(2.6a)
If µi = σi for any index i, then additionally,
e 0 (µi )r i = `T H 0 (µi )r i = θi
`Ti H
i
(2.6b)
is to be satisfied.
3
3.1
Structured Interpolatory Realizations
Standard Loewner Realization
A key tool for our results is the Loewner realization framework introduced in [27]. This
framework uses the Loewner matrix L ∈ Cn×n and the shifted Loewner matrix Lσ ∈ Cn×n ,
whose entries [L]i,j and [Lσ ]i,j for i, j = 1, . . . , n are defined as
[L]i,j =
f Ti r j − `Ti g j
µi − σj
[L]i,i = θi
µi f Ti r j − σj `Ti g j
,
µi − σj
and
[Lσ ]i,j =
and
[Lσ ]i,i = f Ti r i + µi θi ,
5
if
µi 6= σj ,
(3.1)
if
µi = σi .
(3.2)
For SISO systems, L and Lσ are the divided differences matrices corresponding to the
transfer functions H(s) and sH(s), respectively.
Theorem 3.1 (Loewner realization [27]). Let det(s̃L−Lσ ) 6= 0 for all s̃ ∈ {µi }ni=1 ∪{σi }ni=1 .
Then the system
e˙ (t) = −Lσ x
e (t) + F T u(t),
− Lx
e (t) = G x
e (t)
y
(3.3)
is a minimal realization of an interpolant of the data, i. e., its transfer function
e
H(s)
= G(Lσ − sL)−1 F T
satisfies the interpolation conditions (2.6).
The condition det(s̃L − Lσ ) 6= 0 in Theorem 3.1 can be relaxed by means of the short
singular value decomposition (SVD) [1, Remark 3.2.1]. If an s̃ ∈ {µi }ni=1 ∪ {σi }ni=1 violates
the regularity condition, then the short SVD of s̃L − Lσ can be used to truncate the
redundant parts [27].
3.2
Interpolation Conditions
Suppose we are given interpolation data as in (2.5) and for the moment assume that we
e
e K(s)
e −1 B.
e If we can impose conditions on
already have a realization of the form H(s)
=C
e
e
e
e
e
e
e satisfies the interpolation
C, B and the matrix function K such that H(s) = C K(s)−1 B
conditions (2.6), then we can revert the process and use the conditions to construct the
realization. The following observation, which corresponds to an equivalent parametrization
of the interpolation conditions (2.6), suggests how one might proceed. This parametrization
will form the basis for constructing the structured interpolatory realizations.
e
Theorem 3.2. Let K(s)
be a continuously differentiable n × n matrix-valued function of
the complex argument s, which is nonsingular at s = µi and s = σj for i, j = 1, . . . , n. The
e
e K(s)
e −1 B
e satisfies the interpolation conditions (2.6a) if and only if
realization H(s)
=C
e G
G = CP
and
e
F T = PFT B,
(3.4)
where PG , PF ∈ Cn×n are two matrices, whose columns piG := PG ei and piF := PF ei ,
respectively, solve the linear systems
e i )piG = Br
e i
K(σ
and
e i )T piF = C
e T `i ,
K(µ
(3.5)
where ei is the ith column of the n × n identity matrix. If additionally, µi = σi , then
e satisfies the bitangential interpolation condition (2.6b) as well provided
e
e K(s)
e −1 B
H(s)
=C
that
T 0
e (µi )piG = −θi .
piF K
(3.6)
e
e K(s)
e −1 B
e is well-defined at s = µi and s = σi .
Proof. The transfer function H(s)
= C
e i.
Assume (3.4) and (3.5). Multiplying the first equation in (3.4) by ei yields g i = Cp
G
e i ) is invertible, one immediately
Then, using the first equation in (3.5) and the fact that K(σ
e i )r i , i. e., the right tangential interpolation holds. Similarly, using the
obtains g i = H(σ
e i );
second expression in (3.4) and the definition of piF in (3.5), we arrive at f Ti = `Ti H(µ
6
thus (2.6a) holds. The other direction follows directly. Moreover, if µi = σi , then (3.6)
yields
e K(µ
e i )−1 K
e 0 (µi )K(σ
e i )−1 Br
e i = − pi
e 0 (µi )r i = −`Ti C
`Ti H
F
T
e 0 (µi )pi = θi .
K
G
Evidently, in order to satisfy the collected tangent interpolation conditions (2.6a), we can
e
now equivalently require the realization H(s)
to satisfy the conditions of Theorem 3.2. In
e
particular we need K(s) to be nonsingular at the driving frequencies s = µi and s = σj .
P
e
e
For K(s)
= K
k=1 hk (s)Ak , the other conditions (3.4) and (3.5) can be rewritten as
e
F T = PFT B,
(3.7)
ek = LT C,
e
hk (M)PFT A
(3.8)
e G,
G = CP
K
X
K
X
ek PG hk (S) = BR,
e
A
k=1
k=1
where we set hk (M) := diag(hk (µ1 ), . . . , hk (µn )) and hk (S) := diag(hk (σ1 ), . . . , hk (σn )).
To fulfill additionally the bitangential interpolation conditions (2.6b) for the case µi = σi ,
the third condition of Theorem 3.2 needs to be satisfied.
If the matrices PF and PG are nonsingular, then
−1
e
e K(s)
e −1 B
e = G PFT K(s)P
e
H(s)
=C
FT ,
G
and hence the realization is unique up to the state space transformation described by PF
e and C
e are given directly by the data without further
and PG . In this case, the matrices B
computations and the matrices PF and PG capture the non-uniqueness of the realization.
In Section 3.4 we will use these matrices to tailor the realization to interpolate additional
data. In any case, we view equations (3.7) and (3.8) not as a coupled system but as a
e and C
e from (3.7). Then,
staggered process. First, fix matrices PF , PG and determine B
in the second step, use this information to solve (3.8). With this viewpoint, i. e., not
counting PF and PG as unknowns, we have Kn2 unknowns from the coefficient matrices
ek and (m + p)n unknowns from the input and output matrices B
e and C,
e giving a total of
A
2
Kn + (m + p)n unknowns. For these unknowns, (3.7) and (3.8) constitute 2n2 + (m + p)n
equations, leaving (K − 2)n2 degrees of freedom. In particular, we can expect a unique
solution for K = 2.
Remark 3.3. We have (K − 2)n2 degrees of freedom to solve the structured realization
problem, and therefore the K = 1 case does not have enough degrees of freedom to guarantee
a solution in general. Note that the standard rational approximation has K = 2. To
e = b ∈ Rn
further examine the K = 1 case, assume for simplicity that H(s) is SISO, i. e., B
T
n
e = c ∈ R . Then, for K = 1, the reduced model has the form H(s)
e
e−1 b.
and C
= h11(s) cT A
Therefore, the interpolation conditions become
e−1 b = H(σi )h1 (σi )
cT A
e−1 b = H(µi )h1 (µi ),
and cT A
for i = 1, . . . , n.
(3.9)
e−1 b is constant, for the interpolation problem in (3.9) to have a solution, we
Since cT A
need H(σi )h1 (σi ) = H(µi )h1 (µi ) = c where c is a constant for i = 1, . . . , n. This clearly
will not be the case in general and we cannot expect to have a solution. Interestingly, if
e = In , b = e1 and c = c · e1 .
this condition holds, a solution can be found easily by setting A
Based on these considerations, we will focus on K ≥ 2 in the rest of the paper.
♦
7
Remark 3.4. The nonsingularity of the matrices PF and PG is connected to the minimality
of the realization. To see this, assume that we have a SISO standard state space system,
e B
e
e = b ∈ Rn , and C
e T = c ∈ Rn . Note that in this case K
e and its
i. e., K(s)
= sIn − A,
pointwise inverse form a set of commutative matrices. Hence we have
h
i
e 1 )−1 b · · · K(σ
e n )−1 b
rank (PG ) = rank K(σ
h
Q
i
n
e 1 )−1 b K(σ
e 1 )−1 K(σ
e 2 )−1 b · · ·
e i )−1 b
= rank K(σ
K(σ
i=1
h
i
e ··· A
en−1 b
= rank b Ab
such that PG is nonsingular if and only if the realization is controllable. Similarly, PF is
nonsingular if and only if the realization is observable.
♦
Note that the Loewner pencil with h1 (s) ≡ 1 and h2 (s) = −s satisfies the conditions of
e
Theorem 3.2 with K(s)
= Lσ − sL, i. e., the Loewner framework works with matrices PF
and PG being the identity. Indeed, for µi 6= σj , the (i, j) component of the Loewner pencil
is
s − σj
µi − s
T
T e
ei K(s)ej = [Lσ ]i,j − s [L]i,j =
fi rj +
`Ti g j ,
µi − σ j
µi − σj
e i ) = `T G = `T C
e and K(σ
e j )ej = F T r j = Br
e j . Similarly,
so we have immediately, eTi K(µ
i
i
for the case µi = σi , we obtain
e i ) = `Ti C,
e
eTi K(µ
e i,
e i )ei = Br
K(σ
and
e 0 (µi )ei = −Li,i = −θi .
eTi K
The remainder of this section is structured as follows. In Section 3.3 we consider the special
case K = 2 and show its close relation to the Loewner framework. If K ≥ 3 we need a
strategy to fix the remaining degrees of freedom. To this end we propose two approaches
which both provide interpolation of further data while maintaining the dimension of the
matrices in the realization. The first approach uses additional interpolation points for this
(Section 3.4.1), while the second one interpolates additional derivative evaluations of the
transfer functions (Section 3.4.2).
3.3
Structured Loewner Realizations: The case K = 2
Setting PF = PG = In gives 2n2 +(m+p)n equations in (3.7) and (3.8) for the Kn2 +(m+p)n
unknowns such that we can expect (under some regularity) a unique solution for the case
e = FT , C
e = G, and the matrix equations in (3.8) reduce to
K = 2. In this case B
e2 = LT G
e1 + h2 (M)A
h1 (M)A
and
e1 h1 (S) + A
e2 h2 (S) = F T R.
A
To decouple these equations, we multiply the first equation from the right by h2 (S) and
the second equation from the left by h2 (M). Subtracting the resulting systems yields the
Sylvester-like equation
e1 h1 (S) − h1 (M)A
e1 h2 (S) = h2 (M)F T R − LT Gh2 (S).
h2 (M)A
(3.10)
e1 and obtain
Similarly, we can eliminate A
e2 h2 (S) − h2 (M)A
e2 h1 (S) = h1 (M)F T R − LT Gh1 (S).
h1 (M)A
8
(3.11)
Remark 3.5. If the desired model is a generalized state space system as in (2.2), i. e.,
h1 (s) = s and h2 (s) ≡ −1, then (3.10) and (3.11) are given by the Sylvester equations
e1 S − MA
e1 = F T R − LT G
e2 S − MA
e2 = MF T R − LT GS,
A
and
A
(3.12)
respectively. Up to a sign factor, these are exactly the Sylvester equations that define
the Loewner matrix and the shifted Loewner matrix [27]. In particular, if σi 6= µj for
e1 = −L and A
e2 = −Lσ are the unique solutions of (3.10) and (3.11)
i, j = 1, . . . , n, then A
and the Loewner framework is a special case of the general framework presented in this
paper. Similarly, the proportional ansatz for the realization of delay systems introduced
in [32] is covered by our framework.
♦
e1 and A
e2 , for which µi 6= σj , may be obtained by multiplying (3.10)
Those elements of A
T
and (3.11) from left by ei and from right by ej yielding
h i
h i
h1 (µi )f Ti r j − `Ti g j h1 (σj )
h2 (µi )f Ti r j − `Ti g j h2 (σj )
e2
e1
=
A
=
,
A
(3.13)
h2 (µi )h1 (σj ) − h1 (µi )h2 (σj )
h1 (µi )h2 (σj ) − h2 (µi )h1 (σj )
i,j
i,j
under the generic assumption that h1 (µi )h2 (σj ) 6= h2 (µi )h1 (σj ). This is satisfied for all
possible choices of interpolation points (with µi 6= σj ) if the functions h1 and h2 satisfy
the Haar condition [13], see also Section 3.4.1. The components for which µi = σi can be
obtained by translating the conditions in Theorem 3.2 to the K = 2 case. This yields
e1 ]i,i + h2 (µi )[A
e2 ]i,i = `Ti g i ,
h1 (µi )[A
e1 ]i,i + h0 (µi )[A
e2 ]i,i = −θi
h01 (µi )[A
2
and consequently
h i
h2 (µi )θi + h02 (µi )`Ti g i
e1
A
= 0
,
h2 (µi )h1 (µi ) − h01 (µi )h2 (µi )
i,i
h i
e2
A
i,i
=
(3.14)
h1 (µi )θi + h01 (µi )`Ti g i
, (3.15)
h01 (µi )h2 (µi ) − h02 (µi )h1 (µi )
for the components with µi = σi under the generic assumption h02 (µi )h1 (µi ) 6= h01 (µi )h2 (µi ).
Consequently, we have proven the subsequent result.
e1 and A
e2 be as in (3.13) and (3.15) where the denominators are
Corollary 3.6. Let A
assumed nonzero. If
e1 + h2 (s̃)A
e2 6= 0
det h1 (s̃)A
for all s̃ ∈ {µi }ni=1 ∪ {σi }ni=1 ,
−1
e
e1 + h2 (s)A
e2
then the transfer function H(s)
= G h1 (s)A
F T satisfies the interpolation
conditions (2.6).
e1 and A
e2 have a structure similar to the Loewner matrix and the shifted
The matrices A
Loewner matrix. This gives rise to the idea that the result of Corollary 3.6 can be obtained
from the standard Loewner framework using transformed data.
Corollary 3.7. Suppose that h2 (S) and h2 (M) are nonsingular and that the denominators
in (3.13) and (3.15) are not zero. Construct the Loewner matrix L and the shifted Loewner
matrix Lσ for the transformed data
`i
h1 (µi )
left interpolation data:
,
, f , i = 1, . . . , n ,
h2 (µi ) h2 (µi ) i
h1 (σi )
ri
right interpolation data:
,
, g , i = 1, . . . , n ,
(3.16)
h2 (σi ) h2 (σi ) i
h2 (µi ) θi + h02 (µi ) `Ti g i
bitangential derivative data:
i, 0
for µi = σi .
h1 (µi ) h2 (µi ) − h02 (µi ) h1 (µi )
9
If det(h2 (s̃)Lσ − h1 (s̃)L) 6= 0 for all s̃ ∈ {µi }ni=1 ∪ {σi }ni=1 , then the transfer function
e
H(s)
= G(h2 (s)Lσ − h1 (s)L)−1 F T
interpolates the data.
Proof. Simple calculations yield that, when constructing the Loewner pencil with the
transformed interpolation data (3.16), the Loewner matrix and the shifted Loewner matrix
e2 given in (3.13) and (3.15). Corollary 3.6 completes the proof.
e1 and A
coincide with −A
Corollary 3.7 allows one to transfer many results of the standard Loewner framework to
the general framework considered in this subsection. In particular, this allows us to keep
the system matrices real if the interpolation data is closed under complex conjugation.
The details are formulated in Lemma 3.8.
Lemma 3.8. Let the interpolation data be closed under complex conjugation, i. e., there
exist unitary matrices TF , TG ∈ Cn×n with
TF∗ MTF ∈ Rn×n ,
TF∗ LT ∈ Rn ,
TF∗ F T ∈ Rn ,
TG∗ STG ∈ Rn×n ,
RTG ∈ Rn ,
GTG ∈ Rn .
Moreover, assume that the θi ’s (for the case µi = λi ) are closed under complex conjugaT
∗ T
e e
e1 TG , T ∗ A
e
tion. Then, the realization (TF∗ A
F 2 TG , TF F , GTG ) with (A1 , A2 , F , G) from
Corollary 3.6 consists of real-valued matrices and interpolates the data.
Proof. First note that if the interpolation data is closed under complex conjugation, so
is the transformed data in Corollary 3.7. Based on this observation, the proof for the
case that µi 6= λj for all i, j = 1, . . . , n simply follows the lines of [4, section 2.4.4.]. This
can also be comprehended after multiplying the Sylvester-like equations (3.10) and (3.11)
from left by TF∗ and from right by TG . Similar reasoning proves the claim for the µi = λi
case.
Example 3.9. A special case of Lemma 3.8 applies when the interpolation data is sorted
such that the real values have the highest indices, i. e.,
M = diag(µ1 , µ1 , . . . , µ2`−1 , µ2`−1 , µ2`+1 , . . . , µn ),
L = `1 `1 . . . `2`−1 `2`−1 `2`+1 . . . `n ,
F = f 1 f 1 . . . f 2`−1 f 2`−1 f 2`+1 . . . f n ,
S = diag(σ1 , σ1 , . . . , σ2r−1 , σ2r−1 , σ2r+1 , . . . , σn ),
R = r 1 r 1 . . . r 2r−1 r 2r−1 r 2r+1 . . . r n ,
G = g 1 g 1 . . . g 2r−1 g 2r−1 g 2r+1 . . . g n .
In this case possible choices for TF and TG are given by block diagonal matrices
1 1 −ı
1 1 −ı
T• = blkdiag √
, ..., √
, 1, . . . , 1 ,
2 1 ı
2 1 ı
where • ∈ {F, G}. One can also obtain the real realization directly from Theorem 3.2 by
choosing PFT = TF∗ and PG = TG (see discussion after Theorem 3.2).
10
Remark 3.10. The result from Corollary 3.7 can (formally) be obtained by rewriting the
transfer function, as similar to what is done in [17], namely
−1
h1 (s) e
−1 e
e 1 .
e
e
e
e
e
e
B
A1 + A2
H(s) = C(h1 (s)A1 + h2 (s)A2 ) B = C
h2 (s)
h2 (s)
♦
3.4
Structured Realization for the Case K ≥ 3
When K ≥ 3, the conditions in Theorem 3.2 do not provide enough conditions for the
available degrees of freedom (even if PF and PG are fixed). Hence, we have some freedom
ek with k = 1, . . . , n. We can exploit these degrees of freedom,
in choosing the matrices A
for instance, by fitting the transfer function to additional data. For simplicity we assume
{µi }ni=1 ∩ {σi }ni=1 = ∅ for the remainder of this section.
3.4.1
Interpolation at Additional Points
In this subsection we focus on fitting the transfer function to additional data or, equivalently,
match the given data with a smaller state space dimension. To this end, let us assume
that we have (QF − 1)n additional left interpolation points and (QG − 1)n additional
right interpolation points at hand, which we group in sets of n. More precisely, the left
interpolation data is grouped into the matrices
Mq := diag(µq;1 , µq;2 , . . . , µq;n ) ∈ Cn×n ,
Lq := `q;1 `q;2 · · · `q;n ∈ Cp×n ,
Fq := f q;1 f q;2 · · · f q;n ∈ Cm×n ,
(3.17a)
where q = 1, . . . , QF . Here, we set µ1;i := µi , f 1;i := f i , and `1;i := `i , such that we have
M1 = M, L1 = L, and F1 = F. Similarly, we introduce, for q = 1, . . . , QG , the matrices
Sq := diag(σq;1 , σq;2 , . . . , σq;n ) ∈ Cn×n ,
Rq := r q;1 r q;2 · · · r q;n ∈ Cm×n ,
Gq := g q;1 g q;2 · · · g q;n ∈ Cp×n .
(3.17b)
To use the full capacity of available degrees of freedom, we assume K = QF + QG , with
QF , QG ≥ 1. The next result gives us the necessary and sufficient conditions that the
e
matrices in the realization H(s)
must satisfy to interpolate all prescribed information.
P
e
e K(s)
e −1 B
e with K(s)
e
e
e
Theorem 3.11. Let H(s)
=C
= K
k=1 hk (s)Ak and suppose that K(s)
QG
QF
is nonsingular for all s̃ ∈ {µq;i }q=1 ∪ {σq;i }q=1 for all i = 1, . . . , n.
e q;i ) = f T are satisfied for i = 1, . . . , n and
1. The left interpolation conditions `Tq;i H(µ
q;i
q = 1, . . . , QF if and only if there exist matrices PF ,q with q = 1, . . . , QF that satisfy
FqT
=
e
PFT ,q B
K
X
and
ek = LT C.
e
hk (Mq )PFT ,q A
q
(3.18)
k=1
e q;i )r q;i = g q;i are satisfied for i = 1, . . . , n and
2. The right interpolation conditions H(σ
q = 1, . . . , QG if and only if there exist matrices PG,q with q = 1, . . . , QG that satisfy
e G,q
Gq = CP
K
X
and
k=1
11
ek PG,q hk (Sq ) = BR
e q.
A
(3.19)
Proof. The result follows directly from Theorem 3.2. For the sake of completeness we
e=
give the proof of the first statement again. The second identity in (3.18) implies `Tq;i C
P
e
e
eTi PFT ,q K
k=1 hk (µq;i )Ak . Thus, by the first identity and the definition of H we conclude
e q;i ) = eT P T B
e = fT
`Tq;i H(µ
i F ,q
q;i
for i = 1, . . . , n and q = 1, . . . , QF .
Evidently, in order to satisfy the interpolation conditions (2.6a) it will be sufficient to
require that (3.18) and (3.19) hold simultaneously. This gives us the following strategy
ek , B,
e and C.
e Suppose we can find matrices PF ,q
to determine the realization matrices A
and PG,q that satisfy the first identity in (3.18) and (3.19), respectively, i. e., that allow us
e and C.
e Then we can compute the matrices A
ek as follows. Vectorization of the
to fix B
second identity in (3.18) yields
K
X
ek ) = C
e T ⊗ In vec(LTq ),
In ⊗ hk (Mq )PFT ,q vec(A
k=1
where ⊗ denotes the Kronecker product and vec(X) denotes the vector of stacked columns
of the matrix X. Similarly, we obtain from (3.19) the equation
K
X
T
ek ) = In ⊗ B
e vec(Rq ).
hk (Sq )PG,q
⊗ In vec(A
k=1
2
2
All equations together yield the linear algebraic system Aα = β with A ∈ CKn ×Kn ,
2
α, β ∈ CKn given by
In ⊗ h1 (M1 ) PFT ,1
···
In ⊗ hK (M1 ) PFT ,1
..
..
.
.
T
In ⊗ h1 (MQ ) P T
·
·
·
I
⊗
h
(M
)
P
n
K
Q
F
F
F ,QF
F ,QF
,
A :=
T
T
h1 (S1 ) PG,1 ⊗ In
···
hK (S1 ) PG,1 ⊗ In
..
..
. T
.T
h1 SQG PG,QG ⊗ In · · · hK SQG PG,QG ⊗ In
e T ⊗ In vec LT
(3.20)
C
1
..
.
e
T
T
vec(A1 )
e ⊗ In vec L
C
QF
.
..
α :=
and
β :=
.
,
e vec (R1 )
I
⊗
B
e
n
vec(AK )
..
.
e vec RQ
In ⊗ B
G
Note that the solution of the linear equation system Aα = β depends on PF ,q and PG,q
and there is some freedom in choosing these matrices. A simple possibility is given by
G
e := Ip 0 (3.21)
e := Im ,
PFT ,q := FqT ∗ ,
PG,q := q ,
B
and
C
∗
0
12
which satisfies the first identity in (3.18) and (3.19) for any choice of ∗. However, the
trivial choice of setting these blocks to zero makes the system matrix A singular. Instead,
we propose to fill the ∗ part of the matrices PF ,q and PG,q such that they are nonsingular
assuming that Fq and Gq have full row rank. A more specific choice of ∗ may even lead to
real-valued realizations as stated in the following lemma.
Lemma 3.12. Let each of the interpolation data sets be closed under complex conjugation,
i. e., there exist unitary matrices TF ,q , TG,q ∈ Cn×n with
TF∗ ,q Mq TF ,q ∈ Rn×n ,
TF∗ ,q LTq ∈ Rn ,
TF∗ ,q FqT ∈ Rn ,
for
q = 1, . . . , QF ,
∗
TG,q
Sq TG,q ∈ Rn×n ,
Rq TG,q ∈ Rn ,
Gq TG,q ∈ Rn ,
for
q = 1, . . . , QG .
Moreover, let the matrices PF ,q and PG,q be as in (3.21) with free entries ∗ chosen such
e1 , ..., A
eK , B,
e and C
e
that TF∗ ,q PFT ,q ∈ Rn and PG,q TG,q ∈ Rn hold. Then, the matrices A
from Theorem 3.11 are real matrices (if existent).
e and C
e are real matrices. In addition, the second
Proof. From (3.21) it is clear that B
equalities in (3.18) and (3.19) are equivalent to
K
X
ek = TF∗ ,q LTq C
e
TF∗ ,q hk (Mq )TF ,q TF∗ ,q PFT ,q A
k=1
and
K
X
ek PG,q TG,q T ∗ hk (Sq )TG,q = BR
e q TG,q .
A
G,q
k=1
ek are the solutions of these linear matrix equations and since their coefficient
Since the A
ek ’s are also real-valued.
matrices as well as the right hand sides are real-valued, the A
To complete the discussion, we analyze the regularity of A in the SISO case, that is
p = m = 1. Here, we set
1
e := ..
e
PF ,q := diag(Fq ),
PG,q := diag(Gq ),
B
. , and C := 1 . . . 1 . (3.22)
1
With these settings,
of the second
in (3.18) and
PKthe (i, j) components
PK matrix equations
e
e
(3.19) read as f q;i k=1 hk (µq;i )[Ak ]i,j = 1 and g q;j k=1 hk (σq;j )[Ak ]i,j = 1, respectively.
Putting this into matrix notation yields the linear system
f 1;i
h1 (µ1;i ) . . . hK (µ1;i )
[A
..
..
..
e1 ]i,j 1
.
.
.
e2 ]i,j 1
h1 (µQ ;i ) . . . hK (µQ ;i )
[A
f QF ;i
F
F
= . ,
.
.
g 1;j
h1 (σ1;j ) . . . hK (σ1;j ) . ..
e
..
..
..
1
[AK ]i,j
.
.
.
g QG ;j
h1 (σQG ;j ) . . . hK (σQG ;j )
(3.23)
where the system matrix is the product of a diagonal matrix and a generalized Vandermonde matrix. This generalized Vandermonde matrix is also called a Haar matrix and is
nonsingular if the driving frequencies µq;i and σq;j are distinct and the functions hk satisfy
the Haar condition [13]. In particular, the Haar condition is satisfied for monomials, and
13
thus relevant for second-order systems (cf. Table 2.1). The diagonal matrix is nonsingular
if the driving frequencies µq;i and σq;j are distinct from the roots of the original transfer
function. In this case, the system above has a unique solution for each (i, j) combination
and hence, via transformations, we can infer that A is nonsingular.
We illustrate the construction of the realization with additional data with the following
toy example.
Example 3.13. Given scalars a1 , a2 , a3 , b, c ∈ R with bc 6= 0, consider the system
a1 ẋ(t) = a2 x(t) + a3 x(t − 1) + bu(t),
y(t) = cx(t)
with transfer function H(s) =
cb
.
sa1 −a2 −e−s a3
Setting QF = 1 and QG = 2, we pick distinct
e = 1 and C
e = 1 with
interpolation points µ1;1 = µ, σ1;1 = σ, and σ2;1 = λ. We choose B
PF ,1 = H(µ), PG,1 = H(σ), and PG,2 = H(λ). Then the system in (3.23) reads as
e
A1
H(µ)
µ −1 − exp(−µ)
1
e
H(σ)
σ −1 − exp(−σ) A2 = 1 .
(3.24)
e
H(λ)
λ −1 − exp(−λ)
1
A3
The inverse of the Haar matrix is given by
µ σ
e (e − eλ )
−eσ (eµ − eλ )
eλ (eµ − eσ )
1
eµ (σeσ − λeλ ) −eσ (µeµ − λeλ ) eλ (µeµ − σeσ )
µeµ (eσ −eλ )+σeσ (eλ −eµ )+λeλ (eµ −eσ )
−eµ eσ eλ (σ − λ) eµ eσ eλ (µ − λ) −eµ eσ eλ (µ − σ)
h
i
e1 A
e2 A
e3 =
such that the solution of (3.24) is given by A
we recover the original transfer function.
1
cb
a1 a2 a3 . In particular,
Clearly, the realization is real-valued if all quantities in (3.23) are real. If we pick the
driving frequencies on the imaginary axis, then in general the Haar matrix will be complexvalued. The following lemma shows how to obtain real-valued realizations based on complex
interpolation data with P matrices as in (3.22).
Lemma 3.14. Let the interpolation data be closed under complex conjugation and sorted
as in Example 3.9 such that the unitary matrices TF , TG ∈ Cn×n from Example 3.9 satisfy
TF∗ Mq TF ∈ Rn×n ,
TG∗ Sq TG
∈R
n×n
,
TF∗ LTq ∈ Rn ,
n
Rq TG ∈ R ,
TF∗ FqT ∈ Rn ,
n
Gq TG ∈ R ,
for
q = 1, . . . , QF ,
for
q = 1, . . . , QG .
Moreover, let the matrices PF ,q and PG,q be as in (3.22). Then, the realization
e1 TG , . . . , TF∗ A
eK TG , TF∗ B,
e CT
e G ),
(TF∗ A
e1 , . . . , A
eK , B,
e C)
e from Theorem 3.11, consists of real-valued matrices and interpowith (A
lates the data.
Proof. First note that the state space transformation by the unitary matrices TF∗ and TG
does not change the transfer function and thus the interpolation given by Theorem 3.11 is
also valid here. It remains to show that the realization consists of real-valued matrices.
e and C
e are given in (3.22), it is straightforward to see that T ∗ B
e and CT
e G are
Since B
F
14
ek TG by observing
real-valued. As in the proof of Lemma 3.12, we deduce the realness of TF∗ A
that the second equalities in (3.18) and (3.19) are equivalent to
K
X
ek TG = TF∗ LTq CT
e G
TF∗ hk (Mq )TF TF∗ PFT ,q TF TF∗ A
k=1
and
K
X
ek TG T ∗ PG,q TG T ∗ hk (Sq )TG = T ∗ BR
e q TG .
TF∗ A
G
G
F
k=1
Straightforward computations yield that TF∗ PFT ,q TF and TG∗ PG,q TG are real-valued. From
ek or equivalently their transformed
these linear matrix equations we can determine the A
ek TG . In the latter case, we observe that the coefficient matrices as well as
analogues TF∗ A
ek TG are also real-valued.
the right hand sides are real-valued and thus the TF∗ A
e
A
crucial point in Theorem 3.11 is the nonsingularity of the affine structure K(s)
=
PK
e
k=1 hk (s)Ak at the driving frequencies µq;i and σq;i . However, if we add more and more
e
data we expect that at some point the information become redundant, and hence K(s)
might become singular. To remove the redundant part, we suppose that
e1
!
A
K
h
i
X
.
ek = rank A
e1 · · · A
eK = rank
rank
hk (s)A
(3.25)
.. =: r
k=1
eK
A
holds for all s ∈ {µq;i } ∪ {σq;i }. In this case there exist unitary matrices V = V1 V2 and
W = W1 W2 ∈ Cn×n with V1 , W1 ∈ Cn×r and V2 , W2 ∈ Cn×(n−r) such that
ek V2 = 0
A
and
e∗ W2 = 0,
A
k
for all
k = 1, . . . , K.
(3.26)
e
e PK hk (s)A
ek )−1 B
e satisfy the equations
Theorem 3.15. Let the realization H(s)
= C(
k=1
ek ’s satisfy the rank
in Theorem 3.11 with matrices PF ,q and PG,q . Suppose that the A
n×r
assumption (3.25) and let V1 , W1 ∈ C
complete V2 and W2 in (3.26) to unitary matrices.
For k = 1, . . . , K set
ek;r := W ∗ A
e
A
1 k V1 ,
er := W ∗ B,
e
B
1
er := CV
e 1.
C
and
If span{`q;1 , . . . , `q;n } = Cp for all q = 1, . . . , QF and span{r q;1 , . . . , r q;n } = Cm for
e r (s) = C
er (PK hk (s) A
ek;r )−1 B
er interpolates the
all q = 1, . . . , QG , then the realization H
k=1
data.
P
e
Proof. First, bear in mind that by the assumption, the affine structure K
k=1 hk (s)Ak;r
ek V1 V ∗ = A
ek
is nonsingular at the driving frequencies µq;i and σq;i , and observe that A
1
ek = A
ek hold for k = 1, . . . , K by construction of V1 and W1 . Thus, for
and W1 W1∗ A
q = 1, . . . , QF
!
K
K
X
X
T
T
∗
ek;r =
ek V1
hk (Mq )P W1 A
hk (Mq )P W1 W A
F ,q
F ,q
k=1
=
k=1
K
X
k=1
15
1
!
ek
hk (Mq )PFT ,q A
er ,
V1 = LTq C
where the second identity follows from (3.18). Similarly, we obtain for q = 1, . . . , QG
K
X
ek;r V1∗ PG,q hk (Sq ) = W1∗
A
k=1
K
X
ek V1 V1∗ PG,q hk (Sq ) = W1∗
A
k=1
K
X
er Rq .
ek PG,q hk (Sq ) = B
A
k=1
Furthermore, notice
e
LTq C
=
K
X
ek
hk (Mq )PFT ,q A
=
k=1
K
X
ek V1 V1∗ = LTq CV
e 1 V1∗ .
hk (Mq )PFT ,q A
k=1
e = CV
e 1V ∗.
Since the columns of Lq span the whole space Cp , the above identity implies C
1
∗
e = W1 W B.
e Finally, we have
With the same reasoning we obtain B
1
e r (µq;i )
`Tq;i H
=
K
X
er
eTi LTq C
!−1
ek;r
hk (µq;i )A
er
B
k=1
=
eTi
K
X
!
ek;r
hk (Mq )PFT ,q W1 A
k=1
= eTi PFT ,q W1
=
K
X
!−1
ek;r
hk (µq;i )A
er
B
k=1
K
X
!
ek;r
hk (µq;i )A
k=1
T T
e=
ei PF ,q W1 W1∗ B
K
X
!−1
ek;r
hk (µq;i )A
er
B
k=1
f Tq;i
for q = 1, . . . , QF and i = 1, . . . , n. The right interpolation conditions follow analogously.
Example 3.16. If we pick further distinct interpolation points in Example 3.13,
realization is given by the matrices
a1 . . . a 1
a2 . . . a2
a3 . . .
1
1
1
.
.
.
.
.
e1 =
e2 =
e3 =
.. , A
.. , and A
A
.
.
.
cb .
cb .
cb .
a1 . . . a 1
a2 . . . a2
a3 . . .
Clearly, the rank assumption (3.25) is satisfied with r = 1. Setting W1 = 1 0
and V1 = W1T yields the true transfer function.
3.4.2
then the
a3
.. .
.
a3
. . . 0]
Matching Derivative Data
Hermite interpolation provides a well known and robust approach for polynomial approximation that involves the matching of derivative data. When we seek reduced models that
are structurally equivalent to standard first order realizations (that is, when we have in
(2.1) K = 2, h1 (s) = s, and h2 (s) ≡ −1) then first order necessary conditions for optimality
of the reduced order approximant with respect to the H2 norm are known and they require
e
that the reduced transfer function H(s)
must be a Hermite interpolant of the original
H(s) [20]. Even though these necessary conditions do not extend immediately to more
general structured systems as appear in (2.1), it is known for some special cases such
as second order systems with modal damping and port-Hamiltonian systems [9], and for
systems with simple delay structures [16, 17], that Hermite interpolation (in a different
form then for the rational case) still plays a fundamental role in the necessary optimality
16
conditions. Therefore, if derivative information for the transfer function H(s) is accessible
e
then this motivates finding a structurally equivalent realization H(s)
that matches both
the evaluation data and the derivative data. Assume that we have
`Ti H 0 (µi ) = (f 0i )T
H 0 (σi )r i = g 0i
and
for i = 1, . . . , n
(3.27)
d
available, where H 0 denotes the derivative of H, i. e., H 0 := ds
H, and (f 0i )T and g 0i are the
tangential interpolation values of H 0 . These are collected in the matrices
F 0 = f 01 . . . f 0n
and
G 0 = g 01 . . . g 0n .
e interpolates the
In this section, we derive conditions such that the transfer function H
n
n
data (2.5) with {µi }i=1 ∩ {σi }i=1 = ∅ and satisfies in addition the Hermite interpolation
condition (3.27).
e
e PK hk (s)A
ek )−1 B
e and suppose that PK hk (s̃)A
ek is
Theorem 3.17. Let H(s)
= C(
k=1
k=1
n
n
nonsingular for all s̃ ∈ {µi }i=1 ∪ {σi }i=1 .
e i ) = f Ti and the left Hermite interpolation
1. The left interpolation conditions `Ti H(µ
e 0 (µi ) = f 0 T are satisfied for i = 1, . . . , n if and only if there exist
conditions `Ti H
i
matrices PF and PF 0 that satisfy
K
X
e
F T = PFT B,
ek = LT C,
e
hk (M)PFT A
(3.28)
k=1
F
0 T
T
e
= (PF 0 ) B,
K
X
T
ek = −
hk (M) (PF 0 ) A
k=1
K
X
ek .
h0k (M)PFT A
(3.29)
k=1
e i )r i = g i and the right Hermite interpolation
2. The right interpolation conditions H(σ
0
0
e
conditions H (σi )r i = g i for i = 1, . . . , n are satisfied if and only if there exist
matrices PG and PG 0 that satisfy
e G,
G = CP
K
X
ek PG hk (S) = BR,
e
A
(3.30)
k=1
0
e G0 ,
G = CP
K
X
ek PG 0 hk (S) = −
A
k=1
K
X
ek PG h0 (S).
A
k
(3.31)
k=1
Proof. We only prove the first statement; the second statement is proved analogously.
Observe that (3.28) are the left interpolation conditions from Theorem 3.2, `Ti H(µi ) = f Ti
for i = 1, . . . , n. It remains to show that the
interpolation conditions are
P left Hermite
e
ek . The second identity in (3.28)
equivalent to (3.29). As before, let K(s)
= K
h
(s)
A
k
k=1
holds if and only if
e = eT LT C
e = eT P T K(µ
e i)
`Ti C
i
i F
for i = 1, . . . , n.
Similarly from the second identity in (3.29) we obtain
e 0 (µi ) = −eTi
−eTi PFT K
K
X
ek = eTi
h0k (M)PFT A
k=1
K
X
k=1
17
ek = eT (PF 0 )T K(µ
e i ).
hk (M) (PF 0 )T A
i
Thus, for i = 1, . . . , n we have
e 0 (µi ) = −`T C
e K(µ
e i )−1 K
e 0 (µi )K(µ
e i )−1 B
e = −eT P T K
e 0 (µi )K(µ
e i )−1 B
e
`Ti H
i
i F
e = f0 T ,
= eTi (PF 0 )T B
i
where the last identity is nothing else than the first equality in (3.29).
As before, it is sufficient and necessary to satisfy (3.28)-(3.31) simultaneously to satisfy the
interpolation conditions (2.6a) and the Hermite interpolation conditions (3.27). After fixing
the matrices PF , PF 0 , PG , and PG 0 , Theorem 3.17 gives 4n2 equations for Kn2 unknown
variables. In particular for K = 4, we can expect under some regularity conditions that
ek . Hereby, the P matrices can be chosen
there is a unique solution for the matrices A
similarly as in the previous section, for example as
0
h
i
T
G
G
T
T
T
0
∗ ,
PF := F
PF 0 := (F )
PG :=
,
PG 0 :=
,
(3.32)
∗ ,
∗
∗
e = Im 0 T and C
e = Ip 0 . For K = 3, we can either satisfy the left or
yielding B
the right Hermite interpolation conditions. For the sake of completeness, we derive the
equivalent of the system (3.20) for Hermite interpolation for K = 4. Vectorization of the
second equations in (3.28)-(3.31), respectively, yields the system Aα = β with matrix
2
2
2
A ∈ C4n ,4n and vectors α, β ∈ C4n given by
T
T
In ⊗ h1 (M)PF
...
In ⊗ h4 (M)PF
T
. . . In ⊗ h4 (M) (PF 0 )T + h04 (M)PFT
In ⊗ h1 (M) (PF 0 ) + h01 (M)PFT
,
A :=
h1 (S)PGT ⊗ In
...
h4 (S)PGT ⊗ In
...
h4 (S) (PG 0 )T + h04 (S)PGT ⊗ In
h1 (S) (PG 0 )T + h01 (S)PGT ⊗ In
T ⊗I
T
e
e
C
vec
L
n
vec(A1 )
e
vec(
A
)
0
2
.
α :=
and
β :=
,
e
e
vec(A3 )
I
⊗
B
vec
(R)
n
e4 )
vec(A
0
(3.33)
Remark 3.18. Real-valued realizations that accomplish Hermite interpolation may be
obtained in the same manner as in the case of additional interpolation points (cf. Lemmas
3.12 and 3.14). The only additional requirement is that F and F 0 as well as G and G 0 need
to have the same number of complex conjugate pairs such that
TF∗ F T ∈ Rn ,
TF∗ F 0
T
∈ Rn ,
GTG ∈ Rn ,
and G 0 TG ∈ Rn .
♦
e
e K(s)
e −1 B
e
Suppose we have
solved the linear system (3.33) to obtain the realization H(s)
=C
P
K
e
e
with K(s) = k=1 hk (s)Ak . By construction, the matrices satisfy the equations in Thee
orem 3.17. However, K(s)
might be singular at the driving frequencies µi and σi . If
the rank condition (3.25) is satisfied, then we can truncate the redundant data as in
ek;r , B
er , and C
er yielding the reduced realization
Theorem 3.15, i. e., we construct matrices A
P
e r (s) = C
er ( K hk (s)A
ek;r )−1 B
er . This model still satisfies the left and right interpolation
H
k=1
18
conditions by Theorem 3.15 and it suffices to check if the Hermite interpolation conditions
are matched as well. By the same reasoning as in the proof of Theorem 3.15 we can
establish the identity
K
X
T
ek;r = −
hk (M) (PF 0 ) W1 A
k=1
K
X
ek;r
h0k (M)PFT W1 A
k=1
and compute
e r0 (µi )
`Ti H
=
K
X
er
−eTi LT C
!−1
K
X
ek;r
hk (µi )A
k=1
= −eTi PFT W1
K
X
k=1
K
X
!
ek;r
h0k (µi )A
k=1
= eTi
K
X
=
(PF 0 )
K
X
!−1
ek;r
hk (µi )A
er
B
k=1
!−1
ek;r
hk (µi )A
er
B
k=1
!
T
ek;r
hk (M) (PF 0 ) W1 A
k=1
eTi
!
ek;r
h0k (µi )A
K
X
!−1
ek;r
hk (µi )A
er
B
k=1
T
e
W1 W1∗ B
=
eTi
T
e = f 0i
(PF 0 ) B
T
and hence the left Hermite interpolation condition is still satisfied. The proof for the
right Hermite interpolation condition proceeds analogously. We summarize the previous
discussion in the following theorem.
e
e PK hk (s)A
ek )−1 B
e satisfy the equations in
Theorem 3.19. Let the realization H(s)
= C(
k=1
ek ’s satisfy the rank
Theorem 3.17 with matrices PF , PF 0 , PG , and PG 0 . Suppose that the A
n×r
assumption (3.25) and let W1 , V1 ∈ C
be as in Theorem 3.15. If span{`
n} =
PK 1 , . e. . , `−1
p
m
e
e
er
C and span{r 1 , . . . , r n } = C , then the realization Hr (s) = Cr ( k=1 hk Ak;r ) B
interpolates the data and derivative data with
ek;r := W1∗ A
ek V1 ,
A
er := W1∗ B,
e
B
and
er := CV
e 1.
C
Remark 3.20. Evidently, as the number of functions K determining the structure increases,
the number of available degrees of freedom to force interpolation increases as well, and
in particular, when K > 4 there will be sufficient degrees of freedom available to allow
matching of higher order derivatives as well. The calculations involved are both annoyingly
technical and unenlightening, so we choose not to pursue this thread here. In any case for
the applications we have in mind, K ≤ 4, and Hermite interpolation is seen to provide a
satisfactory level of fidelity in the reduced models.
♦
3.5
An Algorithm for Structured Realization
In this section, we synthesize the results of the previous subsections into an algorithmic
format, starting with interpolation data (2.5) and an affine structure given via continuously
e1 , . . . , A
eK ,
differentiable functions hk for k = 1, . . . , K. The P
goal is to construct matrices A
K
−1
e
e
e
e
e
e
B and C such that the realization H(s) = C( k=1 hk (s)Ak ) B associated with the
affine structure interpolates the data. We construct realizations as described in previous
designated subsections, taking advantage of the simplifications available when K = 2.
Before doing so, a pre-processing step is included if the data is closed under complex
conjugation, which facilitates construction of a real-valued realization where appropriate.
19
Although in principle the transformation to a real-valued realization could be performed
after assembling the matrices, it is advisable to enforce this in advance, since rounding errors
tend to break the underlying conjugate symmetry and will cause drift away from a realvalued realization. A post-processing step may also be necessary to truncate redundancies
discovered in the interpolation data. Details are summarized in Algorithm 1.
Algorithm 1 Structured Realization
Input: Interpolation data (2.5), affine structure h1 (s), . . . , hK (s) with
PK K ∈ N.e −1 e
e
e
e
e
e
e
Output: Matrices A1 , . . . , AK , B, and C such that H(s) = C( k=1 h(s)Ak ) B
interpolates the data
if Data is closed under complex conjugation then
. Keep realization real
Transform data as in Lemma 3.8, Lemma 3.12, Lemma 3.14, or Remark 3.18
3: end if
1:
2:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
3.6
if K = 2 then
Transform data as in (3.16)
Construct Loewner matrices according to (3.1) and (3.2) from the transformed data
e1 = −L, A
e2 = Lσ , B
e = F T and C
e=G
Set A
else
if derivative data (3.27) is available then
e C
e and P matrices, for example as in (3.32)
Construct B,
e1 , . . . , A
eK
Assemble system (3.33) and solve for A
else
Partition the data as in (3.17) and adjust n accordingly
e C
e and P matrices, for example as in (3.21)
Construct B,
e1 , . . . , A
eK
Assemble system (3.20) and solve for A
end if
end if
Compute r as in (3.25)
if r < n then
Compute V1 and W1 as in Theorem 3.15
∗e
ek := W ∗ A
e
e
e
e
Set A
1 k V1 , B := W1 B, and C := CV1
end if
. Truncation of redundant data
Connection to Structure-preserving Interpolatory Projections
Although our focus here is on data-driven interpolation, we revisit briefly the structurepreserving interpolatory projection framework introduced in [7] and establish a connection
with realizations arising from Corollary 3.6.
Theorem 3.21 (Structure-preserving interpolatory projection [7]). Consider the generalized realization H(s) = C(s)K(s)−1 B(s) where both C(s) ∈ Cp×N and B(s) ∈ CN ×m are
analytic in the right half plane and K(s) ∈ CN ×N is analytic and full rank throughout
the right half plane. Suppose that the left interpolation points {µ1 , . . . , µn } together with
the left tangential directions {`1 , . . . , `n } and the right interpolation points {σ1 , . . . , σn }
together with the right tangential directions {r 1 , . . . , r n } are given. Define V ∈ CN ×n and
20
W ∈ CN ×n as
W = [K(µ1 )−T C(µ1 )T `1 , · · · , K(µn )−T C(µn )T `n ].
(3.34a)
V = [K(σ1 )−1 B(σ1 )r 1 , · · · , K(σn )−1 B(σn )r n ]
(3.34b)
e
e = W T B(s), and C(s)
e = C(s)V.
K(s)
= W T K(s)V, B(s)
(3.35)
and
Define
e
e K(s)
e −1 B(s)
e
Then the reduced transfer function H(s)
= C(s)
satisfies the interpolation
conditions (2.6).
P
If we use K (s) = K
k=1 hk (s) Ak , B (s) = B, and C (s) = C for the affine structure we
employ here, in Theorem 3.21, then (3.35) leads to a reduced model with
ek = W T Ak V, for k = 1, . . . , K, B
e = W T B, and C
e = CV.
A
The question we want to answer next is how (and if) this projection-based reduced model
is connected to the data-driven one we develop here. The next result provides the link.
Proposition 3.22. The projection
matrices W and V introduced in (3.34), based on the
P
matrix functions K (s) = K
h
(s)
Ak , B (s) = B, and C (s) = C, satisfy the matrix
k=1 k
equations
K
K
X
X
T
T
hk (M)W Ak = L C
and
Ak V hk (S) = BR,
(3.36)
k=1
k=1
as well as
K
X
h0k (µi )[W T Ak V ]i,i = −θi
(3.37)
k=1
for those i with µi = σi .
Proof. Let wi = W ei and v i = V ei denote the columns of the projection matrices W and
V . For i = 1, . . . , n we have
eTi
K
X
hk (M)W T Ak = wTi
k=1
K
X
hk (µi )Ak = `Ti C,
k=1
which proves the first identity. The second identity is proven similarly whereas the third
identity follows from the definitions of W and V and from
!
K
K
X
X
0
T
T
0
hk (µi )[W Ak V ]i,i = wi
hk (µi )Ak v i = −`Ti H 0 (µi )r i = −θi .
k=1
k=1
Proposition 3.22 gives a better understanding of the realization of Corollary 3.6 connecting
it to the projection-based MOR framework. To make this connection more precise, we will
investigate the K = 2 and K ≥ 3 cases separately below.
21
3.6.1
The K = 2 case
Using the identities W T B = F T and CV = G, we can rewrite (3.10), using K = 2, as
e1 h1 (S) − h1 (M)A
e1 h2 (S) = h2 (M)W T BR − LT CV h2 (S).
h2 (M)A
Substituting the expressions for BR and LT C from (3.36) into the right-hand side implies
e1 h1 (S) − h1 (M)A
e1 h2 (S) = h2 (M)W T A1 V h1 (S) − h1 (M)W T A1 V h2 (S),
h2 (M)A
e1 = W T A1 V as long as the interpolation sets {µi }n and
which establishes the relation A
i=1
e2 = W T A2 V is obtained by using (3.11) instead of
{σi }ni=1 are disjoint. The identity A
(3.10). Thus for K = 2, our structured realization approach gives exactly the reduced
model one would obtain via projection if the original system matrices were to be available.
This equivalence of the projected matrices and the matrices obtained by the realizations is
also true if there are overlappings between the left and right interpolation point sets. This
may be comprehended by the observation that the projected matrices also satisfy (3.14)
which is clear due to Proposition 3.22.
3.6.2
The K ≥ 3 case
Consider the second-order model H(s) = C(s2 A1 + sA2 + A3 )−1 B. For simplicity, assume
that H(s) is SISO. Given the 2n interpolation points {µ1 , . . . , µn } and {σ1 , . . . , σn }, one
e
e 2A
e1 + sA
e2 + A
e3 )−1 B
e using
can obtain a projection-based reduced model H(s)
= C(s
Theorem 3.21. This reduced model will interpolate H(s) at 2n interpolation points.
e
However, H(s)
has 3n degrees of freedom1 and should be able to satisfy 3n interpolation
conditions. The projection framework cannot achieve this goal. However, our structured
realization framework with either additional data as in Section 3.4.1 or Hermite interpolation
as in Section 3.4.2 will construct a reduced model that can match this maximum number
of interpolation conditions. In other words, for K ≥ 3, the structured realization cannot
be obtained via projection and indeed satisfies more interpolation conditions than the
projection framework. Next we give a numerical example illustrating this discussion on a
delay example.
Example 3.23. We consider the system with affine structure h1 (s) = s, h2 (s) ≡ −1, and
h3 (s) = − exp(−s) and matrices
1 0
1 0
1
1
A1 =
,
A2 = A3 =
,
b=
,
and
c=
1
1
0 2
0 1
with transfer function H(s) = cT (sA1 − A2 − e−s A3 )−1 b. We set QF = 1 and QG = 2
and pick the driving frequencies µ1;1 = 0, σ1;1 = 1, and σ2;1 = −1. We want to make use
e = 1, C
e = 1, PF ,1 = f 1;1 , PG,1 = g 1;1 , and PG,2 = g 2;1 .
of the system (3.23), i. e., we set B
Altogether, the solution of the system (3.23) is given by
(1− 1e )2
(e−1)(e+2)(e+3)
1
e
e
−
+
+
A1
e
2−e
−5−2e
1
e
1− 1e
(e+2)(e+3)
.
1
=
A
2
1 −e − e − e−2 −
−5−2e
2−e− e
1
e
1− e
(e+2)(e+3)
A3
2 + e−2 + −5−2e
e2 =
e3 ; and hence the realization cannot be obtained via projection.
Clearly, A
6 A
e
e1 + C
e2 )(s2 A
e1 + sA
e2 + A
e3 )−1 B,
e i. e., not only the state x(t) but
A second-order model with H(s)
= (sC
also the velocity ẋ(t) is measured, has 4n degrees of freedom. But here we do not consider this case.
1
22
4
Examples
To illustrate the consequences of the preceding theoretical discussion, we compare various
structured realizations against the standard Loewner realization framework, using in each
case response data that is presented as in (2.5). In all the following examples, H(s),
e L (s), H
e A (s), and H
e H (s) will denote, respectively: the transfer function of the original
H
model, the rational approximation via the standard Loewner realization, the structured
realization interpolating at additional points (section 3.4.1), and the structured realization
satisfying Hermite interpolation conditions(section 3.4.2). In the following plots, we
represent interpolation data with circles. Additional driving frequencies used for the
structured realization interpolating additional points are presented as diamonds.
The results presented in the previous sections are valid for the general MIMO case, however,
for simplicity, we restrict ourselves to SISO examples. Accordingly, the P matrices needed
e A and H
e H have been chosen as in (3.22) and as the
for the realizations corresponding to H
analogue for the Hermite case which is
PF := diag(F),
PG := diag(G),
PF 0 := diag(F 0 ),
and PG 0 := diag(G 0 ).
Example 4.1. We test our approaches with the delay model from [7] given by the N × N
matrices
1 1
1 1
A1 = νIN + T,
A2 =
+ 1 (T − νIN ),
A3 =
− 1 (T − νIN ),
τ ζ
τ ζ
where T is an N × N matrix with ones on the sub- and superdiagonal, at the (1, 1), and at
the (N, N ) position and zeros everywhere else. The hk ’s are given by h1 (s) = s, h2 (s) ≡ −1,
and h3 (s) = −e−τ s . We choose N = 500, τ = 1, ζ = 0.01, and ν = 5. The input matrix
B ∈ RN has ones in the first two components and zeros everywhere else and we choose
C = B T . We pick n = 4 logarithmically equidistant points on the imaginary axis between
1ı and 100ı (indicated as circles in Figure 4.1a) together with their complex conjugates.
For the additional point framework (section 3.4.1) we set QF = 1 and QG = 2, such that
we have two additional interpolation points (diamonds in Figure 4.1a) plus their complex
conjugates. The Bode plots of the transfer functions and of the errors are illustrated for the
different approaches in Figure 4.1a and Figure 4.1b, respectively. Both of our approaches
capture the dynamics of the full model (the graphs are almost on top of that of the original
model) and clearly outperform the Loewner realization. This is supported by the H∞
errors for the different realizations presented in Table 4.1 - given also for other choices of n.
Table 4.1: Example 4.1 - H∞ errors of the different realizations
n
Loewner
Additional points
Hermite
4
6
8
10
2.342312e-01
2.449003e-01
3.397454e-01
5.561860e-01
4.496194e-02
5.100268e-02
4.673353e-02
4.454640e-02
4.011660e-02
4.116856e-02
4.307346e-02
3.694951e-02
Clearly, the choice of the complex driving frequencies µi and σj is important and should
be investigated further, but this is not within the scope of this paper.
23
100
Original
Loewner
Additional points
Hermite
|H(ıω)|
10−1
10−2
10−3
10−4
100
101
102
103
freq (rad/sec)
eL, H
e A , and H
eH .
(a) Bode plot of H, H
|H(ıω)|
10−2
10−5
10−8
Loewner
Additional points
Hermite
10−11
10−14
100
101
102
103
freq (rad/sec)
eL, H
e A , and H
eH .
(b) Bode plot of the absolute errors of H
Figure 4.1: Example 4.1 - Transfer functions of the different realizations with n = 4
Example 4.2 (Example 1.1 continued). We generate data for this model using a model
for acoustic transmission in a duct presented by Curtain and Morris in [14]. Based on
a PDE model, Curtain and Morris derive an analytic transfer function for this problem:
H(s) = ρ0 sinh((L − ξ0 )s/c)/ cosh(Ls/c), where ρ0 is the air density. For our case, we
assign parameter values: L = 1, ξ0 = 1/2, c = 1, and ρ0 = 1 and generate data by
sampling the Curtain-Morris transfer function on the imaginary axis between 0.1ı and
10ı (see Figure 4.2a). To keep the realization real we add the complex conjugate driving
frequencies. We seek structurally equivalent realizations to the hypothesized structure from
Example 1.1 that will interpolate this data. The frequency response of the original transfer
function H(s) together with the different structurally equivalent realizations is presented in
Figure 4.2a. The relative error plot Figure 4.2b shows that structured realizations in this
case outperform the Loewner realization by several orders of magnitude. It is noteworthy
that the exact transfer function can be written in accordance with the hypothesized
24
103
Original
Loewner
Additional points
Hermite
102
|H(ıω)|
101
100
10−1
10−2
10−3 −1
10
100
101
102
freq (rad/sec)
(a) Bode plot of the acoustic transmission model and the structured realizations
103
Loewner
Additional points
Hermite
|H(ıω)|
10−1
10−5
10−9
10−13 −1
10
100
101
102
freq (rad/sec)
(b) Relative error plot for the different realizations
Figure 4.2: Example 4.2 - Bode and relative error plot for n = 16
structure using matrices cT = [ 0 0 0 ρ0 ],
1 0 0 0
0
0 1 0 0
0
A1 =
0 0 1 0 , A2 = 0
0 1 0 1
−1
bT = [ 1 0 0 0 ] and
0 0 0
0
0
0 −1 0
, A3 =
0
0 0 0
0 0 0
−1
0
0
0
0
0 0
0 0
.
0 −1
0 0
Example 4.3. A heated rod with distributed control and homogeneous Dirichlet boundary
conditions, which is cooled by delayed feedback, can be modeled (cf. [10, 29]) as
∂v(ξ, t)
∂ 2 v(ξ, t)
=
+ a1 (ξ)v(ξ, t) + a2 (ξ)v(ξ, t − 1) + u(t)
∂t
∂ξ 2
v(0, t) = v(π, t) = 0
25
in (0, π) × (0, T ],
in [0, T ]
(4.1)
For the coefficient functions we choose a1 (ξ) = −2 sin(ξ) and a2 (ξ) = 2 sin(ξ). Discretization
of (4.1) via centered finite differences with step size h := Nπ+1 yields the system
ẋ(t) = (LN + A1,N )x(t) + A2,N x(t − 1) + Bu(t),
y(t) = Cx(t),
where LN ∈ RN ×N is the discrete Laplacian and A1,N , A2,N ∈ RN ×N are discrete approximations of the functions a1 and a2 , respectively. The input matrix B ∈ RN is a vector of
1
ones. As output we use the average temperature of the rod, i. e, C = kBk
B T . For our tests
101
Original
Loewner
Additional points
Hermite
|H(ıω)|
100
10−1
10−2
10−3
10−1
100
101
freq (rad/sec)
102
103
eL, H
e A , and H
eH .
(a) Bode plot of H, H
10−1
|H(ıω)|
10−3
10−5
10−7
Loewner
Additional points
Hermite
10−9
10−1
100
101
freq (rad/sec)
102
103
eL, H
e A , and H
eH .
(b) Bode plot of the absolute errors of H
Figure 4.3: Example 4.3 - Transfer functions of the different realizations with n = 4
we use N = 100 and n = 4 interpolation points on the imaginary axis between 10−1 ı and
103 ı together with their complex conjugates. For the realization obtained by interpolating
additional data we use the same settings as in Example 4.1. Similarly as in Example 4.1,
our approaches are the only ones that capture the qualitative behavior of the original
system (cf. Figure 4.3). This is true for all tested numbers of interpolation data n and is
26
Table 4.2: Example 4.3 - H∞ errors for the different realizations
n
Loewner
Additional points
Hermite
4
6
8
10
5.863023e-01
7.118732e-01
2.735014e-01
2.110771e-01
1.596379e-01
4.716281e-01
3.020142e-02
1.796065e-01
1.751535e-01
7.580182e-02
3.725486e-02
4.085510e-02
further illustrated by the H∞ errors listed in Table 4.2. For this example the difference is
not as striking as in the two preceding examples, which are much harder to approximate
with a rational transfer function of low degree.
Example 4.4. The full model comes from a finite element discretization of a cantilevered
Euler-Bernoulli beam [22, §1.16], resulting in a second order system having the form
A1 ẍ(t) + A2 ẋ(t) + A3 x(t) = Bu(t),
y(t) = Cx(t).
This is a SISO system (m = 1 and p = 1) with N = 400 internal degrees of freedom. The
input u(t) represents a point force applied to the state x1 (B = e1 ), while the output is the
displacement history at xN (C = eTN ). The damping matrix A2 models light proportional
damping: A2 = α1 A1 + α2 A2 with α1 = α2 = 0.05. The realizations are obtained for
n = 30 complex driving frequencies on the imaginary axis between 10−5 ı and 102 ı together
with their complex conjugates (see upper part of Figure 4.4). Since the transfer function of
the original model is a rational transfer function unlike in the previous example, we expect
the Loewner realization to perform close to our proposed approach here, which is indeed
the case as illustrated in Figure 4.4. The figure shows that both the Loewner realization
and the structured realization with additional interpolation points capture the transfer
function of the original model quite accurately.
driving frequencies
additional interpolation points
|H(iω)|
100
10−1
10−2
10−5
Original
Loewner
Additional points
10−4
10−3
10−2
10−1
freq (rad/sec)
100
101
e L and H
e A with n = 30
Figure 4.4: Example 4.4 - Bode plot of H
27
102
10−1
10−2
|H(ıω)|
10−3
10−4
10−5
10−6
Loewner
Additional points
10−7
10−8 −5
10
10−4
10−3
10−2
10−1
freq (rad/sec)
100
101
102
e L (s) and H
e A (s)
(a) Bode plot of the absolute error of H
101
|H(ıω)|
10−1
Loewner stable
Loewner unstable
Additional points stable
Additional points unstable
10−3
10−5
10−7
10−5
10−4
10−3
10−2
10−1
freq (rad/sec)
100
101
102
e L (s) and H
e A (s)
(b) Bode plot of the stable and unstable part of H
Figure 4.5: Example 4.4 - Transfer functions of the realizations with n = 30.
However, the error plot in Figure 4.5a clearly shows the superior behavior of our approach,
e A (s) is one order of magnitude
especially for higher frequencies; the maximum error due to H
e
smaller than the error due to HL (s). We conclude with a remark on the stability of the
reduced models. As one expects, stability of the reduced model in the Loewner framework
depends on the quality of the interpolation (sampling) points. The Loewner framework does
not guarantee a stable reduced model in general. For a better selection of points (in some
cases, optimal) one can, for example, combine the Loewner framework with interpolatory
H2 optimal methods as done in [8]. For cases where the Loewner model is unstable, [19]
offers various effective post-processing techniques allowing to extract a stable model while
not losing much accuracy. One solution is simply to discard the unstable part of the
resulting model. Indeed, this choice can be shown to be the best solution in minimizing
an H2 -related distance; see, for example, [19, 24, 26] for details. For this beam example,
both the Loewner and our approach yield unstable reduced models. Following [19], we
28
then checked how much the stable and unstable parts of the reduced models contribute
to the approximation. For both models, the unstable part has only a minor, negligible
contribution as illustrated in Figure 4.5b, where the frequency response plots for the
stable and unstable part of the Loewner realization and the structured realization obtained
with additional data are displayed. For this example, simply truncating the anti-stable
part of the reduced models and taking only the stable part as the approximation causes
e A (s), while the L∞ norm of the antistable part is
almost no loss in accuracy. Indeed, for H
−6
8.8882 × 10 , the H∞ norm of the stable part is 1.8799. It appears that this unstable
part is due to near non-minimality of the reduced models. We computed poles and zeros
e A (s), and observed that the unstable poles are very nearly matched by corresponding
of H
zeros as listed in Table 4.3 below:
Table 4.3: Example 4.4 - Near pole-zero cancellation
Poles
Zeros
10−1
2.87548 × 10−1
1.94274 × 10−1 ± ı2.83070 × 10−1
5.39018 × 10−2 ± ı3.51431 × 10−1
2.87550 ×
1.94273 × 10−1 ± ı2.83074 × 10−1
5.39041 × 10−2 ± ı3.51432 × 10−1
Unlike the case for the Loewner framework, we cannot simply take the stable-part of
e A (s) as the approximant, since this truncation is performed after conversion to first-order
H
form and destroys the structure we are seeking to retain. For many examples, including
the previous ones considered here, no equivalent, generic, finite-dimensional, first-order
structure exists. Therefore, one might consider modifying Algorithm 1 so that these near
pole-zero cancellations can be detected during the construction and removed without
destroying structure. This is not the focus of this paper and is deferred to a later work.
5
Conclusion
We have introduced a new framework for structured realizations that are derived from
input/output data obtained by measurements
of an (unknown) transfer function. The
e PK hk (s)A
ek )−1 B,
e which allows for a variety of different
models obtained have the form C(
k=1
structures such as internal delays or second order systems. If the chosen structure is a
generalized state space representation then our framework coincides with the Loewner
realization [27]. In this sense, our work can be seen as an extension of the Loewner framework
to more general system structures. Indeed, for K = 2 we showed that structured realizations
can be obtained directly via the Loewner framework with transformed data. Based on
necessary and sufficient conditions for interpolation, we have offered two strategies for the
more general case K > 2, the first allowing for interpolation at additional interpolation
points and the second allowing for additional interpolation of derivative information of
the transfer function. The remarkable effectiveness of such structured realizations is
demonstrated through several examples.
29
References
[1] A. C. Antoulas. Approximation of Large-Scale Dynamical Systems. SIAM, Philadelphia,
PA, USA, 2005.
[2] A. C. Antoulas, C. A. Beattie, and S. Gugercin. Interpolatory model reduction of
large-scale dynamical systems. In J. Mohammadpour and K. M. Grigoriadis, editors,
Efficient Modeling and Control of Large-Scale Systems, pages 3–58. Springer, New
York, NY, USA, 2010.
[3] A. C. Antoulas, A. C. Ionita, and S. Lefteriu. On two-variable rational interpolation.
Linear Algebra Appl., 436:2889–2915, 2012.
[4] A. C. Antoulas, S. Lefteriu, and A. C. Ionita. A tutorial introduction to the Loewner
framework for model reduction. In P. Benner, A. Cohen, M. Ohlberger, and K. Willcox,
editors, Model Reduction and Approximation for Complex Systems. Birkhäuser, Basel,
Switzerland, to appear.
[5] A. C. Antoulas, D. C. Sorensen, and S. Gugercin. A survey of model reduction methods
for large-scale systems. Contemp. Math., 280:193–219, 2001.
[6] U. Baur, P. Benner, and L. Feng. Model order reduction for linear and nonlinear
systems: A system-theoretic perspective. Arch. Comput. Methods Eng., 21(4):331–358,
2014.
[7] C. Beattie and S. Gugercin. Interpolatory projection methods for structure-preserving
model reduction. Systems Control Lett., 58(3):225–232, 2009.
[8] C. Beattie and S. Gugercin. Realization-independent H2 -approximation. In Proc. 51st
IEEE Conf. Decision Control (CDC), pages 4953–4958, Maui, HI, USA, 2012.
[9] C. A. Beattie and B. Benner. H2 -optimality conditions for structured dynamical
systems. Preprint MPIMD/14-18, Max Planck Institute Magdeburg, Germany, 2014.
Available from http://www.mpi-magdeburg.mpg.de/preprints/.
[10] D. Breda, S. Maset, and R. Vermiglio. Numerical approximation of characteristic values
of partial retarded functional differential equations. Numer. Math., 113(2):181–242,
2009.
[11] Y. Chahlaoui, K. A. Gallivan, A. Vandendorpe, and P. Van Dooren. Model reduction
of second-order systems. In P. Benner, V. Mehrmann, and D. C. Sorensen, editors, Dimension Reduction of Large-Scale Systems, pages 149–172. Springer, Berlin/Heidelberg,
Germany, 2005.
[12] S. Chaturantabut, C. Beattie, and S. Gugercin. Structure-preserving model reduction
for nonlinear port-Hamiltonian systems. To appear in SIAM J. Sci. Comput., 2016.
Available as arXiv:1601.00527.
[13] E. W. Cheney. Introduction to Approximation Theory. Chelsea, New York, NY, USA,
second edition, 1982.
[14] R. Curtain and K. Morris. Transfer functions of distributed parameter systems: A
tutorial. Automatica, 45(5):1101–1116, 2009.
[15] Z. Drmač, S. Gugercin, and C. A. Beattie. Quadrature-based vector fitting for
discretized H2 approximation. SIAM J. Sci. Comput., 37(2):A625–A652, 2015.
30
[16] I. Pontes Duff, S. Gugercin, C. Beattie, C. Poussot-Vassal, and C. Seren. H2 -optimality
conditions for reduced time-delay systems of dimension one. IFAC-PapersOnLine,
49(10):7–12, 2016. Proc. 13th IFAC Workshop Time Delay Systems (TDS) 2016,
Istanbul, Turkey, 22-24 June 2016.
[17] I. Pontes Duff, C. Poussot-Vassal, and C. Seren. Realization independent single
time-delay dynamical model interpolation and H2 -optimal approximation. In Proc.
54th IEEE Conf. Decision Control (CDC), pages 4662–4667, Osaka, Japan, 2015.
[18] R. W. Freund. Structure-preserving model order reduction of RCL circuit equations.
In W. H. A. Schilders, H. A. van der Vorst, and J. Rommes, editors, Model Order Reduction: Theory, Research Aspects and Applications, pages 49–73. Springer,
Berlin/Heidelberg, Germany, 2008.
[19] I. V. Gosea and A. C. Antoulas. Stability preserving post-processing methods applied
in the Loewner framework. In Proc. 20th IEEE Workshop Signal Power Integrity
(SPI), pages 1–4, Turin, Italy, 2016.
[20] S. Gugercin, A. C. Antoulas, and C. Beattie. H2 model reduction for large-scale linear
dynamical systems. SIAM J. Matrix Anal. Appl., 30(2):609–638, 2008.
[21] B. Gustavsen and A. Semlyen. Rational approximation of frequency domain responses
by vector fitting. IEEE Trans. on Power Del., 14(3):1052–1061, 1999.
[22] Thomas JR Hughes. The Finite Element Method: Linear Static and Dynamic Finite
Element Analysis. Courier Corporation, 2012.
[23] A. C. Ionita and A. C. Antoulas. Data-driven parametrized model reduction in the
Loewner framework. SIAM J. Sci. Comput., 36(3):A984–A1007, 2014.
[24] M. Köhler. On the closest stable descriptor system in the respective spaces RH 2 and
RH ∞ . Linear Algebra Appl., 443:34–49, 2014.
[25] S. Lall, P. Krysl, and J. E. Marsden. Structure-preserving model reduction for
mechanical systems. Phys. D, 184(1-4):304–318, 2003.
[26] C. Magruder, C. A. Beattie, and S. Gugercin. Rational Krylov methods for optimal H2
model reduction. In Proc. 49th IEEE Conf. Decision Control (CDC), pages 6797–6802,
Atlanta, GA, USA, 2010.
[27] A. J. Mayo and A. C. Antoulas. A framework for the solution of the generalized
realization problem. Linear Algebra Appl., 425(2-3):634–662, 2007.
[28] D. G. Meyer and S. Srinivasan. Balancing and model reduction for second-order form
linear systems. IEEE Trans. Automat. Control, 41(11):1632–1644, 1996.
[29] W. Michiels, E. Jarlebring, and K. Meerbergen. Krylov-based model order reduction
of time-delay systems. SIAM J. Matrix Anal. Appl., 32(4):1399–1421, 2011.
[30] B. Moore. Principal component analysis in linear systems: Controllability, observability,
and model reduction. IEEE Trans. Automat. Control, 26(1):17–32, 1981.
[31] C. Mullis and R. Roberts. Synthesis of minimum roundoff noise fixed point digital
filters. IEEE Trans. Circuits Syst., 23(9):551–562, 1976.
[32] P. Schulze and B. Unger. Data-driven interpolation of dynamical systems with delay.
Systems Control Lett., 97:125–131, 2016.
31
[33] T.-J. Su and R. R. Craig Jr. Model reduction and control of flexible structures using
Krylov vectors. J. Guid. Control Dynam., 14(2):260–267, 1991.
32
| 3cs.SY
|
Semi-Supervised linear regression
arXiv:1612.02391v1 [math.ST] 7 Dec 2016
David Azriel1 , Lawrence D. Brown2 , Michael Sklar3 , Richard Berk2 , Andreas Buja2 and Linda Zhao2˚
Technion – Israel institute of Technology1 , Wharton – University of Pennsylvania2 and Stanford University3
Abstract
We study a regression problem where for some part of the data we observe both the label variable
(Y ) and the predictors (X), while for other part of the data only the predictors are given. Such a
problem arises, for example, when observations of the label variable are costly and may require a
skilled human agent. If the conditional expectation ErY |Xs is exactly linear in X then typically the
additional observations of the X’s do not contain useful information, but otherwise the unlabeled data
can be informative. In this case, our aim is at constructing the best linear predictor. We suggest
improved alternative estimates to the naive standard procedures that depend only on the labeled data.
Our estimation method can be easily implemented and has simply described asymptotic properties.
The new estimates asymptotically dominate the usual standard procedures under certain non-linearity
condition of ErY |Xs; otherwise, they are asymptotically equivalent. The performance of the new
estimator for small sample size is investigated in an extensive simulation study. A real data example
of inferring homeless population is used to illustrate the new methodology.
1
Introduction
The term “semi-supervised learning” was coined in the machine learning literature to describe a situation
in which some of the data is labeled while the rest of the data is unlabeled (Merz et al., 1992). Such
situations occur when the label variable is difficult to observe and may require a complicated or expensive
procedure. A typical example is web document classification, where the classification is done by a human
agent while there are many more unlabeled on-line documents. Specifically, a sample of n observations
from the joint distribution of pX, Y q is given, where Y is a one-dimensional label variable and X is a
p-dimensional vector of covariates or predictors. Also, an additional sample of size m is observed where
only X is given. The purpose is to study procedures that make use of the additional unlabeled data to
better capture the shape of the underlying joint distribution of the labeled data.
˚
Corresponding author: David Azriel, [email protected]. The work of Lawrence D. Brown and Linda Zhao was
supported in part by NSF grant DMS-1512084
1
A large body of literature focuses on the case that Y takes a small number of values and the problem
reduces to a classification task; see e.g., Zhu (2005) and references therein. Wang et al. (2009) divide
the different methods into two approaches: distributional and margin-based. The distributional approach
relies on an assumption relating the conditional expectation ErY |Xs to the marginal distribution of
X and the margin-based approach uses the extra information on X for estimating the Bayes decision
boundary; see Liang et al. (2008) for a Bayesian perspective within this basic approach. Other works
consider continuous Y ’s and use the unlabeled data to learn the structure of the X’s in order to better
estimate a non-parametric regression (Zhou, 2005; Lafferty and Wasserman, 2008; Johnson and Zhang,
2008). These works could be very helpful in situations where non-parametric regression is useful and
unlabeled data are available.
Here we follow a third methodology. This methodology makes use of possible unmodeled non-linearities
in EpY |Xq and adapts information from the unlabeled data to partially adjust for these. In a recent
manuscript, Chakrabortty and Cai (2016) follow the same methodological plan but the technical details
within their approach are then quite different from ours. Their estimates offer some advantages and some
disadvantages in relation to ours, which we intend to explore in a future study.
We aim at estimating the vector β composed of the parameters of the best linear predictor of Y , but
we do not necessarily assume that ErY |Xs is exactly linear in X. The estimates of β can then be used to
provide predictions. The standard approach in the statistical literature may be best summarized by the
following quote from Little (1992):
The related problem of missing values in the outcome Y was prominent in the early history
of missing-data methods, but is less interesting in the following sense: If the X’s are complete
and the missing values of Y are missing at random, then the incomplete cases contribute no
information to the regression of Y on X1 , . . . , Xp .
But see Cochran (1977, Chapter 7) for a different view more closely in tune with our development.
Buja et al. (2016) show that β does not depend on the distribution of X if and only if ErY |Xs is
linear in X. When the conditional expectation is linear, there is typically at most a limited amount of
additional information in the unlabeled data. Brown (1990) shows that even in this case there may be
some useful information, but it will not provide an asymptotic advantage in the manner we suggest in the
current treatment when the conditional expectation is not linear. The assumption that ErY |Xs is linear
is unrealistic in many situations, and we show that in the absence of such an assumption the unlabeled
sample can be used to provide useful information for the estimation of β.
We consider two scenarios. In the first one the distribution of X is known exactly. This is equivalent
to having infinitely many unlabeled observations (m “ 8). In the second, more frequently encountered,
2
scenario m ă 8 but it is still significant in the sense that m{n is bounded away from 0. We call these
two scenarios total and partial information, respectively. In both situations we provide an asymptotically
better estimate of β than the standard least squares estimator, and hence also a better linear predictor.
Our new procedure is closely related to the semi-supervised estimations of means, which is the topic of a
recent manuscript (Zhang et al., 2016).
To understand our basic methodology, consider the simple linear regression model,
Y “ α ` βX ` δ,
(1)
where α, β are least squares coefficients and δ is a remainder term; these will be defined in more detail
shortly. When EpXq and EpX 2 q are known, we construct an estimator of β that asymptotically dominates
the least squares estimator. This is done by replacing model (1) with a different model where β is the
intercept and also the expectation of the newly defined label variable. For this latter model, Zhang et al.
(2016) show that the intercept estimator dominates the simple empirical estimate in a semi-supervised
setting. We use their ideas to show that the intercept estimator in this new model is better than the least
squares estimator from the original model as explained below.
To present the intercept model and the methodology, begin by stating the general form of the best
linear approximation. The approximation is used in two different models, and therefore we introduce now
the notation U and W instead of X and Y . Suppose that W P R and U P Rd are random variables with
joint distribution G and finite second moments. The best linear predictor is
T
θ “ pθ1 , . . . , θd q “ arg minθ̃PRd EpW ´
d
ÿ
j“1
(´1
EpUW q.
Uj θ̃j q2 “ EpUUT q
Notice that this is a population version of the least squares where
řd
j“1 Uj θ̃j
(2)
minimizes the L2 distance
in the population from W to any linear function of U. It follows that
W “
where the remainder term r “ W ´
řd
j“1 θj Uj
d
ÿ
j“1
θj Uj ` r,
(3)
is orthogonal to U, i.e., EprUq “ 0. Given a sample of n
observations from G, the standard least squares estimate, θ̂ LSE , satisfies asymptotically (White, 1980)
?
´
(´1 ` 2
˘
(´1 ¯
D
.
npθ̂ LSE ´ θq ÝÑ N 0, EpUUT q
E r UUT EpUUT q
(4)
(´1
Unlike the standard fixed X assumption, the asymptotic variance has a “sandwich” form, EpUUT q
`
˘
forming the “bread” and E r 2 UUT the “meat”. See Buja et al. (2016) for further discussion of this
form of the sandwich.
3
We now return to model (1) and assume that EpXq and EpX 2 q are known. We can therefore assume
w.l.o.g that X is standardized (i.e., EpXq “ 0 and EpX 2 q “ 1). In this case we can write model (1) with
α “ EpY q, β “ EpXY q and δ “ Y ´ pα ` βXq; here we consider model (3) with W “ Y , U1 “ 1, U2 “ X
and the remainder term is r “ δ. The standard LSE is a consistent estimate for β and satisfies, according
to (4),
?
(
D
npβ̂LSE ´ βq ÝÑ N 0, EpδXq2 .
Our aim it to better estimate β “ EpXY q. To this end, we multiply (1) by X setting XY to be the
labeled variable W . Furthermore, we also center the varibale U setting EpXY q “ β to be the intercept.
Specifically,
XY “ αX ` βX 2 ` Xδ “ β ` aX ` bpX 2 ´ 1q ` δ̃.
(5)
Now we consider model (3) with W “ XY , U1 “ 1, U2 “ X, U3 “ X 2 ´ 1 and r “ δ̃; here a, b are θ2 , θ3
defined by (2). In setting β “ EpXY q to be the intercept coefficient we used that EX “ EpX 2 q ´ 1 “ 0,
i.e., that we know the first and second moment of X. We define our new estimator, β̂T I (TI for total
information), to be the intercept estimator of (5). This estimate has a simple closed formula, which is
given below in (14). The sandwich theorem (4) implies that
?
D
npβ̂T I ´ βq ÝÑ N p0, E δ̃2 q.
Since δX ´ δ̃ “ Xpa ´ αq ` pX 2 ´ 1qpb ´ βq is a linear function of X, X 2 , then it is orthogonal to the
remainder term of (5), which is δ̃, (i.e., EpδX ´ δ̃qδ̃ “ 0) and therefore,
EpδXq2 “ Etδ̃ ` pδX ´ δ̃qu2 “ E δ̃2 ` EpδX ´ δ̃q2 .
(6)
This implies that the asymptotic variance of β̂T I is smaller than that of β̂LSE with equality iff δX ” δ̃.
The latter occurs iff α “ a and b “ β or equivalently that δ is uncorrelated with both X 2 and X 3 . This
occurs when the non-linear part of ErY |Xs (if exists) is uncorrelated with X 2 and X 3 . In this case,
Models (1) and (5) are essentially the same model and nothing is gained in the new methodology. On
the other hand, when δ is correlated with either X 2 or X 3 , then Models (1) and (5) are different and β̂T I
has smaller asymptotic variance than β̂LSE . We further show that a similar decomposition to (6) holds
for the partial information case and we generalize the results to the p-dimensional case.
The rest of the paper is organized as follows. Section 2 provides the basic setting the and loss functions
that we use. The main results are given in Section 3. Section 4 describes an extensive simulation study and
Section 5 discusses estimation of the asymptotic covariance matrix of the estimates. An implementation
of the new methodology to infer homeless population in Los Angeles is discussed in Section 6. Section 7
concludes with final remarks. The proofs are given in Section 8.
4
2
Preliminaries
Consider a sample of n i.i.d observations pXp1q , Y p1q q, . . . , pXpnq , Y pnq q from a joint distribution G, where
X P Rp , Y P R, and an additional set of m independent observations pXpn`1q . . . , Xpn`mq q from the
marginal distribution of X. We use super-index to denote the number of the observation, and sub-index
to denote coordinates of X. The notation X, Y without super-index denotes a random vector whose
distribution is G.
~ “ p1, X1 , . . . , Xp qT to be a vector X with an additional constant 1 to accommodate an
We write X
´
¯
~X
~ T is invertible.
intercept term. Assume that the second moments of G exists and that the matrix E X
Then, we can define
! ´
¯)´1 ´
¯
T
~
~
~
pα, βq “ arg minα̃PR,β̃PRp EpY ´ α̃ ´ β̃ Xq “ E XX
E XY .
T
2
(7)
In the presence of non-linearity, β is still a meaningful parameter that describes the overall association
between Y and X (Buja et al., 2016). We have in mind two related purposes. The first purpose is just
to better estimate the parameters of interest, while the second purpose pertains to prediction. The latter
is formalized now. Suppose that an independent observation pX˚ , Y ˚ q „ G is given. The optimal linear
predictor is α ` β T X ˚ . We consider the excess loss of an estimator α̃, β̃
¯2 `
´
˘2
T
Lpα̃, β̃q “ Y ˚ ´ α̃ ´ β̃ X˚ ´ Y ˚ ´ α ´ β T X˚ .
We have that (see Lemma 1)
p
ı2
!
)
”
ÿ
β̃j tX̄j ´ EpXj qu ´ EpY q ` E pβ ´ β̃qT Mpβ ´ β̃q “ E L̃pα̃, β̃q,
ELpα̃, β̃q “ E Ȳ ´
j“1
where L̃ is the expression inside the expectation and M is the covariance matrix of X. We further define
the asymptotic risk as
Rpα̃, β̃q “ lim lim E mintnL̃pα̃, β̃q, Bu.
BÑ8 nÑ8
The loss is of order 1{n and therefore we consider expectation of nL̃. This is truncated by an arbitrarily
large number B, since when the loss is large, it makes sense not to penalize any further. Also, the
truncation helps to avoid issues of uniform integrability. This is done for example in Le Cam and Yang
(1990), Chapter 5. The following proposition provides a simple expression for the asymptotic risk.
Proposition 1. Let α̃, β̃ satisfies
α̃ “ Ȳ ´
?
D
p
ÿ
β̃j X̄j ,
j“1
npβ̃ ´ βq ÝÑ N p0, Σq then,
p
ı2
”
ÿ
βj tXj ´ EpXj qu ´ EpY q ` TracepMΣq,
Rpα̃, β̃q “ E Y ´
and assume that β̃ satisfies
(8)
j“1
5
(9)
where M is the covariance matrix of X.
The first term in (9) does not depend on the distribution of β̃. Hence, Proposition 1 shows that
the excess risk is minimized when TracepMΣq is small. Thus, we aim at estimators β̃ with asymptotic
distribution N p0, Σq such that Σ is “smaller” than the covariance matrix of LSE, in the sense that the
difference is positive semidefinite. Such an estimator asymptotically better estimates β and also has
smaller asymptotic excess risk.
For two estimates pα̃p1q , β̃ p1q q and pα̃p2q , β̃ p2q q, Proposition 1 implies that
ELpα̃p1q , β̃
p1q
q ´ ELpα̃p2q , β̃
p2q
!
)
q « Trace MpΣp1q ´ Σp2q q {n,
and therefore the difference of the prediction errors is
"
"
*2
*2
´ p1q ¯T
´ p2q ¯T
!
)
˚
p1q
˚
˚
p2q
˚
E Y ´ α̃ ´ β̃
X
´ E Y ´ α̃ ´ β̃
X
« Trace MpΣp1q ´ Σp2q q {n.
(10)
It follows that the difference of the errors of the prediction of the mean is also
"
"
*2
*2
´ p1q ¯T
´ p2q ¯T
!
)
˚
˚
p1q
˚
˚
˚
p2q
˚
E EpY |X q ´ α̃ ´ β̃
X
´E EpY |X q ´ α̃ ´ β̃
X
« Trace MpΣp1q ´ Σp2q q {n.
(11)
Notice that we estimate the intercept, as in (8), using X̄j rather than the possibly known expectations
EpXj q. This is in accordance with the findings of Zhang et al. (2016).
3
Main results
In this section we provide the main theoretical results of the paper. Before we deal with the p-dimensional
case, we first introduce the results for one-dimensional X. The reason is twofold: first, the presentation
in the one-dimensional case is simpler and captures the main ideas, and second, our results for the pdimensional X are obtained by reducing the regression problem to p problems, each of which is closely
related to the one dimensional X regression.
3.1
One dimensional X
In this section we study the one dimensional case. Summary of the notation used here is presented in
Table 1. When G has finite second moments we can write
Y “ α ` βX ` δ,
where α “ EpY q ´ βEpXq, β “
CovpX,Y q
V arpXq
(12)
and δ “ Y ´ α ´ βX. Equation (12) is the best linear
approximation in the population in the sense that α ` βX minimizes Epα̃ ` β̃X ´ Y q2 over all α̃, β̃ P R
6
and δ is orthogonal to 1, X, i.e., Epδq “ EpδXq “ 0. The regular LSE is
řn
piq ´ X̄qY piq
i“1 pX
.
β̂LSE “ ř
n
piq ´ X̄q2
i“1 pX
This is a consistent and asymptotically unbiased and asymptotically normal estimator for β. See Buja et al.
(2016) for contemporary discussion.
Table 1: Summary of the notation used in Section 3.1.
Basic model
Y “ α ` βX ` δ
Intercept model
W “ β ` aU1 ` bU2 ` δ̃
CovpX,Y q
V arpXq and δ “ Y ´ α ´ βX
X´EpXq
tX´EpXquX
´1
V arpXq , U2 “
V arpXq
α “ EpY q ´ βEpXq, β “
Y tX´EpXqu
V arpXq ,
W “
U1 “
a, b are θ2 , θ3 defined by (2) for model (13)
TI estimator
PI estimator
β̂T I “ W̄ ´ âŪ1 ´ b̂Ū2
¯ ´ âǓ
¯ ´ b̂Ǔ
¯
β̂P I “ W̌
1
2
1
n
W̄ “
and δ̃ is the remainder term
piq
1 řn
piq
i“1 W , Ū1 “ n
i“1 U1 , Ū2 “
řn
â, b̂ are the LSE of model (13)
1
n
řn
piq
i“1 U2
¯ , Ǔ¯ , Ǔ
¯ are the means over the labeled sample,
W̌
1
2
piq ´ĚpXqu
piq
tX piq ´ĚpXquY piq
, Ǔ1 “ tX ~
~
V arpXq
V arpXq
tX piq ´ĚpXquX piq
´ 1, â, b̂ are the LSE
~
V
arpXq
W̌ piq “
piq
Ǔ2 “
of model (13) with W̌ , Ǔ1 , Ǔ2 instead of W, U1 , U2
n
ν “ lim n`m
n (m) the labeled (unlabeled) sample size
2
σLSE
, σP2 I , σT2 I
asymptotic variance of the LSE, PI, TI estimators
Asymptotics
Main results
2
σdiff
2
2
σLSE
“ σT2 I ` σdiff
2
σP2 I “ σT2 I ` νσdiff
as in Theorem 1
ı2
”
´
δ̃
“ E δtX´EpXqu
V arpXq
For the total information estimator (TI) we consider the following regression model, which is obtained
by multiplying (12) by
X´EpXq
V arpXq ,
W “ β ` aU1 ` bU2 ` δ̃,
where W “
Y tX´EpXqu
V arpXq ,
U1 “
X´EpXq
V arpXq ,
U2 “
tX´EpXquX
V arpXq
(13)
´ 1, a, b are the coefficients θ2 , θ3 of the best
linear approximation defined by (2) and δ̃ is the remainder term. Here again β ` aU1 ` bU2 is the best
linear approximation of W in the population. The multiplication term makes the expectation of W to
ı
”
q
“ CovpX,Y
be E Y tX´EpXqu
V arpXq
V arpXq “ β. The covariates U1 , U2 are the predictors 1, X in (12) multiplied by
X´EpXq
V arpXq .
We subtract 1 in U2 in order to set EpU2 q “ 0. Since EpU1 q “ EpU2 q “ 0, then the intercept is
EpW q “ β. Thus, we define the total information estimator to be the intercept estimator of (13), i.e.,
β̂T I “ W̄ ´ âŪ1 ´ b̂Ū2 ,
7
(14)
where â, b̂ are the regular LSE of model (13), and ¯¨ denotes the mean over the supervised sample with n
observations.
For the partial information (PI) estimator we estimate W, U1 , U2 as follows:
tX piq ´ ĚpXquY piq piq tX piq ´ ĚpXqu
tX piq ´ ĚpXquX piq
piq
, Ǔ1 “
and Ǔ2 “
´ 1 , i “ 1, . . . , n,
V}
arpXq
V}
arpXq
V}
arpXq
1 řn`m
1 řn`m
piq }
2
2
2
piq 2
where ĚpXq “ n`m
i“1 X , V arpXq “ ĚpX q ´ tĚpXqu , and ĚpX q “ n`m
i“1 tX u . The
W̌ piq “
partial information estimator is
¯ ´ âǓ
¯ ´ b̂Ǔ
¯ ,
β̂P I “ W̌
1
2
where â, b̂ are the regular LSE of the regression model (13) with W̌ , Ǔ1 , Ǔ2 instead of W, U1 , U2 . (The
estimates â, b̂ are different in the total and partial information cases but the same notation is presented for
simplicity.) We use ¯¨ (respectively, ˇ¨) to denote empirical mean with respect to the labeled n (respectively,
full n ` m) sample. The following theorem states the asymptotic distribution of β̂LSE , β̂T I and β̂P I . The
first part of the theorem is known and is stated here for comparison purposes.
Theorem 1.
(i) Suppose that V arpXq P p0, 8q and that δtX ´ EpXqu has finite second moment, then,
2
where σLSE
“E
”
δ2 tX´EpXqu2
tV arpXqu2
ı
.
?
D
2
npβ̂LSE ´ βq ÝÑ N p0, σLSE
q,
(ii) Suppose further that the vector pW, U1 , U2 q has finite second moments and that the matrix EpUUT q
is invertible for U “ p1, U1 , U2 qT . Then,
?
2
where σT2 I “ Epδ̃2 q and σLSE
“ σT2 I
D
npβ̂T I ´ βq ÝÑ N p0, σT2 I q,
”
ı2
2 “ E δtX´EpXqu ´ δ̃ .
2
where σdiff
` σdiff
V arpXq
n
“ ν, then,
(iii) Suppose further that lim n`m
?
D
npβ̂P I ´ βq ÝÑ N p0, σP2 I q,
2 .
where σP2 I “ σT2 I ` νσdiff
2
2
2 ą 0 then σ 2 ă σ 2
Therefore, if σdiff
TI
LSE and if further ν ă 1 then σP I ă σLSE .
2 V arpXq.
Corollary 1. Theorem 1 and Proposition 1 imply that Rpα̂LSE , β̂LSE q´Rpα̂P I , β̂P I q “ p1´νqσdiff
2 ą 0. The quantity
The improvement of β̂T I and β̂P I over β̂LSE depends on the assumption that σdiff
2
measures the difference between the original regression model (12) and the intercept model (13). If
σdiff
EpδX 2 q “ EpδX 3 q “ 0 then a “ α and b “ β and the two models are essentially the same, in which
2 “ 0, otherwise σ 2 ą 0. In other words, when ErY |Xs is non-linear and the non-linear part is
case σdiff
diff
2 ą 0.
correlated with either X 2 or X 3 then σdiff
8
3.2
Multidimensional X
We now consider the general p-dimensional case as described in the beginning of Section 2. The notation
of this section is summarized in Table 2. The model can be written as
Y “ α ` β1 X1 ` ¨ ¨ ¨ ` βp Xp ` δ,
(15)
where α, β are the coefficients of the best linear predictor in the population defined by (7). The remainder
ř
term δ “ Y ´ pj“1 βj Xj satisfies Epδq “ EpδX1 q “ ¨ ¨ ¨ “ EpδXp q “ 0. Our aim here is at estimating
β. We use the adjustment representation of Buja et al. (2016) that reduces the p-dimensional estimation
procedure to p separate simple regression problems. Correspondingly, we will define our new estimates
by solving p mean-estimation-problems separately, one for each coordinate j.
~ ´j “ p1, X1 , . . . , Xj´1 , Xj`1 , . . . , Xp qT and let
Let X
¯2 ! ´
´
¯)´1 ´
¯
T
~ ´j “ E X
~ ´j X
~T
~ ´j Xj .
β ´j‚ “ arg min E Xj ´ β̃ X
E
X
´j
(16)
β̃
EpY Xj‚ q
2 q .
EpXj‚
p1q
pnq
pXj , . . . , Xj qT
~ T β ´j‚ . Each βj can be written in the one dimensional form βj “
Now, define Xj‚ “ Xj ´ X
´j
The standard LSE can be viewed in a similar manner. Let Y “ pY p1q , . . . , Y pnq qT , Xj “
)´1
!
and let X´j “ p1, X1 , . . . , Xj´1 , Xj`1 , . . . , Xp q. Define also β̂ ´j¯‚ “ XT´j X´j
XT´j Xj and Xj¯‚ “
Xj ´ X´j β̂ ´j¯‚ ; then, tβ̂ LSE uj “
xY,Xj¯‚ y
||Xj¯‚ ||2 .
‚
Recall that ¯¨ denotes mean over the labeled sample; thus, ¯
denotes adjustments over the labeled sample, whereas ‚ denotes adjustments over the population.
The total information estimator is the intercept estimator of the regression model in the population
obtained by multiplying (15) by
Xj‚
2 q,
EpXj‚
that is,
Wj “ βj ` aU1 `
p
ÿ
j 1 “1
bj 1 Uj 1 `1 ` δ̃j ,
(17)
where
Wj “
Xj 1 Xj‚
Y Xj‚
Xj‚
Xj Xj‚
1
, U1 “
, Uj 1 `1 “
2
2
2 q for j ‰ j, and Uj`1 “ EpX 2 q ´ 1
EpXj‚ q
EpXj‚ q
EpXj‚
j‚
are Y, 1, X1 , ¨ ¨ ¨ , Xp multiplied by
Xj‚
2 q
EpXj‚
(18)
and a, b are θ2 , . . . , θp`1 defined by (2) for W “ Wj and U “
pU1 , . . . , Up`1 q. Since EpU1 q “ ¨ ¨ ¨ “ EpUp`1 q “ 0, then the intercept is EpWj q “ βj . In setting EpWj q “
βj and also the expectations of the U ’s to be zero, we exploited the knowledge of the moments of X,
ř
yielding higher efficiency as shown below. The remainder term in (17), δ̃j “ Wj ´ βj ´ aU1 ´ pj1 “1 b1j Uj 1 `1
is orthogonal to pU1 , . . . , Up`1 q. Notice that the vector pU1 , . . . , Up`1 q depends on j but this is suppressed
in the notation.
Specifically, we define the total information (TI) estimator to be
tβ̂ T I uj “ W̄j ´ âŪ1 ´
9
p
ÿ
j 1 “1
b̂j 1 Ūj 1 `1 ,
Table 2: Summary of the notation used in Section 3.2 when j is the chosen coordinate.
Y “α`
Basic model
Adjusted regressor
Intercept model
řp
j“1 βj Xj
`δ
~ T β ´j‚
Xj‚ “ Xj ´ X
´j
Wj “ βj ` aU1 `
β ´j‚
α, β are defined by (7)
! ´
¯)´1 ´
¯
~ ´j X
~T
~ ´j Xj
“ E X
E
X
´j
~ ´j “ p1, X1 , . . . , Xj´1 , Xj`1 , . . . , Xp qT
X
řp
j 1 “1 bj 1 Uj 1 `1
` δ̃j
Wj “
Uj 1 `1 “
Xj 1 Xj‚
2 q
EpXj‚
Y Xj‚
2 q , U1
EpXj‚
“
Xj‚
2 q
EpXj‚
for j 1 ‰ j, Uj`1 “
Xj Xj‚
2 q
EpXj‚
´ 1,
a, b are θ2 , ¨ ¨ ¨ , θp`1 defined
by (2) for model (17),
δ̃j is the remainder term
TI estimator
PI estimator
tβ̂ T I uj “ W̄j ´ âŪ1 ´
řp
j 1 “1 b̂j
1
Ūj 1 `1
¯¨ is the mean over the labeled sample
â, b̂ are the LSE of model (17)
¯ ´ âǓ
¯ ´ řp b̂ 1 Ǔ
¯1
tβ̂ P I uj “ W̌
j
1
j 1 “1 j j `1
¯¨ is the mean over the labeled sample
W̌j “
Ǔj 1 `1 “
Xj 1 Xjˇ‚
2 q
ĚpXjˇ
‚
Y Xjˇ‚
2 q,
ĚpXjˇ
‚
Ǔ1 “
Xjˇ‚
2 q
ĚpXjˇ
‚
for j 1 ‰ j, Ǔj`1 “
Xj Xjˇ‚
2 q
ĚpXjˇ
‚
´1
â, b̂ are the LSE of model (17) with
pW̌j , Ǔ1 , . . . , Ǔp`1 q replacing pWj , U1 , . . . , Up`1 q
Asymptotics
n
ν “ lim n`m
n (m) the labeled (unlabeled) sample size
ΣLSE , ΣP I , ΣT I
asymptotic covariance matrix of LSE, PI, TI
as in Theorem 2
Main results
ΣLSE “ ΣT I ` Σdiff
Σdiff “ CovpδX‚ ´ δ̃q
ΣP I “ ΣT I ` νΣdiff
where â, b̂ are the regular LSE of (17), and ¯¨ denotes the mean over the labeled sample with n observations.
To define the partial information estimator let
! ´
¯)´1 ´
¯
~ ´j X
~ T´j
~ ´j Xj ,
β ´jˇ‚ “ Ě X
Ě X
(19)
where Ě is the the empirical mean based on the full X sample of size n ` m. Now define Xjˇ‚ “
!
)T
~ ´j β ´jˇ‚ and
Xj ´ X
W̌j “
Xj 1 Xjˇ‚
Y Xjˇ‚
Xjˇ‚
Xj Xjˇ‚
, Ǔ1 “
, Ǔj 1 `1 “
for j 1 ‰ j, and Ǔj`1 “
.
2
2
2
2 q
ĚpXjˇ‚ q
ĚpXjˇ‚ q
ĚpXjˇ‚ q
ĚpXjˇ
‚
10
Here ˇ‚ denotes adjustments over the full n ` m sample. The partial information (PI) estimator is
¯ ´ âǓ
¯ ´
tβ̂ P I uj “ W̌
j
1
p
ÿ
¯1 ,
b̂j 1 Ǔ
j `1
j 1 “1
where â, b̂j 1 are the regular LSE of model (17) with pW̌j , Ǔ1 , . . . , Ǔp`1 q replacing pWj , U1 , . . . , Up`1 q.
The following theorem establishes the asymptotic distribution of the estimates and states conditions
under which β̂ T I and β̂ P I , asymptotically dominate β̂ LSE . The asymptotic distribution of β̂ LSE is already
known (White, 1980) and it is presented here so that the comparison to β̂ T I and β̂ P I can be made.
Theorem 2.
(i) Suppose that the vector X‚ “
!
Xp‚
X1‚
2 q , . . . , EpX 2 q
EpX1‚
p‚
)
is well defined (i.e., the projections
in (16) exist and EpXj‚ q2 is positive and finite for j “ 1, . . . , p) and that δX‚ has finite second
moments, then,
?
D
npβ̂ LSE ´ βq ÝÑ N p0, ΣLSE q,
where ΣLSE “ CovpδX‚ q.
(ii) Suppose further that for each j “ 1, . . . , p, the vector pWj , U1 , U2 , . . . , Up`1 q has finite second moments and that the matrix EpUj UTj q is invertible for Uj “ p1, U1 , U2 , . . . , Up`1 qT ; then,
?
D
npβ̂ T I ´ βq ÝÑ N p0, ΣT I q,
where ΣT I “ Covpδ̃q for δ̃ “ pδ̃1 , . . . , δ̃p q. Furthermore, we have that ΣLSE “ ΣT I ` Σdiff where
Σdiff “ CovpVq for V “ δX‚ ´ δ̃.
n
“ ν, then,
(iii) Suppose further that lim n`m
?
D
npβ̂ P I ´ βq ÝÑ N p0, ΣP I q,
where ΣP I “ ΣT I ` νΣdiff .
Therefore, if Σdiff is not the zero matrix, then ΣLSE ´ ΣT I is positive definite, and if further ν ă 1,
then ΣLSE ´ ΣP I is also positive definite.
Corollary 2. Theorem 2 and Proposition 1 imply that Rpα̂LSE , β̂ LSE q ´ Rpα̂P I , β̂ P I q “ Tracetp1 ´
νqΣdiff Mu, where M “ CovpXq.
In short, Theorem 2 states that if the regression models (15) and (17) are well defined and the residuals
have finite second moments, then ΣLSE ´ ΣT I is positive semi-definite and it is strictly positive unless
V ” 0. If further, the unlabeled sample size is not negligible, i.e. lim m{n ą 0, then ΣLSE ´ ΣP I is also
positive definite.
11
As in the one-dimensional case, β̂ T I and β̂ P I improve over β̂ LSE only when Σdiff is not the zero matrix
or equivalently that V is not zero. For each j, V arpVj q measures the difference between the original model
(15) and the intercept model (17). When
2
2
2
EpXj‚
δq “ EpXj‚
X1 δq “ ¨ ¨ ¨ “ EpXj‚
Xp δq “ 0
(20)
then δXj‚ ” δj and the two models are essentially the same, in which case Vj ” 0. Otherwise, if for
some j (20) does not hold, then β̂ P I improves over β̂ LSE when ν ă 1. In the one-dimensional case, (20)
is equivalent to EpX 2 δq “ EpX 3 δq “ 0, i.e., that X 2 and X 3 are uncorrelated with δ. Generally, (20)
implies that certain linear combinations of EpXj Xj 1 δq and EpXj Xj 1 Xj 2 δq are 0. Thus, roughly speaking,
when ErY |Xs is non-linear and the non-linear part is correlated with several second or third moments of
X, then we expect improvement of β̂ P I .
3.3
New methodology (summary)
We now provide a step-by-step description of the new methodology for estimation of β.
For each j “ 1, . . . , p:
)T
!
piq
piq
~ piq
1. (Adjust the regressors) Let β ´jˇ‚ as defined in (19) and define Xjˇ‚ “ Xj ´ X
β ´jˇ‚ for
´j
i “ 1, . . . , n.
2. (Define the intercept model) Define for i “ 1, . . . , n
“
Y piq Xjˇ‚
2 q
ĚpXjˇ
‚
piq
piq
piq
piq
W̌j
piq
, Ǔ1 “
Xjˇ‚
2 q
ĚpXjˇ
‚
piq
, Ǔj 1 `1 “
piq
Xj 1 Xjˇ‚
2 q
ĚpXjˇ
‚
piq
piq
for j 1 ‰ j, and Ǔj`1 “
piq
Xj Xjˇ‚
2 q
ĚpXjˇ
‚
´ 1.
3. (Define the intercept estimator) The partial information (PI) estimator is
p
¯ ´ âǓ
¯ ´ ÿ b̂ 1 Ǔ
¯
tβ̂ P I uj “ W̌
j
1
j j 1 `1 ,
j 1 “1
where â, b̂j 1 are the regular LSE of the regression model
piq
Wj
piq
“ βj ` aǓ1 `
p
ÿ
j 1 “1
piq
piq
bj 1 Ǔj 1 `1 ` δ̃j , for i “ 1, . . . , n.
The methodology is built from standard least squares procedures. An R code that implements the
algorithm and also computes estimates of the variance as in Section 5.1 below, is available at the homepage
of the first author.
4
A simulation study
We compare the performance of the Partial Information (PI) and Total Information (TI) estimators
against the Least Squares Estimate (LSE) across a wide range of settings.
12
4.1
Toy example
We start by studying the following toy model
Y “ αX 2 ` βX ` ε “ α ` βX ` looooooomooooooon
αpX 2 ´ 1q ` ε,
(21)
“δ
where X and ε are i.i.d N(0,1). Under model (21), the linear and non-linear part are determined by β and
α separately and the linear coefficient β does not affect the residual δ. In this case, it easy to calculate
the asymptotic variance of the estimates explicitly
2
2
σLSE
“ 10α2 ` 1, σT2 I “ 6α2 ` 1, σdiff
“ 4α2 .
Considering the excess risk, then by Proposition 1 the ratio of excess risks of β̂P I (β̂T I , respectively) and
β̂LSE converges to
2
2α2 `1`σP
I
2
2
2α `1`σLSE
2α2 `1`σ2
( 2α2 `1`σ2T I , respectively). The limits equal 1 in the linear case (α “ 0)
LSE
and approach 0 when the non-linear part is dominant (α Ñ 8).
Figure 1 summarizes 10,000 simulations of model (21) with β “ 1 and different values of α, n; for the
PI case we used m “ 2n. For all the scenarios we found that β̂T I , β̂P I have smaller excess risk when n is
large enough. As α increases, the departure from linearity is more significant and the ratio of the excess
risks is smaller. For small n, LSE is superior for all scenarios as it is a simpler estimate. When the model
is close to linear (α “ 1{4), the new estimates are better for n ě 200 and for the other values of α the
new estimates are better for n ě 70. For small values of α the limiting excess risk is close to the actual
risk even when n « 300 but this does not hold true for larger values of α. In short, we found that the
new estimates are better for large n, and are much better when the non-linear part is significant.
4.2
Simulation Parameters
We now compare our new estimates to LSE for a broader variety of scenarios. Estimates of the Excess
Risk Ratio (ERR) for PI and TI were obtained for every combination of the following parameter choices:
n (size of labeled dataset)
12, 25, 50, 100, 250, 500
m (size of unlabeled dataset)
n, 2n
p (number of predictors)
Distribution of X
1, 4
Gaussian, Lognormal, Exponential, Cubed Gaussian
Gaussian, N p0, e2}X} q
?
X, eX , X 3 , X
Errors
ErY |Xs
Fixing each of the 768 parameters settings, many sample datasets were generated, along with a large
test dataset (n = 100,000). For each sample dataset, PI, TI, and LSE were fit, (calculating TI using the
13
0.9
1.0
1.1
alpha=1/8
alpha=1/4
alpha=1/2
alpha=1
0.7
0.8
Excess risk: MSE(PI)/MSE(LSE)
1.0
0.9
0.8
0.7
Excess risk: MSE(TI)/MSE(LSE)
1.1
alpha=1/8
alpha=1/4
alpha=1/2
alpha=1
0
100
200
300
400
500
0
100
n
200
300
400
500
n
(a) TI
(b) PI
Figure 1: Estimates of the excess risk ratio for the TI and PI estimates based on 10,000 simulations of model (21) with
β “ 1 and different values of α, n; for the PI case we used m “ 2n. A confidence interval based on two standard deviations
is also plotted. The horizontal lines represents the limiting excess risk.
PI method with m ě 500n.) Next, an estimate of expected MSE was obtained using the test dataset.
Along with the best linear fit for the test dataset (given by least squares regression), these determine the
Excess Risk Ratio for PI (ERRP I ) as follows:
ERRP I “
M SEP I ´ M SEBLF
,
M SELSE ´ M SEBLF
where
M SEP I is the expected MSE of the PI estimator (estimated from mean performance on the test dataset),
M SELSE is the expected MSE of the LSE estimator, and
M SEBLF is the MSE of the best linear fit of the test dataset.
An analagous calculation was performed to calculate ERRT I . Additional datasets were sampled to
improve estimates of M SEP I and M SELSE , until standard errors for ERRP I and ERRT I fell below 1%
(as estimated by the delta method), or a maximum of 100, 000 sample datasets was generated.
4.3
Simulation Results
Table 3 provides the proportions of scenarios where the PI estimate yields a statistically significant
smaller excess risk than the LSE and the proportion where the opposite holds true, for different p and n.
14
Statistical significance is measured by two standard errors. The results demonstrate that the PI estimate
outperforms the LSE across a wide range of scenarios for large n, and the proportion increases with n.
When n “ 500, LSE is significantly better only for about 10% of the scenarios we studied. More detailed
comments follow:
Table 3: Proportions of scenarios where ERRP I is significantly smaller than 1 and where it is significantly larger than 1.
❅
n
❅
p ❅❅
12
25
50
100
250
500
1
0.297
0.344
0.469
0.594
0.656
0.609
4
0.047
0.234
0.391
0.438
0.594
0.547
1
0.344
0.234
0.172
0.125
0.125
0.109
4
0.578
0.391
0.234
0.188
0.188
0.125
ERRP I ă 1 significantly
ERRP I ą 1 significantly
1. A proposed rule of thumb when p “ 1: As indicated by Table 3 across a wide range of parameter
settings, PI performs almost as well as LSE when n ě 100, m ě 100 and linear ErY |Xs, and better
than the LSE for nonlinear ErY |Xs (that is, ERR ă 1). Furthermore, PI’s underperformance is
mild when ErY |Xs is linear. Thus, assuming that the scenarios we studied are representative sample
of reality, we recommend using the PI method when the following conditions are satisfied, as a rule
of thumb:
(a) n ě 100
(b) m ě 100
(c) ErY |Xs could be non-linear
2. If p “ 1, n ą 100, m ą 100, and ErY |Xs is one of the non-linear functions tested, then both PI and
TI appear to outperform LSE; furthermore, the margin of outperformance widens as n increases (See
Figure 2 (a), and Figure 2 (b), respectively). In particular, PI does much better for ErY |Xs “ X 3
?
and eX , as compared to X. Intuitively, this results from the larger non-linear moments of X 3 and
eX , causing Σdiff to be large (as defined in Theorem 2, part ii)
3. Increasing m, the size of the unlabeled dataset, improves performance of PI. Indeed, comparing PI
where m “ 2n in Graph X against TI (m ě 500n) in Graph Y, one observes that greatly increasing
the pool of unlabeled samples can reduce Excess Risk, in some cases on the order of 5 ´ 10 per
cent (Compare Figures 2 (a) and (b)). But, a word of caution: if n is small, then LSE may still
outperform both PI and TI, even when m is large and ErY |Xs is non-linear.
15
0.9
1.0
1.1
|x|
ex
x3
0.6
0.6
0.7
0.8
0.8
0.9
1.0
Excess risk: MSE(TI)/MSE(LSE)
1.1
|x|
ex
x3
0.7
Excess risk: MSE(PI)/MSE(LSE)
x
1.2
1.2
x
0
100
200
300
400
500
0
n
100
200
300
400
500
n
(a) PI
(b) TI
Figure 2: Estimates of the excess risk ratio for the TI and PI estimates, for several possible functions of ErY |Xs. In PI
estimates (a), m “ 2n. In TI estimates (b), the PI method is used with m ě 500n. In all cases, X is Gaussian, errors are
Gaussian, and p “ 1. Confidence intervals are ˘2 standard errors.
4. If n is small, or if ErY |Xs is linear, then LSE performs better than TI and PI. However, ERR Ñ 1
as n Ñ 8, holding other parameters fixed (See Figures 2 (a) and (b)). This result is consistent with
the asymptotic agreement of PI, TI, and LSE when ErY |Xs is linear (See Theorem 2)
5. Extra heteroskedastic noise appears to decrease the relative advantage of PI. See Figure 3 (a).
6. Somewhat similarly, increasing the number of parameters to 4 (where Y is the sum of its relationships
with each Xi ) also decreases the relative advantage of PI. See Figure 3 (b).
7. If X is not Gaussian, PI still does well, especially for large n. See Figure 3 (c) for an example
where X is exponential. (Note that standard errors may be understated, especially in the case
ErY |Xs “ exppXq, partly due to the large moments of Y .)
16
0.9
1.0
1.1
|x|
ex
x3
0.6
0.6
0.7
0.8
0.8
0.9
1.0
Excess risk: MSE(PI)/MSE(LSE)
1.1
|x|
ex
x3
0.7
Excess risk: MSE(PI)/MSE(LSE)
x
1.2
1.2
x
0
100
200
300
400
500
0
n
100
200
300
400
500
n
(b) p “ 4
(a) Heteroskedastic Errors
0.8
0.6
0.4
0.0
0.2
Excess risk: MSE(PI)/MSE(LSE)
1.0
1.2
x
|x|
ex
3
x
0
100
200
300
400
500
n
(c) X Exponential
Figure 3: Estimates of the excess risk ratio of PI estimates, across several possible functions of ErY |Xs, with tweaks to
the previous settings of Figure 2 (a): these were m “ 2n, X Gaussian, p “ 1, and Guassian errors. The following is changed
in each plot: Figure (a): Errors are heteroscedastic, distributed N p0, e2}x} q. Figure (b): p “ 4, where ErY |Xs is the sum of
the given relationship in each X-variable. Figure (c): X is distributed exponentially. Confidence intervals are ˘2 standard
errors.
17
5
Estimating the standard errors of the estimates
In this section we study estimation of the asymptotic covariance matrix. We suggest three estimates for
β P I in Section 5.1 and study their performance by simulation (Section 5.2). Section 5.3 deals with the
estimation of the prediction error (excess risk).
5.1
Three estimates of ΣP I
We proved that under certain conditions,
?
D
npβ̂ LSE ´ βq ÝÑ N p0, ΣLSE q,
?
D
npβ̂ P I ´ βq ÝÑ N p0, ΣP I q and,
?
D
npβ̂ T I ´ βq ÝÑ N p0, ΣT I q.
In these expressions the (asymptotic) variability of the estimates are determined by the matrices ΣLSE ,
ΣP I and ΣT I . These matrices are unknown and therefore need to be estimated. In this section we
consider only estimation of ΣLSE and ΣP I . Estimation of the asymptotic variance yields prediction error
estimates or L2 risk in a standard fashion. In section 5.3 we discuss estimation of the excess risk.
For ΣLSE we follow Buja et al. (2016) who show that the standard estimates are inconsistent when
non-linearity is present and suggest two alternative estimates for the variance. The parametric sandwich
estimator for the asymptotic variance of β̂ LSE is
`
˘
`
˘
ˆ parm pβ̂ LSE q “ XT X ´1 XT D̂X XT X ´1 ,
AV
where X is the design matrix and D̂ is a diagonal matrix with D̂i “ δ̂i2 the standard residual estimator.
The second estimate is derived from a pairs bootstrap where a pair pX˚ , Y ˚ q is sampled NBS times from
the empirical joint distribution of pX, Y q, yielding a sample (under the empirical measure) of size NBS of
ˆ BS pβ̂ LSE q.
β̂ LSE . The resulting estimate of the variance is denoted by AV
For estimating the variance of ΣP I , we consider three estimates described as follows:
1. Parametric: The asymptotic variances of β̂ LSE and β̂ P I are
AVpβ̂ LSE q “ Covpδ̃q ` CovpVq, AVpβ̂ P I q “ Covpδ̃q ` νCovpVq.
(22)
To estimate the variance of Covpδ̃q notice that for j “ 1, . . . , p and i “ 1, . . . , n,
p
ÿ
?
piq
piq
piq
piq
δ̃ˇj “ W̌ piq ´ tβ̂ P I uj ´ âǓ1 ´
b̂j 1 Ǔj 1 `1 “ δ̃j ` op p1{ nq,
j 1 “1
p
ˇ
and hence, Covpδ̃q ÝÑ Covpδ̃q, where Cov is the empirical covariance based on the labeled n
sample. Therefore, using (22) we obtain that a consistent estimator to AVpβ̂ P I q is
ˆ!
˙
)
!
)
ˇ ˇ 1
ˇ ˇ 1
ˆ
ˆ
AVparm ptβ̂ LSE uq
´ Covpδ̃ j , δ̃ j q ,
AVparm pβ̂ P I q
“ Covpδ̃ j , δ̃ j q ` ν
1
1
j,j
j,j
18
for j ‰ j 1 and for j “ 1, . . . , p,
!
)
ˆ parm pβ̂ P I q
AV
j,j
˜
¸
n
n !
)2
!
)
ÿ
1 ÿ ! ˇpiq )2
1
piq
ˆ parm ptβ̂ LSE uq
“
` ν max
,0 .
δ̃
δ̃ˇ
AV
´
n i“1 j
n i“1 j
j,j
!
)
piq
piq n
2. Bootstrap: Let pX˚ , Y˚ q
the labeled observations) and
the unlabeled observations).
be a sample from the empirical distribution of pX, Y q (based on
i“1
piq n`m
be a sample from the empirical distribution of X (based on
tX˚ ui“n`1
¯
´!
)
˚
˚
piq n`m
piq
piq n
is a sample under
, tX˚ ui“n`1
Thus, β̂ P I “ β̂ P I pX˚ , Y˚ q
i“1
the empirical measure of β̂ P I , which can yield an estimate of the variance of the estimator, denoted
ˆ BS pβ̂ P I q.
by AV
3. Variance bootstrap: In the proof of Theorem 2 we showed that for j “ 1, . . . , p,
piq piq
1 řn
n
n`m
ÿ piq
?
?
1
1 ÿ piq
i“1 Xj‚ δ
n
tβ̂ LSE uj “ βj `
nq
and
t
δ̃
`
Vj `op p1{ nq.
`o
p1{
β̂
u
“
β
`
p
j
j
P
I
2
n i“1
n ` m i“1
EpXj‚ q
Hence, the definition of the vector V implies that
β̂ LSE ´ β̂ P I “
n
n`m
ÿ
?
1
1 ÿ
´Vpiq `
Vpiq ` op p1{ nq
n i“1
n ` m i“1
˙
ˆ
n
n`m
ÿ
?
1
n
1 ÿ piq
´1 `
V
Vpiq ` op p1{ nq.
“
n i“1
n`m
n ` m i“n`1
Therefore, dropping the smaller order terms we obtain
nCovpβ̂ LSE
ˆ
˙2
nm
CovpVq
`
pn ` mq2
ˆ
˙
n
“ CovpVq 1 ´
Ñ CovpVq p1 ´ νq “ ΣLSE ´ ΣP I .
n`m
n
´ β̂ P I q « CovpVq 1 ´
n`m
Thus, a a consistent estimator of the difference ΣLSE ´ ΣP I “: ∆ can be obtained by bootstrapping
˚
˚
p A consistent estimate to AV pβ̂ P I q is
nCovpβ̂ P I ´ β̂ LSE q. We denote this estimator by ∆.
p
ˆ V BS pβ̂ P I q “ AV
ˆ BS pβ̂ LSE q ´ ∆.
AV
This estimator has the advantage over the previous suggestions that it is always smaller than the
estimated variance of the LSE (in the sense that the difference is positive definite), in a similar
fashion to the asymptotic distribution.
In the next section we compare the different estimates through a simulation study.
5.2
Simulations for the estimates of the variance
We now investigate the estimates of the variance under the toy model (21). We study the performance
of the estimates for α “ 1, 1{2, 1{4, 1{8, where small α corresponds to little non-linearity. We considered
19
n “ 250 and m “ 1500 similar to the numbers of Section 6. For each α we repeated the simulation 1000
times and also NBS “ 1000.
The simulations are summarized in Table 4 and in Figure 4. We find that the bootstrap estimate
(henceforth BS) is more variable than the parametric (henceforth PARM) and the variance bootstrap
(henceforth VBS). PARM and VBS are comparable but for small α’s for estimating the difference, the
histogram of VBS is narrower around the true value. When α “ 1{8, the estimated difference under BS
(PARM) is negative 32.9% (34.7%) of the simulations, while for VBS it is never negative. However, since
VBS cannot be negative, the mean of the simulations is upwards biased. In short, we find that PARM
ˆ pβ̂LSE q.
and VBS outperform BS and VBS has the advantage of never being smaller than AV
Table 4: The mean (std) of the estimates of the asymptotic variance AV pβ̂P I q and the difference AV pβ̂LSE q ´ AV pβ̂P I q.
Variance
Difference
α
Bootstrap
Parametric
Variance BS
true
Bootstrap
Parametric
Variance BS
true
1
7.65 (3.71)
6.86 (3.11)
6.58 (2.40)
7
2.63 (2.53)
3.51 (1.53)
3.70 (2.13)
3.43
1/2
2.65 (0.97)
2.42 (0.82)
2.36 (0.70)
2.5
0.64 (0.53)
0.88 (0.45)
0.93 (0.46)
0.86
1/4
1.43 (0.40)
1.33 (0.33)
1.31 (0.31)
1.37
0.15 (0.18)
0.24 (0.22)
0.27 (0.16)
0.21
1/8
1.11 (0.23)
1.06 (0.20)
1.03 (0.19)
1.09
0.02 (0.07)
0.06 (0.17)
0.10 (0.06)
0.05
20
Estimates of the difference
2.0
Estimates of the variance
BS
Param.
VBS
Density
2
1.0
0
0.0
1
0.5
Density
3
1.5
4
BS
Param.
VBS
1.0
1.5
2.0
2.5
3.0
−1.0
−0.5
variance
0.0
0.5
1.0
difference
(a) α “ 1{4
Estimates of the variance
Estimates of the difference
10
2.0
Density
6
1.5
0
0.0
2
0.5
4
1.0
Density
BS
Param.
VBS
8
2.5
BS
Param.
VBS
0.5
1.0
1.5
2.0
2.5
−0.2
variance
0.0
0.2
0.4
difference
(b) α “ 1{8
Figure 4: Histograms of the estimates (Bootstrap, Parametric, Variance BS) of the asymptotic variance AV pβ̂P I q and the
difference AV pβ̂LSE q ´ AV pβ̂P I q. The true asymptotic value is illustrated by the vertical gray line.
21
5.3
Estimating the prediction error
The three estimates of ΣP I can be used to estimate the prediction error in a standard fashion. Here,
we describe how the variance bootstrap estimate together with the result of Proposition 1 provide an
estimate of the excess risk.
By Proposition 1, the excess risk of β̂ LSE can be approximated by
p
p
ı2
”
ÿ
ÿ
βj tXj ´EpXj qu `TracepMΣLSE q “ V arpY q`β T Mβ´2
βj CovpY, Xj q`TracepMΣLSE q
E Y ´EpY q´
j“1
j“1
and hence can be consistently estimated by
T
σ̂Y2 ` β̂ LSE M̂β LSE ´ 2
where σ̂Y2 “
1
n
piq
řn
i“1 pY
piq ´ Ȳ q2 ,
M̂ “
p
ÿ
ˆ BS pΣLSE qu,
tβ̂ LSE uj CovpY, Xj q ` TracetM̂AV
j“1
1
n`m
řn`m
i“1
pXpiq ´ X̄q Xpiq ´ X̄
(T
and CovpY, Xj q “
1
n
řn
i“1 pY
piq ´
Ȳ qpXj ´ X̄j q is the empirical covariance. Therefore, a consistent estimate to the excess risk ratio is
T
z PI “
ERR
σ̂Y2 ` β̂ LSE M̂β LSE ´ 2
T
řp
j“1 tβ̂ LSE uj CovpY, Xj q `
σ̂Y2 ` β̂ LSE M̂β LSE ´ 2
řp
ˆ BS pΣLSE q ´ ∆qu
ˆ
TracetM̂pAV
j“1 tβ̂ LSE uj CovpY, Xj q
ˆ BS pΣLSE qu
` TracetM̂AV
,
(23)
ˆ is defined by the Variance bootstrap.
where ∆
6
Inferring homeless population
We now consider the Los Angeles homeless data set (Kriegler and Berk, 2010). Our purpose here is not
to carefully analyze this data set but rather to demonstrate our new method and to compare it to the
standard LSE. For our analysis, we used the same covariates as in Kriegler and Berk (2010); see this
reference for more details about the data.
There are about 2000 census tracts in the Los Angeles county, and a sample was conducted in order
to infer the homeless population. The sample consisted of two parts. First, tracts believed to have large
numbers of homeless people were pre-selected and visited; there are about 240 such tracts. Second, a
sample of about 260 tracts was chosen from the remainder by a random sampling technique and the
homeless population was counted, leaving about 1500 tracts to be imputed.
Leaving aside the preselected tracts this is almost exactly our semi-supervised setting among the
remaining population. We have n “ 261 and m “ 1536. The major difference from the setting in our
introduction here is that sampling from the remaining population is without replacement and hence is not
an i.i.d random sample. Since we have not discussed this type of sampling in detail, we will ignore this
issue in the subsequent discussion which is for illustrative purposes. (The difference in sampling schemes
does not affect the values of the various estimates, but does impact their standard errors.)
22
We compare the LSE of the random sample and the new estimate (PI) where the supervised part is
the random sample (of size 261) and the unsupervised part consists of the 1536 tracts to be imputed. The
resulting estimates, as well as their standard errors (SE) are given in Table 5. The SE of the LSE was
computed using pairs bootstrap (NBS “ 105 ), and the SE of PI was computed using the above variance
bootstrap method (the SE of the intercept of PI was estimated by pairs bootstrap).
We find that the SE’s of PI are smaller by 15%-40% than LSE in the predictors MedianInc, PercVacant,
PercCommercial, PercIndustrial. These predictors also demonstrated a discrepancy between the SE from
linear models to bootstrap SE as reported in Buja et al. (2016) Section 2, indicating non-linearity in these
predictors (a discrepancy was observed also in the predictor PercResidential, but the improvement in this
predictor is relatively small). In summary, we found that the additional information on the distribution
of the predictors in the unsupervised data provides estimates that are more accurate (smaller variance).
Consider the prediction problem of estimating the homeless population for the 1500 tracts. An estimate
z P I “ 0.904, i.e., improvement of about 10%. An estimate of
of the excess risk ratio, using (23), is ERR
ˆ M̂q{n “ 15.2, while an estimate of the
the differences of prediction errors as in (10) and (11) is Tracep∆
prediction error of the mean is
m
1 ÿ~T
~ n`i {n “ 42.8.
X Σ̂LSE X
m i“1 n`i
Since 15.2/42.8 =0.355 the improvement in prediction risk is about 36%. When considering the prediction
of Y itself as in (11) then the improvement is much smaller, about 1.5%. The reason for this difference is
that the predictors we considered are all quite weak, and hence most of prediction error comes from the
variability in the distribution of Y given X rather than the variability in estimating the β’s.
Table 5: The estimates β̂ LSE , β̂ P I and their standard errors.
β̂LSE
ˆ LSE
SE
β̂P I
ˆ PI
SE
SE LSE
SE P I
Intercept
13.138
11.822
14.758
11.184
0.927
MedianInc ($K)
-0.065
0.056
-0.080
0.044
1.284
PercVacant
1.449
0.707
1.583
0.514
1.374
PercMinority
0.060
0.105
0.063
0.099
1.058
PercResidential
-0.070
0.100
-0.078
0.094
1.072
PercCommercial
0.446
0.354
0.335
0.300
1.179
PercIndustrial
0.101
0.188
0.202
0.143
1.317
23
7
Discussion
In this work we showed that β can be better estimated when additional information on the predictors X
is available. The key idea is to replace the regression model
Y “ α ` β1 X1 ` ¨ ¨ ¨ ` βp Xp ` δ,
with p regressions of the form
Wj “ βj ` aU1 `
p
ÿ
j 1 “1
bj 1 Uj 1 `1 ` δ̃j ,
(24)
(25)
for j “ 1, . . . , p, where Wj , U1 , . . . , Up`1 is defined in (18). Regression (25) can be used only when the
first and second moments of X are known. We showed that the intercept estimator of (25) is a better
estimate of βj , in terms of smaller asymptotic variance, than the standrd LSE of (24). Furthermore, we
showed that even if the second moments of X are not known exactly then improvement can be made
over the standard LSE provided that the unlabeled sample is non-negligible with respect to the labeled
sample, i.e., m{n is bounded away from 0.
Here we considered the classical framework where p is fixed and n Ñ 8. Zhang et al. (2016) study the
semi-supervised mean estimation problem where p, n both go to infinity but p grows relatively slowly, i.e.,
?
p “ Op nq. The findings of Zhang et al. (2016) indicate that our results can be extended to the latter
case. However, our method obviously requires that p ăă n and for high-dimensional regression models
where p ą n, a different approach is needed. High-dimensional missspecified regression models are the
topic of a recent manuscript by Bühlmann and van de Geer (2015).
In this work we did not study variable selection and considered the covariates as given. However,
improved asymptotic performance may be achieved by including functions of the existing covariates as
additional new covariates. Section 3 of Zhang et al. (2016) discusses how this can be done for the problem
of semi-supervised estimation of a mean. On the other hand, when the labeled sample, n, is small, then
it may be better not to include all covariates. We hope to study the problem of variable selection in this
context in a future research.
A possible extension of these results is to generalized linear models, where the label variable (Y ) takes
a small number of values and the regression model reduces to a classification problem. Our hope is that
our method can be extended to these cases and improvement can be made over naive classifiers that
consider only the labeled data.
In summary, in this work we considered the framework where the best linear predictor is of interest
even if EpY |Xq is not linear. Under this framework, additional information on the distribution of X is
useful to construct better estimates than the standard estimates. We believe that this framework is of
practical importance and also can lead to interesting statistical issues, which are yet to be studied.
24
8
Proofs
Lemma 1. We have that
p
ı2
”
!
)
ÿ
β̃j tX̄j ´ EpXj qu ´ EpY q ` E pβ ´ β̃qT Mpβ ´ β̃q .
ELpα̃, β̃q “ E Ȳ ´
(26)
j“1
Proof.
First,
¯2
´
`
˘2
T
Rpα̃, β̃q “ E Y ˚ ´ α̃ ´ β̃ X˚ ´ E Y ˚ ´ α ´ β T X˚
)
)!
!
“ E pα ´ α̃q ` pβ ´ β̃qT X˚ 2Y ˚ ´ pα ` α̂q ´ pβ ` β̃qT X˚
)2
)
!
!
(
“ E pα ´ α̃q ` pβ ´ β̃qT X˚ ` 2E pα ´ α̂q ` pβ ´ β̃qT X˚ Y ˚ ´ α ´ β T X˚
)2
!
“ E pα ´ α̃q ` pβ ´ β̃qT X˚ ,
where the last equality holds true since Y ˚ ´ α ´ β T X˚ is orthogonal to X ˚ , 1. We can further write
)2
!
”
ı2
E pα ´ α̃q ` pβ ´ β̃qT X˚ “ E pα ´ α̃q ` pβ ´ β̃qT EpXq ` pβ ´ β̃qT tX˚ ´ EpXqu
!
)2
“ E pα ´ α̃q ` pβ ´ β̃qT EpXq ` Etpβ ´ β̃qT Mpβ ´ β̃qu, (27)
For the first term in (27) notice that α “ EpY q ´
řp
j“1 βj EpXj q
and α̃ “ Ȳ ´
řp
j“1 β̃j X̄j ;
therefore,
p
!
)2
”
ı2
ÿ
T
E pα ´ α̃q ` pβ ´ β̃q EpXq “ E Ȳ ´
β̃j tX̄j ´ EpXj qu ´ EpY q ,
j“1
and (26) is established.
Proof of Proposition 1.
Since for any numbers c, d, B we have that minpc ` d, Bq ď minpc, Bq ` minpd, Bq and minpc ` d, Bq ě
minpc, B{2q ` minpd, B{2q, then
˜
p
ı2
”
ÿ
β̃j tX̄j ´ EpXj qu ´ EpY q , B
Rpα̃, β̃q “ lim lim E mintnL̃pα̃, β̃q, Bu “ lim lim E min n Ȳ ´
BÑ8 nÑ8
BÑ8 nÑ8
j“1
¸
´ !
) ¯
` lim lim E min n pβ ´ β̃qT Mpβ ´ β̃q , B . (28)
BÑ8 nÑ8
We start with the first summand in (28). Since
Ȳ ´
p
ÿ
j“1
?
npβ̃ ´ βq is asymptotically normal then
p
p
”
ı ÿ
ÿ
?
pβj ´ β̃j qtX̄j ´EpXj qu “ op p1{ nq.
β̃j tX̄j ´EpXj qu´EpY q´ Ȳ ´
βj tX̄j ´EpXj qu´EpY q “
j“1
j“1
25
Therefore, by the continuous mapping theorem and since the random variables are bounded, then for any
B,
˜
p
ı2
”
ÿ
β̃j tX̄j ´ EpXj qu ´ EpY q , B
lim E min n Ȳ ´
nÑ8
j“1
˜
¸
p
ı2
”
ÿ
βj tX̄j ´ EpXj qu ´ EpY q , B
“ lim E min n Ȳ ´
nÑ8
j“1
¸
`
˘
“ E min Z12 , B , (29)
”
ı2
ř
where Z1 is normal with mean zero and variance Y ´ pj“1 βj tXj ´ EpXj qu ´ EpY q .
?
D
For the second summand in (28), the asymptotic normality npβ̃ ´ βq ÝÑ N p0, Σq implies that for
any B,
´ !
) ¯
(
lim E min n pβ ´ β̃qT Mpβ ´ β̃q , B “ E min ZT2 Z2 , B ,
nÑ8
(30)
where Z2 is a normal vector with mean zero and variance matrix M1{2 ΣM1{2 . Taking limits as B Ñ 8
in (29) and (30) completes the proof of the proposition since TracepM1{2 ΣM1{2 q “ TracepMΣq.
We will not prove Theorem 1 since it is a special case of Theorem 2.
Proof of Theorem 2.
Part (i) We have that (Buja et al., 2016) tβ̂ LSE uj “ βj `
xδ, Xj¯‚ y
tβ̂ LSE uj “ βj `
“ βj `
||Xj¯‚ ||2
řn
piq piq
i“1 Xj¯
‚δ
)
!
ř
piq 2
n
1
‚
i“1 Xj¯
n
1
n
xδ,Xj¯‚ y
.
||Xj¯‚ ||2
“ βj `
Furthermore,
1
n
řn
piq piq
i“1 Xj‚ δ
2 q
EpXj‚
?
` op p1{ nq,
since
? ÿ
? ÿ
n
n
¯
)T ´
!
n
n
1
p
p
piq
piq
2
~ piq
β̂
´
β
q.
δpiq pXj¯‚ ´ Xj‚ q “
δpiq X
||Xj¯‚ ||2 ÝÑ EpXj‚
´j¯
‚
´j‚ ÝÑ 0, and
´j
n i“1
n i“1
n
The results now follows from the multivariate CLT and Slutsky’s theorem.
Part (ii) For each j,
tβ̂ T I uj “ W̄j ´ âŪ1 ´
p
ÿ
j 1 “1
b̂j 1 Ūj 1 `1 “ W̄j ´ aŪ1 ´
p
ÿ
j 1 “1
?
bj 1 Ūj 1 `1 ` op p1{ nq
“ βj `
Again, the multivariate CLT and Slutsky’s theorem imply the result.
n
?
1 ÿ piq
δ̃j ` op p1{ nq. (31)
n i“1
To prove the furthermore part, notice that for each j,
¯
´
řp
p
1
1
X
Y
´
α
´
β
ÿ
j Xj‚
j 1 “1 j
Xj 1 Xj‚
δXj‚
Y Xj‚
Xj‚
Vj “
´
δ̃
“
´
`
β
´
b
`
a
`
bj 1
j
j
j
2
2
2
2
2 q
EpXj‚ q
EpXj‚ q
EpXj‚ q
EpXj‚ q j 1 “1 EpXj‚
“ βj ´ bj ` pa ´ αq
p
p
ÿ
ÿ
Xj 1 Xj‚
Xj‚
1
1
q
´
β
pb
pbj 1 ´ βj 1 qUj 1 `1 (32)
`
“
pa
´
αqU
`
1
j
j
2 q
2 q
EpXj‚
EpX
1
1
j‚
j “1
j “1
26
is a linear function of U1 , . . . , Up`1 and hence is orthogonal to δ̃j , i.e., EVj δ̃j “ 0 for all j. Therefore,
!
)
ΣLSE “ CovpX‚ δq “ Cov δ̃ ` pX‚ δ ´ δ̃q “ Covpδ̃ ` Vq “ Covpδ̃q ` CovpVq “ ΣT I ` Σdiff .
Part (iii) We show that
tβ̂ P I uj ´ tβ̂ T I uj “
Hence,
?
D
n`m
ÿ piq
?
1
Vj ` op p1{ nq.
n ` m i“1
(33)
npβ̂ P I ´ β̂ T I q ÝÑ N p0, νΣdiff q. Since V is orthogonal to δ̃, then
?
npβ̂ P I ´ βq “
?
npβ̂ T I ´ βq `
?
D
npβ̂ P I ´ β̂ T I q ÝÑ N p0, ΣT I ` νΣdiff q.
We now prove (33) by a somewhat lengthy calculation. We have that
¯ ´ âǓ
¯ ´
tβ̂ P I uj “ W̌
j
1
p
ÿ
j 1 “1
p
ÿ
¯
¯
¯
¯ 1 ` o p1{?nq.
b̂j 1 Ǔj 1 `1 “ Ŵj ´ aǓ1 ´
bj 1 Ǔ
p
j `1
(34)
j 1 “1
Therefore, (34) and (31) yield
p
¯ 1 ´ Ū 1 q ` o p1{?nq.
¯ ´ W̄ ´ apǓ
¯ ´ Ū q ´ ÿ b 1 pǓ
tβ̂ P I uj ´ tβ̂ T I uj “ W̌
p
j
j
1
1
j
j `1
j `1
j 1 “1
We have that
p
ÿ
¯ 1 ´ Ū 1 q ` o p1{?nq,
¯
¯
βj 1 pǓ
W̌j ´ W̄j “ αpǓ1 ´ Ū1 q `
p
j `1
j `1
j 1 “1
and hence,
tβ̂ P I uj ´ tβ̂ T I uj “ pα ´ aqpǓ¯1 ´ Ū1 q `
p
ÿ
¯ 1 ´ Ū 1 q ` o p1{?nq.
pβj 1 ´ bj 1 qpǓ
p
j `1
j `1
(35)
j 1 “1
We now consider the summands in (35). We start with the first summand
˜
¸
ř
~ piq
pβ ´j‚ ´ β ´jˇ‚ qT n1 ni“1 X
1
1
´j
¯
`
´
´
Ǔ1 ´ Ū “
2
2
2 q
2 q
2 q
EpX
q
EpX
q
ĚpXjˇ
ĚpX
ĚpX
j‚
j‚
‚
jˇ
‚
jˇ
‚
piq
1 řn
T
~
pβ ´j‚ ´ β ´jˇ‚ q n i“1 X
?
´j
` op p1{ nq
“
2
EpXj‚ q
´ ř
¯
ř
piq
n
1
1 řn`m ~ piq
T
~
n`m ~ piq
1
T
pβ ´j‚ ´ β ´jˇ‚ q n i“1 X´j ´ n`m i“1 X´j
pβ ´j‚ ´ β ´jˇ‚ q n`m i“1 X´j
?
`
` op p1{ nq
“
2
2
EpXj‚ q
EpXj‚ q
1 řn`m ~ piq
pβ ´j‚ ´ β ´jˇ‚ qT n`m
?
i“1 X´j
“
` op p1{ nq.
2
EpXj‚ q
1
n
řn
piq
i“1 Xjˇ
‚
1
n
řn
piq
i“1 Xj‚
n
1 ÿ piq
X
“
n i“1 j‚
Since
~ ´j q ùñ ĚpXj q “ Ěpβ T X
~
0 “ ĚpXjˇ‚ q “ ĚpXj ´ β T´jˇ‚ X
´jˇ
‚ ´j q
then,
T ~
T
~ ´j q “ Ěpβ T X
~
ĚpXj‚ q “ ĚpXj ´ β T´jˇ‚ X
‚ ´ β ´j‚ q
´jˇ
‚ ´j q ´ Ěpβ ´j‚ X´j q “ pβ ´jˇ
27
n`m
ÿ piq
1
~ ,
X
n ` m i“1 ´j
and therefore,
piq
1 řn`m
´ n`m
?
i“1 Xj‚
¯
Ǔ1 ´ Ū1 “
` op p1{ nq.
2
EpXj‚ q
(36)
We now consider the j 1 -th term in (35) for j 1 ‰ j; then, EpXj 1 Xj‚ q “ 0. We have
¯ 1 ´ Ū 1 “
Ǔ
j `1
j `1
1
n
řn
piq piq
i“1 Xj 1 Xjˇ
‚
2
ĚpXjˇ‚ q
n
1 ÿ piq piq
X 1 Xj‚
“
n i“1 j
#
´
1
n
řn
piq piq
i“1 Xj 1 Xj‚
EpXj‚ q2
1
1
´
2
2
EpX
ĚpXjˇ‚ q
j‚ q
+
“
`
pβ ´j‚ ´ β ´jˇ‚ qT n1
řn
piq ~ piq
i“1 Xj 1 X´j
2 q
ĚpXjˇ
‚
ř
n`m
1
pβ ´j‚ ´ β ´jˇ‚ qT n`m
EpXj‚
i“1
piq ~ piq
Xj 1 X
´j
q2
?
` op p1{ nq.
Since
~ ´j qXj 1 u ùñ ĚpXj Xj 1 q “ Ěpβ T X
~
0 “ ĚpXjˇ‚ Xj 1 q “ ĚtpXj ´ β T´jˇ‚ X
´jˇ
‚ ´j Xj 1 q
then,
T
T ~
~
~ ´j qXj 1 u “ Ěpβ T X
ĚpXj‚ Xj 1 q “ ĚtpXj ´βT´jˇ‚ X
‚ ´ β ´j‚ q
´jˇ
‚ ´j Xj 1 q´Ěpβ ´j‚ X´j Xj 1 q “ pβ ´jˇ
and therefore,
1
¯ 1 ´ Ū 1 “ ´ n`m
Ǔ
j `1
j `1
řn`m
i“1
piq
piq
Xj‚ Xj 1
EpXj‚ q2
n`m
ÿ piq piq
1
~ ,
X1 X
´j
n ` m i“1 j
?
` op p1{ nq.
(37)
For the j-th term in (35), we have that EpXj Xj‚ q “ EpXj‚ q2 ; hence,
¯
Ǔ
j`1 ´ Ūj`1 “
“
#
1
n
řn
piq piq
i“1 Xj Xjˇ
‚
2
ĚpXjˇ‚ q
n
1 ÿ piq piq
X 1 Xj‚ ´ EpXj‚ q2
n i“1 j
+#
´
1
n
řn
piq piq
i“1 Xj Xj‚
EpXj‚ q2
+
ř
piq ~ piq
2
2 q
pβ ´j‚ ´ β ´jˇ‚ qT n1 ni“1 Xj X
1
1
‚
´j ` EpXj‚ q ´ ĚpXjˇ
´
`
2
2
2
EpX
q
ĚpXjˇ‚ q
ĚpXjˇ‚ q
j‚
ř
piq
piq
~ ` EpXj‚ q2 ´ ĚpX 2 q
pβ ´j‚ ´ β ´jˇ‚ qT n1 ni“1 Xj X
?
jˇ
‚
´j
` op p1{ nq.
“
2
EpXj‚ q
Since
2
T ~
2
T ~
2
ĚpXjˇ
‚ Xj q “ ĚtpXj ´ β ´jˇ
‚ q “ ĚpXjˇ
‚ X´j qXj u ùñ ĚpXj q “ Ěpβ ´jˇ
‚ X´j Xj q ` ĚpXjˇ
‚q
then,
2
T ~
~ ´j qXj u “ Ěpβ T´jˇ‚ X
~ ´j Xj q ` ĚpXjˇ
ĚpXj‚ Xj q “ ĚtpXj ´ β T´jˇ‚ X
‚ q ´ Ěpβ ´j‚ X´j Xj q
“ pβ ´jˇ‚ ´ β ´j‚ qT
28
n`m
ÿ piq piq
1
2
~ ` ĚpXjˇ
X X
‚ q.
´j
n ` m i“1 j
Thus, for the j-th term we obtain,
1
´ n`m
¯
Ǔ
´
Ū
“
j`1
j`1
řn`m
i“1
EpXj‚
piq
piq
Xj‚ Xj
q2
?
` 1 ` op p1{ nq.
(38)
Tracking back to (35) through (36),(37), and (38), we obtain
tβ̂ P I uj ´ tβ̂ T I uj “ pβj ´ bj q ` pa ´ αq
řn`m
piq
i“1 Xj‚
2 q
EpXj‚
1
n`m
`
p
ÿ
pbj 1 ´ βj 1 q
1
n`m
j 1 “1
řn`m
i“1
piq
piq
Xj‚ Xj 1
2 q
EpXj‚
?
` op p1{ nq.
Hence, (32) implies that
tβ̂ P I uj ´ tβ̂ T I uj “
n`m
ÿ piq
?
1
Vj ` op p1{ nq,
n ` m i“1
and (33) is established.
References
Brown, L.D. (1990). An Ancillarity Paradox Which Appears in Multiple Linear Regression. Annals of
Statistics, 18, 471–493.
Bühlmann, P., van de Geer, S. (2015). High-dimensional inference in misspecified linear models. Electronic
Journal of Statistics, 9, 1449–1473.
Buja, A., Berk, R., Brown, L., George, E., Pitkin, E., Traskin, M., Zhao, L., Zhang, K. (2016). Models
as Approximations: How Random Predictors and Model Violations Invalidate Classical Inference in
Regression. Submitted; arXiv:1404.1578.
Chakrabortty A., Cai T. (2016). Efficient and Adaptive Linear Regression in Semi-Supervised settings.
Submitted.
Cochran W.G. (1977). Sampling Techniques (3rd ed.), New York: John Wiley.
Johnson, R., Zhang, T. (2008). Graph-based semi-supervised learning and spectral kernel design. IEEE
Transactions on Information Theory. 54, 275–288.
Kriegler, B., Berk, R. (2010). Small area estimation of the homeless in Los Angeles: An application of
cost-sensitive stochastic gradient boosting. Annals of Applied Statistics, 4, 1234–1255.
Lafferty, J., Wasserman, L. (2008). Statistical analysis of semi-supervised regression. Advances in Neural
Information Processing Systems, 20, 801-808.
29
Liang, F., Mukherjee, S., West, M. (2007). The use of unlabeled data in predictive modeling. Statistical
Science, 22, 189-205.
Le Cam, L., Yang, G.L. (1990). Asymptotics in Statistics : Some Basic Concepts. New York: Springer.
Little, R.J.A. (1992). Regression With Missing X’s: A Review. Journal of the American Statistical Association, 87, 1227–1237.
Merz, C.J., Clair, D.C.St., Bond, W.E (1992). Semi-supervised adaptive resonance theory (smart2).
International Joint Conference on Neural Networks, 3, 851-856.
Wang, J., Shen, X., Pan, W. (2009). On efficient large margin semisupervised learning: Method and
theory. The Journal of Machine Learning Research, 10, 719-742.
White, H. (1980). Using Least Squares to Approximate Unknown Regression Functions. International
Economic Review, 21, 149–170.
Zhang, A., Brown L.D., Cai, T. (2016). Semi-supervised Inference: General Theory and estimation of
means. Submitted.
Zhou, Z.H., Li, M. (2005). Semi-supervised regression with co-training. International Joint Conference
on Artificial Intelligence, 5, 908–913.
Zhu, X., (2005). Semi-Supervised Learning Literature Survey. Computer Sciences, University of
Wisconsin-Madison, 1530.
30
| 10math.ST
|
1
Asymptotically optimal codebooks based on
generalized Jacobi sums
Ziling Heng
arXiv:1705.08055v1 [cs.IT] 23 May 2017
Abstract
Codebooks with small inner-product correlation are applied in many practical applications including direct
spread code division multiple access communications, space-time codes and compressed sensing. It is extremely
difficult to construct codebooks achieving the Welch or Levenshtein bound. In this paper, we firstly generalize
Jacobi sums over finite fields and secondly obtain asymptotically optimal codebooks with respect to the Welch or
Levenshtein bound. Our main results generalize those in [11] and the codebooks in this paper have more flexible
parameters than those in [11].
Index Terms
Code division multiple access, codebooks, signal sets, compressed sensing, Welch bound, Levenshtein bound.
I. I NTRODUCTION
ODEBOOKS (also called signal sets) with small inner-product correlation are usually used to
distinguish among the signals of different users in code division multiple access (CDMA) systems.
An (N, K) codebook C is a set {c0 , c1 , ..., cN −1 }, where each codeword cl , 0 ≤ l ≤ N − 1, is a unit norm
1 × K complex vector over an alphabet. The alphabet size is the number of elements in the alphabet. The
maximum cross-correlation amplitude of an (N, K) codebook C is defined by
C
Imax (C) =
max
0≤i<j≤N −1
| ci cH
j |,
where cH denotes the conjugate transpose of a complex vector c. Imax (C) is a performance measure of
a codebook C in practical applications. One important problem is to minimize the codebook’s maximal
cross-correlation amplitude. Minimizing Imax (C) among codewords of a codebook C can approximately
optimize various performance metrics such as outage probability, average signal-to-noise ratio and symbol
error probability for multiple-antenna transmit beamforming from limited-rate feedback [16], [21]. In the
context of unitary space-time modulations, minimizing Imax (C) is equivalent to minimizing the block error
probability [14]. Codebooks are also called frames. A codebook C with minimal Imax (C) is referred to
as a Grassmannian frame. Besides, minimizing Imax (C) of finite frames brings to minimal reconstruction
error in multiple description coding over erasure channels [24].
For a given K, we would like to construct an (N, K) codebook with N being as large as possible
and Imax (C) being as small as possible simultaneously. However, the following Welch and Levenshtein
bounds demonstrate a trade-off among the parameters N, K and Imax (C) of a codebook C.
Lemma I.1 (Welch bound). [28] For any (N, K) codebook C with N ≥ K,
s
N −K
Imax (C) ≥
.
(N − 1)K
In addition, the equality in (I.1) is achieved if and only if
s
N −K
|ci cH
j | =
(N − 1)K
email: [email protected]
(I.1)
2
for all pairs (i, j) with i 6= j.
If a codebook C achieves the Welch bound in (I.1), which is denoted by IW , we call it a maximumWelch-bound-equality (MWBE) codebook [30]. An MWBE codebook is also called an equiangular tight
frame [26]. MWBE codebooks are employed in many applications including CDMA communications [20],
space-time codes [25] and compressed sensing [26]. To our knowledge, only the following constructions
of MWBE codebooks were reported in literature:
(1) In [23], [30], optimal (N, N) and (N, N − 1) codebooks with N > 1 were generated from the
(inverse) discrete Fourier transform matrix or ideal two-level autocorrelation sequences.
(2) In [2], [24], optimal (N, K) codebooks from conference matrices were given when N = 2K = 2d+1
with d being a positive integer and N = 2K = pd + 1 with p being a prime number and d being a
positive integer.
(3) In [3], [4], [30], optimal (N, K) codebooks were constructed with cyclic difference sets in the
Abelian group (ZN , +) or the additive group of finite fields or Abelian groups in general.
(4) In [10], optimal (N, K) codebooks from (2, k, v)-Steiner systems were presented.
(5) In [7], [8], [9], [22], graph theory and finite geometries were employed to study MWBE codebooks.
According to [24], the Welch bound on Imax (C) of a codebook C is not tight when N > K(K + 1)/2
for real codebooks and N > K 2 for all codebooks. The following Levenshtein bound turns out to be
tighter than the Welch bound when N > K 2 .
Lemma I.2 (Levenshtein bound). [18] For any real-valued codebook C with N > K(K + 1)/2,
s
3N − K 2 − 2K
Imax (C) ≥
.
(N − K)(K + 2)
For any complex-valued codebook C with N > K 2 ,
s
2N − K 2 − K
.
Imax (C) ≥
(N − K)(K + 1)
(I.2)
(I.3)
In general, it is very hard to construct codebooks achieving the Levenstein bound, which is denoted
by IL (the right-hand side of (I.2) or (I.3)). There are only a few constructions of codebooks achieving
the Levenshtein bound. These codebooks meeting the Levenstein bound were constructed from Kerdock
codes [1], [31], perfect nonlinear functions [6], bent functions over finite fields [34], and bent functions
over Galois rings [12]. Codebooks achieving the Levenshtein bound are employed in quantum physics
and the design of spreading sequences for CDMA and sets of mutually unbiased bases [6], [29].
Since it is very difficult to construct optimal codebooks, there have been a number of attempts to
construct codebooks asymptotically (or nearly) achieving the Welch bound or the Levenshtein bound, i.e.
Imax (C)
Imax (C)
= 1 or lim
= 1.
K→∞
K→∞
IW
IL
lim
In [23], Sarwate gave some nearly optimal codebooks from codes and signal sets. In [3], [5], [17], [33],
[35], the authors constructed some nearly optimal codebooks based on almost or relative difference sets. In
[32], Yu presented some nearly optimal codebooks from binary sequences. Recently, Hu and Wu proposed
new constructions of nearly optimal codebooks from difference sets and the product of Abelian groups
in [13].
The purpose of this paper is to generalize the main results in [11]. We firstly define a class of character
sums called generalized Jacobi sums which generalize the classical Jacobi sums. Based on generalized
Jacobi sums and related character sums, two classes of asymptotically optimal codebooks with very flexible
parameters are constructed. Our main results contain those in [11] as special cases.
3
II. M ATHEMATICAL F OUNDATIONS
In this section, we recall some necessary mathematical foundations on characters, Jacobi sums and
Gauss sums over finite fields. They will play important roles in our constructions of codebooks.
In this paper, we always assume that p is a prime number and q = pm with m being a positive integer.
Let Fq denote the finite field with q elements. Let α be a primitive element of Fq . Let Trq/p be the trace
function from Fq to Fp defined by
m−1
X j
Trq/p (x) =
xp .
j=0
A. Characters over finite fields
In this section, we recall both additive and multiplicative characters over finite fields.
Definition II.1. An additive character of Fq is a mapping χ from Fq to the set C∗ of nonzero complex
numbers such that χ(x + y) = χ(x)χ(y) for any (x, y) ∈ Fq × Fq .
Every additive character of Fq can be expressed as
Trq/p (ax)
χa (x) = ζp
, x ∈ Fq ,
where ζp is a primitive p-th root of complex unity. If a = 0, we call χ0 the trivial additive character of
Fq . If a = 1, we denote χ = χ1 which is called the canonical additive character of Fq . The orthogonal
relation of additive characters (see [15]) is given by
X
q, if a = 0,
χ(ax) =
0 otherwise.
x∈Fq
Definition II.2. A multiplicative character of Fq is a nonzero function ψ from F∗q to the set C∗ of nonzero
complex numbers such that ψ(xy) = ψ(x)ψ(y) for any x, y ∈ F∗q , where F∗q = Fq \{0}.
The multiplicative characters of Fq can be expressed as follows [15]. Let ζh = e
root of complex unity. For j = 0, 1, · · · , q − 2, the functions ψj defined by
2π
√
h
−1
denote the h-th
jk
ψj (αk ) = ζq−1
, for k = 0, 1, · · · , q − 2,
are all the multiplicative characters of Fq . If j = 0, we have ψ0 (x) = 1 for any x ∈ F∗q . We call ψ0 the
trivial multiplicative character of Fq .
For two multiplicative characters ψ, ψ ′ , we define their multiplication by setting ψψ ′ (x) = ψ(x)ψ ′ (x)
b∗ be the set of all multiplicative characters of Fq . Let ψ denote the conjugate character
for all x ∈ F∗q . Let F
q
of ψ by setting ψ(x) = ψ(x), where ψ(x) denotes the complex conjugate of ψ(x). It is easy to verify that
b∗ forms a group under the multiplication of characters. Furthermore, F
b ∗ is isomorphic
ψ −1 = ψ. Then F
q
q
∗
to Fq .
For a multiplicative character ψ of Fq , the orthogonal relation (see [15]) of it is given by
X
q − 1, if ψ = ψ0 ,
ψ(x) =
0
otherwise.
∗
x∈Fq
4
B. Jacobi sums
We now extend the definition of a multiplicative character ψ by setting
1, if ψ = ψ0 ,
ψ(0) =
0, if ψ 6= ψ0 .
(II.1)
As a result, the property that ψ(xy) = ψ(x)ψ(y) holds for all x, y ∈ Fq . With this definition, we have
then
X
q, if ψ = ψ0 ,
ψ(x) =
(II.2)
0 otherwise.
x∈Fq
Definition II.3. [15, p. 205] Let λ1 , . . . , λk be k multiplicative characters of Fq . The sum
X
J(λ1 , . . . , λk ) =
λ1 (c1 ) · · · λk (ck ),
c1 +···+ck =1
c1 ,··· ,ck ∈Fq
is called a Jacobi sum in Fq .
Jacobi sums are very useful in coding theory, sequence design and cryptography. For any a ∈ F∗q , more
generally, we define the sum
X
Ja (λ1 , . . . , λk ) =
λ1 (c1 ) · · · λk (ck ),
(II.3)
c1 +···+ck =a
where the summation extends over all k-tuples (c1 , . . . , ck ) of elements of Fq satisfying c1 + · · · + ck = a.
Hence, J1 (λ1 , . . . , λk ) = J(λ1 , . . . , λk ). It was shown in [15, p. 205] that
Ja (λ1 , . . . , λk ) = (λ1 · · · λk )(a)J(λ1 , . . . , λk ).
(II.4)
Therefore, |Ja (λ1 , . . . , λk )| = |J(λ1 , . . . , λk )|. In [15], the values of |J(λ1 , . . . , λk )| were determined for
several cases.
C. Gauss sums
Let ψ be a multiplicative character and χ an additive character of Fq . The Gauss sum G(ψ, χ) is defined
by
X
ψ(x)χ(x).
G(ψ, χ) =
x∈F∗q
The explicit value of G(ψ, χ) is very difficult to determine in general. However, its absolute value is
known as follows.
Lemma II.4. [15, Th. 5.11] Let ψ be a multiplicative character and χ an additive character of Fq . Then
G(ψ, χ) satisfies
q − 1, if ψ = ψ0 , χ = χ0 ,
−1,
if ψ = ψ0 , χ 6= χ0 ,
G(ψ, χ) =
0,
if ψ 6= ψ0 , χ = χ0 .
If ψ 6= ψ0 , χ 6= χ0 , then
|G(ψ, χ)| =
√
q.
If we consider the extended definition of ψ in Equation (II.1), then extended Gauss sum can be defined
as
b
G(ψ,
χ) =
X
x∈Fq
ψ(x)χ(x).
5
Lemma II.4 yields the following corollary.
b
Corollary II.5. Let ψ be a multiplicative character and χ an additive character of Fq . Then G(ψ,
χ)
satisfies
q, if ψ = ψ0 , χ = χ0 ,
b
0, if ψ = ψ0 , χ 6= χ0 ,
G(ψ,
χ) =
0, if ψ 6= ψ , χ = χ .
0
0
If ψ 6= ψ0 , χ 6= χ0 , then
b
|G(ψ,
χ)| =
√
q.
III. G ENERALIZED JACOBI SUMS AND RELATED CHARACTER SUMS
In this section, we present a generalization of Jacobi sums.
Let k be any positive integer. For each integer 1 ≤ i ≤ k, let mi be any positive integer, λi a
multiplicative character of Fqmi , χi the canonical additive character of Fqmi , and Trqmi /q the trace function
from Fqmi to Fq . Let χ denote the canonical additive character of Fq .
Now we define the generalized Jacobi sums by
X
Jba (λ1 , . . . , λk ) =
λ1 (c1 ) · · · λk (ck ),
(c1 ,··· ,ck )∈Sb
where
Sb = {(c1 , · · · , ck ) ∈ Fqm1 × . . . × Fqmk : Trqm1 /q (c1 ) + · · · + Trqmk /q (ck ) = a}
b = q m1 +···+mk −1 . If m1 = m2 = · · · = mk = 1 and a = 1, then Jba (λ1 , . . . , λk )
and a ∈ F∗q . Note that |S|
is the usual Jacobi sum.
b∗mi = hφi i
Theorem III.1. Let λi be a multiplicative character of Fqmi for i = 1, 2, . . . , k. Assume that F
q
and λi = φtii , where i = 1, 2, . . . , k and 0 ≤ ti ≤ q mi − 2.
(1) If all the multiplicative characters λ1 , . . . , λk are trivial, then
Jba (λ1 , . . . , λk ) = q m1 +···+mk −1 .
(2) If some, but not all, of λ1 , . . . , λk are trivial, then
Jba (λ1 , . . . , λk ) = 0.
(3) If all the multiplicative characters λ1 , . . . , λk are nontrivial and t1 + · · · + tk ≡ 0 (mod q − 1), then
|Jba (λ1 , . . . , λk )| = q
m1 +···+mk −2
2
.
(4) If all the multiplicative characters λ1 , . . . , λk are nontrivial and t1 + · · · + tk 6≡ 0 (mod q − 1), then
|Jba (λ1 , . . . , λk )| = q
m1 +···+mk −1
2
.
Proof. By the orthogonal relation of additive characters, we have
X
X
1
χ y Trqm1 /q (c1 ) + · · · + Trqmk /q (ck ) − a
Jba (λ1 , . . . , λk ) =
λ1 (c1 ) · · · λk (ck )
q
y∈F
(c1 ,...,ck )∈Fqm1
×···×Fqmk
=
1
q c
X
1 ∈Fq m1
q
λ1 (c1 ) · · ·
X
ck ∈Fqm1
λk (ck )
X
y∈F∗q
χ y Trqm1 /q (c1 ) + · · · + Trqmk /q (ck ) − a
6
X
X
1
+
λ1 (c1 ) · · ·
λk (ck )
q c ∈F m
ck ∈Fqm1
1
q 1
X
X
X
1
χ1 (yc1 ) · · · χk (yck )χ(ay)
λ1 (c1 ) · · ·
λk (ck )
=
q c ∈F m
∗
y∈Fq
ck ∈Fqm1
1
q 1
X
X
1
λ1 (c1 ) · · ·
λk (ck )
+
q c ∈F m
ck ∈Fqm1
1
q 1
X
X
X
1
λk (yck )χk (yck )
λ1 (yc1 )χ1 (yc1 ) · · ·
=
χ(ay)λ1 · · · λk (y)
q y∈F∗
ck ∈Fqmk
c1 ∈Fqm1
q
X
X
1
+
λk (ck )
λ1 (c1 ) · · ·
q c ∈F m
c ∈F m
1
=
k
1
X
q
1
1b
b k , χk )
G(λ1 , χ1 ) · · · G(λ
χ(ay)λ1 · · · λk (y)
q
y∈F∗q
X
X
1
+
λ1 (c1 ) · · ·
λk (ck )
q c ∈F m
c ∈F m
1
=
q
q
1
q
1
k
q
1
q
1
X
1
b 1 , χ1 ) · · · G(λ
b k , χk )
χ(ay)λ1 · · · λk (ay)
(λ1 · · · λk )(a)G(λ
q
y∈F∗q
X
X
1
λ1 (c1 ) · · ·
+
λk (ck ) .
q c ∈F m
c ∈F m
1
k
(III.1)
b∗ = hψi. For
b∗mi = hφi i and λi = φti , where i = 1, 2, . . . , k and 0 ≤ ti ≤ q mi − 2. Let F
Assume that F
q
q
i
y ∈ F∗q , one can deduce that
φi (y) = ψ(y) for i = 1, 2, . . . , k.
Hence, (λ1 · · · λk )(y) = ψ t1 +···+tk (y) where y ∈ F∗q and 0 ≤ ti ≤ q mi − 2 for i = 1, 2, . . . , k. This implies
that
X
X
χ(ay)λ1 · · · λk (ay) =
χ(y)λ1 · · · λk (y)
y∈F∗q
y∈F∗q
=
X
χ(y)ψ
t1 +···+tk
(y)
y∈F∗q
= G(ψ
t1 +···+tk
, χ).
(III.2)
Combining Equations (III.1) and (III.2), we have
1
b 1 , χ1 ) · · · G(λ
b k , χk )G(ψ t1 +···+tk , χ)
(λ1 · · · λk )(a)G(λ
Jba (λ1 , . . . , λk ) =
q
X
X
1
+
λk (ck )
λ1 (c1 ) · · ·
q c ∈F m
c ∈F m
1
q
1
k
q
1
(III.3)
7
where 0 ≤ ti ≤ q mi − 2 for i = 1, 2, . . . , k. In the following, we discuss the absolute values of
Jba (λ1 , . . . , λk ), a ∈ F∗q , in several cases.
(1) If all the multiplicative characters λ1 , . . . , λk are trivial, we have
b = q m1 +···+mk −1 .
Jba (λ1 , . . . , λk ) = |S|
(2) If some, but not all, of λ1 , . . . , λk are trivial, then by Equations (II.2), (III.3) and Corollary (II.5)
we have
Jba (λ1 , . . . , λk ) = 0.
(3) Assume that all the multiplicative characters λ1 , . . . , λk are nontrivial. By Equations (II.2) and (III.3),
we have
1
b 1 , χ1 ) · · · G(λ
b k , χk )G(ψt1 +···+tk , χ).
Jba (λ1 , . . . , λk ) = (λ1 · · · λk )(a)G(λ
q
Now we discuss the absolute values of Jba (λ1 , . . . , λk ) in the following two cases.
• If t1 + · · · + tk ≡ 0 (mod q − 1), then by Lemma II.4 and Corollary II.5 we have
|Jba (λ1 , . . . , λk )| = q
m1 +···+mk −2
2
.
• If t1 + · · · + tk 6≡ 0 (mod q − 1), then by Lemma II.4 and Corollary II.5 we have
The proof is completed.
|Jba (λ1 , . . . , λk )| = q
m1 +···+mk −1
2
.
Now we define another character sum related to generalized Jacobi sums by
X
Jea (λ1 , . . . , λk ) =
λ1 (c1 ) · · · λk (ck ),
(c1 ,··· ,ck )∈Se
where
and a ∈ Fq .
Se = {(c1 , · · · , ck ) ∈ F∗qm1 × . . . × F∗qmk : Trqm1 /q (c1 ) + · · · + Trqmk /q (ck ) = a}
e for a ∈ F∗q and a = 0 are given as follows.
Lemma III.2. The values of |S|
e = 1 (q m1 − 1) · · · (q mk − 1) + (−1)k+1 .
(1) For a ∈ F∗q , |S|
q
e = 1 (q m1 − 1) · · · (q mk − 1) + (−1)k (q − 1) .
(2) For a = 0, |S|
q
Proof. By the orthogonal relation of additive characters, we have
X X
X
e = 1
χ(y(Trqm1 /q (c1 ) + · · · + Trqmk /q (ck ) − a))
···
|S|
q c ∈F∗
∗
y∈F
c
∈F
q
1
k
q m1
q mk
m1
mk
X
X
X
(q − 1) · · · (q − 1) 1
=
χk (yck )
χ1 (yc1 ) · · ·
χ(−ay)
+
q
q y∈F∗
ck ∈F∗mk
c1 ∈F∗qm1
q
q
1
m1
mk
k+1
∗
(q − 1) · · · (q − 1) + (−1)
,
if a ∈ Fq ,
q
=
1
m1
mk
k
(q − 1) · · · (q − 1) + (−1) (q − 1) , if a = 0.
q
The relationship between Jea (λ1 , . . . , λk ) and Jba (λ1 , . . . , λk ), a ∈ F∗q , is established as follows.
Lemma III.3. The relationship between Jea (λ1 , . . . , λk ) and Jba (λ1 , . . . , λk ), a ∈ F∗q , is given as follows.
8
(1) If all of λ1 , . . . , λk are nontrivial, then
Jea (λ1 , . . . , λk ) = Jba (λ1 , . . . , λk ).
(2) If λ1 , . . . , λh are nontrivial and λh+1 , . . . , λk are trivial, 1 ≤ h ≤ k − 1, then
Jea (λ1 , . . . , λk ) = (−1)k−h Jba (λ1 , . . . , λh ).
Proof. The first conclusion is obvious and we only prove the second conclusion. If λ1 , . . . , λh are nontrivial
and λh+1 , . . . , λk are trivial, 1 ≤ h ≤ k − 1, then
X
Jea (λ1 , . . . , λk ) =
λ1 (c1 ) · · · λk (ck )
(c1 ,··· ,ck )∈Se
X
=
(c1 ,··· ,ck )∈Se
λ1 (c1 ) · · · λh (ch ),
where Se = {(c1 , · · · , ck ) ∈ F∗qm1 × . . . × F∗qmk : Trqm1 /q (c1 ) + · · · + Trqmk /q (ck ) = a}. For a fixed h-tuple
(c1 , . . . , ch ) ∈ F∗qm1 × · · · × F∗qmh ,
we deduce that the number of the solutions of the equation
Trqmh+1 /q (ch+1 ) + · · · + Trqmk /q (ck ) = a − (Trqm1 /q (c1 ) + · · · + Trqmh /q (ch ))
equals
(q mh+1 − 1) · · · (q mk − 1) + (−1)k−h+1 ,
if Trqm1 /q (c1 ) + · · · + Trqmh /q (ch ) 6= a,
mh+1
mk
k−h
(q
− 1) · · · (q − 1) + (−1) (q − 1) , if Trqm1 /q (c1 ) + · · · + Trqmh /q (ch ) = a,
1
q
1
q
by Lemma III.2. Hence,
Jea (λ1 , . . . , λk )
1
=
(q mh+1 − 1) · · · (q mk − 1) + (−1)k−h+1
q
+
=
(c1 ,··· ,ch )∈F∗qm1 ×···×F∗qmh
1
(q mh+1 − 1) · · · (q mk − 1) + (−1)k−h+1
q
Since λ1 , . . . , λh are nontrivial, we have
X
(c1 ,··· ,ch )∈F∗qm1 ×···×F∗qmh
λ1 (c1 ) · · · λh (ch ) =
λ1 (c1 ) · · · λh (ch )
Trqm1 /q (c1 )+···+Trqmh /q (ch )6=a
1
(q mh+1 − 1) · · · (q mk − 1) + (−1)k−h (q − 1)
q
+(−1)k−h Jea (λ1 , . . . , λh ).
Thus
X
X
(c1 ,··· ,ch )∈F∗qm1 ×···×F∗qmh
Trqm1 /q (c1 )+···+Trqmh /q (ch )=a
X
(c1 ,··· ,ch )∈F∗qm1 ×···×F∗qmh
X
c1 ∈F∗qm1
λ1 (c1 ) · · · λh (ch )
λ1 (c1 ) · · ·
λ1 (c1 ) · · · λh (ch )
X
ch ∈F∗mh
q
λh (ch ) = 0.
Jea (λ1 , . . . , λk ) = (−1)k−h Jea (λ1 , . . . , λh ) = (−1)k−h Jba (λ1 , . . . , λh ).
Theorem III.1 and Lemma III.3 yield the following theorem.
9
Theorem III.4. Let a ∈ F∗q . Let λi be a multiplicative character of Fqmi for i = 1, 2, . . . , k. Assume that
b∗mi = hφi i and λi = φti , where i = 1, 2, . . . , k and 0 ≤ ti ≤ q mi − 2.
F
q
i
(1) If all the multiplicative characters λ1 , . . . , λk are trivial, then
1
Jea (λ1 , . . . , λk ) =
(q m1 − 1) · · · (q mk − 1) + (−1)k+1 .
q
(2) If λ1 , . . . , λh are nontrivial, λh+1 , . . . , λk are trivial and t1 +· · ·+th ≡ 0 (mod q−1),1 ≤ h ≤ k −1,
then
m1 +···+mh −2
2
|Jea (λ1 , . . . , λk )| = q
.
(3) If λ1 , . . . , λh are nontrivial, λh+1 , . . . , λk are trivial and t1 +· · ·+th 6≡ 0 (mod q−1), 1 ≤ h ≤ k−1,
then
m1 +···+mh −1
2
.
|Jea (λ1 , . . . , λk )| = q
(4) If all the multiplicative characters λ1 , . . . , λk are nontrivial and t1 + · · · + tk ≡ 0 (mod q − 1), then
|Jea (λ1 , . . . , λk )| = q
m1 +···+mk −2
2
.
(5) If all the multiplicative characters λ1 , . . . , λk are nontrivial and t1 + · · · + tk 6≡ 0 (mod q − 1), then
IV. A SYMPTOTICALLY
|Jea (λ1 , . . . , λk )| = q
m1 +···+mk −1
2
.
OPTIMAL CODEBOOKS BASED ON GENERALIZED JACOBI SUMS AND RELATED
CHARACTER SUMS
In this section, we present a construction of codebooks with multiplicative characters of finite fields.
We follow the notations in Section III.
Let Fqm1 , . . . , Fqmk , be any k finite fields, where m1 , . . . , mk are any k positive integers. For an nonempty
set
S ⊆ Fqm1 × · · · × Fqmk ,
let K := |S|.
Let EK denote the set formed by the standard basis of the K-dimensional Hilbert space:
(1, 0, 0, · · · , 0, 0),
(0, 1, 0, · · · , 0, 0),
..
.
(0, 0, 0, · · · , 0, 1).
b∗mi , i = 1, 2, . . . , k, we define a unit-norm codeword of length K by
For any λi ∈ F
q
c(λ1 , . . . , λk ) = √
1
nc̄(λ1 ,...,λk )
(λ1 (c1 ) · · · λk (ck ))(c1 ,...,ck )∈S ,
(IV.1)
where nc̄(λ1 ,...,λk ) denotes the Euclidean norm of the vector
c̄(λ1 , . . . , λk ) := (λ1 (c1 ) · · · λk (ck ))(c1 ,...,ck )∈S .
Now we present a construction of codebooks as
b∗mi for i = 1, 2, . . . , k} ∪ EK .
C = {c(λ1 , . . . , λk ) : λi ∈ F
q
(IV.2)
Q
Then C has N = ki=1 (q mi − 1) + K codewords. If the set S is properly selected, then C may have good
parameters with respect to the Welch or the Levenshtein bound.
10
A. When S = Sb
b where Sb is defined in Section III for
In the following, we investigate Imax (C) if we select S = S,
∗
m1 +···+mk −1
b
a ∈ Fq . Then K = |S| = q
. Now we consider the value of nc̄(λ1 ,...,λk ) defined in Equation
(IV.1). It is easy to verify that
e ≤ nc̄(λ ,...,λ ) ≤ |S|,
b
|S|
(IV.3)
1
k
e = 1 (q m1 − 1) · · · (q mk − 1) + (−1)k+1 by Lemma III.2 and |S|
b = q m1 +···+mk −1 . We remark
where |S|
q
that nc̄(λ1 ,...,λk ) achieves the lower bound of Inequality (IV.3) if all of λ1 , · · · , λk are nontrivial, and achieves
the upper bound of Inequality (IV.3) if all of λ1 , · · · , λk are trivial.
Theorem IV.1. If S = Sb and q ≥ 4, the codebook C in Equation (IV.2) has parameters
!
k
Y
(q mi − 1) + q m1 +···+mk −1 , q m1 +···+mk −1
i=1
and
q
Imax (C) = Qk
m1 +···+mk +1
2
mi − 1) + (−1)k+1
i=1 (q
.
Proof. Let c′ , c′′ ∈ C be any two different codewords. Denote F = C\EK . Now we calculate the correlation
of c′ and c′′ in the following cases.
(1) If c′ , c′′ ∈ EK , we directly have |c′ c′′H | = 0.
(2) If c′ ∈ F , c′′ ∈ EK or c′ ∈ EK , c′′ ∈ F , we have
|c′ c′′H | = √
1
nc̄(λ1 ,...,λk )
|λ1 (c1 ) · · · λk (ck )| = √
1
nc̄(λ1 ,...,λk )
b∗mi , i = 1, 2, . . . , k. By Equation (IV.3), we have
for some (c1 , . . . , ck ) ∈ Sb and λi ∈ F
q
r
1
q
′ ′′H
p
.
≤ |c c | ≤
m
m
1
(q − 1) · · · (q k − 1) + (−1)k+1
q m1 +···+mk −1
Both the lower bound and the upper bound of this inequality can be achieved.
(3) If c′ , c′′ ∈ F , we assume that c′ = c(λ′1 , . . . , λ′k ) and c′′ = c(λ′′1 , . . . , λ′′k ) with (λ′1 , . . . , λ′k ) 6=
b∗mi for i = 1, 2, . . . , k. Denote λi = λ′ λ′′ for all i = 1, 2, . . . , k. Then
(λ′′1 , . . . , λ′′k ) where λ′i , λ′′i ∈ F
i i
q
by Equation (IV.1) we have
X
1
1
c′ c′′H = √
·√
λ1 (c1 ) · · · λk (ck )
nc̄(λ′1 ,...,λ′k )
nc̄(λ′′1 ,...,λ′′k )
(c1 ,...,ck )∈Sb
= √
1
nc̄(λ′1 ,...,λ′k )
·√
1
nc̄(λ′′1 ,...,λ′′k )
Jba (λ1 , . . . , λk ).
Since (λ′1 , . . . , λ′k ) 6= (λ′′1 , . . . , λ′′k ), not all of λi , i = 1, 2, · · · , k are trivial characters. Hence, by
Theorem III.1, we have
(
)
m1 +···+mk −2
m1 +···+mk −1
2
2
q
q
|c′ c′′H | ∈ 0, √
,√
.
√
√
nc̄(λ′1 ,...,λ′k ) · nc̄(λ′′1 ,...,λ′′k )
nc̄(λ′1 ,...,λ′k ) · nc̄(λ′′1 ,...,λ′′k )
Due to the lower bound of Inequality (IV.3), we obtain
m1 +···+mk +1
′ ′′H
|c c
|≤
(q m1
2
q
.
m
k
− 1) · · · (q − 1) + (−1)k+1
(IV.4)
11
In the following, we prove that there indeed exist c′ , c′′ ∈ F such that |c′ c′′H | achieves the upper
bound in Inequality (IV.4). Due to Lemma III.1 and Inequality (IV.3), it is sufficient to prove that
there exist (λ′1 , . . . , λ′k ), (λ′′1 , . . . , λ′′k ) such that
• all of λ′1 , . . . , λ′k are nontrivial,
• all of λ′′1 , . . . , λ′′k are nontrivial,
b∗mi = hφi i
• all of λ1 , . . . , λk are nontrivial and t1 +· · ·+tk 6≡ 0 (mod q −1), where λi = λ′i λ′′i , F
q
ti
and λi = φi for i = 1, 2, . . . , k and 0 ≤ ti ≤ q mi − 2.
In fact, we can choose a positive integer s such that
0 < s ≤ min{q mi − 2 : i = 1, · · · , k} and 0 < s + 1 ≤ min{q mi − 2 : i = 1, · · · , k}.
Firstly, we assume that k = 2t + 1 for some nonnegative integer t. Let
s+1
s
(λ′1 , · · · , λ′t , λ′t+1 , · · · , λ′2t , λ′2t+1 ) = (φs1 , · · · , φst , φs+1
t+1 , · · · , φ2t , φ2t+1 )
and
s+1
(λ′′1 , · · · , λ′′t , λ′′t+1 , · · · , λ′′2t , λ′′2t+1 ) = (φs+1
, φst+1 , · · · , φs2t , φs+1
1 , · · · , φt
2t+1 ).
Then
−1
1
1
−1
(λ1 , · · · , λt , λt+1 , · · · , λ2t , λ2t+1 ) = (φ−1
1 , · · · , φt , φt+1 , · · · , φ2t , φ2t+1 ),
which implies that
t1 + · · · + tk ≡ −1
(mod q − 1).
Secondly, we assume that k = 2t for some positive integer t. Let
s
s
(λ′1 , · · · , λ′t−1 , λ′t , · · · , λ′2t−2 , λ′2t−1 , λ′2t ) = (φs1 , · · · , φst−1 , φs+1
, · · · , φs+1
t
2t−2 , φ2t−1 , φ2t )
and
s+1
s
s
s+1
s+1
(λ′′1 , · · · , λ′′t−1 , λ′′t , · · · , λ′′2t−2 , λ′′2t−1 , λ′′2t ) = (φs+1
1 , · · · , φt−2 , φt−1 , · · · , φ2t−2 , φ2t−1 , φ2t ).
Then
−1
1
1
−1
−1
(λ1 , · · · , λt−2 , λt−1 , · · · , λ2t−2 , λ2t−1 , λ2t ) = (φ−1
1 , · · · , φt−2 , φt−1 , · · · , φ2t−2 , φ2t−1 , φ2t ),
which implies that
t1 + · · · + tk ≡ −2
(mod q − 1).
Thus we have proved that there indeed exist c′ , c′′ ∈ F such that
m1 +···+mk +1
′ ′′H
|c c
|=
(q m1
2
q
.
m
− 1) · · · (q k − 1) + (−1)k+1
Summarizing the conclusions in the three cases above, we obtain
Imax (C) = Qk
q
i=1 (q
m1 +···+mk +1
2
mi
− 1) + (−1)k+1
.
Theorem IV.1 contains Theorem 19 in [11] as a special case. If not all of mi , i = 1, · · · , k, are equal to
each other, then the parameters of C in Theorem IV.1 are different to those of the codebook in Theorem
19 of [11].
Remark IV.2. Now we show that C in Theorem IV.1 are asymptotically optimal.
12
If k = 1 and m1 = 2, then C is a (q 2 + q − 1, q) codebook with Imax (C) =
√
√1 .
q
In this case, we
(C)
limq→∞ Imax
IL
q+2
q+1
by Lemma I.2. Then we have
= 1 which implies that C is
have N > K 2 and IL =
asymptotically achieving the Levenshtein bound.
If k = 1, m1 > 2 or k > 1, we can deduce that N < K 2 and
s
Qk
mi − 1)
i=1 (q
IW =
Qk
( i=1 (q mi − 1) + q m1 +···+mk −1 − 1)q m1 +···+mk −1
due to Lemma I.1. By Theorem IV.1, we can prove that limq→∞
asymptotically achieving the Welch bound.
Imax (C)
IW
= 1 which implies that C is
B. When S = Se
e where Se is defined in Section III for a ∈ F∗ .
In the following, we investigate Imax (C) if we select S = S,
q
e = 1 (q m1 − 1) · · · (q mk − 1) + (−1)k+1 by Lemma III.2. It is clear that nc̄(λ1 ,...,λ ) = K
Then K = |S|
k
q
b∗mi , i = 1, 2, · · · , k.
in Equation (IV.1) for any λi ∈ F
q
e the codebook C in Equation (IV.2) has parameters
Theorem IV.3. If S = S,
!
!!
k
k
k
Y
Y
Y
1
1
(q mi − 1) +
(q mi − 1) + (−1)k+1 ,
(q mi − 1) + (−1)k+1
q
q
i=1
i=1
i=1
and
Imax (C) = Qk
q
i=1 (q
m1 +···+mk +1
2
mi
− 1) + (−1)k+1
.
Proof. Let c′ , c′′ ∈ C be any two different codewords. Denote F = C\EK . Now we calculate the correlation
of c′ and c′′ in the following cases.
(1) If c′ , c′′ ∈ EK , we directly have |c′ c′′H | = 0.
(2) If c′ ∈ F , c′′ ∈ EK or c′ ∈ EK , c′′ ∈ F , we have
1
1
|c′ c′′H | = √ |λ1 (c1 ) · · · λk (ck )| = √
K
K
∗
b mi , i = 1, 2, . . . , k.
for some (c1 , . . . , ck ) ∈ Se and λi ∈ F
q
′ ′′
′
(3) If c , c ∈ F , we assume that c = c(λ′1 , . . . , λ′k ) and c′′ = c(λ′′1 , . . . , λ′′k ) with (λ′1 , . . . , λ′k ) 6=
b∗mi for i = 1, 2, . . . , k. Denote λi = λ′ λ′′ for all i = 1, 2, . . . , k. Then
(λ′′1 , . . . , λ′′k ) where λ′i , λ′′i ∈ F
i i
q
by Equation (IV.1) we have
X
1
c′ c′′H =
λ1 (c1 ) · · · λk (ck )
K
(c1 ,...,ck )∈Se
1 e
Ja (λ1 , . . . , λk ).
K
Since (λ′1 , . . . , λ′k ) 6= (λ′′1 , . . . , λ′′k ), not all of λi , i = 1, 2, · · · , k are trivial characters. Hence, by
Theorem III.4, we have
) ( m1 +···+mh −1
)
( m1 +···+mh −2
[ q
2
2
q
:1≤h≤k
:1≤h≤k .
|c′ c′′H | ∈
K
K
=
Summarizing the conclusions in the three cases above, we obtain
Imax (C) = Qk
q
m1 +···+mk +1
2
mi − 1) + (−1)k+1
i=1 (q
.
13
Theorem IV.2 contains Theorem 15 in [11] as a special case. If not all of mi , i = 1, · · · , k, are equal to
each other, then the parameters of C in Theorem IV.1 are different to those of the codebook in Theorem
15 of [11].
Remark IV.4. Now we show that C in Theorem IV.3 is asymptotically optimal.
If k = 1 and m1 = 2, then C in Theorem IV.3 is the same as that in Theorem IV.1. Hence, by Remark
IV.2, C is asymptotically optimal according to the Levenshtein bound.
If k = 2 and m1 = m2 = 1, then C in Theorem IV.3 is the same as that in [11, Theorem 16]. Hence,
by [11, Theorem 16], C is asymptotically optimal according to the Levenshtein bound.
In other cases, by Theorem IV.3 we have N < K 2 . Combining Lemma I.1 and Theorem IV.3, we can
(C)
easily prove that limq→∞ Imax
= 1. Then C is asymptotically optimal according to the Welch bound.
IW
V. C ONCLUSIONS AND REMARKS
This paper gave two classes of asymptotically optimal codebooks based on generalized Jacobi sums
and related character sums. The main contributions are the following:
• We generalized the classical Jacobi sums over finite fields and defined so-called generalized Jacobi
sums. The absolute values of the generalized Jacobi sums were investigated. Besides, some related
characters sums derived from generalized Jacobi sums were also studied.
• We obtained a class of asymptotically optimal codebooks in Theorem IV.1 based on generalized
Jacobi sums. This result contains that in [11, Theorem 19] as a special case.
• We obtained a class of asymptotically optimal codebooks in Theorem IV.3 based on character sums
related to generalized Jacobi sums. This result contains that in [11, Theorem 15] as a special case.
As pointed out in [16], constructing optimal codebooks with minimal Imax is very difficult in general.
This problem is equivalent to line packing in Grassmannian spaces [2]. In frame theory, such a codebook
with Imax minimized is referred to as a Grassmannian frame [24]. The codebooks presented in this paper
should have applications in these areas. With the framework developed in [19], our codebooks can be
used to obtain deterministic sensing matrices with small coherence for compressed sensing.
R EFERENCES
[1] A. R. Calderbank, P. J. Cameron, W. M. Kantor, and J. J. Seidel, “Z4 -kerdock codes, orthogonal spreads, and extremal Euclidean
line-sets,” Proc. London Math. Soc., vol. 75, no. 3, pp. 436–480, 1997.
[2] J. H. Conway, R. H. Harding, and N. J. A. Sloane, “Packing lines, planes, etc.: Packings in grassmannian spaces,” Exp. Math., vol. 5,
no. 2, pp. 139–159, 1996.
[3] C. Ding, “Complex codebooks from combinatorial designs,” IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 4229–4235, Sept. 2006.
[4] C. Ding and T. Feng, “A generic construction of complex codebooks meeting the Welch bound,” IEEE Trans. Inf. Theory, vol. 53, no.
11, pp. 4245–4250, Nov. 2007.
[5] C. Ding, T. Feng, “Codebooks from almost difference sets,” Des. codes Crypt., vol. 53, no. 11, pp. 4245–4250, Nov. 2007.
[6] C. Ding, J. Yin, “Singal sets from functions with optimal nonlinearity,” IEEE Trans. Commun., vol. 55, no. 5, pp. 936–940, May 2007.
[7] M. Fickus, D. G. Mixon, “Tables of the existence of equiangular tight frames,” arXiv:1504.00253v2, 2016.
[8] M. Fickus, D. G. Mixon, J. Jasper, “Equiangular tight frames from hyperovals,” IEEE Trans. Inf. Theory, vol. 62, no. 9, pp. 5225–5236,
2016.
[9] M. Fickus, J. Jasper, D. G. Mixon, J. Peterson, “Tremain equiangular tight frames,” arXiv:1602.03490v1, 2016.
[10] M. Fickus, D. G. Mixon, and J. C. Tremain, “Steiner equiangular tight frames,” Linear Algebra Appl., vol. 436, no. 5, pp. 1014–1027,
2012.
[11] Z. Heng, C. Ding, Q. Yue, “New constructions of asymptotically optimal codebooks with multiplicative characters,” IEEE Trans. Inf.
Theory, doi: 10.1109/TIT.2017.2693204.
[12] Z. Heng, Q. Yue, “Optimal codebooks achieving the Levenshtein bound from generalized bent functions over Z4 ”, Cryptogr. Commun.,
vol. 9, no. 1 , pp. 41–53, 2017.
[13] H. Hu, J. Wu, “New constructions of codebooks nearly meeting the Welch bound,” IEEE Trans. Inf. Theory, vol. 60, no. 2, pp.
1348–1355, Feb. 2014.
[14] B. M. Hochwald, T. L. Marzetta, T. L. Richardson, W. Sweldens, and R. Urbanke, “Systematic design of unitary space-time
constellations,” IEEE Trans. Inf. Theory, vol. 46, no. 6, pp. 1962–1973, Sept. 2000.
[15] R. Lidl, H. Niederreiter, Finite Fields. Cambridge Univ. Press, Cambridge, 1984.
14
[16] D. J. Love, R.W. Heath, and T. Strohmer, “Grassmannian beamforming for multiple-input multiple-output wireless systems,” IEEE
Trans. Inf. Theory, vol. 49, no. 10, pp. 2735–2747, Oct. 2003.
[17] C. Li, Q. Yue, Y. Huang, “Two families of nearly optimal codebooks,” Des. Codes Crypt., vol. 75, no. 1, pp. 43–57, Apr. 2015.
[18] V. I. Levenshtein, “Bounds for packing of metric spaces and some of their applications,” Probl. Cybern., vol. 40, pp. 43–110, 1983.
[19] S. Li, G. Ge, “Deterministic sensing matrices arising from near orthogonal systems,” IEEE Trans. Inf. Theory, vol. 60, no. 4, pp.
2291–2302, Feb. 2014.
[20] J. L. Massey, T. Mittelholzer, “Welch’s bound and sequence set for code-division muitiple-access systems,” in Sequences II. New York,
NY, USA: Springer-verlag, 1993, pp. 63–78.
[21] K. Mukkavilli, A. Sabharwal, E. Erkip, and B. A. Aazhang, “On beamforming with finite rate feedback in multiple antenna systems,”
IEEE Trans. Inf. Theory, vol. 49, no. 10, pp. 2562–2579, Oct. 2003.
[22] F. Rahimi, “Covering graphs and equiangular tight frames,” Ph.D. Thesis, University of Waterloo, Ontario, 2016 (available at
http://hdl.handle.net/10012/10793).
[23] D. Sarwate, Meeting the Welch bound with equality, Proc. of SETA’98, Springer-Verlag, Berlin, Heidelberg, 1999, pp. 63–79.
[24] T. Strohmer and R. W. Heath Jr., “Grassmannian frames with applications to coding and communication,” Appl. Comput. Harmonic
Anal., vol. 14, no. 3, pp. 257–275, 2003.
[25] V. Tarokh, I. M. Kim, “Existence and construction of noncoherent unitary space-time codes,” IEEE Trans. Inform. Theory, vol. 48, no.
12, pp. 3112–3117, Dec. 2002.
[26] E. V. Tsiligianni, L. P. Kondi and A. K. Katsaggelos, “Construction of incoherent unit norm tight frames with application to compressed
sensing,” IEEE Trans. Inf. Theory, vol. 60, no. 4, pp. 2319–2330, Apr. 2014.
[27] P. Tan, Z. Zhou, D. Zhang, “A construction of codebooks nearly achieving the Levenshtein bound,” IEEE Signal Processing Letters,
vol. 23, no. 10, pp. 1306–1309, Oct. 2016.
[28] L. Welch, “Lower bounds on the maximum cross correlation of signal,” IEEE Trans. Inf. Theory, vol. IT-20, pp. 397–399, 1974.
[29] W. Wooters and B. Fields, “Optimal state-determination by mutually unbiased measurements,” Ann. Phys., vol. 191, no. 2, pp. 363–381,
1989.
[30] P. Xia, S. Zhou, and G. B. Giannakis, “Achieving the Welch bound with difference sets,” IEEE Trans. Inf. Theory, vol. 51, no. 5, pp.
1900–1907, May 2005.
[31] C. Xiang, C. Ding and S. Mesnager, “Optimal codebooks from binary codes meeting the Levenshtein bound,” IEEE Trans. Inf. Theory,
vol. 61, no. 12, pp. 6526–6535, Dec. 2015.
[32] N. Y. Yu, “A construction of codebooks associated with binary sequences,” IEEE Trans. Inf. Theory, vol. 58, no. 8, pp. 5522–5533,
Aug. 2012.
[33] A. Zhang, K. Feng, “Two classes of codebooks nearly meeting the Welch bound,”. IEEE Trans. Inf. Theory, vol. 58, no. 4, pp.
2507–2511, 2009.
[34] Z. Zhou, C. Ding and N. Li, “New families of codebooks achieving the Levenshtein bound,” IEEE Trans. Inf. Theory, vol. 60, no. 11,
pp. 7382–7387, Nov. 2014.
[35] Z. Zhou, X. Tang, “New nearly optimal codebooks from relative difference sets,” Adv. Math. Commun., vol. 5, no. 3, pp. 521–527,
2011.
| 7cs.IT
|
1
Comprehensive Modeling of Three-Phase
Distribution Systems via the Bus Admittance Matrix
arXiv:1705.06782v2 [math.OC] 25 Sep 2017
Mohammadhafez Bazrafshan, Student Member, IEEE, and Nikolaos Gatsis, Member, IEEE
Abstract—The theme of this paper is three-phase distribution
system modeling suitable for the Z-Bus load-flow. Detailed models
of wye and delta constant-power, constant-current, and constantimpedance loads are presented. Models of transmission lines,
step-voltage regulators, and transformers that build the bus
admittance matrix (Y-Bus) are laid out. The Z-Bus load-flow
is then reviewed and the singularity of the Y-Bus in case of
certain transformer connections is rigorously discussed. Based
on realistic assumptions and conventional modifications, the
invertibility of the Y-Bus is proved. Last but not least, MATLAB
scripts that model the components of the IEEE 37-bus, the IEEE
123-bus, the 8500-node feeders, and the European 906-bus lowvoltage feeder are provided.
Index Terms—Three-phase distribution system modeling, ZIP
loads, Z-Bus method, load-flow, bus admittance matrix
I. I NTRODUCTION
N accurately constructed bus admittance matrix (YBus) that captures the unbalanced characteristics of
distribution networks is the basis of several applications
such as 1) three-phase load-flow based on Newton-Raphson
[1], the current injection method [2], or the Z-Bus method
[3]; 2) three-phase optimal power flow (OPF) using interior–
point methods [4], semidefinite relaxations [5], or successive
convex approximations [6]; 3) voltage security assessment [7],
[8] through conditions for solution existence [9]–[13]; 4) optimal system operation by selecting optimal regulator tap
settings [14], [15] and optimal capacitor switch reconfiguration
[16]; and 5) providing real-time voltage solutions by linearizing three-phase power flow equations [17]–[20].
To facilitate three-phase distribution system studies, such
as the above examples, and to bypass single-phase simplifications, this paper brings together models—some previously
available, some novel—for elements of distribution networks
and constructs the Y-Bus matrix. The Y-Bus matrix incorporates models of three-phase transmission lines, transformers,
and step-voltage regulators (SVRs).
In addition, this paper analyzes how each model affects the
invertibility of the Y-Bus matrix and rigorously proves why a
previous proposal in [21] for certain transformer connections
removes the singularity of the Y-Bus. The invertibility of the YBus is especially important since it allows for computation of
load-flow solutions through the Z-Bus method [3] and provides
conditions of solution existence [12], [13].
Three-phase power system modeling is the theme of [22],
although models of voltage-dependent loads and SVRs are
A
The authors are with the Dept. of Electrical & Computer Engineering,
Univ. of Texas at San Antonio. Emails: {mohammadhafez.bazrafshan, nikolaos.gatsis}@utsa.edu. This material is based upon work supported by the
National Science Foundation under Grant No. CCF-1421583.
not included. For the forward-backward sweep load-flow (FBSweep), ABCD matrices of transmission lines, distribution
transformers, and SVRs are derived in [23]. Deriving the corresponding admittance matrices from the ABCD parameters is
not straightforward—especially for transformers and SVRs.
Three-phase nodal admittance matrices of certain transformer connections, such as the delta–delta, are singular [22],
[24]. These singularities are somewhat remedied for the FBSweep in [25] by adding independent KCL and KVL equations
and in [26] by showing that the zero-sequence components
of voltages do not affect the backward sweep calculations;
thus, they render unique non-zero sequence voltage solutions.
Alternative transformer nodal admittances are suggested in
[27] by assembling their symmetrical component circuits.
In the Z-Bus method realm, Ohm’s law is used to obtain
voltage solutions by inverting the Y-Bus at every iteration.
To bypass Y-Bus singularities, [21] advocates adding small
admittances from isolated circuits to the ground and [28]
suggests modifying the LDL factorization to use the limit for
zero by zero divisions. An alternative approach to avoid Y-Bus
singularity is to use equivalent current injections [29].
Despite the extensive literature on distribution network
modeling [21]–[29] and references therein, precise unified YBus modeling accounting for transmission lines with missing
phases and SVRs—as present in IEEE feeders—is lacking.
Moreover, there is a deficit in theoretical studies on the
invertibility of the Y-Bus and how this is affected by nodal
admittances of various distribution system components.
This paper sets out to fill the aforementioned gaps. First,
detailed models of wye and delta constant-power, constantcurrent, and constant-impedance (ZIP) loads, three-phase
transformers, and transmission lines with missing phases are
presented. A novel nodal admittance model for SVRs, with and
without considering the leakage admittance, is also derived.
Specifically, primary-to-secondary gain matrices for line-toneutral voltages are derived from first principles. These are
combined with a series transmission line, which is the typical
configuration in which SVRs are installed in distribution
systems. It is worth noting that the SVR models developed
in this work do not use sequence network matrices (as per
[23]), and can also be utilized in the FB-Sweep algorithm.
Second, this paper precisely explains which distribution
system components cause singularities in the Y-Bus and how
they can be rectified. In particular, typical SVRs do not render
the Y-Bus singular, rather, the Y-Bus singularity is due to
self-admittances of certain transformer connections. Based
on linear algebra, this paper shows how connecting small
shunt admittances, which has been previously proposed for
modification of the Y-Bus [21], aids in restoring its invertibil-
2
ity. The Y-Bus invertibility is then proved for networks that
include arbitrary combinations of very general and practical
component models—including transmission lines with missing
phases—that are found in typical distribution feeders. The
invertibility of the Y-Bus is necessary for computing voltage
solutions via the Z-Bus method.
Third, a set of MATLAB scripts is provided online that takes
as input the data files for the IEEE 37-bus, the IEEE 123-bus,
the 8500-node feeders, and the European 906-bus low voltage
feeder [30], and models the loads, transmission lines, SVRs,
and transformers. These scripts further build the corresponding
Y-Bus for each feeder and implement the Z-Bus method to
compute the voltages. The power flow solutions obtained are
within 0.75% of benchmark solutions.
Paper organization: The notation required for three-phase
distribution system modeling is introduced in Section II.
Models for wye and delta ZIP loads are given in Section III.
Modeling of series elements, i.e., transmission lines, various
types of SVRs, and transformers is taken up in Section IV.
Section V puts the aforementioned models together, constructs
the Y-Bus, reviews the Z-Bus method, and explains the procedures to handle Y-Bus singularities. A rigorous proof of YBus invertibility is detailed in Section VI. Extensions guaranteeing Y-Bus invertibility under more practical considerations
are pursued in Section VII. Section VIII provides numerical
discussions and load-flow results on distribution test feeders.
The paper concludes in Section IX.
II. M ODELING N OTATIONS
Power distribution networks comprise two types of elements: 1) shunt elements such as loads and shunt capacitors;
and 2) series elements such as three-phase transmission lines,
transformers, and SVRs. Mathematically, we model a threephase power distribution network by an undirected graph
(N , E). The set N := {1, 2, . . . , N } ∪ {S} is the set of nodes
and represent the shunt elements, while E := {(m, n)} ⊆
N × N is the set of edges representing the series elements.
Node S is considered to be the slack bus connected to the
substation. Furthermore, we define the set of neighboring
nodes to node n as Nn := {m|(n, m) ∈ E}.
For a series element, i.e., the edge (n, m) ∈ E, let Ωnm =
Ωmn denote its set of phases. Define iφnm as the current flowing
from node n to node m on phase φ ∈ Ωnm
S . Define further
the available phases of a node n as Ωn := m∈Nn Ωnm . Let
inm ∈ C|Ωn | collect the currents on all phases flowing from
node n to node m such that inm (Ωnm ) = {iφnm }φ∈Ωnm and
inm ({φ}) = 0 if φ ∈ Ωn \ Ωnm , that is, notation inm (Ωnm )
picks the indices of inm that correspond to the phases in Ωnm .
We partition N as N = NY ∪N∆ ∪{S} where NY and N∆
collect wye and delta nodes respectively. For wye nodes, i.e.,
n ∈ NY , Ωn may have one, two, or three available phases.
For delta nodes, i.e., n ∈ N∆ , Ωn has at least two available
phases, that is |Ωn | ≥ 2.
For node n and phase φ ∈ Ωn , the complex line to neutral
voltage is denoted by vnφ and the net current injection is
denoted by iφn . Moreover, define vectors in = {iφn }φ∈Ωn ,
vn = {vnφ }φ∈Ωn in C|Ωn | respectively as the vector of
TABLE I
W YE AND DELTA ZIP LOAD MODELS
Current portion
Wye loads
iφ
PQn (vn )
φ ∗
−(sφ
Ln /vn )
φ
vn
iφ
In (vn )
− φ iφ
|v | Ln
iφ
Zn (vn )
φ
φ
vn
−yL
n
Delta loads
−
φ′ ∈Ωn \{φ}
−
n
P
φ′ ∈Ωn \{φ}
−
P
φ′ ∈Ωn \{φ}
φφ′
n
φ
φ′
vn −vn
sL
P
!∗
′
′ v φ −v φ
n
n
φ
φ′
|vn −vn |
′
′
iφφ
Ln
φ
φ
φφ
(vn
− vn
)
yL
n
net current injection and complex line to neutral voltages
at node
P n. Notice that Ohm’s law at each node demands
in = m∈Nn inm . Collect these quantities for all nodes in
the vectors i = {in }n∈N \{S} and v = {vn }n∈N \{S} .
further an index set J := {1, . . . , J} where J =
PDefine
N
|Ω
n |, and j ∈ J is a linear index corresponding to a
n=1
particular pair (n, φ) with n ∈ N \{S} and φ ∈ Ωn . In this
case, denote n = Node[j], and define Jn := {j|Node[j] =
n} as the set of linear indices corresponding to node n.
III. T HREE -P HASE ZIP L OAD M ODELS
Due to existence of loads, the nodal net current injection
i is a function of nodal voltages v. This dependence is
denoted by in (vn ). According to the ZIP load model, in (vn )
is
iPQn =
φcomposed of currents from constant-power loads
φ
iPQn φ∈Ωn , constant-current loads iIn = iIn φ∈Ωn , and
constant-impedance loads iZn = iφZn φ∈Ωn . For n ∈ N \{S}
and φ ∈ Ωn we have that
iφn (vn ) = iφPQn (vn ) + iφIn (vn ) + iφZn (vn )
(1)
where functions iφPQn (vn ), iφIn (vn ), and iφZn (vn ) are given
in Table I for wye and delta connections where sφLn , iφLn ,
φ
and yL
are respectively the nominal constant-power, constantn
current, and constant-impedance portions
of the ZIP model
′
φφ′
φφ′
,
i
for wye nodes n ∈ NY . Quantities sφφ
Ln , and yLn are
Ln
respectively the nominal constant-power, constant-current, and
constant-impedance portions of the ZIP model for nodes n ∈
N∆ and over phases
φ, ′φ′ ∈ Ω′ n . For′ n ∈ N∆ and
φ, φ′ ′ ∈ Ωn ,
φφ′
φ φ φφ
φφ
φφ′
φφ
we have that sLn = sLn , iLn = iLn , and yLn = yL
. Note
n
φ
φ
that a simpler constant-current model is setting iIn = iLn . The
constant-power or constant-current expressions of Table I may
also be used to model distributed generation units connected
to node n with appropriate signs for sφLn and iφLn .
Due to the linear relationship between voltage and currents
of constant-impedance loads in Table I, for all nodes n ∈
NY ∪ N∆ we have
iZn (vn ) = −YLn vn
(2)
where YLn ∈ C|Ωn |×|Ωn | is defined as follows:
φ
, n ∈ NY , φ ∈ Ωn
YLn (φ, φ) = yL
n
X
φφ′
yL
, n ∈ N∆ , φ ∈ Ωn
YLn (φ, φ) =
n
(3a)
(3b)
φ′ ∈Ωn \{φ}
′
φφ
YLn (φ, φ′ ) = −yL
, n ∈ N∆ , φ ∈ Ωn , φ′ ∈ Ωn \ {φ}. (3c)
n
3
Remark 1: The ZIP load model only approximately represents the dependencies of nodal injection currents on voltages.
For example, the shortcomings of the ZIP model in capturing
the high sensitivity of reactive powers on voltages is known
[31]. In such cases, more general models, such as ones where
the active and reactive powers are polynomial or exponential
functions of the voltages, can be used [32, Tables 1.1–1.4].
The ZIP load model presented in this section is in line with
traditional distribution system analysis textbooks such as [23,
Chapter 9]. However, the dependency of nodal injection currents in on voltages vn is technically imposed by various individual loads (e.g., residential electrical appliances) aggregated
at node n. Even in the case of steady-state power flow studies,
determining an accurate in (vn ) requires extensive surveys. In
such setups, the ZIP load model can be incorporated to obtain
more flexible but yet more complicated models for residential
loads; see e.g., [33]. When such complicated load models are
utilized, the relationships provided in Table I can be updated
accordingly. These modifications will not affect the ensuing
discussions on Y-Bus modeling.
Each edge (n, m) represents a series element and is modeled
by the following two equations:
(n)
(m)
inm = Ynm
vn − Ynm
vm
(4a)
(m)
(n)
vn .
imn = Ymn
vm − Ymn
(4b)
(m)
(b)
Fig. 1. (a) Three-phase transmission line. (b) Step-voltage regulator in series
with a transmission line.
For typical multi-phase transmission lines, the series
impedance matrix Znm is symmetric and its real part Re[Znm ]
is positive definite. Based on this property, Lemma 1 proves
−1
that Z−1
nm exists and that Re[Znm ] is also positive definite.
s
The shunt admittance matrix Ynm
is also symmetric and
s
its real part Re[Ynm ] is positive semi-definite. This specific
structure will be shown to play a crucial role in guaranteeing
the invertibility of a network of three-phase transmission lines.
This property will be revisited in Section VI.
B. Step-voltage regulators
IV. M ODELING OF S ERIES E LEMENTS
(n)
(a)
(m)
Matrices Ynm ∈ C|Ωn |×|Ωn | , Ynm ∈ C|Ωn |×|Ωm | , Ymn ∈
(n)
C|Ωm |×|Ωm | , and Ymn ∈ C|Ωm |×|Ωn | are determined based
on the model of the series element while setting to zero those
rows and columns that correspond to missing phases. Matrices
(n)
(m)
(m)
(n)
Ynm and Ymn (respectively, Ynm and Ymn ) are called selfadmittances (mutual admittances).
A. Transmission lines
In general, we have that Ωnm ⊆ Ωn for a transmission line
on edge (n, m), and it is possible to have that |Ωnm | < |Ωn |.
As an example, in the IEEE 13-bus distribution feeder, the
three-phase node 671 is connected to the two-phase node 684
through a two-phase distribution line (config. 604). The πmodel for transmission lines depicted in Fig. 1(a) yields [23]
1 s
−1
inm (Ωnm ) =
Ynm + Z−1
nm vn (Ωnm ) − Znm vm (Ωnm ) (5)
2
where notations inm (Ωnm ) and vn (Ωnm ) pick the indices
of inm and vn that correspond to the phases in Ωnm . For
example, if vn = [vna , vnb , vnc ]T and Ωnm = {b, c}, then
vn (Ωnm ) = [vnb , vnc ]T . Matrix Znm = Zmn ∈ C|Ωnm |×|Ωnm |
s
s
is the series impedance matrix and matrix Ynm
= Ymn
∈
|Ωnm |×|Ωnm |
C
is the shunt admittance matrix of the line.
Comparing (5) with (4) yields
1 s
(n)
(m)
+ Z−1
Ynm
(Ωnm , Ωnm ) = Ymn
(Ωnm , Ωnm )= Ynm
nm (6a)
2
(m)
(n)
−1
Ynm (Ωnm , Ωnm ) = Ymn (Ωnm , Ωnm ) =Znm
(6b)
where the notation Y(Ωnm , Ωnm ) selects rows and columns
of Y corresponding to existing phases in Ωnm .
The next series element to be modeled is the SVR, a device
which is installed either at the substation or along the feeder
to keep nodal voltages within acceptable ranges. In essence,
the SVR is a connection of auto-transformers with adjustable
turns ratios. The turns ratios are dependent on the position
of its taps. The taps are determined through a control circuit
that approximates the voltage drop from the regulator node
to the node whose voltage is to be controlled. In distribution
networks, three-phase SVRs are commonly installed in wye,
closed-delta, or open-delta configurations [23].
The installation of an SVR is in series with a transmission
line similar to Fig. 1(b). The SVR is between nodes n and n′
and the transmission line is between nodes n′ and m. Other
than edges (n′ , n) and (n′ , m), no other edges are connected
to node n′ . With the regulator taps already determined, an
equivalent model for the SVR as a single block between nodes
n and m can be derived. This equivalent model eliminates
node n′ and creates a virtual edge (n, m) in place of (n, n′ ).
Voltages and currents at the primary and the secondary of
a SVR are generally related via
vn = Av vn′ + ZR inn′ , inn′ = Ai in′ m ,
(7)
where matrices Av , Ai , ZR ∈ CΩn ×Ωn will be referred to
as voltage gain, current gain, and impedance matrix of the
SVR respectively. The specific entries of the aforementioned
matrices Av , Ai , and ZR are determined by the configuration
type (wye, closed-delta, or open-delta) as well as the selected
tap positions. A realistic assumption is that Ωn = Ωn′ = Ωnn′ ,
which means the primary and secondary of the SVR have
the same available phases. This assumption holds for realistic
feeders included in the numerical tests of Section VIII.
It will be shown that, for the three common types of SVRs,
matrix ZR is diagonal, matrix Av of (7) is invertible, and the
following property holds:
T
A−1
v = Ai .
(8)
4
Property (8) allows one to obtain nodal admittance models
of SVRs that ultimately guarantee the invertibility of the Y-Bus
in Section VI. Nodal admittance models of SVRs are derived
T
next. From (7), by understanding that A−1
v = Ai , we obtain:
vn′ = ATi (vn − ZR Ai in′ m ) .
(9)
Moreover, the model of (4) for the transmission line (n′ , m)
gives the following for the currents in′ m and imn′ :
′
(m)
(n )
in′ m = Yn′ m vn′ − Yn′ m vm
imn′ =
(m)
Ymn′ vm
−
(10a)
(n′ )
Ymn′ vn′ .
(10b)
Using (9) in (10a) yields
(n′ )
(m)
in′ m = Yn′ m ATi (vn − ZR Ai in′ m ) − Yn′ m vm . (11)
By reorganizing (11), we obtain
(n′ )
(m)
FR in′ m = Yn′ m ATi vn − Yn′ m vm
(12)
where
(n′ )
FR = I|Ωn | + Yn′ m ATi ZR Ai
(13)
and I|Ωn | is the identity matrix in C|Ωn |×|Ωn | . In Lemma 4
we show that FR is invertible so that in′ m is computed as
′
(m)
(n )
−1
T
in′ m = F−1
R Yn′ m Ai vn − FR Yn′ m vm .
(14)
Using the equation for the current in (7), together with (14),
we write inm := inn′ as a function of vn and vm :
inm := inn′ = Ai in′ m
(n′ )
(m)
−1
T
= Ai F−1
R Yn′ m Ai vn − Ai FR Yn′ m vm .
(15)
To write imn in terms of vm and vn , we first use (9) in (10b)
to obtain
(n′ )
(m)
imn := imn′ = Ymn′ vm − Ymn′ vn′
(m)
(n′ )
(m)
(n′ )
= Ymn′ vm − Ymn′ ATi (vn − ZR Ai in′ m )
= Ymn′ vm − Ymn′ ATi vn
(n′ )
+ Ymn′ ATi ZR Ai in′ m .
(16)
Next, we replace in′ m in (16) by its equivalent in (14):
(m)
(n′ )
(m)
imn = Ymn′ − Ymn′ ATi ZR Ai F−1
R Yn′ m vm
(n′ )
(n′ )
T
(17)
Y
− Ymn′ I|Ωn | − ATi ZR Ai F−1
′
n m Ai vn .
R
Fig. 2. Model of a single-phase auto-transformer. For brevity, the series
impedance is not drawn.
rendering ZR = O and FR = I|Ωn | . Thus, we can obtain the
following matrices for SVRs with ideal auto-transformers:
(n′ )
(m)
(m)
(n)
=Ai Yn′ m ,
Ynm
=Ai Yn′ m ATi , Ynm
(n′ )
(n)
Ymn
=Ymn′ ATi .
(m)
(m)
Ymn
=Ymn′ ,
(19a)
(19b)
Notice in (18) and (19) that the matrix sizes conform due to
the initial assumption that Ωn = Ωn′ .
To derive the gain matrices Av and Ai as well as the
impedance matrix ZR for the three SVR configurations (wye,
closed-delta, and open-delta), it is essential to grasp the basic
model of single-phase auto-transformers. A diagram of a
single-phase auto-transformer is given in Fig. 2.1 Given the
taps, the effective regulator ratio aR is determined as follows:
aR = 1 ∓ 0.00625tap.
(20)
Having determined aR , the voltages and currents on the two
sides of the auto-transformer relate as follows:
vS − vN = aR (vL − vN ) + zR iS
1
iS = − iL
aR
(21a)
(21b)
where zR is the series impedance of the auto-transformer.
The voltage and current relationships in (21) are leveraged
to derive the gain matrices Av and Ai for three types of
SVR connections. Table II summarizes the resulting formulas.
Notice Av is invertible and that A−1
= ATi holds for all
v
SVRs; a crucial property for guaranteeing Y-Bus invertibility.
1) Wye-connected SVRs: The diagram of a wye-connected
SVR is given in Fig. 3(a). Using (21a) for the three autotransformers in Fig. 3(a) yields the following equations:
vna = aRa vna′ + zRa iann′
(22a)
vnb
vnc
(22b)
(22c)
=
=
aRb vnb ′
aRc vnc ′
+
+
zRb ibnn′
zRc icnn′
(18d)
where aRa , aRb , and aRc are the effective regulator ratios
determined by the corresponding taps at phases a, b, and c
following (20). Similarly, zRa , zRb , and zRc are per phase
series impedances of each SVR. From (22), one obtains the
SVR voltage gain Av and impedance matrix ZR :
0
0
a Ra
0
0
z Ra
a Rb
0 , ZR = 0
z Rb
0 . (23)
Av = 0
0
0
a Rc
0
0 z Rc
where in (18d) we have used the identity
= I|Ωn | −
′
−1 (n )
T
Ai ZR Ai FR Yn′ m provided from Lemma 4.
In most cases, it turns out that the per unit series impedance
of SVRs significantly depends on the tap position and is zero,
1 Depending on the connection of the shunt winding of auto-transformers,
SVRs can either be of type A or type B. Since type B SVRs are more common
in distribution networks [23], the regulator models developed correspond only
to this type of SVRs. The same procedure can be followed through to obtain
models for type A SVRs. This has been done for the 8500-node feeder of
Section VIII.
Equations (15) and (17) conform to the model of (4) and
yield the following matrices for the SVRs:
(n′ )
−1
(n)
Ynm
= Ai FR
Yn′ m ATi
(m)
Ynm
(m)
Ymn
(n)
Ymn
=
=
=
(18a)
(m)
Ai F−1
R Yn′ m
(m)
(n′ )
(m)
Ymn′ − Ymn′ ATi ZR Ai F−1
R Yn′ m
(n′ )
T
Ymn′ F−T
R Ai
F−T
R
(18b)
(18c)
5
TABLE II
V OLTAGE GAIN , CURRENT GAIN , AND IMPEDANCE MATRICES FOR THE THREE COMMON CONFIGURATIONS OF STEP - VOLTAGE REGULATORS
SVR connection
Voltage gain Av
Current gain Ai
Impedance matrix ZR
wye-connected
0
aRa
0
0
0
aRb
0
1
a Ra
b
0
0
z Rb
0
0
1
aR
0
aRc
0
0
0
zR
a
0
0
closed-delta
0
1
a Rc
0
0
z Rc
0
aRbc
0
1 − aRca
0
a
Rab
1 − aRab
aRbc
0
1 − aRbc
0
zR
ab
0
zRbc
0
0
(a)
Fig. 3.
0
1 − aRbc
aRca
−1
1 − aRca
0
aRca
0
0
zRca
aR
ab
0
0
1
0
0
ab
1
aR
1
1−
ab
0
zR
ab
0
0
0
0
0
0
0
aRcb
1 − aRcb
1
aR
1 −
0
1 − aRab
(b)
aR
cb
1
1
aR
0
cb
0
zRcb
(c)
Three types of step-voltage regulators; (a) wye-connected (b) closed-delta(c) open-delta. For brevity, the series impedances are not drawn.
From (21b), we obtain
iann′
1 − aRab
aRab
open-delta
=−
1
a Ra
ian′ n , ibnn′
1 c
1 b
in′ n , icnn′ = −
i ′ . (24)
=−
a Rb
a Rc n n
Using (24) and recognizing that in′ n = −in′ m holds in (7),
we compute Ai
1
0
0
aR a
1
0
Ai = 0
(25)
.
aR b
1
0
0
aR
c
2) Closed-delta connected SVRs: The diagram of closeddelta connected SVR is drawn in Fig. 3(b). Leveraging (21)
for the three auto-transformers between phases ab, bc and ca
at node n′ , the voltage equations are
vna
vnb
vnc
− vnb ′
⇒ vna
− vnc ′
⇒ vnb
− vna′
⇒ vnc
=
=
=
=
=
=
aRab (vna′ − vnb ′ ) + zRab iann′
aRab vna′ + (1 − aRab )vnb ′ + zRab iann′
aRbc (vnb ′ − vnc ′ ) + zRbc ibnn′
aRbc vnb ′ + (1 − aRbc )vnc ′ + zRbc ibnn′
aRca (vnc ′ − vna′ ) + zRca icnn′
aRca vnc ′ + (1 − aRca )vna′ + zRca icnn′
(26a)
(26b)
(26c)
where aRab ,aRbc , and aRca are the effective regulator ratios
and zRab , zRbc , and zRca are series impedances on phases ab,
bc, and ca at node n′ respectively. From (26), we find
aRab
1 − aRab
0
0
aRbc
1 − aRbc
Av =
(27a)
0
aRca
1 − aRca
zRab
ZR = 0
0
0
0 .
0
zRbc
0
(27b)
zRca
Writing KCL for node n′ at phases a, b, c and considering the
current equation in (21b) for auto-transformers yields:
ian′ n = −aRab iann′ − (1 − aRca )icnn′
(28a)
ibn′ n
icn′ n
(28b)
(28c)
=
=
−aRbc ibnn′
−aRca icnn′
− (1 − aRab )iann′
− (1 − aRbc )ibnn′ .
From (28) and using in′ n = −in′ m in (7), we find
aRab
Ai = 1 − aRab
0
0
aRbc
1 − aRbc
−1
1 − aRca
.
0
aRca
(29)
3) Open-delta connected SVRs: The circuit diagram of the
open-delta SVR is given in Fig. 3(c). In Fig. 3(c), using the
voltage equations in (21) and noticing that vnb = vnb ′ (phase b
of node n is connected to phase b of node n′ ) yields
vna − vnb ′ = aRab (vna′ − vnb ′ ) + zRab iann′
⇒ vna = aRab vna′ + (1 − aRab )vnb ′ + zRab iann′ (30a)
vnc − vnb ′ = aRcb (vnc ′ − vnb ′ ) + zRcb icnn′
⇒ vnc = aRcb vnc ′ + (1 − aRcb )vnb ′ + zRcb icnn′ (30b)
where aRab and aRcb are the effective regulator ratios and zRab
and zRcb are the series impedances on phases ab and cb of node
6
n. Using (30) and the equality vnb = vnb ′ , the gain matrix Av ,
and the impedance matrix ZR are obtained as follows
aRab 1 − aRab
0
1
0
Av = 0
(31a)
0
1 − aRcb aRcb
zRab 0
0
0
0 .
ZR = 0
(31b)
0
0 zRcb
For the two SVRs in Fig. 3(c), from (21b) we obtain
iann′ = −
1
aRab
ian′ n , icnn′ = −
1
aRcb
icn′ n .
(32)
Applying KCL at node K of Fig. 3(c) and subsequently
introducing (32) yields ibnn′ as follows:
ibnn′ = −iann′ − icnn′ − ian′ n − ibn′ n − icn′ n
1
1
− 1 ian′ n − ibn′ n −
− 1 icn′ n . (33)
=
aRab
aRcb
Combining in′ n = −in′ m with (32) and (33) yields Ai :
1
0
0
aRab
1
1
(34)
Ai = 1 − aRab 1 1 − aRcb .
1
0
0
aR
cb
C. Three-phase transformers
In distribution systems, a three-phase transformer can be
appropriately represented by two blocks, namely, a series
block representing the per unit leakage admittance, and a
shunt block modeling transformer core losses. Transformer
core losses can technically be treated as voltage-dependent
loads similar to the ones in Table I, with functions of v given
in [29]; notice though that this information may not be readily
available for many distribution feeders. The most common distribution transformers include delta–wye grounded, wye–delta,
wye–wye, open-wye–open-delta, delta–delta, and open-delta–
open-delta for which the ABCD parameters are listed in [23].
The corresponding nodal admittances of these connections are
given in [29] and [34] and are listed in Table III where
TABLE III
A DMITTANCE MATRICES OF THE MOST COMMON TRANSFORMER
CONNECTIONS IN DISTRIBUTION SYSTEMS .
Transformer connection
1
Y3 = √
3
y
Y5 = t
0
0
yt
0
0
1 2yt −yt −yt
0 , Y2 = −yt 2yt −yt
3 −y −y
2yt
yt
t
t
−yt
yt
0
y
−yt
t
1
0
−yt
yt , Y4 = −yt 2yt
3
yt
0
−yt
0
−yt
1 −yt
0
yt
0
, Y6 = √
.
yt
−yt yt
3 0
(35a)
(m)
(m)
(n)
Node m
Ynm
Ynm
Ymn
Ymn
Wye-G
Wye-G
Y1
Y1
Y1
Y1
Wye-G
Wye
Y2
Y2
Y2
Y2
Wye-G
Delta
Y1
−Y3
Y2
−Y3T
Wye
Wye
Y2
Y2
Y2
Y2
Wye
Delta
Y2
−Y3
Y2
−Y3T
Y2
Delta
Delta
Y2
Y2
Y2
Open-Delta
Open-Delta
Y4
Y4
Y4
Y4
Open-Wye
Open-Delta
Y5
−Y6
Y4
−Y6T
V. Y-B US C ONSTRUCTION
Z-B US M ETHOD
AND THE
The application of KCL at each node n ∈ N leads to
X
inm , n ∈ N .
(36)
in (vn ) =
m∈Nn
Leveraging the series element model in (4), we arrive at the
multidimensional Ohm’s law
v
i
,
(37)
= Ynet
vS
iS
where vS is the voltage at the slack bus and likewise iS is the
current injection at the slack bus. Typically, vS is equal to the
symmetrical voltage {1, 1 −120◦, 1 120◦ }. The matrix Ynet
can be constructed by block matrices leveraging the series
model of transmission lines, SVRs, and transformers given in
(4) as follows:
X
(n)
Ynet (Jn , Jn ) =
Ynm
, n∈N
(38a)
m∈Nn
(m)
Ynet (Jn , Jm ) = −Ynm
,
Partitioning Ynet yields
Y
i
=
YSN
iS
where Y is given by
X
(n)
Y(Jn , Jn ) =
Ynm
,
Y(Jn , Jm ) =
yt
Y1 = 0
0
Matrices
(n)
Node n
m∈Nn
(m)
,
−Ynm
Y(Jn , Jm ) = O,
m ∈ Nn .
YNS
YSS
v
vS
n ∈ N \ {S}
(38b)
(39)
(40a)
n ∈ N \ {S}, m ∈ Nn \ {S} (40b)
n ∈ N \ {S}, m ∈
/ Nn ∪ {S}.
(40c)
The matrix Y defined in (40) is called the bus admittance
matrix of the network, and is also referred to as the Y-Bus.
0
−yt (35b) Matrix Y is used for power flow analysis as we explain next.
yt
From (39), it follows that i(v) = Yv + YNS vS , where the
(35c)
and yt is the per unit leakage admittance. In the next section,
the bus admittance matrix Y-Bus is formulated. It turns out
that rank-deficiency of matrices other than Y1 in (35) are the
only source of Y-Bus singularity. A method is presented to
remedy this issue.
dependence of i on v is made explicit. Due to the ZIP load
model we can decompose i into three parts as follows:
iPQ (v) + iI (v) + iZ (v) = Yv + YNS vS .
(41)
From (2), the constant-impedance currents is iZ (v) = −YL v,
where YL ∈ C J×J is a block diagonal matrix with entries
YL (Jn , Jn ) = YLn ,
n ∈ N \ {S}.
(42)
7
Using iZ (v) = −YL v in (41) and rearranging yields
iPQ (v) + iI (v) = (Y + YL )v + YNS vS .
TABLE IV
A DMITTANCE MATRICES OF SERIES ELEMENTS †
(43)
Edges
If Y + YL is invertible, (43) yields a fixed-point equation for
voltages, rendering the standard Z-Bus method as follows
v[t + 1] = Z [iPQ (v[t]) + iI (v[t])] + w,
(45a)
(45b)
where ǫ′ = ǫ′r − jǫ′i with ǫ′r > 0 is a small shunt admittance
(compared to |yt |). If matrices Y2 and Y4 appear in the mutual
admittances, then the modification is as follows:
Y2′′ = Y2 + ǫ′′ I,
(46a)
Y4′′
(46b)
′′
= Y4 + ǫ I,
where ǫ′′ = ǫ′′r − jǫ′′i with 0 ≤ ǫ′′r < ǫ′r . Doing so yields Ŷ, a
slightly modified version of Y. 2 In the next section, we prove
the invertibility of the matrices Ŷ and Ŷ + YL . The voltages
can then be computed by using Z = (Ŷ + YL )−1 in (44).
A similar approach of modifying the transformer nodal
admittances to handle the delta-side invertibility problem is
employed in OpenDSS, where by default, a high reactance is
connected from the delta terminal to the ground [36]–[38].
VI. I NVERTIBILITY
E1
E2
(44)
where Z = (Y + YL )−1 , w = −ZYNS vS , and t is the
iteration index. The Z-Bus method constitutes initializing v[0]
typically to a flat voltage profile (i.e., vn [0] := vS ) and
running iteration (44). The solution in terms of voltages yields
all power flows in the network. The name of the method is
derived from the inversion of Y + YL which yields Z. In
practice, Y + YL is not inverted, but its LU decomposition is
computed which is unique if Y + YL is invertible.
It is well-known that Y is not always invertible due to
transformer connections other than wye-g–wye-g. For all other
transformer connections, matrices Y2 , Y3 , Y4 , and Y6 in
Table III are rank-deficient, which may lead to a non-invertible
Y, depending on the position of the transformer in the feeder
with respect to other elements. The non-invertibility of Y can
carry over to Y + YL as well. This is indeed the case for all
the test feeders we investigate in Section VIII.
In order to numerically remedy the non-invertibility of Y
for transformers other than wye-g–wye-g, [21] suggests adding
a small shunt admittance from the isolated transformer sides to
the ground. Mathematically, this can be achieved by replacing
Y2 and Y4 appearing in the self-admittances of Table III by
Y2′ = Y2 + ǫ′ I,
Y4′ = Y4 + ǫ′ I,
ETL
OF THE BUS ADMITTANCE MATRIX
Y
This section proves the invertibility of Y, where to avoid
exhaustive notation, the proof is based first on the following
assumptions. The notation S ≻ O (S O) means matrix S
is symmetric and positive (semi-)definite.
2 This modification is only intended to improve the numerical properties of
the matrix Y in order to obtain steady-state power flow solutions. This is
not to be confused with standard earthing practices through connection of a
grounding transformer such as wye-g–delta or zig-zag [35].
EYg∆
EY∆
EO∆
ER
EOY∆
†
(n)
Ynm
s
1
Ynm
2
−1
+Znm
Y1
Y2 + ǫ′ I
Y1
Y2 + ǫ′ I
Y4 + ǫ′ I
Ai 21 Yns ′ m
T
+Z−1
n′ m Ai
Y5
(m)
Ynm
Z−1
nm
Y1
Y2 + ǫ′′ I
−Y3
−Y3
Y4 + ǫ′′ I
Ai Z−1
n′ m
−Y6
(m)
Ymn
s
1
Ynm
2
−1
+Znm
Y1
Y2 + ǫ′ I
Y2 + ǫ′ I
Y2 + ǫ′ I
Y4 + ǫ′ I
1
Yns ′ m
2
+Z−1
n′ m
Y4 + ǫ′ I
(n)
Ymn
Z−1
nm
Y1
Y2 + ǫ′′ I
−Y3T
−Y3T
Y4 + ǫ′′ I
Z−1
AT
n′ m i
−Y6T
In case of missing phases, admittance matrices select Ωnm , cf. (6).
A1) The series elements belong to the following seven
categories (nodal admittances are summarized in Table IV):
(a) The set ETL of transmission lines with nodal admittances given by (6). Two realistic properties verified in
practical distribution feeders are assumed:
s
(i) Matrices Znm and Ynm
are symmetric.
−1
s
(ii) It holds that Re[Znm ] ≻ O and Re[Ynm
] O.
(b) The set E1 of wye-g–wye-g transformers with nodal
admittances given in Table III.
(c) The set E2 of wye-g–wye, wye–wye, and delta–delta
transformers, where in self-admittances the matrix Y2′
in (45a) replaces Y2 while in mutual admittances, the
matrix Y2′′ in (46a) replaces Y2 .
(d) The set EYg∆ of wye-g–delta transformers where Y2′
in (45a) is used in place of Y2 .
(e) The set EY∆ of wye–delta transformers where Y2′
in (45a) is used in place of Y2 .
(f) The set EO∆ of open-delta–open-delta transformers
where in self-admittances, the matrix Y4′ in (45b)
replaces Y4 while in mutual admittances, the matrix
Y4′′ in (46b) replaces Y4 .
(g) The set ER of ideal SVRs with nodal admittances given
in (19) and gain matrices Av and Ai in Table II. Notice
T
that A−1
v = Ai holds, and A1(a) is assumed for line
′
(n , m).
Notice that E = ETL ∪ E1 ∪ E2 ∪ EYg∆ ∪ EY∆ ∪ EO∆ ∪ ER .
A2) The graph (N , E) is connected.
A3) For all transformers it holds that Re[yt ] > 0, i.e., the
transformers are non-ideal.
A4) Nodes and edges have three available phases, that is
Ωn = Ωnm = Ωm = {a, b, c} for all n ∈ N and
(n, m) ∈ E. The extension to the case that includes
transmission lines with missing phases and open-wye–
open-delta transformers is provided afterwards.
A few comments are in order here regarding Assumption
s
A1(a). The symmetricity of Znm and Ynm
holds from electromagnetic considerations, that is, the coupling between phases
a and b is the same as the coupling between b and a. Interestingly, we verified that matrix Re[Znm ] is typically diagonally
8
dominant with positive diagonal entries for all distribution
feeders. Hence, it holds that Re[Znm ] ≻ O, and due to Lemma
−1
1 in Appendix A, Z−1
nm exists, and Re[Znm ] ≻ O. Assumption
A1(a)(ii) is though weaker than diagonal dominance. It is
worth emphasizing that A1(a) allows for untransposed lines.
The invertibility proof relies on Lemma 1 in Appendix A.
To use this lemma, it is first proved that Y is symmetric
(Theorem 1) and that Re[Y] ≻ O (Theorem 2).
Theorem 1: Under A1, Y defined in (40) is symmetric.
(n)
(n)
Proof: It suffices to prove that Ynm = (Ynm )T and
(m)
(n) T
Ynm = (Ymn ) for all n ∈ N \ {S} and m ∈ Nn \ {S}
[cf. (40)] where the admittance matrices are given in Table IV.
The relationships hold for (n, m) ∈ ETL due to A1(a)(i); for
(n, m) ∈ E1 ∪ E2 ∪ EYg∆ ∪ EY∆ ∪ EO∆ due to (35); and for
T
(n, m) ∈ ER due to A1(a)(i) and using A−1
v = Ai .
Let ES collect the edges connected to the slack bus. Such an
edge could be a transmission line, a voltage regulator, or a
three-phase transformer according to Table IV.
Theorem 2: Under A1–A4, it holds that Re[Y] ≻ O.
Proof: The goal is to prove that xT Re[Y]x > 0 for
any nonzero real vector x ∈ RJ . Since x is real, notice that
xT Re[Y]x = Re[xT Yx]. By the definition of Y in (40), and
defining xn as a vector that selects indices of x corresponding
to Jn , xT Re[Y]x is written as follows:
X
xTn
X
xTn Re[YnS ]xn +
X
X
xTn Re[YnS ]xn +
T
x Re[Y]x = Re
n∈N \{S}
=
(n)
X
(n)
Ynm
xn
m∈Nn
−
X
(m)
Ynm
xm
m∈Nn \{S}
h
i
(n)
(m)
Re xTn Ynm
xn − xTn Ynm
xm
n∈N \{S}
m∈Nn \{S}
n∈NS
=
(n)
n∈NS
X
h
(n)
(m)
Re xTn Ynm
xn − xTn Ynm
xm
(n,m)∈E\ES
i
(m)
(n)
+ xTm Ymn
xm − xTm Ymn
xn .
(47)
In (47), replace the second summation using Lemmas 2 and 3
in Appendix A as follows. For ETL \ ES , apply Lemma 2
−1
with u = xn , v = xm , and W = Re[Znm
]. For E1 \ ES ,
apply Lemma 2 with u = xn , v = xm , W = Re[Y1 ]. For
E2 \ ES , apply Lemma 2 with u = xn , v = xm , and W =
Re[Y2 ] + ǫ′′r I. For EYg∆ \ ES apply Lemma 3 with u = xn ,
v = xm , and W = Re[Y1 ], X = −Re[Y3 ], Z = Re[Y2 ].
For EY∆ \ ES , apply Lemma 3 with u = xn , v = xm , and
W = Z = Re[Y2 ], X = −Re[Y3 ]. For E4 \ ES , apply Lemma
2 with u = xn , v = xm , and W = Re[Y4 ] + ǫ′′r I. For
ER \ ES , apply Lemma 2 with u = ATi xn , v = xm , and
W = Re[Z−1
n′ m ]. After the previous replacements, we obtain
xT Re[Y]x =
X
(n)
xTn Re[YnS ]xn
n∈NS
+
X
(xn − xm )T Re[Z−1
nm ](xn − xm )
(n,m)∈ETL \ES
1
1
s
s
]xTn + xTm Re[Ynm
]xTm
+ xTn Re[Ynm
2
2
X
+
(xn − xm )T Re[Y1 ](xn − xm )
(n,m)∈E1 \ES
+
X
(xn − xm )T (Re[Y2 ] + ǫ′′r I)(xn − xm )
(n,m)∈E2 \ES
+ (ǫ′r − ǫ′′r )(xTn xn + xTm xm )
X
xn
+
+ ǫ′r xTm xm
xTn xTm GYg∆nm x
m
(n,m)∈EYg∆ \ES
X
xn
+
xTn xTm GY∆nm x
m
(n,m)∈EY∆ \ES
+ ǫ′ xT xn + ǫ′r xTm xm
Xr n
+
(xn − xm )T (Re[Y4 ] + ǫ′′r I)(xn − xm )
(n,m)∈E4 \ES
+ (ǫ′r − ǫ′′r )(xTn xn + xTm xm )
Xn T
T
(Ai xn − xm )T Re[Z−1
+
n′ m ](Ai xn − xm )
(n,m)∈ER \ES
+
o
1 T
1
xn Ai Re[Yns ′ m ]ATi xn + xTm Re[Yns ′ m ]xm (48)
2
2
h
i
Re[Y ] Re[Y ]
where we set GYg∆nm = Re[YT1 ] Re[Y23 ] , and GY∆nm =
3
h
i
Re[Y2 ] Re[Y3 ]
, GY∆nm O since
.
It
holds
that
G
T
Yg∆
nm
Re[Y3 ] Re[Y2 ]
their nonzero eigenvalues are respectively {Re[yt ], 2Re[yt ]},
and {2Re[yt ]}; and Re[yt ] > 0 for all transformers. Furthermore, Re[Z−1
nm ] ≻ O for transmission lines, and it holds that
Re[Y1 ] ≻ O, Re[Y2 ] O, Re[Y4 ] O, and ǫ′r > ǫ′′r .
(n)
Moreover, YnS is the self-admittance of any of the edges in
(n)
Table IV, and therefore, Re[YnS ] ≻ O holds. The expression
in (48) is thus nonnegative for any real vector x. This proves
that Re[Y] O. At this point, to prove positive definiteness,
we show that whenever xT Re[Y]x = 0 holds, then x = 0.
In particular, if xT Re[Y]x = 0, then every quadratic form in
(48) must be zero and we have the following implications:
(n)
(R1) Re[YnS ] ≻ O ⇒ xn = 0, n ∈ NS .
−1
(R2) Re[Znm ] ≻ O ⇒ xn = xm , (n, m) ∈ ETL \ ES .
(R3) Re[Y1 ] ≻ O ⇒ xn = xm , (n, m) ∈ E1 \ ES .
(R4) ǫ′r − ǫ′′r > 0 ⇒ xn = xm = 0, (n, m) ∈ E2 \ ES .
(R5) ǫ′r > 0 ⇒ xm = 0; and Re[Y1 ] ≻ O ⇒ xn = 0,
(n, m) ∈ EYg∆ \ ES .
(R6) ǫ′r > 0 ⇒ xn = xm = 0, (n, m) ∈ EY∆ \ ES .
(R7) ǫ′r − ǫ′′r > 0 ⇒ xn = xm = 0, (n, m) ∈ E4 \ ES .
T
(R8) Re[Z−1
n′ m ] ≻ O ⇒ xm = Ai xn , (n, m) ∈ EYVR \ ES .
T
Since the graph is connected, x Re[Y]x = 0 implies that
xn = 0 for all n. We conclude that Re[Y] ≻ O.
Remark 2: The concluding part of the proof in Theorem 2
was inspired by a similar argument in [11] that proved the
invertibility of the single-phase Y-Bus.
Theorem 3: Under A1–A4, matrix Y is invertible.
Proof: Combine Theorems 1 and 2 with Lemma 1.
This section proved the invertibility of the Y-Bus under
Assumptions A1–A4. It is worth emphasizing that the proof
is modular, that is, the invertibility of the Y-Bus is shown
to be dependent only on the properties of block admittances
of the edges. In other words, regardless of the number and
order of elements, and of the topology (radial or meshed), the
Y-Bus is invertible, as long as the series elements adhere to
Assumptions A1–A4. The next section relaxes some of these
assumptions.
VII. E XTENSIONS
In this section, extensions to Theorem 3 that guarantee
invertibility of the Y-Bus under more practical considerations
9
are provided. Specifically, Section VII-A covers typical distribution feeders that consist of one-, two-, and three-phase
laterals and open-wye–open-delta transformers. The Z-Bus
method of (44) relies on the invertibility of Y + YL , which is
asserted in Section VII-B. Section VII-C provides a proof of
the Y-Bus invertibility using only the imaginary parts of ǫ′ and
ǫ′′ ; thereby allowing the inclusion of transformers with zero
ohmic losses. This is indeed useful for simplistic models where
transformers are only represented by their series reactances.
Finally, Section VII-D considers the Y-Bus invertibility under
alternative practical transformer connections.
A. Invertibility under transmission lines with missing phases
and open-wye–open-delta transformers
The ensuing Theorem 4 handles circuits that include transmission lines with missing phases and open-wye–open-delta
transformers. To this end, Assumptions A1 and A4 are
modified as follows:
A1′ ) The series elements in Assumption (A1) could also
include the set EOY∆ as the set of open-wye–open-delta
transformers where Y4′ in (45b) is used in place of Y4 .
A4′ ) For any node n ∈ N and phase φ ∈ Ωn , a path exists
from S to n where edges have φ in their phase set.
Typical distribution systems comprise a main three-phase
feeder with one-, two-, or three-phase laterals, and therefore,
A4′ is satisfied in practice.
Theorem 4: The conclusion of Theorem 2 holds under A1′ ,
A2, A3, and A4′ .
Proof: The only modification will occur in the term
(48) where for transmission lines (n, m) ∈ ETL and
SVRs (n, m) ∈ ER , we substitute xn and xm respectively with xn (Ωnm ) and xm (Ωnm ). For open-wye–open-delta
transformers(n, m) ∈ EOY∆ we substitute xn with xn (Ωnm ).
Moreover, we include the following term for open-wye–opendelta transformers(n, m) ∈ EOY∆ in (48)
X
xn (Ωnm )T
(n,m)∈EOY∆ \ES
x (Ω )
xTm GOY∆nm n nm ,
xm
B. Invertibility of Y + YL
Prior to presenting that Y + YL is invertible, the ensuing
Theorem 5 is first presented that brings out some useful
properties of YL .
Theorem 5: Matrix YL defined in (42) is symmetric, and it
holds that Re[YL ] O.
Proof: Since YL is block diagonal [cf. (42)], it suffices
to prove that YLn in (3) is symmetric and that Re[YLn ] O
for n ∈ N \ {S}. For wye loads, YLn is diagonal due to (3a);
φ
and since Re[yL
] ≥ 0, it holds that Re[YLn ] O. For delta
n
loads, symmetricity of YLn follows from (3c); also Re[YLn ]
is symmetric diagonally dominant with nonnegative diagonal
entries, and thus Re[YLn ] O.
Theorem 6: The matrix Y + YL with Y defined in (40) and
YL defined in (42) is invertible under A1′ , A3, A2, A4′ .
Proof: Let Y + YL = G + jB. Theorems 5, 1, and 2
or 4 prove that G = Re[Y + YL ] = Re[Y] + Re[YL ] ≻ O
and B = Im[Y + YL ] = Im[YT + YLT ] = BT . By Lemma 1,
matrix Y + YL is invertible.
The previous analysis reveals that transformers with Y2
or Y4 as their self-admittance compromise the invertibility
of Y + YL . The modification of Y2 or similarly Y4 in
(45) and (46) guarantees invertibility in Theorem 3. It is
alternatively possible to achieve invertibility of Y + YL
when YL automatically adds a constant-impedance load with
positive definite real part to a side of the transformer whose
self-admittance is Y2 or Y4 . This observation is summarized
in the following corollary.
Corollary 1: Matrix Y + YL is invertible if matrices Y2
and Y4 are not modified, but any transformer side n whose
self-admittance is modeled by Y2 or Y4 contains a constantimpedance load YLn with Re[YLn ] ≻ O.
(49)
C. Invertibility using the imaginary parts of ǫ′ and ǫ′′
h
i
Re[Y5 ]
Re[Y6 ]
where GOY∆nm =
. Notice that
Re[Y6T ] Re[Y4 ]+ǫ′r I
Re[Y5 ] = Re[yt ]I2 . Moreover, the non-zero eigen-values of
Re[Y4 ] are { 31 Re[yt ], Re[yt ]} and thus Re[Y4 ] O. Thus,
the Schur complement of GOY∆ is given by
Re[Y4 ] + ǫ′r I − Re[Y6T ]Re[Y5 ]−1 Re[Y6 ].
the slack bus S to node m where all edges must have φ in
their phase set. This path is denoted by P = {S, j, . . . , m}
where j ∈ NS . Based on (R1), xj ({φ}) = 0. Based on (R2′ ),
(R8′ ), (R9′ ), (R3)–(R7), we conclude that xm ({φ}) = 0.
(50)
Furthermore, the following identity holds: Re[Y6T ]Re[Y6 ] =
Re[yt ]Re[Y4 ]. Therefore, if ǫ′r > 0, the matrix Re[Y4 ]+ǫ′r I as
well as the Schur complement of GOY∆ calculated in (50) are
positive definite. This implies that GOY∆ is positive definite
if ǫ′r > 0. Thus, we conclude that
(R2′ ) For (n, m) ∈ ETL \ ES , xn (Ωnm ) = xm (Ωnm )
(R8′ ) For (n, m) ∈ ER \ ES , xm (Ωnm ) = ATi xn (Ωnm ).
(R9′ ) For (n, m) ∈ EOY∆ \ES , since GOY∆ ≻ O ⇒ xm = 0
and xn (Ωnm ) = 0.
Consider an edge (n, m) ∈ (ETL ∪ ER ∪ EOY∆ ) \ ES with
m∈
/ NS and φ ∈ Ωm . Due to A4′ , there exists a path from
In simplistic models of distribution networks, transformer
connections are represented by their series reactances, that
is, the transformer models assume zero ohmic losses. In this
section, we present an alternative to Theorem 3 that provides
a gateway to guarantee invertibility of the Y-Bus under such
transformer models. Concretely, recall that Theorem 3 relies
on Lemma 1. By applying Lemma 1 to matrix jY, a dual
result can be proved: Y−1 exists when Re[Y] = Re[Y]T
and Im[Y] ≺ O hold. The previous result can be used to
provide an alternative to Theorem 3. To this end, the following
modifications to Assumptions A1′ and A3 are considered:
A1′′ ) Instead of A1(a)(ii), it holds that Im[Z−1
nm ] ≺ O and
that shunt admittances of the lines are negligible, that is,
s
kYnm
k = 0. In A1(c)–A1(f) and A1′ , the modification of
Y2 or Y4 on self-admittances should be such that ǫ′i > 0
(added admittance is reactive). In A1(c) and A1(f), the
modification of Y2 or Y4 on mutual admittances should
be such that 0 ≤ ǫ′′i < ǫ′i .
10
A3′′ ) Instead of A3, assume that Im[yt ] < 0, that is, transformers have non-zero leakage inductance.
The ensuing Theorem 7 guarantees invertibility of the Y-Bus
in distribution circuits where transformers are represented only
by their series reactances.
Theorem 7: Under A1′′ , A2, A3′′ , A4′ , Y is invertible.
Proof: The proof is similar to that of Theorem 3, but use
the imaginary operator Im instead of Re, and replace positive
definite arguments with negative definite arguments.
D. Alternative transformer models
Nodal admittances provided in Table III hold true for a bank
of three single-phase transformers. If no information other than
the series admittances for the transformer is available, then the
matrices in Table III provide a reasonable model in unbalanced
distribution system studies. This is indeed the case for the
feeders we tested.
However, transformers with common-core or shell-type
structures are also used in distribution systems. For such
transformer connections, by conducting extensive short-circuit
measurements, we can first obtain primitive admittance parameters and then convert them to nodal admittances of the form
(4) by using connection matrices [22]. The proof of Section
VI provides a template that shows how each of the block
admittances of (4) affects the Y-Bus invertibility and can be
modified accordingly to the individual transformer model.
As an example, consider the model of a three-legged coretype transformer connected in wye-g–delta given in [22, eq.
(40)]. The only difference between that model and the wyeg–delta of Table III is that the corresponding Y1 for the coretype transformer is not diagonal. However, it is mentioned in
[22] that the off-diagonal elements are considerably smaller in
magnitude. Thus, if Re[Y1 ] is diagonally dominant, and hence
positive definite, the wye-g side of the transformer does not
compromise the Y-Bus invertibility. Of course, matrix Y2 on
the delta side, still needs to be modified as per (45a).
VIII. N UMERICAL
RESULTS
In this section, we put together the previously developed
models to build the Y-Bus matrix and run the Z-Bus power
flow for the following distribution feeders: (a) the IEEE
37-bus feeder, (b) the IEEE 123-bus feeder, (c) the 8500-node
feeder, (d) the European 906-bus low voltage feeder (ELV
906-bus feeder). The developed MATLAB scripts to conduct
the numerical tests are available at the following page:
https://github.com/hafezbazrafshan/
three-phase-modeling
For each feeder, the MATLAB script setupYBus imports
the feeder excel files, and builds the required matrices Ynet ,
Y, and YNS . The script setupLoads builds the required
load parameters for the network according to Table I. The
MATLAB script performZBus performs the iterations of
the Z-Bus method given in (44).
The required data for the feeders are obtained from [30].
In order to assess the accuracy of our proposed modeling
approach, computed load-flow voltage solutions are compared
0.7
0.025
0.6
0.02
0.5
0.4
0.015
0.3
0.01
0.2
0.005
0.1
0
0
(a)
(b)
Fig. 4. Maximum difference in computed voltage magnitudes obtained from
load-flow runs on (a) the IEEE 37-bus feeder and (b) the IEEE 123-bus feeder.
The computed voltage solutions in successive load-flow runs converge when
the value of ǫ/|yt | reaches below 10−5 and 10−3 for the IEEE 37-bus feeder
and the IEEE 123-bus feeder respectively.
with benchmark solutions. Specifically, for the IEEE 123-bus
and the 8500-node test feeders, the benchmark is obtained
from the feeder documents in [30].
The benchmark for the ELV 906-bus feeder is not provided
in [30], therefore we use OpenDSS, the EPRI distribution
system simulator [36], [37]. For the IEEE 37-bus feeder, the
benchmark is also obtained from OpenDSS. This is because
OpenDSS assumes a relatively high-impedance for the opendelta SVR and thereby allows us to verify the non-ideal SVR
models of (18) in Section IV-B.
The original data from [30] do not provide the series
impedance for any of the SVRs in any of the feeders. Unlike
the OpenDSS model of the IEEE 37-bus feeder, OpenDSS
models of IEEE 123-bus and the 8500-node feeders assume
very small series impedances for SVRs. For the IEEE 123-bus
and the 8500-node feeders, it will be demonstrated that using
the ideal SVR models in (19) does not affect the accuracy of
power flow solutions.
In the modeling of the delta–delta transformers in IEEE
37-bus and IEEE 123-bus feeders, we use nodal admittances
from Table IV. By setting ǫ′ = 2ǫ′′ = ǫ, we perform a series
of Z-Bus power flow runs with values of ǫ ranging from
10−2 |yt | to 10−10 |yt |. The maximum difference in voltage
magnitudes for successive values of ǫ is depicted in Fig. 4(a)
for the IEEE 37-bus feeder and in Fig. 4(b) for the IEEE
123-bus feeder. It turns out that when ǫ becomes smaller than
a threshold, the computed voltages from successive power
flow runs converge. This threshold as a fraction of ǫ/|yt |
is respectively 10−5 and 10−3 for the IEEE 37-bus and the
IEEE 123-bus feeders. The IEEE 37-bus contains two delta–
delta transformers whereas IEEE 123-bus includes only one,
therefore the required threshold for ǫ is smaller in the IEEE
37-bus than that of IEEE 123-bus.
In the modeling of the delta–wye transformers in 8500node and the ELV 906-bus feeders, we use nodal admittances
of Table IV with ǫ′ = ǫ. Once again, a series of load-flow
computations are performed with values of ǫ ranging from
10−2 |yt | to 10−10 |yt |. The maximum difference in voltage
magnitudes for various values of ǫ are depicted in Fig. 5(a)
for the 8500-node feeder and in Fig. 5(b) for the ELV 906bus feeder. Once the value of ǫ/|yt | reaches below 10−3 , the
voltage solutions converge. In all four feeders, if we do not
11
1.1
1.05
1
0.95
0.9
0.06
0.1
0.05
0.08
0.04
0.06
1
3
5
7
9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
1
3
5
7
9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
1
3
5
7
9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
0.03
0.04
0.02
0.02
0.01
0
0
1.1
1.05
1
0.95
0.9
(a)
(b)
Fig. 5. Maximum difference in computed voltage magnitudes obtained from
load-flow runs on (a) the 8500-node feeder and (b) the ELV 906-bus feeder.
The computed voltage solutions in successive load-flow runs converge when
the value of ǫ/|yt | reaches below 10−3 for both feeders.
TABLE V
M AXIMUM pu DIFFERENCE IN COMPUTED VOLTAGE MAGNITUDES VS .
BENCHMARK
Feeder
IEEE 37-bus
IEEE 123-bus
8500-node
ELV 906-bus
phase a
0.0067
0.0061
0.0010
0.0048
phase b
0.0019
0.0034
0.0013
0.0055
phase c
0.0053
0.0039
0.0034
0.0021
modify Y2 , the resultant Y-Bus is singular.
Table V summarizes the comparison of our results with
benchmark solutions using the fixed value of ǫ = 10−6 |yt |.
The reported values are below 0.75% of the benchmark
solutions. Description of each test feeder is provided next,
along with calculated voltage plots corresponding to Table
V. In the voltage plots, bus labels of each feeder have been
modified to represent successive numbers starting from 1.
A. IEEE 37-bus feeder
The IEEE 37-bus feeder features the following rather distinctive characteristics: (a) a delta–delta substation transformer, rated 2500 kVA, 230 kV/4.8 kV line to line, with
zt = (2 + j8)%, on edge (1, 2); (b) a delta–delta transformer, rated 500 kVA, 4.8 kV/0.48 kV line to line, with
zt = 0.09 + j1.81%, on edge (24, 38); (c) an open-delta
SVR on edge (2, 3), with a relatively high impedance of
zt = j1%; and (d) a variety of delta-connected constant-power,
constant-current, and constant-impedance loads.
The impedance of the SVR was obtained from the OpenDSS
implementation provided in the file OpenDSS/Distrib/
IEEETestCases/37Bus/ieee37.dss. Voltage magnitudes corresponding to the value of ǫ = 10−6 |yt | are plotted
in Fig. 6 for both ideal and non-ideal regulator models where
bus 39 represents the n′ of the SVR. The solution provided by
OpenDSS is also provided for verification. Phase plots have
been omitted here due to space limitations but are available on
our github page. The maximum voltage magnitude difference
between the computed solutions and those of OpenDSS are
tabulated in Table VI. It is inferred from Table VI that when
the series impedance of the SVR is relatively high, the nonideal SVR models in (18) provide more accurate results.
1.1
1.05
1
0.95
0.9
Fig. 6. The IEEE 37-bus feeder voltage profile obtained from the Z-Bus
method using ideal SVRs (black squares) and non-ideal SVRs (blue circles)
in comparison to the voltage profile provided by OpenDSS (red cross). Bus
labels have been modified to represent successive numbers starting from 1.
TABLE VI
I DEAL SVR S VS N ON - IDEAL SVR S : M AX . pu DIFFERENCE IN COMPUTED
VOLTAGE MAGNITUDES VS . O PEN DSS
SVR model
Ideal [eq. (19)]
Non-ideal [eq. (18)]
phase a
0.0136
0.0067
phase b
0.0046
0.0019
phase c
0.0072
0.0053
B. IEEE 123-bus test feeder
The IEEE 123-bus test feeder features (a) three-, two-, and
single-phase laterals; (b) four wye-connected SVRs, namely,
ID1 on edge (1, 2) three-phase gang-operated, ID2 on phase
a of edge (12, 13), ID3 on phases a, c of edge (28, 29), ID4
on three-phases of edge (75, 76); (c) a delta-delta transformer,
rated 150 kVA and 4.16 kV/0.48 kV line to line, on edge
(68, 69). The voltage profile is provided in Fig. 7, where
the buses 127, 128, 129, 130 are the n′ of SVRs with IDs
1, 2, 3, 4 respectively (the last four markers on the plot). The
corresponding phase plots are available on our github page.
C. 8500-node test feeder
The 8500-node test feeder includes approximately 2500
medium voltage buses3 [39]. The network features (a) three-,
two-, and single-phase laterals; (b) a delta–wye substation transformer on edge (1, 2), rated 27.5 MVA,
115 kV/12.47 kV line to line, with zt = 1.344 + j15.51%;
and (c) four wye-connected, individually operated, three-phase
SVRs, namely, ID1 on edge (2, 3), ID2 on edge (201, 202),
ID3 on edge (146, 147), ID4 on edge (1777, 1778).
The voltage profile is provided in Fig. 8, where the buses
2502, 2503, 2504, 2505 are the n′ of SVRs with IDs 1, 2, 3, 4
respectively (the last four markers on the plot). For this feeder,
the SVR models of (19) should correspond to type A SVRs.
In this case, the ∓ sign in (20) is changed to ±, and instead
of Ai , its inverse A−1
is used.
i
3 The feeder documentation distinguishes between the term “bus” and
“node.” A bus can have multiple nodes (phases). The feeder includes 2500
primary buses, each of which may be three-, two-, or one-phase.
12
Proof: The invertibility proof is based on contradiction.
Assume x = xR + jxI ∈ Cn is a nonzero vector in the
nullspace of Y. Then, we have that
1.1
1
0.9
1
10
20
30
40
50
60
70
80
90 100 110 120 130
Yx = (G + jB)(xR + jxI )
= GxR − BxI + j(GxI + BxR ) = 0
1.1
1
(51)
Thus, we conclude that
0.9
1
10
20
30
40
50
60
70
80
90 100 110 120 130
1.1
1
0.9
1
10
20
30
40
50
60
70
80
90 100 110 120 130
GxR − BxI = 0, GxI + BxR = 0.
Multiplying the second equality in (52) by xTI yields xTI GxI +
xTI BxR = 0. In the latter equality, the first term is nonnegative due to the assumption that G ≻ O . Therefore we
must have that xTI BxR ≤ 0. Since B = BT we find that
xTI BxR = xTR BT xI = xTR BxI ≤ 0.
Fig. 7. The IEEE 123-bus feeder voltage profile obtained from the Z-Bus
method (blue circles) in comparison to the voltage profile provided by the
benchmark (red cross). Bus labels have been modified to represent successive
numbers starting from 1.
D. The European 906-bus low voltage feeder
The ELV feeder features (a) three-phase laterals (there are
no missing phases), and (b) a delta–wye substation transformer, rated 800 kVA, 11 kV/416 V line to line, with
zt = (0.4 + j4)%, on edge (1, 2). The obtained voltage profile
is provided in Fig. 9. The phase plots are omitted due to space
limitation, but are available on our github page.
IX. C ONCLUDING
REMARKS
This paper focused on nodal admittance modeling of threephase distribution networks. Models for transmission lines
and most relevant transformer connections are reviewed, while
novel models for step-voltage regulators are derived, explicitly
accounting for their tap positions and their series admittance.
Putting the models for the series elements together yields
the network bus admittance matrix Y-Bus, whose invertibility
is crucial in deriving power flow solutions via the Z-Bus
method. The paper carefully laid out the conditions on each
series element that guarantee the invertibility of the Y-Bus
and proves why previous proposals on modifications of certain
transformer connections restore its invertibility. The conditions
are tailored to practical distribution networks, which can be
radial or meshed, feature any number of transformers and
SVRs, and include missing phases. Comprehensive numerical
tests are presented for the IEEE 37-bus, the IEEE 123-bus, the
8500-node medium-voltage feeders, and the European 906bus low-voltage feeder. The codes that build the Y-Bus and
compute the power flow solutions are provided online.
A PPENDIX A
U SEFUL LINEAR ALGEBRA
RESULTS
Lemma 1: Consider matrix Y = G + jB ∈ Cn×n with
G ≻ O and B = BT . Then Y−1 exists and Re[Y−1 ] ≻ O.
(52)
(53)
Multiplying the first equality in (52) by xTR gives xTR GxR −
= 0. In the latter equality, the first term is non-negative
due to the assumption that G ≻ O, and the second term is
non-negative due to (53). Therefore, both terms must be zero.
Hence xR = 0. Replacing xR = 0 in the second equality
of (52) yields xI = 0 as well. Hence, x = 0 which is a
contradiction, since the vector x was assumed to be nonzero.
To prove the second portion, let Y−1 = R + jX. Then, it
holds that (R + jX)(G + jB) = I. Hence,
xTR BxI
RG − XB = I
(54a)
XG + RB = O.
(54b)
Using (54b), we obtain X = −RBG−1 . The latter expression
can be inserted into (54a) to obtain:
RG + RBG−1 B = I ⇒ R = (G + BG−1 B)−1 .(54c)
Notice that G + BG−1 B is positive definite since G ≻ O.
Hence, it holds that R ≻ O.
Lemma 2: For vectors u and v and matrix W we have that
uT Wu − uT Wv + vT Wv − vT Wu = (u − v)T W(u − v). (55)
Proof: The left-hand side of (55) can be factored
T
u (Wu − Wv) − vT (Wu − Wv) = (u − v)T W(u − v).
Lemma 3: For vectors u, v and matrices W, X, Z we have
uT Wu − uT Xv + vT Zv − vT XT
u=
T
W
−X u
T
u
v
−XT
Z
v
(56)
Proof: The right-hand side of (56) is written as follows
T
Wu − Xv
u
= uT Wu − uT Xv − vT XT u + vT Zv.
T
v
−X u + Zv
Lemma 4: Suppose Assumption A1(a) of Section VI holds
for the line (n′ , m) of the SVR. For the values of Ai and ZR
(n′ )
in Table II, the matrix FR = I|Ωn | + Yn′ m ATi ZR Ai [cf. (13)]
is invertible. Moreover, the following holds for FR :
(n′ )
−1
T
F−T
R = I|Ωn | − Ai ZR Ai FR Yn′ m .
(57)
13
1.1
1
0.9
1
100
200
300
400
500
600
700
800
900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500
1
100
200
300
400
500
600
700
800
900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500
1
100
200
300
400
500
600
700
800
900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500
1.1
1
0.9
1.1
1
0.9
(a)
0
-20
-40
-60
1
100
200
300
400
500
600
700
800
900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500
1
100
200
300
400
500
600
700
800
900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500
1
100
200
300
400
500
600
700
800
900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500
-120
-140
-160
-180
120
100
80
60
(b)
Fig. 8. The 8500-node feeder voltage profile obtained from the Z-Bus method using ideal regulators (blue circles) in comparison to the voltage profile
provided by the benchmark (red cross): (a) voltage magnitudes (b) voltage phases.
1.05
0.95
0.85
1 30
60
90
120 150 180 210 240 270 300 330 360 390 420 450 480 510 540 570 600 630 660 690 720 750 780 810 840 870 900
1 30
60
90
120 150 180 210 240 270 300 330 360 390 420 450 480 510 540 570 600 630 660 690 720 750 780 810 840 870 900
1 30
60
90
120 150 180 210 240 270 300 330 360 390 420 450 480 510 540 570 600 630 660 690 720 750 780 810 840 870 900
1.05
0.95
0.85
1.05
0.95
0.85
Fig. 9. The ELV 906-bus feeder voltage profile obtained from the Z-Bus method (blue circles) in comparison to the voltage profile provided by the benchmark
(red cross). The intensity of the plot is due to the fact that this feeder contains three-phase laterals only.
Proof: Combining (6), Assumption A1(a) of Section VI,
(n)
and Lemma 1, it follows that Ynm is symmetric and invertible,
(n) −1
and Re[(Ynm ) ] ≻ O. Therefore, we can write
It follows from (58) that
i−1
h
(n′ ) −1
(n′ )
T
.
)
+
A
Z
A
=
(Y
F−1
Y
′
′
R i
i
nm
nm
R
(59)
(n′ )
FR =
(n′ )
Yn′ m
i
h
(n′ )
(Yn′ m )−1 + ATi ZR Ai .
(58)
Due to the symmetricity of (Yn′ m )−1 and ATi ZR Ai , it
(n′ )
follows from (59) that F−1
R Yn′ m is symmetric:
(n′ )
(n )
Hence, FR is invertible if and only if (Yn′ m )−1 + ATi ZR Ai
is invertible. For all SVRs, ZR is diagonal and Re[ZR ] O
holds due to ohmic or zero impedances. Furthermore, the
gain matrix Ai is real-valued and ATi ZR Ai is symmetric.
(n)
Also recall that Re[(Ynm )−1 ] ≻ O. Therefore, it holds that
(n′ ) −1
Re[(Yn′ m ) +ATi ZR Ai ] ≻ O. It follows from Lemma 1 that
(n′ )
−1
exists.
(Yn′ m )−1 + ATi ZR Ai is invertible. This proves FR
(n′ )
−T
F−1
R Yn′ m = (Yn′ m )FR .
′
(60)
Multiplying (13) from the left by F−1
R yields:
(n′ )
−1
T
I|Ωn | = F−1
R + FR Yn′ m Ai ZR Ai
⇒
F−1
R
= I|Ωn | −
(61)
(n′ )
T
Yn′ m F−T
R Ai ZR Ai ,
(62)
′
(n )
−1
where in the last equality we replaced FR
Yn′ m with its
equivalent from (60). Transposing (62) yields (57).
14
R EFERENCES
[1] K. A. Birt, J. J. Graffy, J. D. McDonald, and A. H. El-Abiad, “Three
phase load flow program,” IEEE Trans. Power App. Syst., vol. 95, no. 1,
pp. 59–65, Jan. 1976.
[2] P. A. N. Garcia, J. L. R. Pereira, S. Carneiro, V. M. da Costa, and
N. Martins, “Three-phase power flow calculations using the current
injection method,” IEEE Trans. Power Syst., vol. 15, no. 2, pp. 508–
514, May 2000.
[3] T. H. Chen, M. S. Chen, K. Hwang, P. Kotas, and E. A. Chebli,
“Distribution system power flow analysis - a rigid approach,” IEEE
Trans. Power Del., vol. 6, no. 3, pp. 1146–1152, July 1991.
[4] L. R. Araujo, D. R. R. Penido, S. Carneiro, and J. L. R. Pereira, “A threephase optimal power-flow algorithm to mitigate voltage unbalance,”
IEEE Trans. Power Del., vol. 28, no. 4, pp. 2394–2402, Oct. 2013.
[5] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power
flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3, pp.
1464–1475, Sept. 2013.
[6] A. S. Zamzam, N. D. Sidiropoulos, and E. Dall’Anese, “Beyond relaxation and Newton-Raphson: Solving AC OPF for multi-phase systems
with renewables,” IEEE Trans. Smart Grid, vol. PP, no. 99, 2016.
[7] F. Wu and S. Kumagai, “Steady-state security regions of power systems,”
IEEE Trans. Circuits Syst., vol. 29, no. 11, pp. 703–711, Nov. 1982.
[8] T. J. Overbye, “A power flow measure for unsolvable cases,” IEEE Trans.
Power Syst., vol. 9, no. 3, pp. 1359–1365, Aug. 1994.
[9] S. Bolognani and S. Zampieri, “On the existence and linear approximation of the power flow solution in power distribution networks,” IEEE
Trans. Power Syst., vol. 31, no. 1, pp. 163–172, Jan. 2016.
[10] S. Yu, H. D. Nguyen, and K. S. Turitsyn, “Simple certificate of
solvability of power flow equations for distribution systems,” in Proc.
IEEE PES General Meeting, July 2015, pp. 1–5.
[11] C. Wang, A. Bernstein, J. Y. L. Boudec, and M. Paolone, “Explicit
conditions on existence and uniqueness of load-flow solutions in
distribution networks,” IEEE Trans. Smart Grid, vol. PP, no. 99, 2016.
[Online]. Available: http://arxiv.org/abs/1602.08372
[12] ——, “Existence and uniqueness of load-flow solutions in three-phase
distribution networks,” IEEE Trans. Power Syst., vol. PP, no. 99, 2016.
[13] M. Bazrafshan and N. Gatsis, “Convergence of the Z-Bus method for
three-phase distribution load-flow with ZIP loads,” IEEE Trans. Power
Syst., May 2016, to be published. [Online]. Available: https://arxiv.org/
abs/1605.08511
[14] B. A. Robbins, H. Zhu, and A. D. Dominguez-Garcia, “Optimal tap
setting of voltage regulation transformers in unbalanced distribution
systems,” IEEE Trans. Power Syst., vol. 31, no. 1, pp. 256–267, Jan.
2016.
[15] A. R. Baran Jr. and T. S. P. Fernandes, “A three-phase optimal power
flow applied to the planning of unbalanced distribution networks,” Int.
J. of Electrical Power & Energy Systems, vol. 74, pp. 301–309, 2016.
[16] D. Jiang and R. Baldick, “Optimal electric distribution system switch
reconfiguration and capacitor control,” IEEE Trans Power Syst., vol. 11,
no. 2, pp. 890–897, May 1996.
[17] S. Bolognani and F. Dörfler, “Fast power system analysis via implicit
linearization of the power flow manifold,” in 53rd Annu. Allerton Conf.
Communication, Control, and Computing, Sept. 2015, pp. 402–409.
[18] V. Kekatos, L. Zhang, G. B. Giannakis, and R. Baldick, “Voltage
regulation algorithms for multiphase power distribution grids,” IEEE
Trans. Power Syst., vol. 31, no. 5, pp. 3913–3923, Sept 2016.
[19] H. Ahmadi, J. R. Mart, and A. von Meier, “A linear power flow
formulation for three-phase distribution systems,” IEEE Trans. Power
Syst., vol. 31, no. 6, pp. 5012–5021, Nov 2016.
[20] A. Garces, “A linear three-phase load flow for power distribution
systems,” IEEE Trans. Power Syst., vol. 31, pp. 827–828, Jan. 2016.
[21] M. J. Gorman and J. J. Grainger, “Transformer modelling for distribution
system studies part II: Addition of models to YBUS and ZBUS,” IEEE
Trans. Power Del., vol. 7, no. 2, pp. 575–580, Apr. 1992.
[22] M.-S. Chen and W. E. Dillon, “Power system modeling,” Proc. of the
IEEE, vol. 62, no. 7, pp. 901–915, July 1974.
[23] W. H. Kersting, Distribution System Modeling and Analysis, 3rd ed.
CRC Press, 2002.
[24] S. S. Moorthy and D. Hoadley, “A new phase-coordinate transformer
model for Ybus analysis,” IEEE Trans. Power Syst., vol. 17, no. 4, pp.
951–956, Nov. 2002.
[25] W. H. Kersting, W. H. Philips, and W. Carr, “A new approach to
modeling three-phase transformer connections,” IEEE Trans. Ind. Appl.,
vol. 35, no. 1, pp. 169–175, Jan. 1999.
[26] P. Xiao, D. C. Yu, and W. Yan, “A unified three-phase transformer model
for distribution load flow calculations,” IEEE Trans. Power Syst., vol. 21,
no. 1, pp. 153–159, Feb. 2006.
[27] I. Dz̆afić, R. A. Jabr, and H.-T. Neisius, “Transformer modeling for threephase distribution network analysis,” IEEE Trans. Power Syst., vol. 30,
no. 5, pp. 2604–2611, Sept. 2015.
[28] D. M. Anderson and B. F. Wollenberg, “Solving for three phase
conductively isolated busbar voltages using phase component analysis,”
IEEE Trans. Power Syst., vol. 10, no. 1, pp. 98–108, Feb. 1995.
[29] T. H. Chen, M. S. Chen, T. Inoue, P. Kotas, and E. A. Chebli, “Threephase cogenerator and transformer models for distribution system analysis,” IEEE Trans. Power Del., vol. 6, no. 4, pp. 1671–1681, Oct. 1991.
[30] IEEE PES Distribution System Analysis Subcommittee’s Distribution
Test Feeder Working Group, “Distribution test feeders.” [Online].
Available: http://ewh.ieee.org/soc/pes/dsacom/testfeeders/index.html
[31] “Standard load models for power flow and dynamic performance simulation,” IEEE Trans. Power Syst., vol. 10, no. 3, pp. 1302–1313, 1995.
[32] “Bibliography on load models for power flow and dynamic performance
simulation,” IEEE Trans. Power Syst., vol. 10, no. 1, pp. 523–538, 1995.
[33] K. McKenna and A. Keane, “Open and closed-loop residential load
models for assessment of conservation voltage reduction,” IEEE Trans.
Power Syst., pp. 1–1, 2016.
[34] T. H. Chen and J. D. Chang, “Open wye-open delta and open deltaopen delta transformer models for rigorous distribution system analysis,”
Proc. Inst. Elect. Eng. C - Generation Transmission Distribution, vol.
139, no. 3, pp. 227–234, May 1992.
[35] A. A. Sallam and O. P. Malik, Electric distribution systems. John Wiley
& Sons, 2011, vol. 68.
[36] R. C. Dugan, OpenDSS Manual, EPRI, March 2016.
[37] EPRI, “The Open Distribution System SimulatorTM , OpenDSS Version
7.6.5.” [Online]. Available: https://sourceforge.net/projects/electricdss/
[38] OpenDSS Wiki, “Technote transformers,” OpenDSS/Doc/TechNotes.
[39] R. C. Dugan and R. F. Arritt, “The IEEE 8500-node test feeder,”
https://ewh.ieee.org/soc/pes/dsacom/testfeeders/, EPRI, April 2010.
| 3cs.SY
|
arXiv:1311.6093v6 [cs.DS] 3 Aug 2016
A New Algorithm for Updating and Querying
Sub-arrays of Multidimensional Arrays
Pushkar Mishra
Computer Laboratory,
University of Cambridge
[email protected]
Abstract
Given a d-dimensional array A, an update operation adds a given constant C to each element
within a continuous sub-array of A. A query operation computes the sum of all the elements within a
continuous sub-array of A. The one-dimensional update and query handling problem has been studied
intensively and is usually solved using segment trees with lazy propagation technique. In this paper,
we present a new algorithm incorporating Binary Indexed Trees and Inclusion-Exclusion Principle to
accomplish the same task. We extend the algorithm to update and query sub-matrices of matrices
(two-dimensional array). Finally, we propose a general form of the algorithm for d-dimensions which
achieves O(4d ∗ logd n) time complexity for both updates and queries. This is an improvement over
the previously known algorithms which utilize hierarchical data structures like quadtrees and octrees
and have a worst-case time complexity of Ω(nd−1 ) per update/query.
Keywords: Algorithm; Data Structure; Multidimensional Array; Binary Indexed Tree; Range-update;
Range-query.
1
Introduction
The problem of updating and querying sub-arrays of multidimensional arrays is of consequence to several fields including data management, image processing and geographical information systems. The
one-dimensional version of this problem has conventionally been solved using segment trees with lazy
propagation technique. We show in this paper that a d-dimensional segment tree (d > 1) supports lazy
propagation only along one out of the d dimensions. Consequently, the worst-case time complexity for
updates and queries becomes O(nd−1 log n) instead of O(logd n).
Space-partitioning hierarchical data-structures like the quadtree — and its generalization to higher
dimensions, e.g., octree — perform better than multidimensional segment trees.[5] These structures work
by recursively dividing the given d-dimensional space (d > 1) into smaller axis-parallel hyper-rectangles.
Such trees have a worst-case time complexity of Ω(nd−1 ) for updates and queries.
The algorithm that we propose is based on Binary Indexed Trees[1] (BITs) and inclusion-exclusion
principle. It has the same space complexity as the aforementioned algorithms; but the worst-case time
complexity for updates and queries is O(4d ∗ logd n).
For the purpose of this paper, we define three sets of operations.
1. Point-update range-query
Point-update range-query refers to the set of operations wherein updates are performed only on unitsized sub-arrays (i.e., individual elements) while queries are performed on sub-arrays of arbitrary
sizes.
2. Range-update point-query
Range-update point-query refers to the set of operations wherein updates are performed on subarrays of arbitrary sizes while queries are performed only on unit-sized sub-arrays.
1
3. Range-update range-query
Range-update range-query refers to the set of operations wherein both updates and queries are
performed on sub-arrays of arbitrary sizes. Note that the first two sets of operations are subsets of
this set.
Problem Definition. Given is a d-dimensional array A of size N = n1 × n2 × · · · × nd , where nk (1 ≤
k ≤ d) is the length of the k th dimension. For the sake of simplicity, we assume that n1 , n2 , . . . , nd = n.
The goal of the problem is to handle on-line range-update range-query operations on A. By on-line we
imply that a particular operation must be performed before the next one can be handled, and there is
no prior knowledge of the order of operations.
Using the known algorithms for the first two sets of operations, we devise an efficient algorithm for the
aforementioned problem. Note that throughout this paper, we assume a unit-cost word Random Access
Machine (RAM) model with word size Θ(log n). On such a model, standard arithmetic and boolean
bitwise operations on word-sized operands can be performed in Θ(1) time.
Paper organization. In section 2, we discuss the motivation behind this work. In Section 3, we
summarize our contributions. In section 4, we define the terms and notations used in this paper. We
give the formal problem statement in section 5. In section 6, we discuss the existing algorithms for
handling range-update range-query operations on multidimensional arrays. In section 7, we provide a
concise description of Binary Indexed Trees and associated algorithms. We describe the new algorithm
and its time and memory complexities in section 8. In section 9, we provide the data obtained from
experimental comparison between execution times of the old and new algorithms. Lastly, we offer brief
conclusions in section 10.
2
Background and motivation
Representation of a d-dimensional space or object as a d-dimensional array has numerous applications
in several fields of Computer Science. For example, images can be represented as 2D arrays where
individual cells can represent pixels. Geographical regions can also be represented as 2D arrays with
individual cells denoting unit area. Space-partitioning, hierarchical data structures like the quadtree
(and its generalization to higher dimensions, e.g., octree) are used perform a variety of operations on
such representation of space and objects[9].
Motivation for this paper stems from the need to process range-update range-query operations with
better time complexity. Geographical Information Systems(GIS) are an example where faster processing
can be of significance. In their paper[2], Samet et al. describe the development of a GIS which uses
quadtree as the underlying data structure. Consider the case when different regions of a country have
received different amounts of rainfall, and total rainfall needs to be calculated for the season. Consider
another case when different regions in a country have received different amounts of water in supply for a
month, and the net water supplied is to be calculated for the country. These are cases of Range-update
and Range-query in a GIS respectively. Directly visiting each unit cell of a particular region to add or
query certain a parameter can prove to be time taking, specially if the unit cells represent very small
areas, i.e., the GIS is detailed. Therefore, Samet uses a quadtree for processing updates and queries. We
show in this paper that these types of operations can be performed with better time complexity using
Binary Indexed Trees.
The need for processing updates and queries is not limited to one or two dimensions. Thus, it is
important to have an algorithm which can be extended to any number of dimensions, without much
effort.
3
Our Contributions
The main contributions of this paper are as follows.
2
1. Discussion and analysis of known algorithms.
2. A new algorithm for processing on-line range-update range-query operations on multidimensional
arrays.
3. Generalization of the algorithm to arbitrary number of dimensions.
4. Experimental comparison between execution times of the old and new algorithms for two and three
dimensional arrays.
4
Preliminaries
Let us first provide some general definitions for clarity. By log n, we mean log2 n. An array refers to a
1D array unless stated otherwise. Both sub-array and ‘range’ refer to a contiguous portion of an array.
A sub-array of a d-dimensional array A is depicted as A[a1 , b1 , c1 , d1 , ... : a2 , b2 , c2 , d2 , ...], where a1 ,
a2 are coordinates along the 1st dimension, b1 , b2 are coordinates along the 2nd dimension, and so on. It
consists of all the elements arr[a][b][c][d][...] such that a1 ≤ a ≤ a2 , b1 ≤ b ≤ b2 , c1 ≤ c ≤ c2 , d1 ≤ d ≤ d2 ,
and so on.
A general sub-array is denoted by [a1 , b1 , c1 , d1 , ... : a2 , b2 , c2 , d2 , ...].
Throughout this paper, we use rsum(a1 , b1 , ... : a2 , b2 , ...) to refer to the cumulative sum of all the
elements of the sub-array [a1 , b1 , ... : a2 , b2 , ...].
We denote a point in d-dimensional space by (x1 , x2 , x3 , . . . , xd ). A point in 1D space is denoted
without parentheses.
bitAnd(a, b) refers to the bitwise and of integers a and b. For example, bitwise(2, 3) will return 2
since 2 and 3 are represented as 10 and 11 in binary.
We recommend that the reader be familiar with Binary Indexed Trees and basic operations on them[1]
in order to understand the algorithms and concepts being put forward.
5
Problem statement
In this section, we put forward the formal problem statement.
Given is a 1D array A of length n with all the elements initially set to 0. Two types of operations
need to be performed on this array:
1. Given x1 and x2 , add a constant c to all the elements of the sub-array A[x1 : x2 ].
2. Given x1 and x2 , output the sum of all the elements of the sub-array A[x1 : x2 ].
We refer to this as the ‘1D version’ of the problem.
The problem can be extended to any number of dimensions. Consider the 2D version. Given is a
matrix M with side-length n with all the elements initially set to 0. Again, two types of operations need
to be performed on this matrix:
1. Given (x1 , y1 ) and (x2 , y2 ), add a constant c to all the elements of the sub-matrix M [x1 , y1 : x2 , y2 ].
2. Given (x1 , y1 ) and (x2 , y2 ), output the sum of all the elements of the sub-matrix M [x1 , y1 : x2 , y2 ].
In this paper, we describe our algorithm for 1D and 2D versions of the problem, and then generalize the
idea to d-dimensions.
3
6
Previous work
In this section of the paper, we discuss the previously known algorithms and data structures for handling
on-line range-update range-query operations. We analyze their time and space complexities.
6.1
For 1D version
1D version of the problem has conventionally been solved using segment trees. De Berg et al., in their
book [4, p. 226], prove that any interval [i : j] can be constructed by using a maximum of O(log N )
nodes of the segment tree. The lazy propagation technique ensures that no more than O(log n) nodes are
to be visited to perform the required update or query. In this technique, the node in the segment tree,
which contains the range to be updated, is marked (or ‘flagged’), and the update is propagated down to
its children only when a query is to be performed on that node, or on its children.
Since, a maximum of O(log n) nodes must be visited for performing update or query, the running time
of this algorithm is O(log n) per update/query operation. The space required to execute it is O(n).[4, p.
226-227]
6.2
For Higher Dimensions
The segment tree can be generalized to any number of dimensions in the form of multilevel segment trees.
For example, for a d-dimensional array with side-length n, we first build one-dimensional segment trees
along the dth dimension (also referred to as ‘last dimension’). We then build segment trees along (d − 1)th
dimension, and so on.
In multidimensional versions, the segment tree stores a collection of axis-parallel hyper-rectangles.
Point-update range-query or range-update point-query operations can be implemented in O(logd n) complexity. However, we show that a multidimensional segment tree does not produce an optimal solution
in the case of range-update range-query operations.
Proposition 1. A d-dimensional segment tree does not support Lazy Propagation technique on more
than one of the dimensions.
Proof. We consider a 2D segment tree, built on a matrix mat. As described earlier, in building the 2D
segment tree, we first build segment trees along each row, and then along columns. Each of the node
in this 2D segment tree stores sum of elements of a sub-matrix of mat. From the manner in which it
is build, every individual element (x, y) of mat is contained in O(log2 n) nodes of the 2D segment tree.
Of these O(log2 n), we take two nodes l and r. We choose l and r such that the sub-matrix contained
by l does not completely lie inside the sub-matrix contained by r and vice versa. We now perform a
lazy-update on node l. Since l and r have at least one common element, the value contained at r is also
affected by the update. But, when value at r is queried, it would not return the updated value. This is
because, in a segment tree, a query always moves towards the root. The path from node r to the root of
the segment tree would not contain l, since the sub-matrix associated with l does not fully contain the
sub-matrix associated with r. Consequently, the update ‘flag’ at l is never encountered during query on
r. Therefore, Lazy Propagation technique does not produce the correct result.
As a result of Proposition 1, the running time for each operation (update or query) on a d-dimensional
segment tree becomes O(nd−1 log n). This is because, for each of the d − 1 dimensions that does not
support lazy propagation, we need to recursively visit individual indexes in the range to be updated or
queried. As proved by Berg et. al, a segment tree possesses 2n nodes [4, p. 226-227]. Consequently,
memory requirement of a d-dimensional segment tree is given by O((2n)d ).
As stated in the first section, quadtrees and their generalization to higher dimensions perform better
than multidimensional segment trees. Below, we analyze their time and space complexity.
A quadtree[6] recursively divides the given 2D space into four quadrants. Quadtrees can be used to
perform range-update range-query operations on matrices in the same manner as segment trees can be
4
used for 1D arrays. For simplicity, let us assume that the side-length of the matrix to be updated/queried
is a power of two. Each node of the quadtree built for handling operations on this matrix stores data for
a square sub-matrix.[7] Accordingly, the root node stores the data for the whole matrix and the leaves
store the individual elements. Once the quadtree has been built, range-update range-query operations
can be performed by incorporating lazy propagation.
Proposition 2. For an n × n matrix, the worst case time complexity for query/update using a quadtree
is Ω(n).
Proof. We prove this by example. Consider a matrix M with side-length n (n = 2k ). Quadtree recursively
divides M into square sub-matrices. Now, consider querying/updating a row in this matrix. No part of
this row forms a square region except for the individual elements. This implies that no internal node in
the quadtree stores data specifically for this row, or a part of it. Therefore, to update/query, we need
to visit all the leaves associated with the elements of the row to be updated/queried. There are n such
leaves. Since we traverse each edge between the node and the leaves only once, the height of the tree is
traversed only once. Thus, the time complexity of such an update/query becomes O(n). As there exists
at least one operation with worst case time complexity O(n), the worst case running time of a quadtree
is Ω(n).
In his paper[9, p. 240], Samet states that the worst-case memory requirement for building a quadtree
occurs when the region concerned corresponds to checker-board pattern. We encounter this case when
building quadtree over a 2D array. According to Samet, the number of nodes in such cases is a function
of r, where r is height (also known as resolution) of the quadtree. Since, each node in a quadtree has
r
four children, therefore, the total number of nodes in a tree with height r is 4 3−1 .[8] Hence, the memory
r
requirement is O(4 ).
For an n × n matrix, the height of the quadtree is log2 n + 1 (height of a quadtree is same as the
depth of recursion[8, p. 2–3]). Therefore, the memory requirement is O(4log2 n+1 ). On simplifying this
expression, we get (2n)2 .
The 3D version of the quadtree — known as ‘octree’ — recursively divides a 3D space into octants.
By similar analysis as for quadtree, we can say that an octree has worst case running time of Ω(n2 ).
The analysis of memory requirement can be done on the same lines as for quadtree, and is given by
O(8log2 n+1 ). This simplifies to O((2n)3 ).
In conclusion, for a quadtree generalized to d-dimensions, the worst-case time and space complexities
are Ω(nd−1 ) and O((2n)d ) respectively.
7
7.1
Binary Indexed Trees and associated algorithms
Description of the data structure
Binary Indexed Tree (BIT) or Fenwick Tree[1] is a data structure commonly used for calculating prefix
sums efficiently. A BIT works by storing partial cumulative sums. For example, index 8 of BIT contains
the cumulative sum of elements from 1 to 8, i.e., rsum(1 : 8). Similarly, index 6 stores rsum(5 : 6), and
so on. For answering queries, BIT combines these stored partial sums. A segment tree can also be used
to perform the functions of a BIT, but BITs are easier to code and have a lower constant of complexity.
Hence, we base our algorithm on them. A BIT can handle range-update point-query operations or rangeupdate point-query operations, but not both simultaneously. Range-update range-query includes both the
sets of operations. By this, we mean that range-update range-query is a general case of both range-update
point-query and point-update range-query.
7.2
Point-update range-query on BIT
This is the standard set of operations that a BIT can handle. As explained in the paper by Fenwick[1],
a BIT is capable of updating the value at a particular point, and querying the cumulative sum up till
5
a particular point. Throughout this paper, we use updatep(bit, x, c) to denote a point-update operation
on a 1D BIT bit which adds c to element at position x. Similarly, we use queryr(bit, x) to denote
the operation which returns rsum(1 : x), i.e., the cumulative sum of elements up till position x. Any
arbitrary range [x : y] can be queried by querying [1 : x − 1] and [1 : y].
We give pseudo-codes for both the functions. The algorithms were devised by Fenwick[1].
Point-update on 1D BIT
1:
2:
3:
4:
5:
6:
7:
function updatep(bit, x, c)
i←x
while i ≤ n do
bit[i] ← bit[i] + c
i ← i + bitAnd(i, −i)
end while
end function
Range-query on 1D BIT
1:
2:
3:
4:
5:
6:
7:
8:
9:
function queryr(bit, x)
sum ← 0
i←x
while i > 0 do
sum ← sum + bit[i]
i ← i − bitAnd(i, −i)
end while
return sum
end function
A 2D BIT uses a 2D array to store values. Just as each index in a 1D BIT stores the cumulative sum of
a particular sub-array, similarly, each index in a 2D BIT stores the cumulative sum of a particular submatrix. For example, the index (8, 8) stores the cumulative sum of all elements in the range [1, 1 : 8, 8],
i.e., rsum(1, 1 : 8, 8). Similarly, index (6, 4) stores rsum(5, 1 : 6, 4), and so on. It can be observed that
a 2D BIT can be treated as a BIT of 1D BITs. Throughout this paper, we use updatep(bit, (x, y), c) to
denote a point-update operation on 2D BIT named bit, which adds constant c to the element at position
(x, y). Similarly, we use queryr(bit, (x, y)) to denote the operation which returns rsum(1, 1 : x, y).
Further using the inclusion-exclusion principle, any sub-matrix can be queried.
For clarity, we present pseudocodes for updatep(bit, (x, y), c) and queryr(bit, (x, y)) operations on a 2D
BIT of side-length n:
Point-update on 2D BIT
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
function updatep(bit, (x, y), c)
i←x
while i ≤ n do
j←y
while j ≤ n do
bit[i][j] ← bit[i][j] + c
j ← j + bitAnd(j, −j)
end while
i ← i + bitAnd(i, −i)
end while
end function
6
Range-query on 2D BIT
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
function queryr(bit, (x, y))
sum ← 0
i←x
while i > 0 do
j←y
while j > 0 do
sum ← sum + bit[i][j]
j ← j − bitAnd(j, −j)
end while
i ← i − bitAnd(i, −i)
end while
return sum
end function
The algorithms can be generalized to BITs of any number of dimensions on the same lines. The time
complexity of the functions for a d-dimensional BIT is O(logd n).
7.3
Range-update point-query on BIT
We take an array arr. Each of its values is initially set to 0. We wish to add a constant c to all the
elements in the sub-array arr[i : j]. We want to do this multiple times for arbitrary i, j and c. After some
of the update operations, we want to know the value of an arbitrary element arr[k]. This is the simplest
case of range-update point-query. Using the algorithms and functions of point-update range-query, a BIT
can be made to handle such operations.
Throughout this paper, we use updater(bit, x1 , x2 , c) to denote an update operation on a 1D BIT bit,
which adds constant c to each element in the sub-array [x1 : x2 ]. We use queryp(bit, x) to denote
the operation that returns the value of the element at position x. We first give the pseudocode for the
functions, and then explain their working.
Range-update on 1D BIT
1:
2:
3:
4:
function updater(bit, x1 , x2 , c)
updatep(x, c)
updatep(y + 1, −c)
end function
Point-query on 1D BIT
1:
2:
3:
4:
5:
6:
7:
8:
9:
function queryp(bit, x))
val ← 0
i←x
while i > 0 do
val ← sum + bit[i]
i ← i − bitAnd(i, −i)
end while
return val
end function
By using two point-update operations, we can perform a range-update. The algorithm for query operation
remains the same as in the first set of operations. As can be noticed, the update algorithm works on the
inclusion-exclusion principle. Whenever we have to update a range [x1 : x2 ] with a constant c, the value
at position x of the bit is increased by c. Due to this increase, rsum(1 : x), where x1 ≤ x, increases by
7
c. The value of rsum(1 : x), where x > x2 , does not increase since all the elements after x2 are updated
with −c. Thus, an updater operation only increases the values of the elements in the specified range.
The query algorithm remains the same. However, since an update operation only increases the value of
the elements in a specific range, and initially bit is set to 0, thus, queryp returns the sum of updates that
affected position x, i.e., the updated value of element at x.
We extend this algorithm to two dimensions. Henceforth, we use updater(bit, (x1 , y1 ), (x2 , y2 ), c) to
denote an update operation on a 2D BIT bit, which adds constant c to each element in the sub-array
[x1 , y1 : x2 , y2 ]. We use queryp(bit, (x, y)) to denote the operation that returns the value of the element
at position x, y.
Below, we give the pseudocode for the functions.
1:
2:
3:
4:
5:
6:
Range-update on 2D BIT
function updater(bit, (x1 , y1 ), (x2 , y2 ), c)
updatep(bit, (x1 , y1 ), c)
updatep(bit, (x2 + 1, y1 ), −c)
updatep(bit, (x1 , y2 + 1), −c)
updatep(bit, (x2 + 1, y2 + 1), c)
end function
Point-query on 2D BIT
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
function queryp(bit, (x, y))
val ← 0
i←x
while i > 0 do
j←y
while j > 0 do
val ← sum + bit[i][j]
j ← j − bitAnd(j, −j)
end while
i ← i − bitAnd(i, −i)
end while
return val
end function
The algorithms can be generalized to BITs of any number of dimensions on the same lines. The time
complexity of the functions for a d-dimensional BIT is O(logd n).
8
Proposed algorithm
By using the functions of range-update point-query described in section 7.3, we can make a BIT handle
range-update range-query operations. In this section, we put forward the algorithm for 1D version and
2D version, and then present a generalization for arbitrary number of dimensions.
8.1
For 1D arrays
An important observation to be made here, is that if rsum(1 : x − 1) and rsum(1 : y) can be computed
efficiently, then rsum(x : y) can be calculated in constant time. This is because,
rsum(x : y) = rsum(1 : y) − rsum(1 : x − 1)
(1)
We define an operation query(i) which returns the value rsum(1 : i), and an operation update(x, y, c)
which updates the sub-array [x : y] with constant c.
8
8.1.1
Range-update
Let there be an array arr, indexed from 1 to n, with all elements initially set to 0. At this point, query(i)
returns 0 for any index i. An update operation update(3, 5, 4) is performed.
Now, query(x) is called for an arbitrary index x. There are three possible cases:
i. i < 3
ii. 3 ≤ i ≤ 5
iii. i > 5
For the first case, the value returned by the function query(i) is 0.
For the second case, the value returned should be 4 if i = 3, 8 if i = 4 and 12 if i = 5.
For the third case, query(i) should return 12 for all i.
From the example above, the following inferences can be made:
When an operation update(x1 , x2 , c) is performed, there is no change in the value of rsum(1 : x) if
x < x1 . For indexes x, where x1 ≤ x ≤ x2 , the value of rsum(1 : x) changes by (c ∗ x − c ∗ (x1 − 1)).
For x > x2 , rsum(1 : x) changes by c ∗ (x2 − x1 + 1).
Thus, after a range-update operation update(x1 , x2 , c), the change in the value of rsum(1 : x) varies
with index x linearly. For indexes which are less that x, there is no change. For an index x, which is
between x1 and x2 , the change is given by the function (c ∗ x − c ∗ (x1 − 1)). For all indexes x > x2 , the
change is given by the function (0 ∗ x + c ∗ (x2 − x1 + 1)).
As linear functions can be added, if the sum of all the functions (referred to as ‘net function’) for an index
x is known, the total change to the initial value of rsum(1 : x) can be calculated by putting x into the
net function. To define a linear function at an index, we need to know the coefficient of the variable x
(henceforth referred to as ‘co-efficient’), and the value of the term independent of x (henceforth referred
to as ‘independent term’). The variable term and the independent terms in addition form the required
linear function.
If the sum of co-efficients and the sum of independent terms at an index are known, then the net function
is also known. Hence, we keep two BITs. One BIT stores the co-efficients and the other stores the
independent term. We name these BITs bitc and biti respectively. Whenever a function ax + b (where
x is the index) is to be added to a range [x1 : x2 ], we update the range [x1 : x2 ] of bitc with a, and
the range [x1 : x2 ] of biti with b. When we need to calculate rsum(1 : x), for some x, we can perform
point-query operation at position x in both the BITs to get the co-efficient and independent term of the
net function. Thus, by performing range-update point-query on two BITs, we can keep track of the net
function for each index.
For an operation update(x1 , x2 , c), the functions which need to be added, along with the ranges, are as
follows.
1. Function cx − c(x1 − 1) to all indexes x such that x1 ≤ x ≤ x2 .
2. Function c(x2 − x1 + 1) to all indexes x such that x2 < x ≤ n.
Below, we present the pseudocode of the update function. Note that we have used the updater function
from section 7.3 to perform the required range-updates on the two BITs.
1:
2:
3:
4:
5:
Range-update on 1D BIT
function update(x1 , x2 , c)
updater(bitc, x1 , x2 , c)
updater(biti, x1 , x2 , −c(x1 − 1))
updater(biti, x2 + 1, n, c(x2 − x1 + 1))
end function
The complexity of the range-update operation is O(log n).
9
8.1.2
Range-query
Whenever the value of rsum(1 : x), for an arbitrary index x, is needed, we call queryp(bitc, x) on bitc
to get the coefficient, and on queryp(biti, x) to get independent term of the net function. Multiplying
the value obtained from bitc with x, and adding the value obtained from biti gives the net change in
the value of rsum(1 : x). Since, we assumed all values to be 0 initially, hence, we get the actual value
rsum(1 : x).
Below, we present the pseudocode of the query function. Note that we have used the queryp function
from section 7.3 to get the values of co-efficients and independent terms at the required positions in the
BITs.
Range-query on 1D BIT
1:
2:
3:
4:
5:
function query(x)
a ← queryp(bitc, x)
b ← queryp(biti, x)
return a ∗ x + b
end function
The complexity of the range-query operation is 2 ∗ O(log n), since 2 BITs are involved.
8.2
Extension to 2D arrays
We extend the algorithm devised above to matrices (2D arrays). An important observation based on
inclusion-exclusion principle is that if rsum(1, 1 : x, y) can be computed efficiently for arbitrary x and y,
then rsum(x1 , y1 : x2 , y2 ), for arbitrary x1 , y1 , x2 , y2 , can be computed efficiently too. This is because,
rsum(x1 , y1 : x2 , y2 ) = rsum(1, 1 : x2 , y2 ) − rsum(1, 1 : x2 , y1 − 1)
− rsum(1, 1 : x1 − 1, y2 ) + rsum(1, 1 : x1 − 1, y1 − 1)
(2)
We assume an operation query((x, y)) which returns the value of rsum(1, 1 : x, y), and an operation
update((x1 , y1 ), (x2 , y2 ), c) which adds c to each element of the sub-matrix [x1 , y1 : x2 , y2 ]. Further, any
range can be queried by using the inclusion-exclusion principle.
8.2.1
Range-update
We begin by analyzing the change in values of rsum(1, 1 : x, y), for arbitrary (x, y), after an
update((x1 , y1 ), (x2 , y2 ), c) is performed.
1. For all (x, y) where x1 ≤ x ≤ x2 and y1 ≤ y ≤ y2, change in rsum(1, 1 : x, y) is given by
cxy − c(y1 − 1)x − c(x1 − 1)y + c.
2. For all (x, y) where x2 < x ≤ n and y1 ≤ y ≤ y2, change in rsum(1, 1 : x, y) is given by
c(x2 − x1 + 1)y − c(y1 − 1)(x2 − x1 + 1).
3. For all (x, y) where x1 ≤ x ≤ x2 and y2 < y ≤ n, change in rsum(1, 1 : x, y) is given by
c(y2 − y1 + 1)x − c(x1 − 1)(y2 − y1 + 1).
4. For all (x, y) where x2 < x ≤ n and y2 < y ≤ n, change in rsum(1, 1 : x, y) is given by c(x2 − x1 +
1)(y2 − y1 + 1).
We have to maintain functions with 4 terms, i.e., xy, y, x and an independent term. Thus, we maintain
4 2D BITs, namely bitxy, bitx, bity and biti. We use the function updater, devised in section 7.3 for
performing the required range-update operations in order to maintain the functions.
Below we give the pseudocode for the update function for 2D case.
10
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
Range-update on 2D BIT
function update((x1 , y1 ), (x2 , y2 ), c)
updater(bitxy, (x1 , y1 ), (x2 , y2 ), c)
updater(bitx, (x1 , y1 ), (x2 , y2 ), −c(y1 − 1))
updater(bitx, (x1 , y2 + 1), (x2 , n), c(y2 − y1 + 1))
updater(bity, (x1 , y1 ), (x2 , y2 ), −c(x1 − 1))
updater(bity, (x2 + 1, y1 ), (n, y2 ), c(x2 − x1 + 1))
updater(biti, (x1 , y1 ), (x2 , y2 ), c)
updater(biti, (x2 + 1, y1 ), (n, y2 ), −c(y1 − 1)(x2 − x1 + 1))
updater(biti, (x1 , y2 + 1), (x2 , n), −c(x1 − 1)(y2 − y1 + 1))
updater(biti, (x2 + 1, y2 + 1), (n, n), c(x2 − x1 + 1)(y2 − y1 + 1))
end function
Since, there are 4 2D BITs and operation on each 2D BIT takes O(log2 n), the net complexity per update
is given by 4 ∗ O(log2 n).
8.2.2
Range-query
In order to query the net change in the value fo rsum(1, 1 : x, y), we need to know the net function at
(x, y). We use the queryp function devised in section 7.3 to perform the point queries at (x, y).
We present the pseudocode for the query operation below.
Range-query on 2D BIT
1:
2:
3:
4:
5:
6:
7:
function query((x, y))
a ← queryp(bitxy, (x, y))
b ← queryp(bitx, (x, y))
c ← queryp(bity, (x, y))
d ← queryp(biti, (x, y))
return a ∗ x ∗ y + b ∗ x + c ∗ y + d
end function
Since, there are 4 2D BITs, the net complexity per query is given by 4 ∗ O(log2 n).
8.3
Generalization to higher dimensions
It is evident from the algorithm’s extension to matrices that it can be generalized to any number of
dimensions. The algorithm for the d-dimensional version can be constructed from the one for (d − 1)
dimensions.
Proposition 3. For handling range-update range-query operations on a d-dimensional array, 2d d-dimensional BITs are required.
Proof. We prove this by mathematical induction. We have shown that 2 BITs are required for handling
range-update range-query operations on a 1D array. We assume that 2k BITs are required to handle
operations on a k-dimensional array, for some whole number k > 1. We now take a (k + 1)-dimensional
array arr. The array arr can be understood as a 1D array, each of whose elements is a k-dimensional
array. As per our assumption, for updating an element of arr, we further need 2k k-dimensional BITs
per element (since each element is a k-dimensional array). Having a k dimensional BIT for each element
is same as having one (k + 1)-dimensional BIT. This implies that we need 2k (k + 1)-dimensional BITs
to perform update an element of arr. As shown previously, we need two BITs to perform range-update
range-query operations on a 1D array. As a result, in total, we need 2k+1 (k + 1)-dimensional BITs for
arr.
By principle of mathematical induction, since the assertion is true for 1, and also true for some whole
number k + 1 whenever its true for k, therefore, its true for the entire set of whole numbers. Therefore,
11
we need 2d d-dimensional BITs to perform range-update range-query on a d-dimensional array.
Alternatively, the conclusion can also be reached by realizing that a multi-linear function with d variables
has 2d coefficients. Therefore, 2d d-dimensional BITs are required.
An update or query on a d-dimensional BIT, with n elements along each dimension, requires O(logd n)
time. By Proposition 3, it is known that 2d such BITs are required to handle operations. Therefore, the
overall time complexity for each update/query operation is O(4d ∗ logd n).
9
Experimental comparison of running times
In this section, we provide an experimental comparison between the execution times of the old and new
algorithms. In the table below, we give the total time taken (in milliseconds) by the two algorithms to
perform 100, 000 update/query operations on 2D and 3D arrays of different side-lengths (n).
n
10
50
100
150
200
500
1000
4000
Old Algorithm
2D (Quadtree) 3D (Octree)
81
522
680
53051
888
219512
1387
488673
3184
910796
4420
8996
39126
New Algorithm
2D
3D
93
693
238
1923
309
3186
400
3992
431
6357
393
525
1644
Table 1: Comparison of running times
For the purpose of this experiment, we used tree-based implementation of quadtree and octree. The
parameters for updates and query operations — the sub-arrays to be updated/queried and the constants
for the update operations — were produced using the rand() function in the stdlib.h header file of C
Library. The parameters that we used were taken uniformly at random in the space of all parameters.
The implementations were done on a machine with Ubuntu 13.10 64-bit as the operating system, 8 GB
RAM and Intel Core i3 3.1 GHz Sandy Bridge processor.
Below, we plot the graph for number of instructions (T ) vs. number of dimensions (d) for arrays of
side-lengths (n) equal to 100, 1000 and 10000.
15
×10 4
Number of instructions (T)
10
T = 100d−1
T = 4d · logd2 100
5
0
0
0.5
1
1.5
2
Dimensions (d)
(a) n = 100
12
2.5
3
3.5
4
10000
9000
8000
Number of instructions (T)
7000
6000
T = 1000d−1
T = 4d · logd2 1000
5000
4000
3000
2000
1000
0
0
0.5
1
1.5
2
2.5
3
2
2.5
3
Dimensions (d)
(a) n = 1000
10000
9000
8000
Number of instructions (T)
7000
6000
T = 10000d−1
T = 4d · logd2 10000
5000
4000
3000
2000
1000
0
0
0.5
1
1.5
Dimensions (d)
(a) n = 10000
Figure 1: Number of instructions (T ) vs. Number of dimensions (d)
From the graphs and the table of experimental running times, the following inferences can be made.
• The graph for the old algorithm rises more steeply as compared to the graph for the new algorithm.
• The new algorithm outperforms the old algorithm in the number of instructions per second and,
consequently, in time of execution.
• Efficiency of the new algorithm gets more pronounced as n increases.
10
Conclusion and future scope
The algorithm that this paper proposes significantly reduces the time required for handling on-line rangeupdate range-query operations on sub-arrays of multidimensional arrays. However, the memory needed
to execute the algorithm — as in the case with previously known algorithms — increases exponentially
with the number of dimensions. We still believe that our algorithm is of practical significance since the
number of dimensions rarely exceeds three in real world applications.
13
References
[1] Peter M. Fenwick. A New Data Structure for Cumulative Frequency Tables. In Software: Practice and
Experience, Volume 24, pages 327–336, 1994.
[2] Hanan Samet et al. A Geographic Information System Using Quadtree. In Pattern Recognition, Volume
17, pages 647–656, 1984.
[3] A. Klinger. Patterns and search statistics, Optimizing Methods in Statistics, J. S. Rustagi, ed., pp.
303-337. Academic Press, New York (1971).
[4] Mark de Berg, Marc van Kreveld, Mark Overmars, Otfried Schwarzkopf. Computational Geometry:
algorithms and applications (2nd ed.). Springer-Verlag Berlin Heidelberg New York, ISBN 3-54065620-0, 2000.
[5] Hanan Samet. An Overview of Quadtrees, Octrees, and Related Hierarchical Data Structure. In Theoretical Foundations of Computer Graphics and CAD, Volume F40, 1988.
[6] R.A. Finkel, J.L. Bentley. Quadtrees: a data structure for retrieval on composite keys. In Acta Informatica, Volume 4, pages 1–9, 1974.
[7] Pinaki Mazumder. Planar Decomposition for Quadtree Data Structure. In Computer Vision, Graphics,
and Image Processing, Volume 38, pages 258–274, 1987.
[8] Sriram V. Pemmaraju, Clifford A. Shaffer. Department of Computer Science Virginia Polytechnic
Institute and State University, USA. Analysis of the Worst-Case Space Complexity of a PR Quadtree,
1992
[9] Hanan Samet. Using Quadtrees to Represent Spatial Data. In Computer Architectures for Spatially
Distributed Data, Volume F18, pages 229–247, 1985.
14
| 8cs.DS
|
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS:
PREDICTIONS USING STATISTICAL PHYSICS
arXiv:1803.11132v1 [stat.ML] 29 Mar 2018
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
In memory of Amelia Perry, and her love for learning and teaching.
Abstract. In these notes we describe heuristics to predict computational-tostatistical gaps in certain statistical problems. These are regimes in which the
underlying statistical problem is information-theoretically possible although
no efficient algorithm exists, rendering the problem essentially unsolvable for
large instances. The methods we describe here are based on mature, albeit
non-rigorous, tools from statistical physics.
These notes are based on a lecture series given by the authors at the Courant
Institute of Mathematical Sciences in New York City, on May 16th , 2017.
1. Introduction
Statistics has long studied how to recover information from data. Theoretical
statistics is concerned with, in part, understanding under which circumstances such
recovery is possible. Oftentimes recovery procedures amount to computational
tasks to be performed on the data that may be computationally expensive, and so
prohibitive for large datasets. While computer science, and in particular complexity
theory, has focused on studying hardness of computational problems on worst-case
instances, time and time again it is observed that computational tasks on data can
often be solved far faster than the worst case complexity would suggest. This is
not shocking; it is simply a manifestation of the fact that instances arising from
real-world data are not adversarial. This illustrates, however, an important gap
in fundamental knowledge: the understanding of “computational hardness of
statistical estimation problems”.
For concreteness we will focus on the case where we want to learn a set of parameters from samples of a distribution, or estimate a signal from noisy measurements
(often two interpretations of the same problem). In the problems we will consider,
there is a natural notion of signal-to-noise ratio (SNR) which can be related to the
variance of the distribution of samples, the strength of the noise, the number of
samples or measurements obtained, the size of a hidden planted structure buried
in noise, etc. Two “phase transitions” are often studied. Theoretical statistics
and information theory often study the critical SNR below which it is statistically
impossible to estimate the parameters (or recover the signal, or find the hidden
ASB was partially supported by NSF DMS-1712730 and NSF DMS-1719545.
AP was supported in part by NSF CAREER Award CCF-1453261 and a grant from the MIT
NEC Corporation.
ASW received Government support under and awarded by DoD, Air Force Office of Scientific
Research, National Defense Science and Engineering Graduate (NDSEG) Fellowship, 32 CFR
168a.
1
2
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
structure), and we call this threshold SNRStat . On the other hand, many algorithm
development fields propose and analyze efficient algorithms to understand for which
SNR levels different algorithms work. Despite significant effort to develop ever better algorithms, there are various problems for which no efficient algorithm is known
to achieve recovery close to the statistical threshold SNRStat . Thus we are interested in the critical threshold SNRComp ≥ SNRStat below which it is fundamentally
impossible for an efficient (polynomial time) algorithm to recover the information
of interest.
There are many problems believed to exhibit computational-to-statistical gaps.
Examples include community detection [HLL83, DKMZ11, AS15, BMNN16], planted
clique [AKS98, DM15, BHK+ 16], sparse principal component analysis [BR12, BR13,
LKZ15b], structured spiked matrix models [LKZ15a, PWBM16b, KXZ16, BDM+ 16,
LM16], spiked tensor models [RM14, HSS15, PWB16, LML+ 17, KBG17], and synchronization problems over groups [Sin11, PWBM16b, PWBM16a].
In these notes we will be concerned with predicting the locations of the thresholds SNRStat and SNRComp for Bayesian inference problems. In particular, we will
focus on a couple of heuristics borrowed from statistical physics and illustrate them
on two example problems: the Rademacher spiked Wigner problem (Example 2.1)
and the related problem of community detection in the stochastic block model (Example 2.2). While we focus on these problems, we will try to cover the techniques
in a way that conveys how they are broadly applicable.
At first glance, it may seem surprising that statistical physics has anything to do
with Bayesian inference problems. The connection lies in the Gibbs (or Boltzmann)
distribution that is widely used in statistical physics to model disordered systems
such as magnets. It turns out that in many Bayesian inference problems, the
posterior distribution of the unknown signal given the data also follows a Gibbs
distribution, and thus many techniques from statistical physics can be applied.
More specifically, many inference problems follow similar equations to spin glasses,
which are physical systems in which the interaction strength between each pair of
particles is random. The techniques that we borrow from statistical physics are
largely non-rigorous but yield extremely precise predictions of both the statistical
and computational limits. Furthermore, the predictions made by these heuristics
have now been rigorously verified for many problems, and thus we have good reason
to trust them on new problems. See the survey [ZK16] for more on the deep
interplay between statistical physics and inference.
Many techniques have been developed in order to understand computational-tostatistical gaps. We now give a brief overview of some of these methods, both the
ones we will cover in these notes and some that we will not.
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
3
Reductions. A natural approach to arguing that a task is computationally hard is
via reductions, by showing that a problem is computationally hard conditioned on
another problem being hard. This technique is extremely effective when studying
the worst-case hardness of computational problems (a famous example being the
list of 21 NP-hard combinatorial problems of Karp [Kar72]). There are also some
remarkable successes in using this idea in the context of average-case problems
(i.e. statistical inference on random models), starting with the work of Berthet
and Rigollet on sparse PCA [BR12, BR13] and including also some conditional
lower bounds for community detection with sublinear sized communities [MW13,
HWX14]. These works show conditional hardness by reduction to the planted clique
problem, which is widely believed to be hard in certain regimes. Unfortunately this
method of reductions has so far been limited to problems that are fairly similar to
planted clique.
Sum-of-squares hierarchy. Sum-of-squares [Las01, Par00, Nes00, Sho87, BS14] is
a hierarchy of algorithms to approximate solutions of combinatorial problems, or
more generally, polynomial optimization problems. For each positive integer d, the
algorithm at level d of the hierarchy is a semidefinite program that relaxes the
notion of a distribution over the solution space by only keeping track of moments
of order ≤ d. As you go up the hierarchy (increasing d), the algorithms get more
powerful but also run slower: the runtime is nO(d) . The level-2 relaxation coincides
with the algorithms in the seminal work of Goemans and Williamson [GW95] and
Lovasz [Lov79]. The celebrated unique games conjecture of Khot implies that the
level-2 algorithm gives optimal worst-case approximation ratio for a wide class
of problems [Kho02, Rag08, Kho10]. Sum-of-squares algorithms have also seen
many success stories for average-case inference problems such as planted sparse
vector [BKS13, HSSS16], dictionary learning [BKS15], tensor PCA [HSS15], tensor
decomposition [BKS15, GM15, HSSS16, MSS16], and tensor completion [BM16,
PS17]. One way to argue that an inference problem is hard is by showing that the
sum-of-squares hierarchy fails to solve it at a particular level d (or ideally, at every
constant value of d). Such lower bounds have been shown for many problems such as
planted clique [BHK+ 16] and tensor PCA [HSS15]. There is also recent work that
gives evidence for computational hardness by relating the power of sum-of-squares
to the low-degree moments of the posterior distribution [HS17].
Belief propagation, approximate message passing, and the cavity method. Another
important heuristic to predict computational thresholds is based on ideas from
statistical physics and is often referred to as the cavity method [MPV86]. It is
based on analyzing an iterative algorithm called belief propagation (BP) [Pea86], or
its close relative approximate message passing (AMP) [DMM09]. Specifically, BP
has a trivial fixed point wherein the algorithm fails to perform inference. If this
fixed point is stable (attracting) then we expect inference to be computationally
hard. In these notes we will cover this method in detail. For further references, see
[MM09, ZK16].
Replica method and the complexity of the posterior. Another method borrowed from
statistical physics is the replica method (see e.g. [MM09]). This is a mysterious nonrigorous calculation from statistical physics that can produce many of the same
predictions as the cavity method. One way to think about this method is as a
way to measure the complexity of the posterior distribution. In particular, we
4
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
are interesting in whether the posterior distribution resembles one big connected
region or whether it fractures into disconnected clusters (indicating computational
hardness). We will cover the replica method in Section 5 of these notes.
Complexity of a random objective function. Another method for investigating computational hardness is through the lens of non-convex optimization. Intuitively, we
expect that “easy” optimization problems have no “bad” local minima and so an
algorithm such as gradient descent can find the global minimum (or at least a point
whose objective value is close to the global optimum). For Bayesian inference problems, maximum likelihood estimation amounts to minimizing a particular random
non-convex function. One tool to study critical points of random functions is the
Kac-Rice formula (see [AT07] for an introduction). This has been used to study
optimization landscapes in settings such as spin glasses [ABAČ13], tensor decomposition [GM17], and problems arising in community detection [BBV16]. There
are also other methods to show that there are no spurious local minima in certain
settings, e.g. [GJZ17, BVB16, LV18].
2. Setting and vocabulary
Throughout, we’ll largely focus on Bayesian inference problems. Here we have a
signal σ ∗ ∈ Rn viewed through some noisy observation model. We present two examples, and examine them simultaneously through the parallel language of machine
learning and statistical physics.
Example 2.1 (Rademacher spiked Wigner). The signal σ ∗ is drawn uniformly at
random from {±1}n . We observe the n × n matrix
Y =
1
λ ∗ ∗ >
σ (σ ) + √ W,
n
n
where λ is a signal-to-noise parameter, and W is a GOE matrix1. We wish to
approximately recover σ ∗ from Y , up to a global negation (since σ ∗ and −σ ∗ are
indistinguishable).
This problem is motivated by the statistical study of the spiked Wigner model
from random matrix theory (see e.g. [PWBM16b]). This model has also been
studied as a Gaussian variant of community detection [DAM16] and as a model for
synchronization over the group Z/2 [JMRT16].
Example 2.2 (Stochastic block model). The signal σ ∗ is drawn uniformly at random from {±1}n . We observe a graph G with vertex set [n] = {1, . . . , n}, with
edges drawn independently as follows: for vertices u, v, we have u ∼ v with probability a/n if σu σv = 1, and probability b/n if σu σv = −1. We will restrict ourselves
to the case a > b. We imagine the entries σu∗ as indicating membership of vertex u
in either the +1 or −1 ‘community’; thus vertices in the same community are more
likely to share an edge. We wish to approximately recover the community structure
σ ∗ (up to global negation) from G.
This is a popular model for community detection in graphs. See e.g. [Abb17,
Moo17] for a survey. Here we consider the sparse regime, but other regimes are
also considered in the literature.
1Gaussian orthogonal ensemble: symmetric with the upper triangle drawn i.i.d. as N (0, 1).
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
5
There is a key difference between the two models above. The Rademacher spiked
Wigner model is dense in the sense that we are given an observation for every pair
of variables. On the other hand, the stochastic block model is sparse (at least in the
regime we have chosen) because essentially all the useful information comes from
the observed edges, which form a sparse graph. We will see that different tools are
needed for dense and sparse problems.
2.1. Machine learning view. We are interested in inferring the signal σ ∗ , so
it is natural to write down the posterior distribution. For the Rademacher spiked
Wigner problem (Example 2.1), we can compute the posterior distribution explicitly
as follows:
2 !
Y
λ
n
Yij − σi σj
exp −
Pr[σ | Y ] ∝ Pr[Y | σ] ∝
2
n
i<j
Y
n
λ2 2 2
=
exp − Yij2 + λYij σi σj −
σi σj
2
2n
i<j
Y
∝
exp (λYij σi σj ) .
i<j
(HereP∝ hides a normalizing constant which depends on Y but not σ; it is chosen so
that σ∈{±1}n Pr[σ | Y ] = 1.) The above factorization over edges defines a graphQ
ical model: a probability distribution factoring in the form Pr[σ] = S⊂[n] ψS (σS )
into potentials ψS that each only depend on a small (constant-size) subset S of
the entries of σ. (For instance, in our example above, S ranges over all subsets of
size 2.)
2.2. Statistical physicsPview. The observation Y defines a Hamiltonian, or
energy function, H(σ) = i<j Yij σi σj , consisting of two-spin interactions; we refer
to each entry of σ as a spin, and to σ as a state. A Hamiltonian together with
a parameter T = β1 , called the temperature, defines a Gibbs distribution (or
Boltzmann distribution):
Pr[σ] ∝ e−βH(σ) .
Thus low-energy states are more likely than high-energy states; moreover, at low
temperature (large β), the distribution becomes more concentrated on lower energy states, becoming supported entirely on the minimum energy states (ground
states) in the limit as β → ∞. On the other hand, in the high-temperature limit
(β → 0), the Gibbs distribution becomes uniform.
Connecting the ML and physics languages, we observe that the posterior distribution on σ is precisely the above Gibbs distribution, at the particular inversetemperature β = λ. (This is often referred to as lying on the Nishimori line, or
being at Bayes-optimal temperature.)
2.3. Optimization and statistical physics. A common optimization viewpoint
on inference is maximum likelihood estimation, or the maximization task of
finding the state σ that maximizes the posterior likelihood. This optimization
problem is frequently computationally hard, but convex relaxations or surrogates
may be studied. To rephraze this optimization task in physical terms, we wish to
6
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
minimize the energy H(σ) over states σ, or equivalently sample from (or otherwise
describe) the low-temperature Gibbs distribution in the limit β → ∞.
This viewpoint is limited, in that the MLE frequently lacks any a priori guarantee of optimality. On the other hand, the Gibbs distribution at the true temperature
β = λ enjoys optimality guarantees at a high level of generality:
Claim 2.3. Suppose we are given some observation Y leading to a posterior distribution on σ. For any estimate σ
b = σ
b(Y ), define the (expected) mean squared
error (MSE) Ekb
σ − σk22 . The estimator that minimizes the expected MSE is given
by σ
b = E[σ | Y ], the posterior expectation (and thus the expectation under the Gibbs
distribution at Bayes-optimal temperature).
Remark 2.4. In the case of the Rademacher spiked Wigner model, there is a
caveat here: since σ ∗ and −σ ∗ are indistinguishable, the posterior expectation is
zero. Our objective is not to minimize the MSE but to minimize the error between
σ and either σ
b or −b
σ (whichever is better).
Thus the optimization approach of maximum likelihood estimation aims for too
low a temperature. aggregate likelihood. Intuitively, MLE searches for the single
state with highest individual likelihood, whereas the optimal Bayesian approach
looks for a large cluster of closely-related states with a high aggregate likelihood.
Fortunately, the true Gibbs distribution has an optimization property of its own:
Claim 2.5. The Gibbs distribution with Hamiltonian H and temperature T > 0 is
the unique distribution minimizing the (Helmholtz) free energy
F = EH − T S,
where S denotes the Shannon entropy S = −Eσ log Pr(σ).
Proof. Entropy is concave with infinite derivative at the edge of the probability
simplex, and the expected Hamiltonian is linear in the distribution, so the free
energy is convex and minimized in the interior of the simplex. We find the unique
local (hence global) minimum with a Lagrange multiplier:
X
const · 1 = ∇F = ∇
p(σ)(H(σ) + T log p(σ))
σ
const = H(σ) + T log p(σ)
const · e−H(σ)/T = p(σ),
which we recognize as the Gibbs distribution.
This optimization approach is willing to trade off some energy for an increase in
entropy, and can thus detect large clusters of states with a high aggregate likelihood,
even when no individual state has the highest possible likelihood. Moreover, the
free energy is convex, but it is a function of an arbitrary probability distribution
on the state space, which is typically an exponentially large object.
We are thus led to ask the question: is there any way to reduce the problem of
free energy minimization to a tractable, polynomial-size problem? Can we get a
theoretical or algorithmic handle on this problem?
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
7
3. The cavity method and belief propagation
3.1. BP as an algorithm for inference. Belief propagation (BP) is a general
algorithm for inference in graphical models, generally credited to Pearl [Pea86] (see
e.g. [MM09] for a reference). As we’ve seen above, the study of graphical models
is essentially the statistical physics of Hamiltonians consisting of interactions that
each only depend on a few spins. Quite often, we care about the average case
study of random graphical models that describe a posterior distribution given some
noisy observation of a signal, such as in the Rademacher spiked Wigner example
discussed above. Much of statistical physics is concerned with disorder and random
systems, and indeed the concept of belief propagation appeared in physics as the
cavity method—not only as an algorithm but as a theoretical means to make
predictions about systems such as spin glasses [MPV86].
To simplify the setting and notation, let us consider sparse graphical models
with only pairwise interactions:
Y
Pr[σ] ∝
ψuv (σu , σv ),
u∼v
where each vertex v has only relatively few “neighbors” u (denoted u ∼ v).
Belief propagation is an iterative algorithm. We think of each spin σu as a
vertex and each pair of neighbors as an edge. Each vertex tracks a “belief” about
its own spin (more formally, an estimated posterior marginal). These beliefs are
often initialized to something like a prior distribution, or just random noise, and
then iteratively refined to become more consistent with the graphical model. This
refinement happens as follows: each vertex u transmits its belief to each neighbor,
and then each vertex updates its belief based on the incoming beliefs of its neighbors.
If we let mu→v denote the previous beliefs sent from neighbors u to a vertex v, we
can formulate a new belief for v in a Bayesian way, assuming that the incoming
influences of the neighbor vertices are independent (more on this assumption below):
YX
mv (σv ) ∝
ψuv (σu , σv )mu→v (σu )
u∼v σu
Each message mu→v is a probability distribution (over the possible values for σu ),
with the proportionality constant being determined by probabilities summing to 1
over all values of σu .
This is almost a full description of belief propagation, except for one detail. If
the belief from vertex v at time t−2 influences the belief of neighbor u at time t−1,
then neighbor u should not parrot that influence back to neighbor v, reinforcing its
belief at time t without any new evidence. Thus we ensure that the propagation of
messages does not immediately backtrack:
YX
(t−1)
(1)
m(t)
ψuv (σu , σv )mu→v
(σu ).
v→w (σv ) ∝
u∼v σu
u6=w
This formula is the iteration rule for belief propagation.
The most suspicious aspect of the discussion above is the idea that neighbors of a
vertex v exert probabilistically independent influences on v. If the graphical model
is a tree, then the neighbors are independent after conditioning on v, and in this
setting it is a theorem (see e.g. [MM09]) that belief propagation converges to the
exact posterior marginals. On a general graphical model, this independence fails,
and belief propagation is heuristic. In many sparse graph models, neighborhoods of
8
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
most vertices are trees, with most loops being long, so that independence might approximately hold. BP certainly fails in the worst case; outside of special cases such
as trees it is certainly only suitable in an average-case setting. However, on many
families of random graphical models, belief propagation is a remarkably strong approach; it is general, efficient, and often yields a state-of-the-art statistical estimate.
It is conjectured in many models that belief propagation achieves asymptotically
optimal inference, either among all estimators or among all polynomial-time estimators, but most rigorous results in this direction are yet to be established.
To connect to the previous viewpoint of free energy minimization: belief propagation is intimately connected with the Bethe free energy, a heuristic proxy
for the free energy which may be described in terms of the messages mu→v (see
[ZK16], Section III.B). It can be shown that the fixed points of BP are precisely
the critical points of the Bethe free energy, justifying the view that BP is roughly a
minimization procedure for the free energy. Again, rigorously the situation is much
worse: the Bethe free energy is non-convex, and BP is not guaranteed to converge,
let alone guaranteed to find the global minimum.
3.2. The cavity method for the stochastic block model. The ideas of belief
propagation above appear as the cavity method in statistical physics, owing to
the idea that the Bethe free energy is believed to be essentially an accurate model
for the true (Helmholtz) free energy on a variety of models of interest. In passing
to the Bethe free energy, we can pass from studying a general distribution (an
exponentially complicated object) to studying node and edge marginals, which are
theoretically much simpler objects and, crucially, can be studied locally on the
graph. Local neighborhoods of sparse graphs as in the SBM (stochastic block
model) look like trees, and so we are drawn to studying message passing on a tree.
Much as for the Rademacher spiked Wigner model above, we derive a Hamiltonian from the block model posterior:
X
X
H(σ) =
θ + σi σj +
θ − σi σj ,
i∼j
i6∼j
where u ∼ v denotes adjacency in the observed graph, and θ+ > 0 > θ− are
constants depending on a and b; θ+ is of constant order, while θ− is of order 1/n.
In expressing belief propagation, we will make a small notational simplification:
instead of passing messages m that are distributions over {+, −}, it suffices to pass
the expectation m(+) − m(−). The reader can verify that rewriting the belief
propagation equations in this notation yields
X
X
(t−1)
m(t)
atanh(θ+ mw→u
)+
atanh(θ− m(t−1)
u→v = tanh
w→u )
w∼u
w6=v
w6∼u
w6=v
where tanh is the hyperbolic tangent function tanh(z) = (ez − e−z )/(ez + e−z ), and
atanh is its inverse.
The first term inside the tanh represents strong, constant-order attractions with
the few graph neighbors, while the second term represents very weak, low-order
repulsions with the multitude of non-neighbors. The value of the second term thus
depends very little on any individual spin, but rather on the overall balance of
positive and negative spins in the graph, with the tendency to cause the global
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
9
spin configuration to become balanced. As we are only interested in a local view
of message passing, we will assume here that the global configuration is roughly
balanced and neglect the second term:
X
(t−1)
m(t)
atanh(θ+ mw→u
) .
u→v ≈ tanh
w∼u
w6=v
As this message-passing only involves the graph edges, it now makes sense to
study this on a tree-like neighborhood. We now discuss a generative model for
(approximate) local neighborhoods under the stochastic block model.
Model 3.1 (Galton–Watson tree). Begin with a root vertex, with spin + or −
chosen uniformly. Recursively, each vertex gives birth to a Poisson number of child
nodes: Pois((1 − ε)k) vertices of the same spin and Pois(εk) vertices of opposite
spin, up to a total tree depth of d.
As shown in [MNS12], the Galton–Watson tree with k = (a + b)/2 and ε =
b/(a + b) is distributionally very close to the radius-d neighborhood of a vertex in
the SBM with its true spins, so long as d = o(log n). Thus we will study belief
propagation on a random Galton–Watson tree.
Let us consider only the BP messages passing toward the root of the tree. The
upward message from any vertex v is computed as:
!
X
(2)
mv = tanh
atanh((1 − 2ε)mu )
u
where u ranges over the children of v. We now imagine that the child messages
(t−1)
for children with spin
mu are independently drawn from some distribution D+
(t−1)
(t−1)
for children
= −D+
+, and (leveraging symmetry) from the distribution D−
with spin −; this distribution represents the randomness of our BP calculation below
each child, over the random subtree hanging off each one. Then, from equation (2),
together with the fact that there are Pois((1 − ε)k) same-spin children and Pois(εk)
(t)
opposite-spin children, the distribution D± of the parent message m is determined!
(t)
Thus we obtain a distributional recurrence for D+ .
The calculation above is independent of n, and the radius of validity of the tree
approximation grows with n, so we are interested in the behavior of the recurrence
above as t → ∞, i.e. fixed points of the distributional recurrence above and their
stability.
Typically one initializes BP with small random messages, a perturbation of the
trivial all-0 fixed point that represents our prior. For
P small messages, we can
linearize tanh and atanh, and write mv ≈ (1 − 2ε) u mu . Then if the child
(t−1)
distribution D+
has mean µ and variance σ 2 , it is easily computed that the
(t)
parent distribution D+ has mean k(1 − 2ε)2 µ and variance k(1 − 2ε)2 σ 2 . Thus
2
if k(1 − 2ε) < 1, then perturbations of the all-0 fixed point decay, or in other
words, this fixed point is stable, and BP is totally uninformative on this typical
initialization. If k(1 − 2ε)2 > 1, then small perturbations do become magnified
under BP dynamics, and one imagines that BP might find a more informative fixed
point (though this remains an open question!).
10
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
This transition is known as the Kesten–Stigum threshold [KS66], and calculations of this form are loosely conjectured to describe the computational threshold beyond which no efficient algorithm can perform inference, for sparse models
such as the SBM. In this particular form of the SBM, this idea has been rigorously vindicated using techniques slightly different from BP: inference is known
to be statistically impossible when k(1 − 2ε)2 < 1 (meaning that any estimator
has zero correlation with the truth as n → ∞) [MNS12], and efficiently possible
when k(1 − 2ε)2 > 1 (meaning that asymptotically nonzero correlation is possible)
[Mas14, MNS13].
One might also endeavor to study the other fixed points of BP, not just the
trivial fixed points. This is a difficult undertaking in most situations, as the BP
recurrence lacks convexity properties, but it is expected to give an understanding of
the statistical threshold of the problem, i.e. the limit below which even inefficient
inference techniques fail. This has been rigorously proven for some variants of
the stochastic block model [COKPZ16]. Intuitively, exploring the BP landscape
by brute force for the best (in terms of Bethe free energy) BP fixed point is a
statistically optimal inference technique. For more general stochastic block models
with 4 or more communities, there exists a gap between the statistical threshold
and the analogous Kesten–Stigum bound [DKMZ11, AS15, BMNN16].
4. Approximate message passing
4.1. AMP as a simplification of BP. Our cavity analysis of the block model
above was well-adapted to sparse models, in which the analysis localizes onto a
tree of constant average degree. But many models, such as the Rademacher spiked
Wigner model, are dense and their analysis cannot be local. Thankfully, many of
these models are amenable to analysis for different reasons: as each vertex is acted
on by a large number of individually weak influences, the quantities of interest
in belief propagation are subject to central limit theorems and concentration of
measure. In this section we will demonstrate this on the Rademacher spiked Wigner
example.
P
Recall the Hamiltonian H = i<j Yij σi σj and inverse temperature λ. As in
the SBM discussion above, we can summarize BP messages by the expectation
m(+) − m(−). Then BP for this model reads as
X
(t−1)
m(t)
atanh(λYwu mw→u
) .
u→v = tanh
w6=v
(t−1)
We next exploit the weakness of individual interactions. Note that the values mw→u
lie in [−1, 1], while Ywu is of order n−1/2 in probability. Taylor-expanding atanh,
we simplify:
X
(t−1)
m(t)
λYwu mw→u
+ O(n−1/2 ) w.h.p.
u→v = tanh
w6=v
We next simplify the non-backtracking nature of BP. Naı̈vely, one might expect
that we can simply drop the condition w 6= v from the sum above, as the contribution from vertex v in the above sum should be only of size n−1/2 . As our formula
(t)
for mu→v would then no longer depend on v, we could write down messages indexed
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
11
by a single vertex:
!
m(t)
u
= tanh
X
(t−1)
λYwu mw
,
w
or in vector notation,
(3)
m(t) = tanh(λY m(t−1) ),
where tanh applies entrywise. This resembles the “power iteration” iterative algorithm to compute the leading eigenvector of Y :
m(t) = Y m(t−1) ,
but with tanh(λ • ) providing some form of soft projection onto the interval [−1, 1],
exploiting the entrywise ±1 structure.
Unfortunately, the non-backtracking simplification above is flawed, and equation
(3) does not accurately summarize BP or provide as strong an estimator. The problem is that the terms we have neglected add up constructively over two iterations.
(t−2)
on each neighbor
Specifically: consider that vertex v exerts an influence λYvu mv
(t−1)
(scaled
u; this small perturbation translates directly to a perturbation of mu
(t)
by a derivative of tanh). At the next iteration, vertex u influences mv according
(t−2)
(t−1)
2
,
mv
; the total contribution from backtracking here is thus λ2 Yvu
to λYvu mu
scaled through some derivatives of tanh. This influence is a random, positive, order
(t−2)
. Summing over all neighbors u, we realize that the aggregate
1/n multiple of mv
contribution of backtracking over two steps is in fact of order 1.
Thankfully, this contribution is also a sum of small random variables, and exhibits concentration of measure. The solution is thus to subtract off this aggregate
backtracking term in expectation, adding a correction called the Onsager reaction
term:
(4)
m(t) = tanh Y m(t−1) − λ2 (1 − kmk22 /n)m(t−2) .
This iterative algorithm is known as approximate message passing. The
simplifications above to BP first appeared in the work of Thouless, Anderson, and
Palmer [TAP77], who used it to obtain a theoretical handle on spin glasses at
high temperature. These techniques later appeared as an algorithm for compressed
sensing [DMM09]. The AMP algorithm (4) for this problem was introduced (stated
slightly differently) in [DAM16], and AMP has been applied to many other problems
such as sparse PCA [DM14], non-negative PCA [MR16], planted clique [DM15], and
synchronization over groups [PWBM16a] (just to name a few).
4.2. AMP state evolution. In contrast to belief propagation, approximate message passing (AMP) algorithms tend to be amenable to exact analysis in the limit
n → ∞. Here we introduce state evolution, a simple heuristic argument for the
analysis of AMP that has been proven correct in many settings. The idea of state
evolution was first introduced by [DMM09], based on ideas from [Bol12]; it was
later proved correct in various settings [BM11, JM13].
We will focus again on the Rademacher spiked Wigner model: we observe
Y =
λ >
1
xx + √ W
n
n
12
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
where x ∈ {±1}n is the true signal (drawn uniformly at random) and the n × n
noise matrix W is symmetric with the upper triangle drawn i.i.d. as N (0, 1). In
this setting, the AMP algorithm and its analysis are due to [DAM16].
We have seen above that the AMP algorithm for this problem takes the form
v t+1 = Y f (v t ) + [Onsager]
where f (v) denotes entrywise application of the function f (v) = tanh(λv). (Here
we abuse notation and let f refer to both the scalar function and its entrywise
application to a vector.) The superscript t indexes timesteps of the algorithm (and
is not to be confused with an exponent). The details of the Onsager term, discussed
previously, will not be important here.
The state evolution heuristic proceeds as follows. Postulate that at timestep t,
AMP’s iterate v t is distributed as
v t = µt x + σt g
(5)
where g ∼ N (0, I).
t
This breaks down v into a signal term (recall x is the true signal) and a noise
term, whose sizes are determined by parameters µt ∈ R and σt ∈ R≥0 . The idea
of state evolution is to write down a recurrence for how the parameters µt and σt
evolve from one timestep to the next. In performing this calculation we will make
two simplifying assumptions that will be justified later: (1) we drop the Onsager
term, and (2) we assume the noise W is independent at each timestep (i.e. there
is no correlation between W and the noise g in the current iterate). Under these
assumptions we have
1
λ >
xx + √ W f (v t )
v t+1 = Y f (v t ) =
n
n
1
λ
= hx, f (v t )i x + √ W f (v t )
n
n
which takes the form of (5) with a signal term and a noise term. We therefore have
λ
λ
µt+1 = hx, f (v t )i = hx, f (µt x + σt g)i
n
n
≈ λ E [Xf (µt X + σt G)] with scalars X ∼ Unif{±1}, G ∼ N (0, 1)
X,G
= λ E[f (µt + σt G)]
since f (−v) = −f (v).
G
For the noise term, think of f (v t ) as fixed and consider the randomness over W .
Each entry of the noise term √1n W f (v t ) has mean zero and variance
X1
X1
(σ t+1 )2 =
f (vit )2 =
f (µt xi + σt gi )2
n
n
i
i
≈ λ E [f (µt X + σt G)2 ]
with scalars X, G as above
X,G
= E[f (µt + σt G)2 ]
again by symmetry of f .
G
We now have “state evolution” equations for µt+1 and σt+1 in terms of µt and σt .
Since we could arbitrarily scale our iterates v t without adding or losing information,
we really only care about the parameter γ , (µ/σ)2 . It is possible (see [DAM16])
to reduce the state evolution recurrence to this single parameter:
√
(6)
γt+1 = λ2
tanh(γt + γt G)
E
G∼N (0,1)
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
13
(where we have substituted the actual expression for f ).
We can analyze AMP as follows. Choose a small positive initial value γ0 and
iterate (6) until we reach a fixed point γ∞ . We then expect the output v ∞ of AMP
to behave like
(7)
v ∞ = µ∞ x + σ∞ g
2
where g ∼ N (0, I), µ∞ = γ∞ /λ, and σ∞
= γ∞ /λ2 . For the Rademacher spiked
Wigner model, this has in fact been proven to be correct in the limit n → ∞
[BM11, JM13]. Namely, when we run AMP (with the Onsager term and without
fresh noise W at each timestep), the output behaves like (7) in a particular formal
sense.
State evolution reveals a phase transition at λ = 1: when λ ≤ 1 we have γ∞ = 0
(so AMP has zero correlation with the truth as n → ∞) and when λ > 1 we have
γ∞ > 0 (so AMP achieves nontrivial correlation with the truth). Furthermore, from
(7) we can deduce the value of any performance metric (e.g. mean squared error)
at any signal-to-noise ratio λ. It has in fact been shown (for Rademacher spiked
Wigner) that the mean squared error achieved by AMP is information-theoretically
optimal [DAM16].
It is perhaps surprising that state evolution is correct, given the seeminglyquestionable assumptions we made in deriving it. This can be understood as follows.
Recall that we eliminated the Onsager term and assumed independent noise W at
each timestep. Also recall that the Onsager term is a correction that makes the
update step non-backtracking: a message sent across an edge at one iteration does
not affect the message sent back across the edge (in the opposite direction) at the
next iteration. It turns out that to leading order, using fresh noise at each timestep
is equivalent to using a non-backtracking update step. This is because the largest
effect of fresh noise is to terms where a particular noise entry Wij is used twice in a
row, i.e. backtracking steps. So the two assumptions we made actually cancel each
other out! Note that both of the two assumptions are crucial in making the state
evolution analysis tractable, so it is quite spectacular that we are able to make both
of these assumptions for free (and still get the correct answer)!
One caveat in the rigorous analysis of AMP is that it assumes an initialization
that has some nonzero correlation with the truth [DAM16]. In other words, we need
to assume that we start with some nonzero γ because if we start with γ = 0 we will
remain there forever. In practice this is not an issue; a small random initialization
suffices.
4.3. Free energy diagrams. In this section we will finally see how to predict
computational-to-statistical gaps (for dense problems)! Above we have seen how
to analyze a particular algorithm: AMP. In various settings it has been shown
that AMP is information-theoretically optimal. More generally, it is believed that
AMP is optimal among all efficient algorithms (for a wide class of problems). We
will now show how to use AMP to predict whether a problem should be easy,
(computationally) hard or (statistically) impossible. The ideas here originate from
[LKZ15a, LKZ15b].
Recall that the state of AMP is described by a parameter γ, where larger γ
indicates better correlation with the truth and γ = 0 means that AMP achieves
zero correlation with the truth. Also recall that the Bethe free energy is the quantity
that belief propagation (or AMP) is locally trying to minimize. It is possible to
14
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
analytically write down the function f (γ) which gives the (Bethe) free energy of
the AMP state corresponding to γ; in the next section, we will see one way to
compute f (γ). AMP can be seen as starting near γ = 0 and naively moving in
the direction of lowest free energy until it reaches a local minimum; the γ value at
this minimum characterizes AMP’s output. The information-theoretically optimal
estimator is instead described by the global minimum of the free energy (and this
has been proven rigorously in various cases [BDM+ 16, LM16]); this corresponds to
the inefficient algorithm that uses exhaustive search to find the AMP state which
globally minimizes free energy. Figure 1 illustrates how the free energy landscape
f (γ) dictates whether the problem is easy, hard, or impossible at a particular λ
value.
−0.2
−1.845
−0.4
−0.6
−1.850
f
f
−0.8
−1.855
−1.0
−1.2
−1.860
−1.4
−1.6
0.0
−1.865
0.5
1.0
1.5
2.0
γ
2.5
3.0
0.0
0.2
0.4
(a) impossible
0.6
0.8
γ
1.0
1.2
1.4
1.6
(b) impossible
−2.6
−1.896
−2.7
−1.898
−2.8
−1.902
f
f
−1.900
−2.9
−1.904
−3.0
−1.906
−3.1
−1.908
−1.910
0.0
0.2
0.4
0.6
0.8
γ
(c) hard
1.0
1.2
1.4
1.6
−3.2
0
1
2
γ
3
4
5
(d) easy
Figure 1. (a) The global minimizer is γ = 0 so no estimator achieves
nontrivial recovery. (b) A new local minimum in the free energy has appeared, but the global minimum is still at γ = 0 and so nontrivial recovery
remains impossible. (c) AMP is stuck at γ = 0 but the (inefficient) statistically optimal estimator achieves a nontrivial γ (the global minimum).
AMP is not statistically optimal. (d) AMP achieves nontrivial (in fact
optimal) recovery. The above image is adapted from [PWBM16a] (used
with permission).
For Rademacher spiked Wigner, we have phase (a) (from Figure 1) when λ ≤ 1
and phase (d) when λ > 1, so there is no computational-to-statistical gap. However,
for some variants of the problem (for instance if the signal x is sparse, i.e. only a
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
15
small constant fraction of entries are nonzero) then we see phases (a),(b),(c),(d)
appear in that order as λ increases; in particular, there is a computational-tostatistical gap during the hard phase (c).
Although many parts of this picture have been made rigorous in certain cases,
the one piece that we do not have the tools to prove is that no efficient algorithm
can succeed during the hard phase (c). This is merely conjectured based on the
belief that AMP should be optimal among efficient algorithms.
There are a few different ways to compute the free energy landscape f (γ). One
method is to use the replica method discussed in the next section. Alternatively,
there is a direct formula for Bethe free energy in terms of the BP messages, which
can be adapted to AMP (see e.g. [LKZ15a]).
5. The replica method
The replica method is an alternative viewpoint that can derive many of the same
results shown in the previous section. We will again use the example of Rademacher
spiked Wigner to illustrate it. A general introduction to the replica method can
be found in [MM09]. The calculations of this section are carried out in somewhat
higher generality in Appendix B of [PWB16].
Recall again the setup: we observe Y = nλ xx> + √1n W with x ∈ {±1}n and
Wij = Wji ∼ N (0, 1).
The posterior distribution of x given Y is
2 !
Y
X
n λ
Pr[x | Y ] ∝
exp −
xi xj − Yij
∝ exp λ
Yij xi xj
2 n
i<j
i<j
and so we are interested in the Gibbs distribution over σ ∈ P
{±1}n given by
Pr[σ | Y ] ∝ exp(−βH(σ)) with energy (Hamiltonian) H(σ) = − i<j Yij σi σj and
inverse temperature β = λ.
1
E log Z where
The goal is to compute the free energy density, defined as f = − βn
X
Z=
exp(−βH(σ)).
σ∈{±1}n
(This can be shown to coincide with the notion of free energy introduced earlier.)
The idea of the replica method is to compute the moments E[Z r ] of Z for r ∈ N
and perform the (non-rigorous) analytic continuation
1
log E[Z r ].
r
Note that this is quite bizarre – we at first assume r is a positive integer, but then
take the limit as r tends to zero! This will require writing E[Z r ] in an analytic form
that is defined for all values of r. An informal justification for the correctness of
(8) is that when r is close to 0, Z r is close to 1 and so we can interchange log and
E on the right-hand side.
The moment E[Z r ] can be expanded in terms of r ‘replicas’ σ 1 , . . . , σ r with
a
σ ∈ {±1}n :
r
X
X
X
Yij
σia σja .
E[Z r ] =
E exp β
(8)
E[log Z] = lim
r→0
{σ a }
i<j
a=1
16
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
After applying the definition of Y and the Gaussian moment-generating function
(to compute expectation over the noise W ) we arrive at
E[Z r ] =
X
{σ a }
2 X
2 X
λ
λ
2
exp n
c2 +
qab
2 a a
4
a,b
P
where qab = n1 i σia σib is the correlation between replicas a and b, and ca =
P
1
a
i σi xi is the correlation between replica a and the truth.
n
Without loss of generality we can assume (by symmetry) the true spike is x = 1
(all-ones). Let Q be the (r + 1) × (r + 1) matrix of overlaps (qab and ca ), including
x as the zeroth replica. Note that Q is the average of n i.i.d. matrices and so
by the theory of large deviations (Cramér’s Theorem in multiple dimensions), the
number of configurations {σ a } corresponding to given overlap parameters qab , ca is
asymptotically
(9)
X
X
X
X
X
1
1
inf exp n −
µab qab + log
exp
µab σa σb .
ν a ca −
νa σa +
µ,ν
2
2
r
a
a
a6=b
a6=b
σ∈{±1}
We now apply the saddle point method: in the large n limit, the expression for
E[Z r ] should be dominated by a single value of the overlap parameters qab , ca . This
yields
1
∗
log E[Z r ] = −G(qab
, c∗a , µ∗ab , νa∗ )
n
∗
where (qab
, c∗a , µ∗ab , νa∗ ) is a critical point of
X
λ2 X 2 λ2 X 2
1X
ca −
qab +
νa ca +
µab qab
2 a
4
2
a
a,b
a6=b
X
X
X
1
− log
exp
νa σa +
µab σa σb .
2
r
a
G(qab , ca , µab , νa ) = −
σ∈{±1}
a6=b
We next assume that the dominant saddle point takes a particular form: the socalled replica symmetric ansatz. The replica symmetric ansatz is given by qaa = 1,
ca = c, νa = ν, and for a 6= b, qab = q and µab = µ for constants q, c, µ, ν. This
yields
(10)
1
λ2
λ2 λ2
1
√
lim G(q, c, µ, ν) = − c2 − + q 2 +νc− µ(q−1)− E
log(2 cosh(ν+ µz))
r→0 r
2
4
4
2
z∼N (0,1)
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
17
where the last term is handled as follows:
X
X
X
1
1
lim log
exp
νa σa +
µab σa σb
r→0 r
2
a
a6=b
σ∈{±1}r
X
X
X
1
µ
= lim log
exp ν
σa +
σa σb
r→0 r
2
a
a6=b
σ∈{±1}r
X
X
X
1
µ
rµ
= lim log
exp ν
σa +
σa σb −
r→0 r
2
2
a
a,b
σ∈{±1}r
!2
X
X
X
1
µ
= lim log
exp(−rµ/2) exp ν
σa +
σa
r→0 r
2
a
a
σ∈{±1}r
!2
X
X
X
1
µ
µ
exp ν
σa +
σa
= − + lim log
2 r→0 r
2
a
a
σ∈{±1}r
!
X
X
µ
1
√ X
(a)
= − + lim log
exp ν
σa + µz
σa
E
2 r→0 r
z∼N (0,1)
r
a
a
σ∈{±1}
!
X
X
1
µ
√
exp (ν + µz)
σa
= − + lim log
E
2 r→0 r
z∼N (0,1)
r
a
σ∈{±1}
µ
1
√
√
r
= − + lim log
[exp(ν + µz) + exp(−(ν + µz))]
E
2 r→0 r
z∼N (0,1)
1
µ
√
r
= − + lim log
E (2 cosh(ν + µz))
2 r→0 r
z∼N (0,1)
µ
√
(b)
=− +
log(2 cosh(ν + µz))
E
2 z∼N (0,1)
where (a) uses the Gaussian moment-generating function and (b) uses the replica
trick (8).
We next find the critical points by setting the derivatives of (10) (with respect
to all four variables) to zero, which yields
√
√
ν = λ2 c,
µ = λ2 q,
c = Ez tanh(ν + µz),
q = Ez tanh2 (ν + µz).
Recall that the replicas are drawn from the posterior distribution Pr[x | Y ] and
so the truth x behaves as if it is a replica; therefore we should have c = q. Using
√
√
the identity Ez tanh(γ + γz) = Ez tanh2 (γ + γz) (see e.g. [DAM16]), we obtain
the solution c = q and ν = µ where q and µ are solutions to
√
(11)
µ = λ2 q,
q=
tanh(µ + µz).
E
z∼N (0,1)
The solution q to this equation tells us about the structure of the posterior distribution; namely, if we take two independent draws from this distribution, their
overlap will concentrate about q. (Equivalently, the true signal x and a draw from
the posterior distribution will also have overlap that concentrates about q.) Note
18
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
that (11) exactly matches the state evolution fixed-point equation (6) with µ in
place of γ and q = γ/λ2 .
The free energy density of a solution to (11) is given by
2
1
1
λ 2
1
1
√
− (q + 1) + µ(q + 1) − Ez log(2 cosh(µ + µz)) .
f = lim G(q, c, µ, ν) =
β r→0 r
λ
4
2
This is how one can derive the free energy curves such as those shown in Figure 1.
If there are multiple solutions to (11), we should take the one with minimum free
energy.
Above, we had a Gibbs distribution corresponding to the posterior distribution
of a Bayesian inference problem. In this setting, the replica symmetric ansatz is
always correct; this is justified by a phenomenon in statistical physics: “there is no
static replica symmetry breaking on the Nishimori line” (see e.g. [ZK16]).
More generally, one can apply the replica method to a Gibbs distribution that
does not correspond to a posterior distribution (e.g. if the ‘temperature’ of the
Gibbs distribution does not match the signal-to-noise of the observed data). This
is important when investigating computational hardness of random non-planted or
non-Bayesian problems. In this case, the optimal (lowest free energy) saddle point
can take various forms, which are summarized below; the form of the optimizer
reveals a lot about the structure of the Gibbs distribution. An important property
of a Gibbs distribution is its overlap distribution: the distribution of the overlap
between two independent draws from the Gibbs distribution (in the large n limit).
• RS (replica symmetric): The overlap matrix is qaa = 1 and qab = q for
some q ∈ [0, 1]. The overlap distribution is supported on a single point
mass at value q. The Gibbs distribution can be visualized as having one
large cluster where any two vectors in this cluster have overlap q. This
case is “easy” in the sense that belief propagation can easily move around
within the single cluster and find the true posterior distribution.
• 1RSB (1-step replica symmetry breaking): The r × r overlap matrix takes
the following form. The r replicas are partitioned into blocks of size m.
We have qaa = 1, qab = q1 if a, b are in the same block, and qab = q2
otherwise (for some q1 , q2 ∈ [0, 1]). The overlap distribution is supported
on q1 and q2 . The Gibbs distribution can be visualized as having a constant
number of clusters. Two vectors in the same cluster have overlap q1 whereas
two vectors in different clusters have overlap q2 . This case is “hard” for
belief propagation because it gets stuck in one cluster and cannot correctly
capture the posterior distribution.
• 2RSB (2-step replica symmetry breaking): Now we have “clusters of clusters.” The overlap matrix has sub-blocks within each block. The overlap
distribution is supported on 3 different values (corresponding to “same subblock”, “same block (but different sub-block)”, “different blocks”). The
Gibbs distribution has a constant number of clusters, each with a constant
number of sub-clusters. This is again “hard” for belief propagation.
• FSRB (full replica symmetry breaking): We can define kRSB for any k as
above (characterized by an overlap distribution supported on k + 1 values);
FRSB is the limit of kRSB as k → ∞. Here the overlap distribution is a
continuous distribution.
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
19
• d1RSB (dynamic 1RSB): This phase is similar to RS and (unlike kRSB for
k ≥ 1) can appear in Bayesian inference problems. The overlap matrix is
the same as in the RS phase (and so the replica calculation proceeds exactly
as in the RS case). However, the Gibbs distribution has exponentially-many
small clusters. The overlap distribution is supported on a single point mass
because two samples from the Gibbs distribution will be in different clusters
with high probability. This phase is “hard” for BP (or AMP) because it
cannot easily move between clusters. For a Bayesian inference problem, you
can tell whether you are in the RS (easy) phase or 1dRSB (hard) phase by
looking at the free energy curve; 1dRSB corresponds to the “hard” phase
(c) in Figure 1.
Acknowledgements. The authors would like to thank the engaging audience at
the Courant Institute when this material was presented, and their many insightful
comments. The authors would also like to thank Soledad Villar for feedback on an
earlier version of this manuscript.
References
[ABAČ13]
[Abb17]
[AKS98]
[AS15]
[AT07]
[BBV16]
[BDM+ 16]
[BHK+ 16]
[BKS13]
[BKS15]
[BM11]
[BM16]
[BMNN16]
[Bol12]
[BR12]
Antonio Auffinger, Gérard Ben Arous, and Jiřı́ Černỳ. Random matrices and
complexity of spin glasses. Communications on Pure and Applied Mathematics,
66(2):165–201, 2013.
Emmanuel Abbe. Community detection and stochastic block models: recent developments. arXiv preprint arXiv:1703.10146, 2017.
N. Alon, M. Krivelevich, and B. Sudakov. Finding a large hidden clique in a random
graph. SODA, 1998.
Emmanuel Abbe and Colin Sandon. Detection in the stochastic block model
with multiple clusters: proof of the achievability conjectures, acyclic bp, and the
information-computation gap. arXiv preprint arXiv:1512.09080, 2015.
R. J. Adler and J. E. Taylor. Random Fields and Geometry. Springer, 2007.
A. S. Bandeira, N. Boumal, and V. Voroninski. On the low-rank approach for semidefinite programs arising in synchronization and community detection. COLT, 2016.
Jean Barbier, Mohamad Dia, Nicolas Macris, Florent Krzakala, Thibault Lesieur,
and Lenka Zdeborová. Mutual information for symmetric rank-one matrix estimation: A proof of the replica formula. In Advances in Neural Information Processing
Systems, pages 424–432, 2016.
B. Barak, S. B. Hopkins, J. Kelner, P. K. Kothari, A. Moitra, and A. Potechin. A
nearly tight sum-of-squares lower bound for the planted clique problem. Available
online at arXiv:1604.03084 [cs.CC], 2016.
B. Barak, J. Kelner, and D. Steurer. Rounding sum-of-squares relaxations. Available
online at arXiv:1312.6652 [cs.DS], 2013.
Boaz Barak, Jonathan A Kelner, and David Steurer. Dictionary learning and tensor
decomposition via the sum-of-squares method. In Proceedings of the forty-seventh
annual ACM symposium on Theory of computing, pages 143–151. ACM, 2015.
Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense
graphs, with applications to compressed sensing. IEEE Transactions on Information
Theory, 57(2):764–785, 2011.
Boaz Barak and Ankur Moitra. Noisy tensor completion via the sum-of-squares
hierarchy. In Conference on Learning Theory, pages 417–445, 2016.
Jess Banks, Cristopher Moore, Joe Neeman, and Praneeth Netrapalli. Informationtheoretic thresholds for community detection in sparse networks. In Conference on
Learning Theory, pages 383–416, 2016.
Erwin Bolthausen. An iterative construction of solutions of the tap equations for the
sherrington-kirkpatrick model. arXiv preprint arXiv:1201.2891, 2012.
Q. Berthet and P. Rigollet. Optimal detection of sparse principal components in
high dimension. Annals of Statistics, 2012.
20
[BR13]
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
Q. Berthet and P. Rigollet. Complexity theoretic lower bounds for sparse principal
component detection. Conference on Learning Theory (COLT), 2013.
[BS14]
B. Barak and D. Steurer. Sum-of-squares proofs and the quest toward optimal algorithms. Survey, ICM 2014, 2014.
[BVB16]
N. Boumal, V. Voroninski, and A. S. Bandeira. The non-convex burer-monteiro
approach works on smooth semidefinite programs. NIPS, 2016.
[COKPZ16] Amin Coja-Oghlan, Florent Krzakala, Will Perkins, and Lenka Zdeborova.
Information-theoretic thresholds from the cavity method. arXiv preprint
arXiv:1611.00814, 2016.
[DAM16]
Yash Deshpande, Emmanuel Abbe, and Andrea Montanari. Asymptotic mutual information for the binary stochastic block model. In Information Theory (ISIT),
2016 IEEE International Symposium on, pages 185–189. IEEE, 2016.
[DKMZ11] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zdeborová. Inference and phase transitions in the detection of modules in sparse networks. Physical
Review Letters, 107(6):065701, 2011.
[DM14]
Yash Deshpande and Andrea Montanari. Information-theoretically optimal sparse
PCA. In Information Theory (ISIT), 2014 IEEE International Symposium on, pages
2197–2201. IEEE, 2014.
p
[DM15]
Yash Deshpande and Andrea Montanari. Finding hidden cliques of size
N/e in
nearly linear time. Foundations of Computational Mathematics, 15(4):1069–1128,
2015.
[DMM09]
David L Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for compressed sensing. Proceedings of the National Academy of Sciences,
106(45):18914–18919, 2009.
[GJZ17]
Rong Ge, Chi Jin, and Yi Zheng. No spurious local minima in nonconvex low rank
problems: A unified geometric analysis. arXiv preprint arXiv:1704.00708, 2017.
[GM15]
Rong Ge and Tengyu Ma. Decomposing overcomplete 3rd order tensors using sumof-squares algorithms. arXiv preprint arXiv:1504.05287, 2015.
[GM17]
R. Ge and T. Ma. On the optimization landscape of tensor decompositions. Available
online at arXiv:1706.05598 [cs.LG], 2017.
[GW95]
M. X. Goemans and D. P. Williamson. Improved approximation algorithms for maximum cut and satisfiability problems using semidefine programming. Journal of the
Association for Computing Machinery, 42:1115–1145, 1995.
[HLL83]
P. W. Holland, K. Blackmond Laskey, and S. Leinhardt. Stochastic blockmodels:
First steps. Social networks, 5(2):109–137, 1983.
[HS17]
Samuel B Hopkins and David Steurer. Bayesian estimation from few samples: community detection and related problems. arXiv preprint arXiv:1710.00264, 2017.
[HSS15]
S. B. Hopkins, J. Shi, and D. Steurer. Tensor principal component analysis via sumof-squares proofs. Available at arXiv:1507.03269 [cs.LG], 2015.
[HSSS16]
Samuel B Hopkins, Tselil Schramm, Jonathan Shi, and David Steurer. Fast spectral
algorithms from sum-of-squares proofs: tensor decomposition and planted sparse
vectors. In Proceedings of the forty-eighth annual ACM symposium on Theory of
Computing, pages 178–191. ACM, 2016.
[HWX14]
B. Hajek, Y. Wu, and J. Xu. Computational lower bounds for community detection
on random graphs. Available online at arXiv:1406.6625, 2014.
[JM13]
Adel Javanmard and Andrea Montanari. State evolution for general approximate
message passing algorithms, with applications to spatial coupling. Information and
Inference: A Journal of the IMA, 2(2):115–144, 2013.
[JMRT16]
Adel Javanmard, Andrea Montanari, and Federico Ricci-Tersenghi. Phase transitions in semidefinite relaxations. Proceedings of the National Academy of Sciences,
113(16):E2218–E2223, 2016.
[Kar72]
R. M. Karp. Reducibility among combinatorial problems. Complexity of Computer
Computation, 1972.
[KBG17]
C. Kim, A. S. Bandeira, and M. X. Goemans. Community detection in hypergraphs,
spiked tensor models, and sum-of-squares. SampTA 2017: Sampling Theory and
Applications, 12th international conference, 2017.
[Kho02]
S. Khot. On the power of unique 2-prover 1-round games. Thiry-fourth annual ACM
symposium on Theory of computing, 2002.
NOTES ON COMPUTATIONAL-TO-STATISTICAL GAPS
[Kho10]
[KS66]
[KXZ16]
[Las01]
[LKZ15a]
[LKZ15b]
[LM16]
[LML+ 17]
[Lov79]
[LV18]
[Mas14]
[MM09]
[MNS12]
[MNS13]
[Moo17]
[MPV86]
[MR16]
[MSS16]
[MW13]
[Nes00]
[Par00]
[Pea86]
[PS17]
[PWB16]
21
S. Khot. On the unique games conjecture (invited survey). In Proceedings of the
2010 IEEE 25th Annual Conference on Computational Complexity, CCC ’10, pages
99–121, Washington, DC, USA, 2010. IEEE Computer Society.
Harry Kesten and Bernt P Stigum. A limit theorem for multidimensional galtonwatson processes. The Annals of Mathematical Statistics, 37(5):1211–1223, 1966.
Florent Krzakala, Jiaming Xu, and Lenka Zdeborová. Mutual information in rankone matrix estimation. In Information Theory Workshop (ITW), 2016 IEEE, pages
71–75. IEEE, 2016.
J. B. Lassere. Global optimization with polynomials and the problem of moments.
SIAM Journal on Optimization, 11(3):796–817, 2001.
Thibault Lesieur, Florent Krzakala, and Lenka Zdeborová. Mmse of probabilistic
low-rank matrix estimation: Universality with respect to the output channel. In
53rd Annual Allerton Conference on Communication, Control, and Computing,
pages 680–687. IEEE, 2015.
Thibault Lesieur, Florent Krzakala, and Lenka Zdeborová. Phase transitions in
sparse pca. In Information Theory (ISIT), 2015 IEEE International Symposium
on, pages 1635–1639. IEEE, 2015.
Marc Lelarge and Léo Miolane. Fundamental limits of symmetric low-rank matrix
estimation. arXiv preprint arXiv:1611.03888, 2016.
T. Lesieur, L. Miolane, M. Lelarge, F. Krzakala, and L. Zdeborová. Statistical
and computational phase transitions in spiked tensor estimation. arXiv preprint
arXiv:1701.08010, 2017.
L. Lovasz. On the shannon capacity of a graph. IEEE Trans. Inf. Theor., 25(1):1–7,
1979.
J. Bruna L. Venturi, A. S. Bandeira. Neural networks with finite intrinsic dimension
have no spurious valleys. arXiv:1802.06384 [math.OC], 2018.
Laurent Massoulié. Community detection thresholds and the weak ramanujan property. In Proceedings of the forty-sixth annual ACM symposium on Theory of computing, pages 694–703. ACM, 2014.
Marc Mezard and Andrea Montanari. Information, physics, and computation. Oxford University Press, 2009.
Elchanan Mossel, Joe Neeman, and Allan Sly. Stochastic block models and reconstruction. arXiv preprint arXiv:1202.1499, 2012.
Elchanan Mossel, Joe Neeman, and Allan Sly. A proof of the block model threshold
conjecture. arXiv preprint arXiv:1311.4115, 2013.
Cristopher Moore. The computer science and physics of community detection: Landscapes, phase transitions, and hardness. arXiv preprint arXiv:1702.00467, 2017.
Marc Mézard, Giorgio Parisi, and MA Virasoro. SK model: The replica solution
without replicas. EPL (Europhysics Letters), 1(2):77, 1986.
Andrea Montanari and Emile Richard. Non-negative principal component analysis:
Message passing algorithms and sharp asymptotics. IEEE Transactions on Information Theory, 62(3):1458–1484, 2016.
Tengyu Ma, Jonathan Shi, and David Steurer. Polynomial-time tensor decompositions with sum-of-squares. In Foundations of Computer Science (FOCS), 2016
IEEE 57th Annual Symposium on, pages 438–446. IEEE, 2016.
Z. Ma and Y. Wu. Computational barriers in minimax submatrix detection. Available
online at arXiv:1309.5914, 2013.
Y. Nesterov. Squared functional systems and optimization problems. High performance optimization, 13(405-440), 2000.
P. A. Parrilo. Structured semidefinite programs and semialgebraic geometry methods
in robustness and optimization. PhD thesis, California Institute of Technology, 2000.
Judea Pearl. Fusion, propagation, and structuring in belief networks. Artificial intelligence, 29(3):241–288, 1986.
Aaron Potechin and David Steurer. Exact tensor completion with sum-of-squares.
arXiv preprint arXiv:1702.06237, 2017.
Amelia Perry, Alexander S Wein, and Afonso S Bandeira. Statistical limits of spiked
tensor models. arXiv preprint arXiv:1612.07728, 2016.
22
AFONSO S. BANDEIRA, AMELIA PERRY, AND ALEXANDER S. WEIN
[PWBM16a] Amelia Perry, Alexander S Wein, Afonso S Bandeira, and Ankur Moitra.
Message-passing algorithms for synchronization problems over compact groups.
Communications on Pure and Applied Mathematics (to appear). arXiv preprint
arXiv:1610.04583, 2016.
[PWBM16b] Amelia Perry, Alexander S Wein, Afonso S Bandeira, and Ankur Moitra. Optimality
and sub-optimality of PCA for spiked random matrices and synchronization. arXiv
preprint arXiv:1609.05573, 2016.
[Rag08]
P. Raghavendra. Optimal algorithms and inapproximability results for every CSP?
In Proceedings of the Fortieth Annual ACM Symposium on Theory of Computing,
STOC ’08, pages 245–254. ACM, 2008.
[RM14]
Emile Richard and Andrea Montanari. A statistical model for tensor PCA. In Advances in Neural Information Processing Systems, pages 2897–2905, 2014.
[Sho87]
N. Shor. An approach to obtaining global extremums in polynomial mathematical
programming problems. Cybernetics and Systems Analysis, 23(5):695–700, 1987.
[Sin11]
A. Singer. Angular synchronization by eigenvectors and semidefinite programming.
Appl. Comput. Harmon. Anal., 30(1):20 – 36, 2011.
[TAP77]
David J Thouless, Philip W Anderson, and Robert G Palmer. Solution of ‘solvable
model of a spin glass’. Philosophical Magazine, 35(3):593–601, 1977.
[ZK16]
Lenka Zdeborová and Florent Krzakala. Statistical physics of inference: Thresholds
and algorithms. Advances in Physics, 65(5):453–552, 2016.
(Bandeira) Department of Mathematics and Center for Data Science, Courant Institute of Mathematical Sciences, New York University
E-mail address: [email protected]
(Perry) Department of Mathematics, Massachusetts Institute of Technology
(Wein) Department of Mathematics, Massachusetts Institute of Technology
E-mail address: [email protected]
| 8cs.DS
|
Reducing the Number of Annotations
in a Verification-oriented Imperative Language
arXiv:1011.3407v1 [cs.PL] 15 Nov 2010
Guido de Caso, Diego Garbervetsky, and Daniel Gorín
Departamento de Computación, FCEyN, Universidad de Buenos Aires
{gdecaso,diegog,dgorin}@dc.uba.ar
Abstract. Automated software verification is a very active field of research which has made enormous progress both in theoretical and practical aspects. Recently, an important amount of research effort has been
put into applying these techniques on top of mainstream programming
languages. These languages typically provide powerful features such as
reflection, aliasing and polymorphism which are handy for practitioners but, in contrast, make verification a real challenge. In this work we
present Pest, a simple experimental, while-style, multiprocedural, imperative programming language which was conceived with verifiability
as one of its main goals. This language forces developers to concurrently
think about both the statements needed to implement an algorithm and
the assertions required to prove its correctness. In order to aid programmers, we propose several techniques to reduce the number and complexity
of annotations required to successfully verify their programs. In particular, we show that high-level iteration constructs may alleviate the need
for providing complex loop annotations.
Key words: Annotations, language design, verifiability, high-level iteration constructs.
1
Introduction
Formal automated software verification regained in recent years the attention of
the community. There are at least two reasons behind this resurgent success: on
the one hand, there were crucial developments in automated theorem proving
in the last fifteen years, with SAT-solvers finally reaching industrial strength;
on the other, the focus was shifted to partial specifications which somehow overcomes many of the objections raised in [20]. Verification of partial specifications
must be regarded as an error-detection procedure and, as such, akin to traditional forms of testing.
We will center on the particular form of verification where the source code
is annotated with special assertions. These normally take the form of method
pre and postconditions, loop and class invariants, etc. Special tools then read
these annotated sources, generate verification conditions from them and feed
these into automated provers [15]. Spec# [4] and ESC/Java [14] are two of the
best-known examples. The former is based on a dialect of C# while the latter
takes Java code with JML [17] annotations.
2
de Caso, Garbervetsky, Gorín
Apart from the abovementioned, there are ongoing research efforts in automated verification for almost every major programming language in use [6,24,23].
The rationale is to lower the adoption barrier by giving practitioners tools for
verifying the code they are writing today. Now, while this is an undeniably sensible plan, we perceive that, in almost every case, the resulting “programminglanguage-with-annotations” regarded as a whole ends-up being not entirely satisfying. We consider next some of the reasons for this.
Lack of cohesion. Annotations are usually introduced as a “patch” to the language. Most of the time, this is done in a way such that regular compilers and
IDE-tools regard them as mere comments. Moreover, most programming languages provide a way to perform optional run-time assertion checks. These are
usually used to validate pre and postconditions or invariants and are, thus, the
run-time counterparts of the verification annotations. But despite their dual
nature, both mechanisms have normally no syntactical relation whatsoever.
Redundancy. In statically typed languages, the type of the input and output
variables of a function are clearly part of its contract. But this means one endsup with two completely unrelated ways of specifying contracts: one enforced by
compilers (types) and the other by static checkers (the additional annotations).
Missed optimization opportunities. Optimizing compilers cannot leverage on program annotations in the same way they currently do on type information.
Inadequate semantics. We can most certainly exclude “to ease automated verifiability” from the list of goals that have driven the design of most modernday programming languages. We cannot know for sure if today’s mainstream
languages would have been as popular without features such as complex inheritance mechanisms, uncontrolled method reentrancy or unrestricted aliasing.
Nevertheless, the fact that the designers of Spec# already had to diverge in
slight ways from C#’s semantics [4] is indicating, in our opinion, a new driver
for the programming languages to come.
In this paper we report on an ongoing experiment in language design. Our
language forces programmers to introduce annotations specifying their intention
when writing code. We will show that by coupling tightly the annotations with
the language semantics we can rely on a verifying compiler to infer many of the
annotations. Moreover, we find that high level iteration constructs can effectively
be used as means of reducing the annotation burden of invariants.
2
A walk-through of the Pest language
In this section we will briefly present the Pest1 programming language by way
of examples. For a formal description the reader is referred to [8].
1
After the eastern side of the Danube river in Budapest, pronounced ["pESt].
Reducing Annotations in a Verification-oriented Language
3
Pest is an multi-procedural, non-recursive, structured, while-style, imperative programming language whose syntax natively incorporates various kinds
of annotations. Its main objective is to provide a test-bed for exploring new
concepts and ideas.
max(a,b,c)
:? true
:! (a ≥ b ⇒ c = a) ∧ (a < b ⇒ c = b)
:! a = a@pre ∧ b = b@pre
{
if a ≥ b then
c ←a
else
c ←b
fi
}
Fig. 1. Simple Pest procedure definition.
Figure 1 shows the definition of a simple procedure in Pest. Keywords :?
and :! introduce pre and postconditions respectively. In the postcondition a@pre
and b@pre denote the value of a and b at the beginning of the procedure execution. Therefore, this clause states that the value of a and b does not change;
this is necessary since in Pest all values are copied-in and then copied-out.
There are currently only three data types in Pest: booleans, integers and
arrays of integers. Variables are monomorphic and their type is inferred from
use. For example, in Figure 1, all variables are integers, since the ≥ operator
takes integers as arguments. Apart from classical boolean operators, boolean
expressions may include bounded first-order quantification.
Figure 2 lists a Pest procedure containing a while-loop iterating over an
array and a procedure call. Loop invariants are introduced using the :?! keyword
and exactly one loop variant must be provided, using the :# keyword. Only
variables are allowed as arguments on procedure calls and they are syntactically
enforced to be distinct. Observe that since, additionally, values are copied in
variable assignments, we impose a strict control over aliasing.
Operationally, annotations in Pest are interpreted as assertions. Roughly
speaking, preconditions are evaluated prior to procedure calls and postconditions
on return; invariants are checked before evaluating the loop guard. Since most
of the semantic rules include the evaluation of annotations, a program execution
gets stuck when the interpretation of one of them fails.
To enforce this semantics, annotations can be checked at run-time, but this
can be very expensive. Alternatively, a Pest compiler can simply remove an
assertion if it can be statically verified. In a way, this is reminiscent of type
4
de Caso, Garbervetsky, Gorín
arrayMax(A, m)
:? |A| > 0
:! ∀k / 0 ≤ k < |A| : m ≥
:! ∃k / 0 ≤ k < |A| : m =
:! A = A@pre
{
m ← A[0]
local i ← 1
while i < |A|
:?! 1 ≤ i ≤ |A|
:?! ∀k / 0 ≤ k < i :
:?! ∃k / 0 ≤ k < i :
:?! A = A@pre
:# |A| − i
do
local t ← 0
local e ← A[i]
call max(e, m, t)
m←t
i ← i +1
od
}
A[k]
A[k]
m ≥ A[k]
m = A[k]
Fig. 2. A Pest procedure containing a loop.
erasure. Of course, the programmer will normally want to be aware of which
assertions failed to be statically verified.
A Pest compiler can also infer pre and postconditions of a procedure. If only
the precondition is given a postcondition can be obtained by way of a symbolic
computation; on the other hand, starting from a postcondition, a precondition
can be obtained using a variation of the notion of weakest precondition. In simple
cases, like the procedure in Figure 1, one can remove the annotation altogether
and rely solely on inference. Details are given in §4.
Eliminating the need for explicitly given loop invariants is, of course, highly
desirable. A lot of research has been done in loop invariant inference (for instance, [13,18]). In §5 we will show that higher-level iteration constructs can be
alternatively regarded as encapsulating “directives” on how to build a proper
invariant from the loop body.
Using the ideas sketched above, arrayMax can be rewritten in Pest as shown
in Figure 3 which requires no annotations for the loop and still enforces its
contract. Notice that, in addition to the invariant, in this case we are also able
to remove the precondition and part of the postcondition since they can be
inferred by the Pest compiler using the rules presented in §4.
Reducing Annotations in a Verification-oriented Language
5
easyArrayMax(A,m)
:! ∀k / 0 ≤ k < |A| : m ≥ A[k]
:! ∃k / 0 ≤ k < |A| : m = A[k]
{
m ← A[0]
for i from 1 to |A| do
local t ← 0
local e ← A[i]
call max(e, m, t)
m←t
od
}
Fig. 3. Using a for construct to remove annotations.
3
Formal semantics of Pest
In this section we introduce Pest semantics. In §3.1 we briefly comment on
Pest (big-step) operational semantics. As we already mentioned in the preceding
section, annotations are interpreted as assertions and a computation gets stuck
whenever one of these fails. In §3.2, we present a static formulation based on
Hoare-style clauses. This characterizes those programs that cannot get stuck
according to the operational semantics. In this sense, it is reminiscent of a type
system; but, observe that it relies on a semantic entailment relation over formulas
(|=) that is undecidable in the general case.
In this setting, automatic verification of a Pest program can be performed by
checking conformance to its static semantics using a computable approximation
(⊢) of the |=-relation. Later, in §4 we will derive some calculi from the static
semantics to cope with several inference tasks.
3.1
Operational semantics
Pest operational semantics is based on the standard notion of while-style programming languages semantics. Annotations are handled by incorporating conditions (i.e. assertions) into the semantic rules that disallow the execution of a
program that violates them.
The semantic rules are given in terms of state transformations. A state σ is
a function that maps program variables to concrete values of the proper type.
With σ{v 7→ n} we denote a state that coincides with σ except, possibly, in the
value for v which, in the former, is n; σ ⊖V is the restriction of σ to a domain
that does not contain any variable in V . We use JeKσ to denote the value of an
expression e under σ.
We write σ ⊲ s ⊲ σ ′ to express that a Pest sentence s, when run from state
σ, finishes correctly and arrives in state σ ′ . Some rules are depicted in Figure 4;
for the complete set of rules refer to [8].
6
de Caso, Garbervetsky, Gorín
JgKσ = false JinvKσ = true
(O-WHILE-F)
σ ⊲ while g :?! inv :# var do s od ⊲ σ
JgKσ = true JinvKσ = true
JvarKσ > 0 σ ⊲ s ⊲ σ ′
JvarKσ′ < JvarKσ
σ′ ⊖ locals(s) ⊲ while g :?! inv :# var do s od ⊲ σ ′′
σ ⊲ while g :?! inv :# var do s od ⊲ σ′′
(O-WHILE-T)
Jpre(proc)Kρ = true
ρ ⊲ body(proc) ⊲ ρ′
Jpost(proc)Kρ′ = true
(O-CALL)
σ ⊲ call proc(cp1 , . . . , cpk ) ⊲ σ{cp 1 7→ ρ′ (p1 ), . . .}
def
def
where ρ(pi ) = σ(cpi ) and ρ(pi @pre) = σ(cpi )
Fig. 4. Pest operational semantics (fragment)
Consider the rules for the while statement. Observe first that if JinvKσ 6= true
then no rule applies and, thus, a computation will stick in that case. When the
guard is false, the state is not affected (the language syntax guarantees that
guards are free of side-effects). Alternatively, when the guard is true, the variant
must be above zero in order to continue. Observe that σ ′ is the state after an
execution of the loop body, including locally-defined variables; if the variant
didn’t decrease the computation will stuck.
For procedure calls, state ρ binds formal and actual parameters; the precondition of the called procedure must hold for this state. The state ρ′ , if defined, is
the result of executing the procedure body from ρ; the postcondition must hold
in ρ′ . At the end, actual parameters are updated with the final values assigned
to the formal parameters (recall that all parameters are in/out in Pest).
3.2
Hoare-style static semantics
We first need to introduce some preliminary definitions. For σ a state and b a
Pest boolean expression2, we write σ |= b if JbKσ = true. We will write b1 |= b2
to indicate that, for all σ, whenever σ |= b1 holds, then σ |= b2 must hold too
(i.e. b1 is stronger than b2 ).
We also require a notion of “safeness” for the evaluation of expressions. Given
an expression e, we want safe(e) to be a boolean expression such that σ |= safe(e)
implies that JeKσ is defined. For example, we expect safe(a[ i ] / y) to be the
expression “0 ≤ i ∧ i < |a| ∧ y 6= 0”. A formal definition of these conditions
is straightforward. Finally, e1 ⌊e2 7→ e3 ⌋ denotes the expression that results from
replacing every occurrence of e2 by e3 in e1 . Observe that e2 and e3 must be of
the same type.
2
Throughout this section, boolean expressions are augmented with unbounded existential quantification.
Reducing Annotations in a Verification-oriented Language
7
Instead of using states like in the operational case, for the static semantics we
will use predicates (i.e. boolean expressions) that describe a (possibly infinite)
set of states. For boolean expressions p and q and s a Pest sentence, {p} s {q}
must be read “after executing s from a σ such that σ |= p, we obtain a σ ′ such
that σ ′ |= q”.
Figure 5 lists some of the static rules (for the complete list, refer to [8]).
Consider the rule for assignments: the first premise guarantees that the program
won’t get stuck when evaluating e; the second one states that q is a consequence
of what was known prior to the assignment (p⌊v 7→ v ′ ⌋) and its effect (v = e⌊v 7→
v ′ ⌋). The existentially quantified variable v ′ stands for the value of v before the
assignment (this requires unbounded quantification).
p |= safe(e)
∃ v ′ (p⌊v 7→ v ′ ⌋ ∧ v = e⌊v 7→ v ′ ⌋) |= q
{p} v ← e {q}
p |= safe(g)
p ∧ g |= p1 {p1 } s1 {q1 } q1 |= q
p ∧ ¬g |= p2 {p2 } s2 {q2 } q2 |= q
{p} if g then s1 else s2 fi {q}
(S-ASSIGN)
(S-IF)
true |= safe(inv) inv |= safe(var)
inv |= safe(g) p |= inv inv ∧ g |= p′
p′ |= var > 0 {p′ } var0 ← var s {q ′ }
q ′ |= inv q ′ |= var < var0 inv ∧ ¬g |= q
{p} while g :?! inv :# var do s od {q}
(S-WHILE)
Fig. 5. Pest static semantics (fragment)
The premises of the rule for while can be seen as both a proof of the Fundamental Invariance Theorem for Loops [11] and a proof of termination using the
loop variant. The predicate p′ represents any state where the invariant and the
condition of the while hold; the loop body is augmented with an initial assignment to a fresh variable var 0 that is used to prove that the variant decreases.
There is a clear correlation between Pest’s operational and static semantics.
Using the latter, we can give a notion of safe program. In what follows, if π is a
program and p a procedure, then π, p is the program obtained by appending p
to π.
Definition 1 (Safe programs) The set Safe of programs is inductively defined as follows:
(SAFE-EMPTY)
π ∈ Safe
∅ ∈ Safe
{pre(p)} body(p) {post(p)}
(SAFE-EXTEND)
π, p ∈ Safe
8
de Caso, Garbervetsky, Gorín
Theorem 1 (Safe programs execute normally) Let π ∈ Safe and let p be
a procedure in π. For each σ such that σ |= pre(p), there exists a state σ ′ such
that σ ⊲ body(p) ⊲ σ ′ and σ ′ |= post(p).
Proof. It is a longish yet straightforward induction on the length of a derivation
of {pre(p)} body(p) {post(p)}. See [8].
4
Reducing the annotation burden
In the previous section we presented the Pest programming language, including
a soundness result showing that programs conforming to Pest static semantics
do not go wrong. Still, providing Pest annotations is a heavy and complex task.
In this section we discuss two techniques that assist the programmer by inferring
and completing annotations.
4.1
Inference of procedure contracts
We will describe a technique for synthesizing procedure pre and postconditions.
To do that we will specialize the static semantic rules and turn them into inference rules. This will somehow resemble the process of developing type inference
rules from a type system.
Figure 6 shows some of the rules to compute a postcondition post(s, p) from
given p and s. The complete calculus is given in [8].
p |= safe(e)
(Q-ASSIGN)
post(v ← e, p) = ∃ v ′ (p⌊v 7→ v ′ ⌋ ∧ v = e⌊v 7→ v ′ ⌋)
p |= safe(g)
(Q-IF)
post(if g then s else t fi, p) =
Cl∃ locals(s) (post(s, p ∧ g)) ∨ Cl∃ locals(t) (post(t, p ∧ ¬g))
true |= safe(inv) inv |= safe(var)
inv |= safe(g) p |= inv inv ∧ g |= var > 0
post(var0 ← var s, inv ∧ g) = q ′
q ′ |= inv q ′ |= var < var0
(Q-WHILE)
post(while g :?! inv :# var do s od, p) = inv ∧ ¬g
Fig. 6. Pest postcondition calculus (fragment)
Observe the way in which these rules specialize those in Figure 5. For example, in the specialized rule for assignments, post(v ← e, p) is simply the strongest
Reducing Annotations in a Verification-oriented Language
9
q satisfying the original assignment rule. The rule for if eliminates (using an existential closure3 ) the local variables defined in each of the branches from their
respective inferred postcondition.
It is worth noticing that the inferred postcondition may contain unbounded
existential quantifications. The main drawback of this is that, in that case, it
is not valid to literally include it in the code. Observe also that an assertion
with unbounded quantification cannot be checked at runtime; nevertheless, the
inferred postconditions are correct by construction and the compiler can omit
its associated runtime checks.
A similar approach can be used to infer procedure preconditions (denoted
pre(s, q)). In this case, we use weakest-preconditions to find a suitable predicate
p, except for while sentences where we simply use the provided invariant. The
rules for computing pre(s, q) can be found in [8].
Interestingly, we can infer non-trivial pre and postconditions even if the programmer provides no procedure annotations (other than loop annotations). First,
we infer a predicate that guarantees the normal execution of a procedure body
s:
def
P = pre(s, true)
This gives us a proper precondition for the procedure. Next, we strengthen P
by giving a symbolic initial value to each parameter (pi = pi @pre) and use it
to infer a postcondition Q:
def
Q = post(s, P ∧ p1 = p1 @pre ∧ . . . ∧ pk = pk @pre)
The reader should verify that if the pre and postconditions of the procedures
in Figures 1 and 2 are omitted, this procedure will infer logically equivalent ones.
4.2
Strengthening annotations
We will now show how we can take advantage of Pest’s restrictions on variable
aliasing to incorporate inexpensive enhancements of invariants and postconditions. In a nutshell, we can propagate known facts about variables in certain
scopes if we know that their contents are not altered. For example, let us say we
know that before entering a loop it is always the case that variable j has value
e; then if we can determine that the loop body does not update j, we can add
j = e to the invariant.
In the absence of aliasing, a simple, sound approximation of the set of unmodified variables is to compute first the set of variables potentially modified and
then take its complement. We call modVars(s) the set of (potentially) modified
variables; its definition is straightforward and can be found in [8].
What follows is a definition of our annotation strengthening function which
takes a sentence s and an entry point p and produces a sentence s′ with possibly
3
The existential closure of a boolean expression b with respect to a set of variables
{v1 , . . . , vn } is ∃v1 , . . . , vn (b)
10
de Caso, Garbervetsky, Gorín
1 ≤ i ≤ |A| ∧ ∀k / 0 ≤ k < i :
m ≥ A[k] ∧ ∃k / 0 ≤ k < i :
m = A[k] ∧ A = A@pre
∀k / 0 ≤ k < |A| : m ≥ A[k] ∧ ∃k / 0 ≤ k < |A| : m = A[k] ∧ A = A@pre
Fig. 7. Invariant and postcondition of the main loop in the arrayMax procedure.
Differences are highlighted.
strengthened annotations. The translation does not alter assignments, local variable definitions or procedure calls. In the case of sentence sequence, we define it
as:
def
trS (s1 s2 , p) = s′1 trS s2 , post(s′1 , p)
where s′1 = trS (s1 , p). That is, we first translate s1 and use its postcondition to
translate s2 .
When dealing with conditional statements we proceed by translating each
branch using the conjunction of p with the guard g:
def
trS (if g then s1 else s2 fi, p) =
if g then trS (s1 , p ∧ g)
else trS (s2 , p ∧ ¬g) fi
Finally, in the presence of a loop we strengthen its invariant. We existentially
close p over the variables that are potentially modified by the loop body:
def
trS (while g :?! inv :# var do s od, p) =
while g :?! I :# var
do trS (s, I ∧ g) od
where I = inv ∧ Cl∃ modVars(s) (p).
Using this technique we can strengthen both invariants and postconditions.
For the latter case we simply take the postcondition of the last instruction of
the procedure and existentially eliminate the local variables, in order to leave
the postcondition only in terms of procedure parameters.
5
High-level iteration constructs as annotations
In this section we focus on what we call “high-level iteration constructs”. These
capture recurrent iteration patterns and are frequently included in programming
languages to reduce error-prone boilerplate code. Examples of such constructs
are Pascal-style for -loop and C#’s foreach. We will see next that this kind
of constructs can be seen as implicitly carrying their own proof obligations,
reducing the need for explicit annotations.
5.1
The for construct
We first consider the Pascal-style for loop and take as motivating example the
arrayMax procedure in §2. In Figure 7 we compare the postcondition of the loop
with its invariant. Clearly, they are syntactically very close. In fact, if we are given
Reducing Annotations in a Verification-oriented Language
11
arrayInc(A)
{
local i ← 0
while i < |A|
:?! 0 ≤ i ≤ |A|
:?! ∀k / 0 ≤ k < i : A[k] = A@pre[k] + 1
:?! ∀k / i ≤ k < |A| : A[k] = A@pre[k]
:# |A| − i
do
A ← update A on i with A[i] + 1
i ← i +1
od
}
Fig. 8. Pest procedure that increases each element in an array.
a postcondition and a for -loop, we can simply try to guess a candidate invariant
for the loop without even considering the loop body. Of course, the correctness
of the invariant will have to be statically verified. On the other hand, a correct
variant can be trivially obtained.
Formally, given a sentence of the form for i from l to h do s od (without
annotations) and its postcondition in the form of a predicate Qf , we can macroexpand the sentence into the following lower-level code:
local i← l
while i < h
:?! l ≤ i ≤ h ∧ Qf ⌊h 7→ i⌋
:# h − i
do
s
i ← i+1
od
In order to apply this expansion we are required to provide a loop postcondition Qf . Nevertheless, this can be accomplished using the precondition calculus
of §4.1 from the procedure postcondition up to the for -sentence we need to expand. If s contains a nested for -construct, a postcondition for it can be derived
from the inferred invariant and the expansion can be recursively applied.
Observe that in the easyArrayMax procedure of Figure 3, Qf is simply the
postcondition of the procedure, since the for construct is located at the very end
of the procedure body.
5.2
Declarative iteration constructs
In §4.1 we saw that if the loop-invariants are provided, then postconditions can
be omitted; conversely, in the previous section we showed that in the case of for loops invariants can be traded for postconditions. In this section we will show
12
de Caso, Garbervetsky, Gorín
that when more specific iteration constructs are used, both can be dispensed
with. We will only consider here the map-construct (reminiscent of the map
function over lists in functional languages) but the idea is easily extensible to
other constructs.
Consider the procedure of Figure 8 (notice that the procedure contract was
omitted and left for automatic inference). The loop iterates over an array A using
an indexing variable i. On each iteration, only the element of A at position i is
updated. Furthermore, this is done taking into account only the value of i and
A[i]. The loop condition is not affected by changes to A. We call this iteration
pattern, in which array elements are updated independently of the others, a map.
The invariant for this iteration pattern states that the already visited array
elements were updated whereas the remaining elements are unchanged. The
variant reflects the fact that the array is traversed in a forward direction.
Using our proposed map construct, the procedure in Figure 8 can be written
like this:
arrayInc(A)
{
map
A[i ] ← A[i] + 1
in A[.. i ..]
}
Formally, this works as follows. Let A be an array, i a fresh variable and s a
sentence such that modVars(s) r locals(s) = {A} and A is accessed only indexed
by i. In that case, map s in A[.. i ..] is well-formed and gets expanded as follows:
local i ← 0
while i < |A|
:?! 0 ≤ i < |A|
:?! ∀k / 0 ≤ k < i : posts ⌊i 7→ k⌋
:?! ∀k / i ≤ k < |A| : A[k] = A@pre[k]
:# |A| − i
do
s
i ← i +1
od
where post s = post(s, 0 ≤ i < |A|). It can easily be seen that the inferred
invariant is always correct. Thus, not only can the compiler erase the assertion
checks, it doesn’t even need to statically verify them.
Notice that the invariant inferred for the map high-level construct can be
further strengthened using the technique presented in §4.2.
Reducing Annotations in a Verification-oriented Language
6
13
Experience
In order to test these ideas we developed a tool called BudaPest which is
available online4 as an Eclipse plug-in.
The tool takes Pest programs, statically verifies them and compiles them to
Java code. For the verification step the code is translated into an intermediate
assume/assert style language similar to BoogiePL [9]. Unlike the Boogie verifier,
our tool leverages the structure of the original program to discharge verification
conditions one by one to SMT-solvers, such as CVC3 [5], Yices [12] and Z3 [2].
We split verification conditions in several pieces in order to allow the provers to
work in parallel, so we can leverage each of the provers’ power and combine their
results. Splitting verification conditions also has the advantage of enabling early
detection (during the verification process) of unsatisfiable conditions since once
a subformula fails there is no need to continue with the rest of the subformulas.
We tested the usability of this tool in a first-year Computer Science course.
The curriculum of the course includes proving correctness and termination of
simple imperative programs specified using contracts written in a fragment of
first-order logic and proven by way of Hoare axiomatic semantics [16]. With the
aid of the tool they were able to automatically verify the correctness of their
implementations of algorithms such as bubble and insertion sorting, linear and
binary search, etc., which they previously had to do by hand.
The experience was very encouraging for the students since they found that
concepts they had learnt during the course (preconditions, postconditions, invariants, variant functions) could be applied in “real” applications and produced
code with guarantees of (partial) correctness and termination.
Finally, it is worth mentioning that the integrated nature of the Pest language forced the students to concurrently think about both the statements
needed to implement an algorithm and the assertions required to prove its correction.
7
Related Work
There is a plenty of research on the automatic verification of programs. Due to
lack of space we will just mention some of those we consider closer to our work.
There are several languages and tools that incorporate Hoare-style specifications to automatically prove partial correctness in imperative languages. Some
well known examples are ESC/Java [14], JML [17], ESC/Modula-3 [10],
Spec# [4] and SPARK/Ada [3]. These systems enrich the language with user
provided annotations which can be checked on runtime or statically analyzed by
generating verification conditions that are discharged to a theorem prover.
The approach followed by SPARK/Ada is closer to ours in the sense that
they impose limitations to the language in order to make verification possible. We
believe they go too far since they limit syntactically to programs where assertions
4
http://lafhis.dc.uba.ar/budapest
14
de Caso, Garbervetsky, Gorín
are decidable, making them too restrictive. On the contrary, the problem with
the other tools is that they deal with complex languages that provide many
features such as polymorphism, concurrency, aliasing or reflection, which are
comfortable for program development but complicate verification.
Our work follows the philosophy that a compiler (or interpreter) should try
to reject misbehaving programs. This is similar the approach followed by several languages that provide expressive type systems such as Deputy [7], which
uses dependent types [19] to type-check low level imperative programs. For functional languages there are many proposals; just to name one, Cayenne [1] extends Haskell with dependent types. Recently, [21] presented a type theory for
higher order functional programs which incorporates Hoare style specifications
into types, making it possible to enforce correct use of side effects.
8
Final thoughts
In this paper we presented Pest, a simple imperative programming language,
which was designed with verifiability in mind. We proposed a series of techniques that aim at mitigating the annotation burden required to verify programs.
Firstly, we showed an inference mechanism for procedure pre and postconditions
and then a simple method to strengthen loop invariants and postconditions.
As a distinguishing contribution we extended Pest with high-level iteration
constructs that allow programmers to dramatically reduce annotations and yet
maintain the ability to prove correctness of relatively complex pieces of software. Our first experiences showed that an integrated approach such as ours
was gracefully adopted by young students with no pre-concepts with respect to
programming languages or tools.
We would like to explore, in the near future, the possibility of extending
Pest incorporating features that would increase its expressiveness without sacrificing verifiability. We plan to allow the programmer to define and use her own
data types and provide means to reason about representation invariants. Adding
dynamic memory support is another priority but, in order to keep a verifiable
language, we believe we must enforce an alias control mechanism such as [22].
Finally, we are interested in providing the means for programmers to define their
own language constructs by feeding the BudaPest tool with syntax and macro
expansion definitions, and relying on the base tools (i.e. pre and postcondition
calculi) that we provided in this paper to ensure correctness.
References
1. L. Augustsson. Cayenne-a language with dependent types. ACM SIGPLAN Notices, 34(1):239–250, 1999.
2. T. Ball, S. K. Lahiri, and M. Musuvathi. Zap: Automated theorem proving for
software analysis. In LPAR, 2005.
3. J.G.P. Barnes and Praxis Critical Systems Limited. High Integrity Ada: The
SPARK Approach. Addison-Wesley, 1997.
Reducing Annotations in a Verification-oriented Language
15
4. M. Barnett, K.R.M. Leino, and W. Schulte. The Spec# Programming System: An
Overview. In CASSIS. Springer, 2005.
5. C. Barrett and S. Berezin. CVC Lite: A new implementation of the cooperating
validity checker. In Proceedings of the 16th CAV, 2004.
6. S. Chatterjee, S.K. Lahiri, S. Qadeer, and Z. Rakamaric. A reachability predicate
for analyzing low-level software. 2007.
7. J. Condit, M. Harren, Z. Anderson, D. Gay, and G.C. Necula. Dependent Types
for Low-Level Programming. Lecture Notes in Computer Science, 4421:520, 2007.
8. G. de Caso, D. Garbervetsky, and D. Gorín. Pest formal specification. Technical report, Universidad de Buenos Aires, http://lafhis.dc.uba.ar/budapest/
Theory section, 2008.
9. R. DeLine and K.R.M. Leino. BoogiePL: A typed procedural language for checking
object-oriented programs. Technical report, Technical Report MSR-TR-2005-70,
Microsoft Research, 2005.
10. D. L. Detlefs, K.R.M. Leino, G. Nelson, and J.B. Saxe. Extended static checking.
Technical Report #159, Palo Alto, USA, 1998.
11. E.W. Dijkstra. A Discipline of Programming. Prentice Hall PTR Upper Saddle
River, NJ, USA, 1997.
12. B. Dutertre and L. de Moura.
The Yices SMT solver.
Available at
http://yices.csl.sri.com/, August, 2006.
13. M.D. Ernst, J.H. Perkins, P.J. Guo, S. McCamant, C. Pacheco, M.S. Tschantz, and
C. Xiao. The Daikon system for dynamic detection of likely invariants. Science of
Computer Programming, 2007.
14. C. Flanagan, K.R.M. Leino, M. Lillibridge, G. Nelson, J.B. Saxe, and R. Stata.
Extended static checking for Java. In PLDI ’02, 2002.
15. D.I. Good, R.L. London, and WW Bledsoe. An interactive program verification
system. In Proceedings of the international conference on Reliable software table
of contents, pages 482–492. ACM New York, NY, USA, 1975.
16. C. A. R. Hoare. An axiomatic basis for computer programming. Commun. ACM,
12(10):576–580, 1969.
17. G.T. Leavens, A.L. Baker, and C. Ruby. JML: A notation for detailed design.
In Behavioral Specifications of Businesses and Systems, pages 175–188. Kluwer
Academic Publishers, 1999.
18. K. Rustan M. Leino and Francesco Logozzo. Loop invariants on demand. In
APLAS, pages 119–134, 2005.
19. J. McKinna. Why dependent types matter. Proc. ACM Symp. on Principles of
Programming Languages (POPL 2006), 2006.
20. R.A. De Millo, R.J. Lipton, and A.J. Perlis. Social processes and proofs of theorems
and programs. Commun. ACM, 22(5):271–280, 1979.
21. A. Nanevski, G. Morrisett, and L. Birkedal. Polymorphism and separation in hoare
type theory. In ICFP ’06, 2006.
22. J. Noble, J. Vitek, and J. Potter. Flexible alias protection. In ECOOP ’98: Proceedings of the 12th European Conference on Object-Oriented Programming, pages
158–185, London, UK, 1998. Springer-Verlag.
23. N. Norwitz. PyChecker. SourceForge project http://pychecker. sourceforge. net.
24. D.N. Xu. Extended static checking for Haskell. In Proceedings of the 2006 ACM
SIGPLAN workshop on Haskell, pages 48–59. ACM New York, NY, USA, 2006.
| 6cs.PL
|
arXiv:1711.01130v1 [math.AC] 3 Nov 2017
ON COMBINATORIAL ASPECTS ARISING FROM MODULES
OVER COMMUTATIVE RINGS
RAMEEZ RAJA
School of Mathematics, Harish-Chandra Research Institute, HBNI,
Chhatnag Road, Jhunsi, Allahabad 211019, India
[email protected]
Abstract. Let R be a commutative ring with unity, M be an unitary Rmodule and Γ be a simple graph. This research article is an interplay of
combinatorial and algebraic properties of M . We show a combinatorial object
completely determines an algebraic object and characterize all finite abelian
groups. We discuss the correspondence between essential ideals of R, submodules of M and vertices of graphs arising from M . We examine various types
\
\
\
of equivalence relations on objects A\
f (M ), As (M ) and At (M ), where Af (M )
\
is an object of full-annihilators, As (M ) is an object of semi-annihilators and
A\
t (M ) is an object of star-annihilators in M . We study essential ideals corresponding to elements of an object A\
f (M ) over hereditary and regular rings.
Further, we study isomorphism of annihilating graphs arising from M and
tensor product M ⊗R T −1 R, where T = R\C(M ), where C(M ) = {r ∈ R
: rm = 0 f or some 0 6= m ∈ M }, and show that annf (Γ(M ⊗R T −1 R)) ∼
=
annf (Γ(M )) for every R-module M .
AMS Mathematics Subject Classification: 13A99, 05C78, 05C12
Key words: Module, ring, essential ideal, tensor product, annihilator
1. Introduction
The subject of algebraic combinatorics deals with the study of combinatorial
structures arising in an algebraic context, or applying algebraic techniques to a
combinatorial problem. One of the areas in algebraic combinatorics introduced by
Beck [7] is to study the interplay between graph theoretical and algebraic properties of an algebraic structure. Continuing the concept of associating a graph to an
algebraic structure another combinatorial approach of studying commutative rings
was given by Anderson and Livingston in [1]. They associated a simple graph to a
commutative ring R with unity called a zero-divisor graph denoted by Γ(R) with
vertices as Z ∗ (R) = Z(R)\{0}, where Z(R) is the set of zero-divisors of R. Two
distinct vertices x, y ∈ Z ∗ (R) of Γ(R) are adjacent if and only if xy = 0. The
zero-divisor graph of a commutative ring has also been studied in [2, 21, 23, 25]
and has been extended to non-commutative rings and semigroups in [17, 24].
The combinatorial properties of zero-divisors discovered in [7] has also been
studied in module theory. Recently in [19], the elements of a module M has been
1
2
RAMEEZ RAJA
classified into full-annihilators, semi annihilators and star-annihilators, see Definition 2.1 in section 2. For M = R, these elements are the zero-divisors of a ring R,
so the three simple graphs annf (Γ(M )), anns (Γ(M )) and annt (Γ(M )) corresponding to full-annihilators, semi annihilators and star-annihilators in M are natural
generalizations of a zero-divisor graph introduced in [1].
On the other hand, the study of essential ideals in a ring R is a classical problem.
For instance, Green and Van Wyk in [11] characterized essential ideals in certain
class of commutative and non-commutative rings. The authors in [6, 16] also studied essential ideals in C(X) and topologically characterized the scole and essential
ideals. Moreover, essential ideals also have been investigated in rings of measurable
functions [18] and C ∗ - algebras [14]. For more on essential ideals see [12, 13, 22].
We call any subset of M as an object. By combinatorial object we mean an
object which can be put into one-to-one correspondence with a finite set of integers and by an algebraic object we mean a combinatorial object which is also an
algebraic structure. The main objective of this paper is to study combinatorial
objects, objects arising from modules and the graphs with vertex set as objects and
combinatorial objects.
For an R-module M and x ∈ M , set [x : M ] = {r ∈ R | rM ⊆ Rx}, which
clearly is an ideal of R and an annihilator of the factor module M/Rx, where as
the annihilator of M is [0 : M ]. In section 2, we study the correspondence of ideals
in R, submodules of M and the elements of an object A\
f (M ), and characterize all
finite abelian groups (Proposition 2.2). We further show (Theorem 2.11) that an
R-module [x : M ] is injective if and only if R is non-singular and the radical of
R/[x : M ] is zero. In section 3, we examine two different equivalence relations on
the elements of an object A\
f (M ) and discuss (Theorem 3.6) the adjacencies of vertices in the graph annf (Γ(M )). Furthermore, we explore the equivalence relations
(Theorem 3.8, Theorem 3.10) to establish the structure of a module M and the
annihilating graph annf (Γ(M )). Finally, in section 4, we study some applications
of annihilating graphs (Theorem 4.2) and extract certain module theoretic properties from these graphs. Further, we discuss the annihilating graphs arising from the
tensor product (Theorem 4.7) and show that annf (Γ(M ⊗R T −1 R)) ∼
= annf (Γ(M ))
for every R-module M .
We conclude this section with some notations, which are mostly standard and
will be used throughout this research article.
Throughout, R is a commutative ring (with 1 6= 0) and all modules are unitary
unless otherwise stated. A submodule N is said to be an essential submodule of
M if it intersects non-trivially with every nonzero submodule of M . [N : M ] =
{r ∈ R | rM ⊆ N } denotes an ideal of ring R. The symbols ⊆ and ⊂ has usual set
theoretic meaning as containment and proper containment of sets. We will denote
the ring of integers by Z, positive integers by N and the ring of integers modulo n
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS3
by Zn . For basic definitions from graph theory we refer to [[9, 27]], and for ring
theory and module theory we refer to [[4, 8, 15, 28]].
2. Essential ideals determined by elements of an object A\
f (M)
In this section, we discuss the correspondence of essential ideals in R, submodules of M and the elements of an object A\
f (M ). We characterize essential ideals
corresponding to Z-modules. We show that if M is not simple R-module, then
an ideal [x : M ], x ∈ A\
f (M ) considered as an R-module is injective. We study
essential ideals corresponding to the vertices of graph annf (Γ(M )) over hereditary
and regular rings.
We recall a definition concerning full-annihilators, semi annihilators and starannihilators of a module M .
Definition 2.1. An element x ∈ M is a,
(i) full-annihilator, if either x = 0 or [x : M ][y : M ]M = 0, for some nonzero
y ∈ M with [y : M ] 6= R,
(ii) semi-annihilator, if either x = 0 or [x : M ] 6= 0 and [x : M ][y : M ]M = 0, for
some nonzero y ∈ M with 0 6= [y : M ] 6= R,
(iii) star-annihilator, if either x = 0 or ann(M ) ⊂ [x : MR ] and [x : M ][y :
M ]M = 0, for some nonzero y ∈ M with ann(M ) ⊂ [y : M ] 6= R.
We denote by Af (M ), As (M ) and At (M ) respectively the objects of full-annihilators,
semi-annihilators and star-annihilators for any module M over R. We set A\
f (M ) =
\
Af (M )\{0}, A\
(M
)
=
A
(M
)\{0}
and
A
(M
)
=
A
(M
)\{0}.
s
s
t
t
In [19] authors introduced annihilating graphs arising from modules over commutative rings called as full-annihilating, semi-annihilating and star-annihilating
graphs denoted by annf (Γ(M )), anns (Γ(M )) and annt (Γ(M )) respectively. The
\
\
vertices of annihilating graphs are elements of objects A\
f (M ), As (M ) and At (M ),
and two vertices x and y are adjacent if and only if [x : M ][y : M ]M = 0. By
Definition 2.1, we see that there is a correspondence of ideals in R, submodules
\
\
of M and the elements of objects A\
f (M ), As (M ) and At (M ). Furthermore, the
containment annt (Γ(M )) ⊆ anns (Γ(M )) ⊆ annf (Γ(M )) as induced subgraphs is
clear, so our main emphasis is on the object A\
f (M ) and the full-annihilating graph
annf (Γ(M )). However, one can study these objects and graphs separately for any
module M . Note that if M is a finite module over R or the graph annf (Γ(M )) is
\
\
\
finite, then the objects A\
f (M ), As (M ) are combinatorial with |Af (M )| = |As (M )|
and the annihilating graphs annf (Γ(M )), anns (Γ(M )) coincide, where as the graph
annt (Γ(M )) with vertex set as combinatorial object A\
t (M ) may be different.
Let G be any finite Z-module. Clearly, G is a finite abelian group. Below, we
discuss the correspondence of ideals in Z and the elements of an object A\
f (G).
We study cases of finite abelian groups where the essential ideals corresponding to
4
RAMEEZ RAJA
the submodules generated by the vertices of graph annf (Γ(M )) are same and the
submodules determined by these vertices are isomorphic.
The following is an interesting result in which a combinatorial object completely
determines an algebraic object.
Proposition 2.2. Let G be a finite Z-module. Then for each x ∈ A\
f (G), [x : M ]
is an essential ideal if and only if G is a finite abelian group without being simple.
Proof. For all x ∈ A\
f (G), we have [x : G] = nZ, n ∈ N. It is clear that nZ
intersects non-trivially with any ideal mZ, m ∈ N in Z.
For the converse, observe that among all finite abelian groups A\
f (G) = ∅ if and
only if G is simple.
Recall that a graph Γ is said to be a complete if there is an edge between every
pair of distinct vertices. A complete graph with n vertices is denoted by Kn .
Remark 2.3. For a finite abelian group Zp ⊕ Zp , where p ≥ 2 is prime, the essen\
tial ideals [x : M ], x ∈ Af (Z
p ⊕ Zp ) corresponding to the submodules of Zp ⊕ Zp
\
generated by elements of Af (Z
p ⊕ Zp ) are same. In fact [x : M ] = ann(Zp ⊕ Zp )
\
for all x ∈ Af (Zp ⊕ Zp ). Furthermore, the abelian group Zp ⊕ Zp is a vector space
over field Zp and all one dimensional subspaces are isomorphic. So, the submodules
\
generated by elements of Af (Z
p ⊕ Zp ) are all isomorphic. For a finite abelian group
Zp ⊕ Zq , where p and q are any two prime numbers, the essential ideals determined
\
by each x ∈ Af (Z
p ⊕ Zq ) are either pZ or qZ.
Remark 2.4. Proposition 2.2 is not true for all Z-modules. Consider a Z-module
M = Z ⊕ Z ⊕ · · · ⊕ Z, which is a direct sum of n copies of Z. It is easy to verify
c
that A\
f (M ) = M with [x : M ][y : M ]M = 0 for all x, y ∈ M , which implies
annf (Γ(M )) is a complete graph. The cyclic submodules generated by the vertices
of annf (Γ(M )) are simply the lines with integral coordinates passing through the
origin in the hyper plane R ⊕ R ⊕ · · ·⊕ R and these lines intersect at the origin only.
It follows that for each x ∈ M , [x : M ] is not an essential ideal in Z, in fact [x : M ]
is a zero-ideal in Z.
Using the description given in Remark 2.4, we now characterize all the essential
ideals corresponding to Z-modules determined by elements of A\
f (M ).
Proposition 2.5. If M is any Z-module, then [x : M ] is an essential ideal if and
only if [x : M ] is non-zero for all x ∈ A\
f (M ).
Proof. [x : M ] is an ideal in Z for each x ∈ A\
f (M ).
For any R-module M , it would be interesting to characterize essential ideals
[x : M ], x ∈ A\
f (M ) corresponding to submodules determined by elements of
\
Af (M ) (or vertices of the graph annf (Γ(M ))) such that the intersection of all
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS5
essential ideals is again an essential ideal. It is easy to see that a finite intersection
of essential ideals in any commutative ring is an essential ideal. But an infinite
intersection of essential ideals need not to be an essential ideal, even a countable
intersection of essential ideals in general is not an essential ideal as can be seen in
[6]. If annf (Γ(M )) is a finite graph, then by [Theorem 3.3, [19]] M is finite over R,
so the submodules determined by the vertices of graph are finite and therefore the
ideals corresponding to submodules are finite in number and we conclude the intersection of essential ideals [x : M ], x ∈ A\
f (M ) in R is an essential ideal. Motivated
by [6], we have have the following question regarding essential ideals corresponding
to submodules M determined by vertices of the graph annf (Γ(M )).
Question 2.6. Let M be an R-module. For x ∈ A\
f (M ), characterize essential
ideals [x : M ] in R such that their intersection is an essential ideal.
The Question 3.13 is true if every submodule of M is cyclic with nonzero intersection.
Proposition 2.7. Let M be an R-module such that every submodule of M is cyclic
over R. For x ∈ A\
f (M ), if the submodule generated by x intersects non-trivially
with every other submodule of M , then [x : M ] is an essential ideal in R.
T
Rx 6= 0. If A\
Proof. Assume
f (M ) = φ, then M is simple, a contradiction.
x∈M
Let x ∈ A\
f (M ) and let Rx be the submodule generated by x. Since Rx intersects non-trivially with every other submodule, so there exist y ∈ A\
f (M ) such
that Rx ∩ Ry 6= 0. It suffices to prove the result for Rx ∩ Ry. Let z ∈ Rx ∩ Ry
and let [x : M ], [y : M ], [z : M ] be ideals of R corresponding to submodules
Rx, Ry and Rz. Then [z : M ] ⊆ [x : M ] ∩ [y : M ] 6= 0, which implies [x : M ]
intersects non-trivially with every nonzero ideal corresponding to the submodule
generatedP
by an element of A\
f (M ). For any other ideal I of R, it is clear that
am : a ∈ I, m ∈ M } = Ra for some a ∈ M . Thus I corresponds
IM = {
f inite
to the cyclic submodule generated by a ∈ M . It follows that [x : M ] ∩ I 6= 0, for
every nonzero ideal of R and we conclude that [x : M ] is an essential ideal for each
x ∈ A\
f (M ).
The converse of Proposition 2.7 is not true in general. We can easily construct
examples from Z-modules such that an ideal corresponding to the submodule generated by some element of an object A\
f (M ) is an essential ideal, but the intersection
of all submodules determined by elements of A\
f (M ) is empty. However, if every
\
ideal [x : M ], x ∈ Af (M ) corresponds to an essential submodule of M , then we
have the nonzero intersection.
Corollary 2.8. Let M be an R-module. For x ∈ A\
f (M ), if the cyclic submodule
Rx intersects with every other cyclic submodule of M non-trivially, then [x : M ] is
6
RAMEEZ RAJA
an essential ideal in R.
Corollary 2.9. The intersection
T
[x : M ] is an essential ideal in R if and
x∈A\
f (M)
only if every submodule of M is essentially cyclic over R.
For an R-module M , let Z(M ) denote the following,
Z(M ) = {m ∈ M : ann(m) is an essential ideal in R}.
If Z(M ) = M , then M is said to be singular and if Z(M ) = 0, then M is said to be
non-singular. By rad(M ), we denote the intersection of all maximal submodules of
M . So, rad(R) is the Jacobson radical of a ring R. The socle of an R-module M
denoted by Soc(M ) is the sum of simple submodules or equivalently the intersection
of all essential submodules. To say that Soc(M ) is an essential socle is equivalent
to saying that every cyclic submodule of M contains a simple submodule of M . An
essential socle of M is denoted by essoc(M ).
In the following result, we consider singular simple R-modules (ideals) which
are injective, and obtain some properties of essential ideals corresponding to the
submodules generated by elements of A\
f (M ).
Lemma 2.10. For x ∈ A\
f (M ), if [x : M ] is an essential ideal, then R/[x : M ] is a
singular R-module.
Proof. The proof simply follows by using definition of a singular module.
Theorem 2.11. Let M be an R-module with essoc(M ) 6= 0 and
T
Rx 6= 0.
x∈M
Then every singular simple R-module [x : M ], x ∈ A\
f (M ) is injective if and only if
Z(R) = 0 and rad(R/[x : M ]) = 0.
T
Rx 6= 0, so A\
Proof. We have essoc(M ) 6= 0 and
f (M ) 6= 0. Therefore
x∈M
corresponding to every cyclic submodule generated by elements of A\
f (M ) we have
an ideal in R. Suppose all singular simple R-modules [x : M ], x ∈ A\
f (M ) are
\
injective. If for some z ∈ Af (M ), I = [z : M ] ⊆ Z(R) is a simple R-module, then
Z(I) = I. This implies I is injective and thus a direct summand of R. However,
the set Z(R) is free from nonzero idempotent elements. Therefore I = 0 and it
follows that Z(R) = 0. For x ∈ A\
f (M ), A = [x : M ] is an essential ideal of R.
Thus by Lemma 2.10, R/A is a singular module and so is every submodule of R/A.
Therefore every simple submodule of R/A is injective, which implies that every
simple submodule is excluded by some maximal submodule. Thus we conclude
that rad(R/A) = 0.
For the converse, we again consider the correspondence of cyclic submodules of
M and ideals of R. Let I˜ be a singular simple R-module corresponding to the
submodule of M . In order to show that I˜ is injective, we must show that for every
essential ideal A in R corresponding to the submodule determined by an element
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS7
˜
˜
x ∈ A\
f (M ), every ϕ ∈ HomR (A, I) has a lift ψ ∈ HomR (R, I) such that the
following diagram commutes
A
f
ϕ
R
ψ
I˜
Let K = ker(ϕ). We claim, K is an essential ideal of R, for if K ∩ J = 0 for some
nonzero ideal J of R. Then,
I ∗ = J ∩ A 6= 0 and I ∗ ∩ K = 0.
˜ a contradiction, since I˜ is a singular simple submodule
This implies I ∗ ⊆ ϕ(I ∗ ) ⊆ I,
˜
and Z(R) = 0. It is clear that if µ 6= 0, ϕ induces an isomorphism µ : A/K −→ I.
So, A/K is a simple R-submodule of R/K. By our assumption rad(R/K) = 0,
so there is a maximal submodule M/K such that R/K = A/K ⊕ M/K. Let
g : R −→ R/K be a canonical map and let p : R/K −→ A/K be a projection map.
Then we have pg : R −→ A/K. Therefore the composition h = µpg is the required
lift such that the above diagram commutes.
In the rest of this section, we discuss some interesting consequences of preceding
theorem.
T
Rx 6= 0 and
Theorem 2.12. Let M be an R-module with essoc(M ) 6= 0,
x∈M
every singular simple R-module [x : M ], x ∈ A\
f (M ) is injective. Then every ideal
[x : M ] is an intersection of maximal ideals, J(R)2 = 0 and [x : M ]2 = [x : M ].
Proof. For any x ∈ A\
f (M ), [x : M ] is an essential ideal in R. Therefore,
J(R) ⊆ [x : M ], since J(R) is contained in every essential ideal of R. On the other
hand intersection of all essential ideals in R is scole of R, therefore J(R) ⊆ Soc(R).
This implies J(R)2 = 0 and [x : M ] is an intersection of maximal ideals in R.
Suppose that [x : M ]2 6= [x : M ], for an essential ideal [x : M ] of R. By Theorem
2.11, Z(R) = 0 and therefore for every essential ideal I, I ⊆ I 2 . In particular,
2
[x : M ] ⊆ [x : M ]2 for each x ∈ A\
f (M ). It follows that [x : M ] is an essential ideal and is the intersection of maximal ideals in R. Finally, if y ∈ [x : M ]2 ,
y ∈
/ [x : M ], there is some maximal ideal P of R such that [x : M ] ⊆ P , y ∈
/ P.
Then R = Ry + P , that is, 1 = ry + m. This implies y = yry + ym ∈ P , a
contradiction and hence we conclude that [x : M ]2 = [x : M ].
Corollary 2.13. Let M be an R-module, where R is hereditary. For x ∈ A\
f (M ),
if [x : M ] is an essential ideal of R and J(R)2 = 0, then every singular simple
R-module [x : M ] is injective.
Proof. Let R be hereditary. Then from [8], the exact sequence,
0 −→ ann(x) −→ R −→ Rx −→ 0
splits for any x ∈ R. Since J(R)2 = 0 and R/J(R) is an artinian ring. Therefore,
J(R) ⊆ Soc(R). But any essential ideal of R contains Soc(R). So, J(R) ⊆ [x : M ].
This implies R/[x : M ] is completely reducible R-module and therefore rad(R/[x :
8
RAMEEZ RAJA
M ]) = 0. Thus by Theorem 2.11, every singular simple R-module [x : M ] is injective.
Next, we consider modules over regular rings.
Theorem 2.14.
T Let M be an R-module such that every submodule of M is cyclic
Rx 6= 0. Then the following are equivalent.
over R and
x∈M
(i) R is regular
(ii) A2 = A for each ideal A of R
(iii) [x : M ]2 = [x : M ] for each x ∈ A\
f (M )
Proof. The equivalence of (i) and (ii) is clear and certainly (ii) implies (iii). Thus
we just need to show that (iii) implies (ii). By Theorem 2.11, [x : M ] is an essential
2
ideal for each x ∈ A\
f (M ). Suppose [x : M ] = [x : M ]. If A is non-essential ideal
of A, choose J to be maximal ideal of R such that A ∩ J = 0, then A + J is an
essential ideal of R. Therefore again by Theorem 2.11, A + J correspond to some
submodule of M and we have A + J = [z : M ] for some z ∈ M . So,
If x ∈ A, then x =
P
f inite
(A + J)2 = A2 + J 2 = A + J.
P
mn, where a, b ∈ A and m, n ∈ J. Therefore,
ab +
x−
f inite
X
f inite
ab =
X
mn ∈ A ∩ J = 0.
f inite
This implies x ∈ A2 and we conclude that A = A2 .
Corollary 2.15. Let M be an R-module with essoc(M ) 6= 0 and
T
Rx 6= 0.
x∈M
Then every singular simple R-module [x : M ], x ∈ A\
f (M ) is injective if and only if
R is regular.
Proof. By Theorem 2.12, if every singular simple R-module [x : M ] is injective,
2
then for x ∈ A\
f (M ), [x : M ] = [x : M ]. Therefore by Theorem 2.14, R is regular.
If R is regular, then by [Theorem 6, [26]] every singular simple R-module is injective.
\
\
3. Equivalence relations on the elements of A\
f (M), As (M), At (M)
This section is devoted for the study of equivalence relations defined on elements
\
\
of objects A\
f (M ), As (M ) and At (M ). We define two equivalence relations on
\
Af (M ) called as neighbourhood similar relation (combinatorial relation) and the
submodule similar relation (algebraic relation). We investigate the equivalence of
these two relations on A\
f (M ) and study the conditions for any two elements of
\
Af (M ) to be adjacent in annf (Γ(M )). We explore the neighbourhood similar relation on A\
f (M ) to establish the structure of module M and the full-annihilating
graph annf (Γ(M )).
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS9
Definition 3.1. For an R-module M , two distinct elements m1 , m2 ∈ A\
f (M ) are
submodule similar (∼M ) denoted by m1 ∼M m2 if and only if the submodules
ann(m1 )M and ann(m2 )M of M coincide, that is,
m1 ∼M m2 if and only if ann(m1 )M = ann(m2 )M .
Clearly, ∼M is an equivalence relation on A\
f (M ) and the equivalence classes of any
m ∈ A\
(M
)
is
denoted
by,
f
[m]M = {m∗ ∈ M : m∗ ∼M m}.
Analogously, we can define the submodule relation on the elements of objects A\
s (M )
\
and At (M ).
Let Γ be any connected graph. A neighbour of any vertex v in Γ is a vertex
adjacent to v. N (v) denotes the set all neighbours of v and N [v] = N (v) ∪ {v}.
The study of neighbourhoods of vertices in a connected graph Γ is related to the
symmetry of that graph. There is a close relationship, which is being discussed
in the following definition between the neighbourhoods and the distance similar
classes of vertices defined in [20] .
Definition 3.2. For a connected graph Γ, two distinct vertices u, v ∈ V (Γ) on a
vertex set V (Γ) are neighbourhood similar (∼nbd ) denoted by u ∼nbd v if and only
if N (u) = N (v).
It can be easily checked that ∼nbd is an equivalence relation on V (Γ). The
neighbourhood similar equivalence class of a vertex v is,
[v]Γ = {w ∈ V (Γ) : N (v) = N (w)}.
Two distinct vertices a, b ∈ V (Γ) which are not neighbourhood similar can be
identified as the vertices for which ab ∈ E(Γ) with N (a) 6= N (b) or ab ∈
/ E(Γ) with
N (a) 6= N (b), where E(Γ) denotes the edge set of Γ.
If Γ is a finite connected graph, then the neighbourhood similar relation on V (Γ)
is a distance similar relation (∼d ) defined in [20] with two vertices u, v ∈ V (Γ) are
distance similar denoted by u ∼d v if d(u, w) = d(v, w) for all w ∈ V (Γ) \ {u, v}.
Clearly, vertices u and v are distance similar if either uv ∈
/ E(Γ) and N (u) = N (v)
or uv ∈ E(Γ) and N [u] = N [v].
The following example illustrates neighbourhood similar relation on a connected
graph Γ.
Example 3.3. Consider the graph Γ of order n = 10, shown in Figure 1. Two
of the five neighbourhood similar equivalence classes are V1 = {v6 , v5 } and V2 =
{v7 , v8 , v9 , v10 }. Each of the three remaining classes {v1 }, {v4 } and {v6 } consists
of a single vertex.
10
RAMEEZ RAJA
v2
v3
v7
v5
v8
v1
v6
v4
v9
v10
.
Now, we explain the connection between neighbourhood similar and submodule
similar equivalence relations on an object A\
f (M ). In fact, we will see that the
neighbourhood similar relation ∼nbd , which is a combinatorial relation implies an
algebraic relation ∼M and conversely. Moreover, we investigate the condition for
any two elements of A\
f (M ) to be adjacent in annf (Γ(M )). We start with following
two lemmas.
Lemma 3.4. Let M be an R-module. Then for x, z, y ∈ A\
), [x : M ][z :
f (M T
M ]M = 0 if and only if z ∈ ann(x)M or z ∈ ann(y)M with ann(x)M ann(y)M =
{0}.
Proof. The proof follows because of the fact that each element of an ideal ann(y)
in R annihilate whole of Ry.
Lemma 3.5. If M is an R-module and x, y ∈ A\
f (M ) with x ∼nbd y, then either
[x : M ]2 M = [y : M ]2 M = 0 or both [x : M ]2 M 6= 0 and [y : M ]2 M 6= 0.
Proof. Let M be an R-module and let x ∼nbd y for x, y ∈ A\
f (M ). Assume that
[x : M ]2 M = 0 and [y : M ]2 M 6= 0. We consider the following cases.
Case 1. x and y are adjacent in annf (Γ(M )), that is [x : M ][y : M ]M = 0. Then
it is easy to check that [x + y : M ][x : M ]M = 0, which implies x + y ∈ N (x). Note
here that x + y 6= x and x + y 6= y. However, x + y ∈
/ N (y), since [x + y : M ][y :
M ]M 6= 0. Therefore N (x) 6= N (y), a contradiction.
Case 2. x and y are not adjacent in annf (Γ(M )), that is [x : M ][y : M ]M 6= 0.
Then by case 1, we can easily find an element z ∈ A\
f (M ) such that [z : M ][x :
M ]M = 0, where as [z : M ][y : M ]M 6= 0. Therefore, N (x) 6= N (y), again a
contradiction.
Theorem 3.6. Let M be an R-module with |A\
f (M )| being at least 3. Then for
x, y ∈ A\
(M
)
the
following
hold.
f
(i) If M is a multiplication module, then x ∼nbd y if and only if x ∼M y.
(ii) If x ∼nbd y, then [x : M ][y : M ]M = 0 if and only if [x : M ]2 M = [y :
M ]2 M = 0.
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
11
Proof. (i) Directly follows from Lemma 3.4 and the fact that every submodule of
a multiplication module M is of the form IM , for some ideal I of R.
(ii) Let x ∼nbd y and suppose [x : M ]2 M 6= 0. Then by Lemma 3.5, [y : M ]2 M 6=
0. Assume that [x : M ][y : M ]M = 0. There must be without loss of generality
some m ∈ N (x) with m 6= x, such that m ∈ N (y). Then [y : M ][x + m : M ]M = 0,
which implies either x + m ∈ N (x) or x + m = x or x + m = y. But, x + m 6= x as
[x : M ]2 M 6= 0, also x + m ∈
/ N (x), since [x : M ][x + m : M ]M 6= 0 and x + m 6= y,
since y ∈ N (x). Lastly, if x + m = 0, then [x : M ]2 M = 0. Thus, we have a
contradiction and therefore, [x : M ][y : M ]M 6= 0.
Now, suppose [x : M ]2 M = 0. Therefore by Lemma 3.5, [y : M ]2 M = 0. Assume
that [x : M ][y : M ]M 6= 0. Then by Lemma 3.4, x ∈
/ ann(x)M and x ∈
/ ann(y)M .
\
Since |A\
(M
)|
≥
3,
so
there
is
some
z
∈
A
(M
)
such
that
[x
:
M
][z
:
M ]M = 0.
f
f
This implies z = x or z = y, since x ∼nbd y. But, z 6= y as y ∈
/ N (x). Therefore
z = x and we have for all r ∈ [x : M ], x + ry ∈ ann(x)M or x + ry ∈ ann(y)M .
However, x + ry ∈
/ ann(y)M , since [x + ry : M ][y : M ]M 6= 0. On the other hand,
x + ry ∈ N (y), since x ∼nbd y, a contradiction. If x + ry = x or x + ry = y, then
[x : M ][y : M ]M = 0, which is again a contradiction.
\
The preceding theorem is also true for the elements of objects A\
s (M ) and At (M ),
that is, the same adjacency relations hold for the vertices of semi-annihilating
anns (Γ(M )) and the star-annihilating graph annt (Γ(M )).
Recall that an element a ∈ R is said to be nilpotent if for some n ∈ N, an = 0.
We denote the set of all nilpotent elements of R by nil(R) called the nil radical of
R which is contained in every prime ideal of R. For a module M , we denote by
nil(M ) the sub module which is contained in every prime submodule of M that is,
T
N = {x ∈ M | [x : M ]x = 0},
nil(M ) =
N ∈Spec(M)
where N = {x ∈ M | [x : M ]x = 0} and Spec(M) is the set of all prime submodules
of M . Also, note that a vertex v ∈ V (Γ) is said to be a pendant vertex if there is
only one vertex adjacent to it. That is, if degree of vertex a v, denoted by deg(v) is 1.
In the remaining section, we study the neighbourhood classes of elements in M .
Consider the elements of M which are not neighbourhood similar in annf (Γ(M )),
we investigate about the structure of a module M if any two elements in annf (Γ(M ))
are not neighbourhood similar. We make use of neighbourhood similar elements to
study the nature of all annihilating graphs arising from M . Before, we discuss the
results regarding neighbourhood variant and invariant classes we have the following
key lemmas.
Lemma 3.7. For any R-module M , annf (Γ(M )) is not an n − gon for n ≥ 5.
Proof. Suppose annf (Γ(M )) is the graph with vertices {x1 , x2 , · · · , xn } such that
x1 − x2 , x2 − x3 , · · · , xn−2 − xn−1 , xn−1 − xn , xn − x1 are the only adjacencies in
annf (Γ(M )). Then, we have [x1 : M ][x2 : M ]M = 0 = [x1 : M ][xn : M ]M . This
implies [x1 : M ][x2 +xn : M ]M = 0. It follows that x2 +xn is either x1 , x2 , · · · , xn−1
or xn . A simple check yields that x2 + xn = x1 is the only possibility. Similarly,
12
RAMEEZ RAJA
x1 + xn−1 = xn . Therefore, xn = x2 + xn + xn−1 . So,
[x1 : M ][xn : M ]M = [x1 : M ][x2 : M ]M +[x1 : M ][xn : M ]M +[x1 : M ][xn−1 : M ]M.
This implies [x1 : M ][xn−1 : M ]M = 0, a contradiction, and hence we conclude
that annf (Γ(M )) is not an n − gon for n ≥ 5.
Lemma 3.8. Let M be an R-module. Then for all x ∈ M , [x : M ]x = 0 if and
only if [x : M ]2 M = 0.
Proof. The proof follows because of the fact that an ideal [x : M ] of R annihilate
every multiple of x.
Lemma 3.9. For an R-module M with nil(M ) 6= 0, let annf (Γ(M )) be the full
annihilating graph such that annf (Γ(M )) contains no cycle of length 3. Then the
following hold.
(i) If for every pair y, z ∈ A\
f (M ) with y ∼nbd z there exists an element x ∈
\
Af (M ) such that [x : M ][z : M ]M = 0 = [x : M ][y : M ]M , then either 8 ≤ |M | ≤
16 or |M | ≥ 17 and nil(M ) = {0, x}.
\
(ii) If for each x ∈ A\
f (M ), there exists a pair z, y ∈ Af (M ) \ {x} such that
y ∼nbd z and [x : M ][y : M ]M = 0 = [x : M ][z : M ]M , then any a ∈ A\
f (M ) with
[x : M ][a : M ]M = 0 is a pendant vertex in annf (Γ(M )).
Proof. (i) Let nil(M ) 6= 0. We first show that for all r ∈ [x : M ], rn x = 0 for some
n ∈ N, where x ∈ nil(M ). Consider a set S = {rn x : n ∈ N}, we must show that
0 ∈ S. Suppose that 0 ∈
/ S and let N be a submodule of M such that N ∩ S = ∅.
Then by Zorn’s Lemma the collection Σ = {N : N ∩ S = ∅} contains a maximal
element. Let K be a maximal member of Σ. We claim that [N : M ] is a prime
ideal of R. Clearly, [N : M ] ⊂ R. Let r1 , r2 ∈ R and suppose r1 r2 ∈ [N : M ] with
r1 , r2 ∈
/ [N : M ]. Then (r1 M + K) ∈
/ Σ and (r2 M + K) ∈
/ Σ. So r1n x ∈ S ∩ r1 M + K
n
and r2 x ∈ S ∩ r2 M + K for some n1 , n2 ∈ N. Therefore rn1 +n2 x ∈ K ∩ S, a
contradiction. Thus [N : M ] is a prime ideal and by [Corollary 2.11, [10]], K is a
prime submodule of M . Therefore we have rx ∈ N ∩ S, since x ∈ [N : M ], which
is a contradiction and consequently rn x = 0. By well ordering principle choose n
to be smallest such that rn x = 0. Then for n ≥ 1, rn−1 x 6= 0.
Claim: n ≤ 3. Suppose to the contrary that n > 3. Clearly, rn−1 x, rx ∈ A\
f (M ),
since [rn−1 x : M ][rx : M ]M = 0. So, there exist a vertex say y ∈ A\
(M
)
such
f
that [rn−1 x : M ][y : M ]M = 0 = [y : M ][rx : M ]M , which implies [rn−1 x : M ][y :
M ]M = 0 = [rn−1 x : M ][rx : M ]M . Therefore rn−1 x = y is the only possibility,
since rn−1 x ∼nbd rx. Similarly for each i, 1 ≤ i ≤ n − 2, we have [ri x : M ][rn−1 x :
M ]M = 0. For m = rn−2 x + rn−1 x, we see that annf (Γ(M )) contains a cycle
of length 3, since [rn−1 x : M ][m : M ]M = 0 = [rn−2 x : M ][m : M ]M = 0 with
m∈
/ {0, rn−1 x, rn−2 x}. Thus n ≤ 3.
We consider the following cases for n ≤ 3.
Case 1. n = 3. We show that ann(r2 x)M is the unique maximal submodule of
M and |M | = 8 or |M | = 16. We first show that |M | = 8 or |M | = 16. By our claim
above, [r2 x : M ][rx : M ]M = 0. If 0 6= z ∈ ann(x)M , then [z : M ][x : M ]M = 0.
This implies annf (Γ(M )) contains a cycle of length 3 on vertices z, rx and r2 x.
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
13
Therefore z = r2 x and ann(x)M ⊆ {0, r2 x}. In fact, ann(x)M ⊆ Rr2 x, since
for all s ∈ R, [sr2 x : M ][rx : M ]M = 0 = [sr2 x : M ][r2 x : M ]M . Therefore
sr2 x ∈ {0, rn−1 x, rn−2 x}. If sr2 x = rx, then r2 x = 0, a contradiction. Thus
Rr2 x = {0, r2 x}.
Moreover, ann(r2 x)M ⊆ {0, x, rx, r2 x, x + rx, x + r2 x, rx + r2 x, x + rx + r2 x}.
If z ∈ ann(r2 x)M , then r2 z ∈ ann(x)M ⊆ {0, r2 x}. So, either r2 z = 0 or r2 z =
r2 x. Therefore [rz : M ][rx : M ]M = 0 = [rz : M ][r2 x : M ]M or [rz − rx :
M ][rx : M ]M = 0 = [rz − rx : M ][r2 x : M ]M . This implies rz ∈ {0, rx, r2 x} or
rz − rx ∈ {0, rx, r2 x}. Let r2 z = 0. So, rz 6= rx and therefore either rz = 0 or
r(z − rx) = 0, which implies [z : M ][rx : M ]M = 0 = [z : M ][r2 x : M ]M or [z − rz :
M ][rx : M ]M = 0 = [z − rz : M ][r2 x : M ]M . Thus z ∈ {0, rx, r2 x, rx + r2 x}.
Therefore we may assume that r2 z = r2 x, which implies rz − rx 6= rx. But,
rz − rx ∈ {0, rx, r2 x}. So, either rz − rx = 0 or rz − rx = r2 x and by similar
argument as above z ∈ {x, r2 x, x + rx, x + rx + r2 x}.
If [x : M ][r2 x : M ]M = 0, then
(1)
ann(r2 x)M = {0, x, rx, r2 x, x + rx, x + r2 x, rx + r2 x, x + rx + r2 x}.
If [x : M ][r2 x : M ]M 6= 0, then
(2)
ann(r2 x)M = {0, rx, r2 x, rx + r2 x}
For (1), |M | = 16 and for (2), |M | = 8, since [x : M ]r2 M 6= 0, so there are
t ∈ [x : M ] and m ∈ M such that r2 tm 6= 0. It is clear that r2 tm = r2 x. Let
m∗ ∈ M . Then r2 tm∗ ∈ Rr2 x = {0, r2 x}. If r2 tm∗ = 0, then m∗ ∈ ann(r2 x)M
and if r2 tm∗ = r2 x, then m∗ − m ∈ ann(r2 x)M .
Let K = ann(r2 x)M . Clearly, K ⊂ M and Rr2 x ∼
= R/ann(r2 x). Therefore
ann(r2 x) is a maximal ideal, since Rr2 x = {0, r2 x}. Thus it follows by [Theorem
2.5, [10]] that K is a maximal submodule. Further, K ⊆ Rx ⊆ nil(M ) ⊆ K.
Therefore K = nil(M ) is the unique maximal submodule of M . If Af (M ) ⊆ K,
then Af (M ) = K, which implies annf (Γ(M )) is a star graph.
Case 2. n = 2. We show that |M | ≤ 12. If [x : M ]2 x = [x : M ]3 M 6= 0, then
there exist two elements r, s ∈ [x : M ] such that rsx 6= 0. Further, there are m ∈ M
and t ∈ [x : M ] such that rstm 6= 0. However, r2 x = s2 x = t2 x = 0 and there is
some a ∈ A\
f (M ) such that [a : M ][rx : M ]M = 0. It is easy to check that Rrx ⊆
{0, rx, a} and a = rsx. This implies that [srx : M ][rx : M ]M = 0 and therefore
Rrx = {0, rx, rsx} and ann(rx)M = {0, rx, rsx}. Clearly, [rstm : M ][s1 x :
M ]M = 0, [rstm : M ][s1 s2 x : M ]M = 0 with rstm 6= s1 x. Thus rstm = rsx and
it is clear that [rsm : M ][rx : M ]M = 0 and [rsm : M ][rsx : M ]M = 0. But,
rsm 6= s1 x and rsm 6= rsx, a contradiction. Therefore [x : M ]2 x = 0.
Moreover, we have rsm 6= 0, for some s ∈ [x : M ] and m ∈ M , since r2 x = 0,
\
rx 6= 0. For x, y ∈ A\
f (M ) with [x : M ][y : M ]M = 0, there is rx ∈ Af (M )
such that [x : M ][rx : M ]M = 0 and [rx : M ][y : M ]M = 0. So, rx = x
or rx = y. If rx = x, then rx = 0, which is not possible. Therefore rx = y
and we have [x : M ][rx : M ]M = 0. Let z ∈ ann(x)M . Then z ∈ {0, x, rx},
since [x : M ][rx : M ]M = 0. If z = x, then x[x : M ] = [x : M ]2 M = 0,
which is a contradiction. Thus ann(x)M = {0, rx} and in fact Rrx = {0, rx}.
\
On the other hand, rm ∈ A\
f (M ), so there exist b ∈ Af (M ) such that [rm :
14
RAMEEZ RAJA
M ][b : M ]M = 0. But, [rsm : M ][rm : M ]M = 0 and [rsm : M ][b : M ]M = 0,
therefore rsm = b is the only possibility, so [rsm : M ][rm : M ]M = 0. Also,
[rsm : M ][rx : M ]M = 0 and [rsm : M ][x : M ]M = 0. Thus by the same reasoning
as above we have rsm = rx. Let t ∈ ann(rx)M . Therefore rt ∈ ann(x)M =
{0, rx}. If rt = 0, then [t : M ][rsm : M ]M = 0 = [t : M ][rm : M ]M . If
rt = rx, then [t − x : M ][rsm : M ]M = 0 = [t − x : M ][rsm : M ]M . Therefore
ann(rx)M = {0, rm, rx, x+rm, x+rx}. By similar argument as in case 1, it can be
shown that |M | ≤ 12 and ann(rx)M = nil(M ) is the unique maximal submodule
of M .
Case 3. n = 1. If [x : M ]x = [x : M ]2 M 6= 0, then by cases 1 and 2 we
have 8 ≤ |M | ≤ 16. Assume that [x : M ]x = 0, we show that either
|M | = 9 or
P
ri mi , where Λ
nil(M ) = {0, x} with 2x = 0. Let x ∈ [x : M ]M . Then x =
i∈Λ
is finite, ri ∈ [x : M ] and mi ∈ M with 1 ≤ i ≤ |Λ|. For x ∈ A\
f (M ), there is
\
rx ∈ Af (M ) such that [x : M ][y : M ]M = 0. So, Rx ⊆ {0, x, y}. If x 6= ri mi for
all i, 1 ≤ i ≤ |Λ|, then ri mi ∈ Rx. Therefore ri mi = y for all i, 1 ≤ i ≤ |Λ|. Thus
it follows that x = rm for some r ∈ [x : M ] and m ∈ M with rm 6= 0. Clearly,
x + x ∈ Rx ⊆ {0, x, y}. If x + x 6= 0, then Rx = {0, x, 2x}, [x : M ][2x : M ]M = 0
and ann(x)M = {0, x, 2x}. Thus for all m∗ ∈ M , rm∗ ∈ Rx and we have,
[m∗ : M ][x : M ]M = 0 = [m∗ : M ][2x : M ]M ,
or
[m∗ − m : M ][x : M ]M = 0 = [m∗ − m : M ][2x : M ]M ,
or
[m∗ − 2m : M ][x : M ]M = 0 = [m∗ − 2m : M ][2x : M ]M .
By a similar argument as in case 1 it can be shown that |M | = 9 and ann(x)M is
the unique maximal submodule of M . Let |M | 6= 9. Then by the above argument
we must have 2x = 0. We show that nil(M ) = {0, x}. If 0 6= z ∈ nil(M )
with z 6= x, then z[z : M ] = [z : M ]2 M = 0. Therefore z = sm1 for some
s ∈ [z : M ] and m1 ∈ M with s2 m1 6= 0. It is clear that [x : M ][x1 : M ]M = 0
\
and [z : M ][z1 : M ]M = 0 with x1 , z1 ∈ A\
f (M ), since x, z ∈ Af (M ). Thus
Rx ⊆ {0, x, x1 } and Rz ⊆ {0, z, z1}. If 0 6= rsm ∈ Rx and rsm ∈ Rz, say
rsm = x ∈ Rz, then x = z1 which implies [x : M ][z : M ]M = 0. Therefore
rsm = 0, a contradiction. If rsm = x1 , then Rx = {0, x, rsm} = ann(x)M
and by similar argument as above we see that |M | = 9, again a contradiction.
Thus rsm = 0 and similarly rsm1 = 0. Clearly, x + z 6= x, x + z 6= z and
\
x + z ∈ A\
f (M ), so there is t ∈ Af (M ) such that [x + t : M ][t : M ]M = 0. Also,
rt ∈ Rx ⊆ {0, x, x1 }. If rt = 0, then [x : M ][t : M ]M = 0 = [x : M ][x + z : M ]M .
That is, we have a cycle on vertices x, t and x + z, a contradiction. If rt = x1 , then
Rx = {0, x, rt} = ann(x)M and by the same argument as above we have |M | = 9,
again a contradiction. Therefore we may assume that x = rt and z = st. Then
[x + z : M ][x : M ]M = 0 = [x + z : M ][t : M ]M .
This implies x + z = 0. Thus we have a contradiction in every possible case and
hence we conclude that nil(M ) = {0, x} with 2x = 0.
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
15
(ii) By (i), we see that nil(M ) = {0, x} for some 0 6= x ∈ M with 2x = 0,
\
[x : M ]2 M = 0 and |M | ≥ 17. Let x ∈ A\
f (M ). Then, there is y ∈ Af (M ) such
\
that [x : M ][y : M ]M = 0. Clearly, there is another x + y ∈ Af (M ) with x + y 6= x
and x + y 6= y such that [x + y : M ][y : M ]M = 0. Therefore, x + y ∼nbd y, that is
N (x + y) = N (y). If there is some t ∈ A\
f (M ) such that [x + y : M ][t : M ]M = 0,
then [y : M ][t : M ]M = 0. This implies [x : M ][t : M ]M = 0, a contradiction, since
annf (Γ(M )) contains no cycle of length 3. Thus, x + y = t or y = t. But, y 6= t,
otherwise x = 0, a contradiction, since nil(M ) = {0, x} with 0 6= x. Therefore
x + y = t and consequently t ∈ A\
f (M ) is a pendant vertex in annf (Γ(M )).
Using the neighbourhood similar relation on the elements of objects A\
f (M ),
\
\
As (M ) and At (M ), we now reveal the structure of a module M .
Theorem 3.8. Let M be an R-module and let annf (Γ(M )) be the full annihilating
graph of M which contains no cycle of length 3. If for each x ∈ A\
f (M ), there exists
\
a pair z, y ∈ Af (M ) \ {x} such that y 6∼nbd z and [x : M ][y : M ]M = 0 = [x :
M ][z : M ]M , then M = M1 ⊕ M2 , for some submodules M1 and M2 of M .
\
Proof. Let x ∈ A\
f (M ). There are z, y ∈ Af (M ) \ {x} such that y 6∼nbd z and
[x : M ][y : M ]M = 0 = [x : M ][z : M ]M . So, there is some t ∈ A\
f (M ) such that
[t : M ][y : M ]M = 0, where as [t : M ][z : M ]M 6= 0. Clearly, for all r ∈ [z : M ],
0 6= rt ∈ Af (M ) and we have [rt : M ][y : M ]M = 0. This implies rt = x or rt = y,
since [x : M ][y : M ]M = 0. Therefore [y : M ]2 M = 0 = y[y : M ] or [x : M ]2 M =
0 = x[x : M ]. That is, x ∈ nil(M ) or y ∈ nil(M ). By Lemma 3.7, nil(M ) = {0, m}
for some 0 6= m ∈ M . Suppose that x = m. Then [m : M ][y : M ]M = 0 and
[m : M ][t : M ]M 6= 0, otherwise annf (Γ(M )) contains a cycle of length three on
t, m and y. So, there is some r1 ∈ [t : M ] such that r1 m 6= 0. It is clear that
[r1 m : M ]2 M = 0, since [m : M ]2 M = 0. Therefore for all r1 ∈ [m : M ], r1 m = m.
Let s = r1 t − t. Then [s : M ][y : M ]M = 0, since [t : M ][y : M ]M = 0. For r1 ∈
[m : M ], we see that us = ur1 t − ut = ut − t = 0. Therefore [m : M ][s : M ]M = 0.
This implies s ∈ {0, y, m}, since [m : M ][y : M ]M = 0. We consider the following
cases.
Case 1. s = y. Then y ∈ nil(M ), which is a contradiction.
Case 2. s = 0. Then r1 t = t, that is, (r1 − 1)t = 0, which implies r1 − 1 ∈ ann(t)
and therefore, M = Rt ⊕ ann(t)M .
Case 3. s = m. Then r1 t − t ∈ nil(M ). Let r2 = r12 − r1 . Therefore r2 ∈ [r1 t − t :
M ] and by Lemma 3.7, we have r2n (r1 t − t) = 0, for some n ∈ N. Thus r2n+1 t = 0.
For some suitable choice s2 in terms of r2 we see that (s22 − s2 )(1 + 4r2 )t + r2 t = 0.
Therefore for v = r1 + s2 (1 − 2r1 ), we have vt = v 2 t and v ∈ [t : M ]. It follows
that there is some w ∈ R such that w ∈ ann(r1 t). Thus by a similar argument as
in case 2 above we conclude that M = Rr1 t ⊕ ann(r1 t)M .
16
RAMEEZ RAJA
Remark 3.9. Theorem 3.8 is also true for the vertices of annihilating graphs
anns (Γ(M )) and annt (Γ(M )). That is, if we have the information regarding the el\
ements of objects A\
s (M ) and At (M ), we can establish the structure of a module M .
Recall that a complete bipartite graph is one whose each vertex of one partite
set is joined to every vertex of the other partite set. We denote the complete bipartite graph with partite sets of size m, n ∈ N by Km,n . More generally a complete
r-partite graph is one whose vertex set can be partitioned into r subsets so that
no edge has both ends in any one subset and each vertex of a partite set is joined
to every vertex of the another partite sets. A complete bipartite graph of the form
K1,n is called a star graph.
In the following result, we discuss the neighbourhood relation for the elements
x, y ∈ M with [x : M ][y : M ]M 6= 0 in annf (Γ(M )). We show that the neighbourhood similar vertices determines the nature of the annihilating graphs arising from
M.
Theorem 3.10. Let M be an R-module and let annf (Γ(M )) be the full annihilating graph of M which contains no cycle of length 3. If for every pair z, y ∈ A\
f (M )
\
with [z : M ][y : M ]M 6= 0 there exists x ∈ Af (M ) \ {y, z} such that y ∼nbd z and
[x : M ][y : M ]M = 0 = [x : M ][z : M ]M , then annf (Γ(M )) is a star graph.
Proof. By Lemma 3.7 (i), we have either, 8 ≤ |M | ≤ 16 or |M | ≥ 17 and
nil(M ) = {0, x} with 2x = 0, 0 6= x ∈ M . If 8 ≤ |M | ≤ 16, then again by Lemma
3.7 (i), annf (Γ(M )) is star graph with at most 5 edges.
Claim 1: annf (Γ(M )) is an infinite graph, for |M | ≥ 17, with nil(M ) = {0, x}.
Let y ∈ A\
f (M ) such that [x : M ][y : M ]M = 0. Then for all r ∈ [y : M ], we have
n
P
ri , ri ∈ [y : M ] with 1 ≤ i ≤ n. If z ∈ A\
[x : M ][ry : M ]M = 0, where r =
f (M )
i=1
is any other element such that [x : M ][z : M ]M = 0 = [ry : M ][z : M ]M .
Then annf (Γ(M )) contains a cycle on x, z and ry, a contradiction. Similarly,
it can be shown that for all p, with 1 ≤ p ≤ m, [x : M ][rp y : M ]M = 0 and
consequently by Lemma 3.7 (ii), rp y is a pendant vertex in annf (Γ(M )). Note,
that all rp y are distinct. If not, then rp y = rq y, with 1 ≤ p < q ≤ m. Therefore
y(rp − rq )P
= yrp (1 − rq−p ) = 0. This implies (1 − rq−p ) ∈ ann(rp y). By Lemma 3.7
sm = s∗ m∗ for some s∗ ∈ [x : M ] and m∗ ∈ M . Thus, (1 − rq−p )m∗ ∈
(i), x =
f inite
ann(rp y)M = {0, x}. So, either m∗ − rq−p m∗ = 0 or m∗ − rq−p m∗ = x. If
m∗ = rq−p m∗ , then
x = s∗ m∗ = s∗ rq−p m∗ ∈ s∗ rq−p−1 )m∗ Rr ⊆ [x : M ][rq−p−1 y : M ]M = 0,
which is a contradiction. If m∗ − rq−p m∗ = x, then
x − rq−p s∗ m∗ = s∗ m∗ − rq−p s∗ m∗ = s∗ x ⊆ [x : M ]2 M = 0,
again a contradiction. Thus all rp y are distinct and consequently, annf (Γ(M )) is
an infinite graph.
Claim 2: annf (Γ(M )) is a star graph with center x.
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
17
Let a, b ∈ A\
f (M ) such that [a : M ][b : M ]M 6= 0 and [x : M ][a : M ]M = 0 =
[x : M ][b : M ]M . Suppose that annf (Γ(M )) is not a star graph. There is some
c ∈ A\
f (M ) \ {x, a} such that [a : M ][c : M ]M = 0, that
Pis a is not a pendant vertex
tα , tα ∈ [a : M ] for each
in annf (Γ(M )). Consider an element tc, where t =
α∈Λ
α ∈ Λ, Λ is any finite index set. We show that tc 6∈ {0, a, x, c, b}. If tc = 0, then for
\
a ∈ A\
f (M ) we have x, c ∈ Af (M ) with x ∼nbd c. Therefore by Lemma 3.7 (ii), a
is a pendant vertex in annf (Γ(M )), which contradicts our supposition. If tc = x,
then [x : M ][tc : M ]M = 0. Therefore c ∈ ann(at)M = {0, x}, a contradiction.
If tc = c, then [x : M ][tc : M ]M = 0, a contradiction, since annf (Γ(M )) contains
no cycle of length 3. Finally, if tc = b, then [c : M ][tc : M ]M = 0, which implies,
c ∈ nil(M ) = {0, x}, again a contradiction. Thus tc ∈ A\
f (M ) \ {x, a, c, b}. So,
there is some a1 ∈ A\
(M
)
such
that
[a
:
M
][tc
:
M
]M
=
0. It is easy to verify
f
1
that a1 6∈ {0, tc, a, x, c, b}. Moreover, it is clear that [x : M ][a1 : M ]M 6= 0,
P othλβ ,
erwise annf (Γ(M )) contains cycle of length 3. Let a1 = λm1 , where λ =
β∈Λ
λβ ∈ [a1 : M ] for each β ∈ Λ and m1 ∈ M . Then [λx : M ][a : M ]M = 0,
which implies λx ∈ ann(a)M . Thus λx = x, note that λx 6= 0, otherwise
[x : M ][a1 : M ]M = 0. On the other hand [a1 : M ][tc : M ]M = 0. Therefore,
[λc : M ][a : M ]M = 0 = [ta1 : M ][c : M ]M , which implies λc ∈ ann(a)M . Thus
λc = 0 or λc = x. So, [x : M ][c : M ]M = 0 = [x : M ][b : M ]M , a contradiction. Thus we have a contradiction in every possible case and hence it follows that
annf (Γ(M )) is a star graph with center x.
We conclude this section with the following remark.
Remark 3.11. Theorem 3.10 is not true for general simple connected graphs.
Consider the graph Γ with vertex set V = {a1 , a2 , a3 , a4 } and edge set E = {a1 −
a2 , a2 − a3 , a3 − a4 , a4 − a1 }. Clearly, a1 ∼nbd a3 with N (a1 ) = N (a3 ) = {a2 , a4 },
but Γ is not a star graph in fact Γ is a cycle graph on four vertices.
4. Graph isomorphism and graphs arising from tensor product
In this section, we discuss isomorphism of annihilating graphs. We exhibit certain module theoretic properties of modules M and N which they share if their
full annihilating graphs annf (Γ(M )) and annf (Γ(N )) are isomorphic. Moreover,
we consider the annihilating graph arising from tensor product M ⊗R T −1 R, where
T = R\C(M ), where C(M ) = {r ∈ R : rm = 0 f or some 0 6= m ∈ M }. We investigate the case when M is a multiplication module and show that annf (Γ(M )) ∼
=
annf (M ⊗R T −1 R) for every module M .
Definition 4.1. Let x ∈ M be a vertex in annf (Γ(M )). We say that x is primitive
vertex if the submodule generated by x is cyclic over R.
It can be easily checked that an element a ∈ Z ∗ (R), where R is a von Neumann
regular ring is a primitive vertex in the zero-divisor graph [1] if and only if the ring
18
RAMEEZ RAJA
Ra is a field. We define the order of vertex x ∈ A\
f (M ) by Θ(x) = |Rx|. Clearly, if
\
x ∼nbd y for y ∈ Af (M ), then |Rx| = |Ry|, which implies Θ(x) = Θ(y). Thus we
can talk about the order of equivalence classes.
The following result is one of the entanglement for modules M and N if their
full annihilating graphs are isomorphic.
Theorem 4.2. Let M and N be two R-modules such that annf (Γ(M )) ∼
= annf (Γ(N )).
If Soc(M ) is a sum of finite simple cyclic submodules, then Soc(M ) ∼
= Soc(N ).
Proof. Let η : annf (Γ(M )) −→ annf (Γ(N )) be an isomorphism of graphs. It
is clear that if x ∼nbd y, then η(x) ∼nbd η(y) and Θ(x) = Θ(η(x)). Also, the
primitive elements of annf (Γ(M )) are in bijection with the primitive elements of
annf (Γ(N )). Thus if x is primitive, so is Θ(x). Suppose that Soc(M ) contains
the sum of simple cyclic submodules δ times. Then there are δ many equivalence
classes of primitive vertices in annf (Γ(M )). This implies that annf (Γ(N )) must
have δ equivalence classes of primitive vertices. It follows that the same number of
copies of each simple cyclic submodule must occur in both Soc(M ) and Soc(N ).
Thus, Soc(M ) and Soc(N ) are isomorphic.
Corollary 4.3. Let M =
Q
Mi and N =
Q
Ni , where Mi , Ni are finite simple
i∈I
i∈I
cyclic modules for all i ∈ I and I is an index set. If annf (Γ(M )) ∼
= annf (Γ(N )),
then M ∼
= N.
P
P
Ni . Thus the result follows.
Mi and Soc(N ) =
Proof. Soc(M ) =
i∈I
i∈I
Corollary 4.4. Let M and N be two R-modules such that annf (Γ(M )) ∼
=
annf (Γ(N )). If M has an essential socle, then so does N .
Let Zm ⊗ Zn be tensor product of two finite abelian groups. It is easy to verify
that if g.c.d of m, n ∈ Z is 1, then Zm ⊗ Zn = 0 and in general Zm ⊗ Zn ∼
= Zd , where
V
d is g.c.d of m and n. It follows that if g.c.d of m and n is 1, then Af (Zm ⊗ Zn ) = φ.
However, if g.c.d of m and n is d, d > 1 and Zd is not a simple finite abelian group,
V
then Af (Zm ⊗ Zn ) 6= φ, in fact the graphs annf (Γ(Zm ⊗ Zn )) and annf (Γ(Zd ))
are isomorphic. Furthermore, if Zp , Zq and Zr are any three finite simple abelian
groups, where p, q, r ∈ Z are primes, then we have the following equality between
the combinatorial objects,
V
V
Af (Zp ⊕ Zq ⊗ Zp ⊕ Zr ) = Af (Zp ⊕ Zr ).
It follows that the full-annihilating graph arising from the tensor product Zp ⊕ Zq ⊗
V
Zp ⊕ Zr with vertex set Af (Zp ⊕ Zq ⊗ Zp ⊕ Zr ) is same as the full-annihilating
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
19
graph arising from the direct sum Zp ⊕ Zr .
Now, we study the annihilating graph structures arising from the tensor product
M ⊗R T −1 R. The following result proved in [3] is perhaps the first result which
establishes a connection between the graph structure of R and its localization S −1 R
(total quotient ring) at S, where S = R \ Z(R).
Theorem 4.5. Let R be a commutative ring with unity and let S −1 R be the
localization of R at S. Then the graphs Γ(R) and Γ(S −1 R) are isomorphic.
The equivalence class corresponding to equivalence relation ∼M for m ⊗
M ⊗R T −1 R is denoted by,
[m ⊗ 1s ]M⊗R T −1 R = {m∗ ⊗
1
s
1
s
∈
∈ M ⊗R T −1 R : m∗ ∼M m}.
For any module M , it can be easily seen that the localization T −1 M at T is the
special case of tensor product with T −1 M ∼
= M ⊗R T −1 R. We first investigate the
cardinalities of equivalence classes corresponding to the equivalence relation ∼M on
sets Af (M\
⊗R T −1 R) and A\
f (M ).
Lemma 4.6. Let M be an R-module. Then the equivalence classes corresponding
to equivalence relation ∼M on sets Af (M\
⊗R T −1 R) and A\
f (M ) have the same
cardinality.
Proof. It suffices to prove the result for sets Af (M\
⊗R T −1 R) and A\
f (M ). First
\
note that for all m ∈ Af (M ) and for all modules M, N we have,
annT −1 R (m ⊗ 1s ) = ann(m) ⊗ T −1 R
[N : M ] ⊗ T
−1
R M ⊗T
−1
R = [N ⊗ T
−1
and
R : M ⊗ T −1 R]M ⊗ T −1 R.
Moreover,
−1
Af (M\
⊗ T −1 R) = A\
R
f (M ) ⊗ T
[m]M ⊗ T
−1
R=
[m
1 ]M⊗T −1 R
⊗T
and
−1
R.
This implies,
S
A\
[mλ ]M and
f (M ) =
λ∈Λ
Af (M\
⊗ T −1 R) =
S
[mλ ⊗ 11 ] =
λ∈Λ
S
λ∈Λ
[ m1λ ]M⊗T −1 R .
Here, Λ is an index set with [mλ ] = [mµ ] = ∅, for λ, µ ∈ Λ. We show that for all
m
m ∈ A\
f (M ), | [ 1 ]M⊗T −1 R | = | [m]M |.
We consider the following cases.
Case 1. [x : M ] 6= ann(M ). Then by [Lemma 3.2, [19]], M is a multiplication
module.
Subcase 1.1. | [m]M | is finite.
∗
Let ms ∈ [m ⊗ 1s ]M⊗T −1 R with m∗ ∈ [m]M , s ∈ T . Then ann(m∗ )M =
ann(m)M and {sn m∗ : n ≥ 1} ⊆ [m]M . Therefore,
m∗ ⊗
1
s
=
m∗
s
=
m∗ sk
sk+1
= m∗ ∈ [m]M .
20
RAMEEZ RAJA
Since | [m]M | is finite, so there exists some k ∈ N such that sk m∗ = sk+1 m∗ . Thus
m
Af (M\
⊗ T −1 R) ⊆ A\
f (M ) which implies [ 1 ]M⊗T −1 R ⊆ [m]M and we have,
| [m
1 ]M⊗T −1 R | = | [m]M |.
Subcase 1.2. | [m]M | is infinite.
We define an equivalence relation ∼ on T by s ∼ x if and only if sm = xm. Define
a map
−1
ϕ : M × T / ∼ −→ [ m
1 ]M⊗T R
(a, [s]) −→ a ⊗ 1s
We first check that the map ϕ is well defined. Let m1 = m2 and [s] = [x]. Then,
(s − x)M ⊆ ann(m)M = ann(m1 )M = ann(m2 )M .
Since M is a multiplication module therefore we have,
(s − x)M = ann(m)M = ann(m1 )M = ann(m2 )M .
Thus sm1 = xm1 and sm2 = xm2 which implies,
m1
m2
s = x .
It is clear that the map ϕ is surjective. Thus,
| [m
1 ]M⊗T −1 R | ≥ | [m]M || T / ∼|.
Further, the map
T / ∼ −→ [m]M
[s] −→ sx
is well defined and injective. Therefore,
2
| [m
1 ]M⊗T −1 R | ≤| [m]M | = | [m]M |.
m
Since | [m]M | is infinite we conclude that | [ 1 ]M⊗T −1 R | = | [m]M |.
Case 2. [x : M ] = ann(M ), for some 0 6= x ∈ M . Then for all 0 6= y ∈ M , we
\ −1 R) = M \
c
have [x : M ][y : M ]M = 0. Thus A\
⊗ T −1 R.
f (M ) = M and Af (M ⊗ T
Using above subcases 1.1 and 1.2, it can be shown that for each α ∈ Λ, the cardinality of equivalence classes [mα ]M and [ m1α ]M⊗T −1 R are same.
Theorem 4.7. Let M be an R-module. Then annf (Γ(M ⊗T −1 R)) ∼
= annf (Γ(M )).
Proof. For each α ∈ Λ, Λ an index set, we see by Lemma 4.6 that there is a
bijection
ϕα : [mα ]M −→ [ m1α ]M⊗T −1 R
Define a map,
\ −1 R)
ψ : A\
f (M ) −→ Af (M ⊗R T
m −→ [mα ]M
Clearly ψ is a bijective map. To show the required isomorphism of graphs we just
need to show that m∗1 and m∗2 are adjacent in annf (Γ(M )) if and only if ψ(m∗1 )
and ψ(m∗2 ) are adjacent in annf (Γ(M ⊗ T −1 R)). That is,
[m∗1 : M ][m∗2 : M ]M = 0 if and only if
[ψ(m∗1 ) : M ][ψ(m∗2 ) : M ]M = 0.
It suffices to show that for m∗1 ∈ [m]M , m∗1 ∈ [r]M , x1 ∈ [ m
1 ]M⊗R T −1 R and x2 ∈
[ 1r ]M⊗R T −1 R
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
21
[m∗1 : M ][m∗2 : M ]M = 0
[ x11
if and only if
: M ⊗ T −1 R][ x12 : M ⊗ T −1 R]M ⊗ T −1 R = 0.
Therefore,
[m∗1 : M ][m∗2 : M ]M = 0
⇐⇒ m∗1 ∈ ann(m∗2 )M = ann(r)M
m∗
m∗
⇐⇒ 11 ∈ ann( 12 )M ⊗ T −1 R = ann( r1 )M ⊗ T −1 R = ann( x12 )M ⊗ T −1 R
∗
m
⇐⇒ [ 11 : M ⊗ T −1 R][ x12 : M ⊗ T −1 R]M ⊗ T −1 R = 0
m∗
x∗
−1
R = ann( x11 )M ⊗ T −1 R
⇐⇒ 12 ∈ ann( 11 )M ⊗ T −1 R = ann( m
1 )M ⊗ T
x1
x2
−1
−1
−1
⇐⇒ [ 1 : M ⊗ T R][ 1 : M ⊗ T R]M ⊗ T R = 0.
Thus it follows that,
annf (Γ(M ⊗ T −1 R)) ∼
= annf (Γ(M )).
The following result is an immediate consequence of preceding theorem.
Corollary 4.8. Let M be an R-module. Then the following hold.
(i) anns (Γ(M ⊗ T −1 R) ∼
= anns (Γ(M ))
(ii) annt (Γ(M ⊗ T −1 R) ∼
= annt (Γ(M ))
Remark 4.9. Consider a Z-module Z2 ⊕ Z2 . It is easy to verify that
annf (Γ(Z2 ⊕ Z2 )) ∼
= anns (Γ(Z2 ⊕ Z2 )) ∼
= K3 ,
By Theorem 4.7, we see that,
annf (Γ(Z2 ⊕ Z2 ⊗ T −1 Z)) ∼
= annf (Γ(Z2 ⊕ Z2 )) ∼
= K3 ,
and
anns (Γ(Z2 ⊕ Z2 ⊗ T −1 Z)) ∼
= anns (Γ(Z2 ⊕ Z2 )) ∼
= K3 ,
where as,
annt (Γ(Z2 ⊕ Z2 ⊗ T −1 Z)) ∼
= annt (Γ(Z2 ⊕ Z2 )) ∼
= ∅.
−1
It follows that annihilating graphs annt (Γ(Z2 ⊕ Z2 ⊗ T Z)) and annf (Γ(Z2 ⊕
Z2 ⊗ T −1 Z)) are not isomorphic. That is, all annihilating graphs for a same tensor
product cannot be similar. However, if M is multiplication R-module, then by
[Theorem 3.9, [19]] all three annihilating graphs arising from M are same. Therefore by Theorem 4.7, all three annihilating graphs arising from the tensor product
M ⊗ T −1 R are same. In fact, the three annihilating graphs arising from M ⊗ T −1 R
coincides with the three annihilating graphs arising from M .
P
Q
Mi
Mi /
We conclude this paper with some discussion on factor modules M =
i∈I
i∈I
P
Q
Ni , where I is an index set, Mi , Ni are finite simple modules
Ni /
and N =
i∈I
i∈I
not equal to Z2 (Z) for any i ∈ I. Consider the equivalence relation on subsets of I
given by J is equivalent to K if symmetric difference of J and K is finite. Pick
P one
element J from each equivalence class and for each such set J, let XJ = {x+ Mi :
i∈I
P
Ni : y(j) = 0 ⇐⇒ j ∈ J} ⊂ N .
x(j) = 0 ⇐⇒ j ∈ J} ⊂ M and YJ = {y +
i∈I
22
RAMEEZ RAJA
Since a finite simple module Z2 (Z) is a component of neither M nor N , therefore
sets XJ and YJ have same cardinality. Let ϕJ be some bijection between sets XJ
and YJ . It is clear that the graphs annf (Γ(M )), annf (Γ(N )) are complete with
c \
b
A\
f (M ) = M , Af (N ) = N and for each a ∈ M , we have the associated zero set J.
Thus the map a to ϕJ (a) defines a bijection between the vertices of annf (Γ(M )),
annf (Γ(N )) and we conclude that annf (Γ(M )) ∼
= annf (Γ(N )).
References
[1] D. F. Anderson and P. S. Livingston, The zero-divisor graph of a commutative ring, J. Algebra
217 (1999) 434 − 447.
[2] D. F. Anderson, A. Frazier, A. Lauve and P. S. Livingston, The Zero-Divisor Graph of a
Commutative Ring, II, Lecture Notes in Pure and Applied Mathematics, Vol. 220, Marcel
Dekker, Newyork, Basel, (2001) 61 − 72.
[3] D. F. Anderson, R. Levy and J. Shapiro, Zero-divisor graphs, von Neumann regular rings, and
Boolean algebras. J. Pure Appl. Algebra 180 (2003) 221 − 241.
[4] M. F. Atiyah and I. G. MacDonald, Introduction to Commutative Algebra, Addison-Wesley,
Reading, MA, (1969).
[5] F. Azarpanah, Essential ideals in C(X), Period. Math. Hungar. 31 (1995) 105 − 112.
[6] F. Azarpanah, Intersection of essential ideals in C(X), Proc. Amer. Math. Soc. 125 (1997)
2149 − 2154.
[7] I. Beck, Coloring of commutative rings, J. Algebra 116 (1988) 208 − 226.
[8] H. Cartan and S. Eilenberg, Homological algebra, Princeton, (1956).
[9] R. Diestal, Graph Theory, Springer-Verlag, New York, (1997).
[10] Z. A. El-bast and P. F. Smith, Multiplication modules, Commun. Algebra 16 (4) (1988)
755 − 779.
[11] B. W. Green and L. Van Wyk, On the small and essential ideals in certain classes of rings,
J. Austral. Math. Soc. (ser. A), 46 (1989) 262 − 271.
[12] F. Hongin and P. Stewart, Graded rings and essential ideals, Acta Math. Sinica, 9 (4) (1993)
344 − 351.
[13] S. K. Jain, S. R. Lopez-permouth and S. T. Rizvi, Continous rings with acc on essentials is
left artinian, Proc. Amer. Math. Soc. 108 (3) (1990) 192 − 195.
[14] M. Kaneda and V. I. Paulsen, Characterization of essential ideals as operator modules over
C ∗ − algebras, J. Operator Theory, 49 (2003) 245 − 262.
[15] I. Kaplansky, Commutative Rings, rev. ed., Univ. of Chicago Press, Chicago, (1974).
[16] O.A. S. Karamzadeh and M. Rostami, On the intrinsic topology and some related ideals of
C(X), Proc. Amer. Math. Soc., 93 (1) (1985), 179 − 184.
[17] F. D. Meyer and L. D. Meyer, Zero-divisor graphs of semigroups, J. Algebra 283 (2005)
190 − 198.
[18] E. Momathan, Essential ideals in rings of measurable functions. Commun. Algebra, 38 (2010)
4739 − 4746.
[19] S. Pirzada, Rameez Raja, On graphs associated with modules over commutative rings, J.
Korean. Math. Soc., 53 (5) (2016) 1167 − 1182.
[20] S. Pirzada, Rameez Raja, On the metric dimension of a zero-divisor graph, Commun. Algebra,
45 (4) (2017) 1399 − 1408.
[21] S. Pirzada, Rameez Raja and S. P. Redmond, Locating sets and numbers of graphs associated
to commutative rings, J. Algebra Appl. 13 (7) (2014) 1450047 18 pp.
[22] E. R. Puczylowski, On essential extensions of rings, Bull. Austral. Math. Soc., 35 (1987)
379 − 386.
[23] Rameez Raja, S. Pirzada and S. P. Redmond, On Locating numbers and codes of zero-divisor
graphs associated with commutative rings, J. Algebra Appl. 15 (1) (2016) 1650014 22 pp.
[24] S. P. Redmond, The zero-divisor graph of a non-commutative ring, Internat. J. Commutative
Rings 1 (4) (2002) 203 − 211.
[25] S. P. Redmond, An ideal-based zero-divisor graph of a commutative ring, Commun. Algebra
31 (2003) 4425 − 4443.
ON COMBINATORIAL ASPECTS ARISING FROM MODULES OVER COMMUTATIVE RINGS
23
[26] A. Resenberg and D. Zelinsky, Finiteness of the injective hull, Math. Z., 70 (1959) 372 − 380.
[27] D. B. West, Introduction to Graph Theory, 2nd ed. USA: Prentice Hall, (2001).
[28] R. Wisbauer, Foundations of Modules and Ring Theory, Gordon and Breach Reading (1991).
| 0math.AC
|
Graph-based Clustering under Differential Privacy
Rafael Pinot ∗1,2 , Anne Morvan †1,2 , Florian Yger2 , Cédric Gouy-Pailler1 , and Jamal Atif2
1
arXiv:1803.03831v1 [cs.DS] 10 Mar 2018
2
CEA, LIST, 91191 Gif-sur-Yvette, France
Université Paris-Dauphine, PSL Research University, CNRS, LAMSADE, 75016 Paris,
France
March 10, 2018
Abstract
In this paper, we present the first differentially private clustering method for arbitrary-shaped
node clusters in a graph. This algorithm takes as input only an approximate Minimum Spanning Tree
(MST) T released under weight differential privacy constraints from the graph. Then, the underlying
nonconvex clustering partition is successfully recovered from cutting optimal cuts on T . As opposed to
existing methods, our algorithm is theoretically well-motivated. Experiments support our theoretical
findings.
1
Introduction
Weighted graph data is known to be a useful representation data type in many fields, such as bioinformatics or analysis of social, computer and information networks. More generally, a graph can always
be built based on the data dissimilarity where points of the dataset are the vertices and weighted edges
express “distances” between those objects. For both cases, graph clustering is one of the key tools for
understanding the underlying structure in the graph [Schaeffer, 2007]. These clusters can be seen as
groups of nodes close in terms of some specific similarity.
Nevertheless, it is critical that the data representation used in machine learning applications protects
the private characteristics contained into it. Let us consider an application where one wants to identify
groups of similar web pages in the sense of traffic volume i.e. web pages with similar audience. In that
case, the nodes stand for the websites. The link between two vertices represents the fact that some people
consult them both. Edge weights are the number of common users and thus, carry sensitive information
about individuals. During any graph data analysis, no private user surfing behavior should be breached
i.e. browsing from one page to another should remain private. As a standard for data privacy preservation,
differential privacy [Dwork et al., 2006b] has been designed: an algorithm is differentially private if, given
two close databases, it produces statistically indistinguishable outputs. Since then, its definition has been
extended to weighted graphs. Though, machine learning applications ensuring data privacy remain rare,
in particular for clustering which encounters severe theoretical and practical limitations. Indeed, some
clustering methods lack of theoretical support and most of them restrict the data distribution to convexshaped clusters [Nissim et al., 2007, Blum et al., 2008, McSherry, 2009, Dwork, 2011] or unstructured
data [Ho and Ruan, 2013, Chen et al., 2015]. Hence, the aim of this paper is to offer a theoretically
motivated private graph clustering. Moreover, to the best of our knowledge, this is the first weight
differentially-private clustering algorithm able to detect clusters with an arbitrary shape for weighted
graph data.
∗ [email protected]
† [email protected].
Partly supported by the Direction Générale de l’Armement (French Ministry of Defense).
1
Our method belongs to the family of Minimum Spanning Tree (MST)-based approaches. An MST
represents a useful summary of the graph, and appears to be a natural object to describe it at a lower cost.
For clustering purposes, it has the appealing property to help retrieving non-convex shapes [Zahn, 1971,
Asano et al., 1988, Grygorash et al., 2006, Morvan et al., 2017]. Moreover, they appear to be well-suited
for incorporating privacy constraints as will be formally proved in this work.
Contributions: Our contributions are threefold: 1) we provide the first theoretical justifications
of MST-based clustering algorithms. 2) We endow DBMSTClu algorithm [Morvan et al., 2017], an
MST-based clustering algorithm from the literature, with theoretical guarantees. 3) We introduce a
differentially-private version of DBMSTClu and give several results on its privacy/utility tradeoff.
2
Preliminaries
2.1
Notations
Let G = (V, E, w) be a simple undirected weighted graph with a vertex set V , an edge set E, and a weight
function w := E → R. One will respectively call the edge set and the node set of a graph G using the
applications E(G) and V (G). Given a node set S ⊂ V , one denotes by G|S the subgraph induced by S.
We call G = (V, E) the topology of the graph, and WE denotes the set of all possible weight functions
mapping E to weights in R. For the remaining of this work, cursive letter are use to represent weighted
graphs and straight letters refer to topological arguments. Since graphs are simple, the path Pu−v between
two vertices u and v is characterized as the ordered sequence of vertices {u, . . . , v}. We also denote VPu−v
the unordered set of such vertices. Besides, edges eij denote an edge between nodes i and j. Finally, for
all positive integer K, [K] := {1, . . . , K}.
2.2
Differential privacy in graphs
As opposed to node-differential privacy [Kasiviswanathan et al., 2013] and edge-differential privacy [Hay
et al., 2009], both based on the graph topology, the privacy framework considered here is weight-differential
privacy where the graph topology G = (V, E) is assumed to be public and the private information to
protect is the weight function w := E → R. Under this model introduced by Sealfon [2016], two graphs
are said to be neighbors if they have the same topology, and close weight functions. this framework allows
one to release an almost minimum spanning tree with weight-approximation error of O (|V | log |E|) for fixed
privacy parameters. Differential privacy is ensured in that case by using the Laplace mechanism on every
edges weight to release a spanning tree based on a perturbed version of the weight function. The privacy
of the spanning tree construction is thus provided by post-processing (cf. Th. 5). However, under a similar
privacy setting, Pinot [2018] recently manages to produce the topology of a tree under differential privacy
without relying on the post-processing of a more general mechanism such as the “Laplace mechanism”.
Their algorithm, called PAMST, privately releases the topology of an almost minimum spanning tree
thanks to an iterative use of the “Exponential
mechanism” instead. For fixed privacy parameters, the
|V |2
weight approximation error is O |E| log |V | , which outperforms the former method from Sealfon [2016]
on arbitrary weighted graphs under weak assumptions on the graph sparseness. Thus, we keep here
privacy setting from Pinot [2018].
Definition 2.1 (Pinot [2018]). For any edge set E, two weight functions w, w0 ∈ WE are neighboring,
denoted w ∼ w0 , if ||w − w0 ||∞ := max |w(e) − w0 (e)| ≤ µ.
e∈E
µ represents the sensitivity of the weight function and should be chosen according to the application and
the range of this function. The neighborhood between such graphs is clarified in the following definition.
Definition 2.2. Let G = (V, E, w) and G 0 = (V 0 , E 0 , w0 ), two weighted graphs, G and G 0 are said to be
neighbors if V = V 0 , E = E 0 and w ∼ w0 .
The so-called weight-differential privacy for graph algorithms is now formally defined.
2
Definition 2.3 (Sealfon [2016]). For any graph topology G = (V, E), let A be a randomized algorithm
that takes as input a weight function w ∈ WE . A is called (, δ)-differentially private on G = (V, E) if for
all pairs of neighboring weight functions w, w0 ∈ WE , and for all set of possible outputs S, one has
P [A(w) ∈ S] ≤ e P [A(w0 ) ∈ S] + δ.
If A is (, δ)-differentially private on every graph topology in a class C, it is said to be (, δ)-differentially
private on C.
One of the first, and most used differentially private mechanisms is the Laplace mechanism. It is
based on the process of releasing a numerical query perturbed by a noise drawn from a centered Laplace
distribution scaled to the sensitivity of the query. We present here its graph-based reformulation.
Definition 2.4 (reformulation Dwork et al. [2006b]). Given some graph topology G = (V, E), for any
fG : WE → Rk , the sensitivity of the function is defined as ∆fG = max
||fG (w) − fG (w0 )||1 .
0
w∼w ∈WE
Definition 2.5 (reformulation Dwork et al. [2006b]). Given some graph topology G = (V, E), any function
fG : WE → Rk , any > 0, and w ∈ WE , the graph-based Laplace mechanism is ML (w, fG , ) =
fG (w) + (Y1 , . . . , Yk ) where Yi are i.i.d.
Lap(∆f
G /), and Lap(b) denotes
random variables drawn from
the Laplace distribution with scale b i.e probability density
1
2b
exp − |x|
b
.
Theorem 1 (Dwork et al. [2006b]). The Laplace mechanism is -differentially private.
We define hereafter the graph-based Exponential mechanism. In the sequel we refer to it simply as
Exponential mechanism. The Exponential mechanism represents a way of privately answering arbitrary
range queries. Given some range of possible responses to the query R, it is defined according to a utility
function uG := WE ×R → R, which aims at providing some total preorder on the range R according to the
total order in R. The sensitivity of this function is denoted ∆uG := max max
|uG (w, r) − uG (w0 , r)| .
0
r∈R w∼w ∈WE
Definition 2.6. Given some graph topology G = (V, E), some output range R ⊂ E, some privacy
parameter > 0, some utility function uG := WE ×R → R, and some w ∈ WE the graph-based Exponential
mechanism
M
Exp (G, w, uG , R, ) selects and outputs an element r ∈ R with probability proportional to
exp
uG (w,r)
2∆uG
.
The Exponential mechanism defines a distribution on a potentially complex and large range R. As
the following theorem states, sampling from such a distribution preserves -differential privacy.
Theorem 2 (reformulation McSherry and Talwar [2007]). For any non-empty range R, given some
graph topology G = (V, E), the graph-based Exponential mechanism preserves -differential privacy, i.e if
w ∼ w0 ∈ WE ,
P [MExp (G, w, uG , R, ) = r]
≤ e P [MExp (G, w0 , uG , R, ) = r] .
Further, Th 3 highlights the trade-off between privacy and accuracy for the Exponential mechanism
when 0 < |R| < +∞. Th 4 presents the ability of differential privacy to comply with composition while
Th 5 introduces its post-processing property.
Theorem 3 (reformulation Dwork and Roth [2013]). Given some graph topology G = (V, E), some
w ∈ WE , some output range R, some privacy parameter > 0, some utility function uG := WE × R → R,
and denoting OP TuG (w) = max uG (w, r), one has ∀ t ∈ R,
r∈R
uG (G, w, MExp (w, uG , R, ))
2∆uG
≤ OP TuG (w) −
(t + ln |R|)
with probability at most exp(−t).
3
Theorem 4 (Dwork et al. [2006a]). For any > 0, δ ≥ 0 the adaptive composition of k (, δ)-differentially
private mechanisms is (k, kδ)-differentially private.
Theorem 5 (Post-Processing Dwork and Roth [2013]). Let A : WE → B be a randomized algorithm that
is (, δ)-differentially private, and h : B → B 0 a deterministic mapping. Then h ◦ A is (, δ)-differentially
private.
2.3
Differentially-private clustering
Differentially private clustering for unstructured datasets has been first discussed in Nissim et al. [2007].
This work introduced the first method for differentially private clustering based on the k-means algorithm.
Since then most of the works of the field focused on adaptation of this method [Blum et al., 2008, McSherry,
2009, Dwork, 2011]. The main drawback of those works is that they are not able to deal with arbitrary
shaped clusters. This issue has been recently investigated in Ho and Ruan [2013] and Chen et al. [2015].
They proposed two new methods to find arbitrary shaped clusters in unstructured datasets respectively
based on density clustering and wavelet decomposition. Even though both of these works allow one to
produce non-convex clusters, they only deal with unstructured datasets and thus are not applicable to
node clustering in a graph. Our work focuses on node clustering in a graph under weight-differential
privacy. Graph clustering has already been investigated in a topology-based privacy framework [Mülle
et al., 2015, Nguyen et al., 2016], however, these works do not consider weight-differential privacy. Our
work is, to the best of our knowledge, the first attempt to define node clustering in a graph under weight
differential privacy.
3
Differentially-private tree-based clustering
We aim at producing a private clustering method while providing bounds on the accuracy loss. Our
method is an adaptation of an existing clustering algorithm DBMSTClu. However, to provide theoretical
guarantees under differential privacy, one needs to rely on the same kind of guarantees in the non-private
setting. Morvan et al. [2017] did not bring them in their initial work. Hence, our second contribution is
to demonstrate the accuracy of this method, first in the non-private context.
In the following we present 1) the theoretical framework motivating MST-based clustering methods,
2) accuracy guarantees of DBMSTClu in the non-private setting, 3) PTClust our private clustering
algorithm, 4) its accuracy under differential privacy constraints.
3.1
Theoretical framework for MST-based clustering methods
MST-based clustering methods, however efficient, lack of proper motivation. This Section closes this gap
by providing a theoretical framework for MST-based clustering. In the sequel, notations from Section 2.1
are kept. The minimum path distance between two nodes in the graph is defined which enables to explicit
our notion of Cluster.
Definition 3.1 (Minimum path distance). Let be G = (V, E, w) and u, v ∈ V . The minimum path
distance between u and v is
X
d(u, v) = min
w(e)
Pu−v
e∈VPu−v
with Pu−v a path from u to v in G, and VPu−v the set of vertices contained in Pu−v .
Definition 3.2 (Cluster). Let be G = (V, E, w), 0 < w(e) ≤ 1 ∀e ∈ E a graph, (V, d) a metric space based
on the minimum path distance d defined on G and D ⊂ V a node set. C ⊂ D is a cluster iff. |C| > 2 and
∀C1 , C2 s.t. C = C1 ∪ C2 and C1 ∩ C2 = ∅, one has:
argmin{ min d(z, v) } ⊂ C2
z∈D\C1
v∈C1
4
Assuming that a cluster is built of at least 3 points makes sense since singletons or groups of 2 nodes
can be legitimately considered as noise. For simplicity of the proofs, the following theorems hold in the
case where noise is neglected. However, they are still valid in the setting where noise is considered as
singletons (with each singleton representing a generalized notion of cluster).
Theorem 6. Let be G = (V, E, w) a graph and T a minimum spanning tree of G. Let also be C a cluster
in the sense of Def. 3.2 and two vertices v1 , v2 ∈ C. Then, VPv1 −v2 ⊂ C with Pv1 −v2 a path from v1 to v2
in G, and VPv1 −v2 the set of vertices contained in Pv1 −v2 .
Proof. Let be v1 , v2 ∈ C. If v1 and v2 are neighbors, the result is trivial. Otherwise, as T is a tree, there
exist a unique path within T between v1 and v2 denoted by Pv1 −v2 = {v1 , . . . , v2 }. Let now prove by
reductio ad absurdum that VPv1 −v2 ⊂ C. Suppose there is h ∈ VPv1 −v2 s.t. h ∈
/ C. We will see that
it leads to a contradiction. We set C1 to be the largest connected component (regarding the number of
vertices) of T s.t. v1 ∈ C1 , and every nodes from C1 are in C. Because of h’s definition, v2 ∈
/ C1 . Let be
C2 = C\C1 . C2 6= ∅ since v2 ∈ C2 . Let be z ∗ ∈ argmin{ min d(z, v) } and e∗ = (z ∗ , v ∗ ) an edge that
∗
z∈V \C1
∗
v∈C1
reaches this minimum. Let us show that z ∈
/ C. If z ∈ C, then two possibilities hold:
1. There is an edge ez∗ ∈ T , s.t. ez∗ = (z ∗ , z 0 ) with z 0 ∈ C1 . This is impossible, otherwise by definition
of a connected component, z ∗ ∈ C1 . Contradiction.
2. For all ez∗ = (z ∗ , z 0 ) s.t z 0 ∈ C1 , one has ez∗ ∈
/ T . In particular e∗ ∈
/ T . Since h is the neighbor of
0
C1 in G there is also eh ∈ T , s.t. eh = (h, h ) with h0 ∈ C1 . Once again two possibilities hold:
(a) w(ez∗ ) = min { min d(z, v) } < w(eh ). Then, if we replace eh by ez∗ in T , its total weight
z∈V \C1
v∈C1
decreases. So T is not a minimum spanning tree. Contradiction.
/ C, one gets that
(b) w(ez∗ ) = w(eh ), therefore h ∈ argmin{ min d(z, v) }. Since h ∈
z∈V \C1
v∈C1
argmin{ min d(z, v) } 6⊂ C2 . Thus, C is not a cluster. Contradiction.
z∈V \C1
v∈C1
We proved that z ∗ ∈
/ C. In particular, z ∗ ∈
/ C2 . Then, argmin{ min d(z, v) } 6⊂ C2 . Thus, C is not a
cluster. Contradiction. Finally h ∈ C and VPv1 −v2 ⊂ C.
z∈V \C1
v∈C1
This theorem states that, given a graph G, an MST T , and any two nodes of C, every node in the
path between them is in C. This means that a cluster can be characterized by a subtree of T . It justifies
the use of all MST-based methods for data clustering or node clustering in a graph. All the clustering
algorithms based on successively cutting edges in an MST to obtain a subtree forest are meaningful in
the sense of Th.6. In particular, this theorem holds for the use of DBMSTClu [Morvan et al., 2017]
presented in Section 3.2.1.
3.2
Deterministic MST-based clustering
This Section introduces DBMSTClu [Morvan et al., 2017] that will be adapted to be differentially-private,
and provide accuracy results on the recovery of the ground-truth clustering partition.
3.2.1
DBMSTClu algorithm
Let us consider T an MST of G, as the unique input of the clustering algorithm DBMSTClu. The
clustering partition results then from successive cuts on T so that a new cut in T splits a connected
component into two new ones. Each final connected component, a subtree of T , represents a cluster.
Initially, T is one cluster containing all nodes. Then, at each iteration, an edge is cut if some criterion,
called Validity Index of a Clustering Partition (DBCVI) is improved. This edge is greedily chosen to
locally maximize the DBCVI at each step. When no improvement on DBCVI can be further made, the
algorithm stops. The DBCVI is defined as the weighted average of all cluster validity indices which are
based on two positive quantities, the Dispersion and the Separation of a cluster:
5
Definition 3.3 (Cluster Dispersion). The Dispersion of a cluster Ci (DISP) is defined as the maximum
edge weight of Ci . If the cluster is a singleton (i.e. contains only one node), the associated Dispersion is
set to 0. More formally:
(
max wj if |E(Ci )| =
6 0
j, ej ∈Ci
∀i ∈ [K], DISP(Ci ) =
0
otherwise.
Definition 3.4 (Cluster Separation). The Separation of a cluster Ci (SEP) is defined as the minimum
distance between the nodes of Ci and the ones of all other clusters Cj , j 6= i, 1 ≤ i, j ≤ K, K 6= 1 where
K is the total number of clusters. In practice, it corresponds to the minimum weight among all already
cut edges from T comprising a node from Ci . If K = 1, the Separation is set to 1. More formally, with
incCuts(Ci ) denoting cut edges incident to Ci ,
(
min
wj if K 6= 1
j, ej ∈incCuts(Ci )
∀i ∈ [K], SEP(Ci ) =
1
otherwise.
Definition 3.5 (Validity Index of a Cluster). The Validity Index of a cluster Ci is defined as:
VC (Ci ) =
SEP(Ci ) − DISP(Ci )
∈ [−1; 1]
max(SEP(Ci ), DISP(Ci ))
Definition 3.6 (Validity Index of a Clustering Partition). The Density-Based Validity Index of a Clustering partition Π = {Ci }, 1 ≤ i ≤ K, DBCVI(Π) is defined as the weighted average of the Validity Indices
of all clusters in the partition where N is the number of vertices.
K
X
|Ci |
DBCVI(Π) =
VC (Ci ) ∈ [−1, 1]
N
i=1
DBMSTClu is summarized in Algorithm 1: evaluateCut(.) computes the DBCVI when the cut in
parameter is applied to T . Initial DBCVI is set −1. Interested reader could refer to [Morvan et al., 2017]
Section 4. for a complete insight on this notions.
3.2.2
DBMSTClu exact clustering recovery proof
In this section, we provide theoretical guarantees for the cluster recovery accuracy of DBMSTClu. Let us
first begin by introducing some definitions.
Definition 3.7 (Cut). Let us consider a graph G = (V, E, w) with K clusters, T an MST of G. Let denote
(Ci∗ )i∈[K] the set of the clusters. Then, CutG (T ) := {ekl ∈ T | k ∈ Ci∗ , l ∈ Cj∗ , i, j ∈ [K]2 , i 6= j}. In
the sequel, for simplicity, we denote e(ij) ∈ CutG (T ) the edge between cluster Ci∗ and Cj∗ .
CutG (T ) is basically the set of effective cuts to perform on T in order to ensure the exact recovery
of the clustering partition. More generally, trees on which CutG (.) enables to find the right partition are
said to be a partitioning topology.
∗
Definition 3.8 (Partitionning topology). Let us consider a graph G = (V, E, w) with K clusters C1∗ , . . . , CK
.
A spanning tree T of G is said to have a partitioning topology if ∀i, j ∈ [K], i 6= j, |{e = (u, v) ∈
CutG (T ) | u ∈ Ci∗ , v ∈ Cj∗ }| = 1.
Def. 3.7 and 3.8 introduce a topological condition on the tree as input of the algorithm. Nevertheless,
conditions on weights are necessary too. Hence, we define homogeneous separability which expresses the
fact that within a cluster the edge weights are spread in a controlled manner.
6
Algorithm 1 DBMSTClu(T )
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
Input: T , the MST
dbcvi ← −1.0
clusters ← ∅
cut list ← {E(T )}
while dbcvi < 1.0 do
cut tp ← ∅
dbcvi tp ← dbcvi
for each cut in cut list do
newDbcvi ← evaluateCut(T , cut)
if newDbcvi ≥ dbcvi tp then
cut tp ← cut
dbcvi tp ← newDbcvi
if cut tp 6= ∅ then
clusters← cut(clusters, cut tp)
dbcvi ← dbcvi tp
cut list ← cut list\{cut tp}
else
break
return clusters, dbcvi
Definition 3.9 (Homogeneous separability condition). Let us consider a graph G = (V, E, w), s ∈ E and
T a tree of G. T is said to be homogeneously separable by s, if
max w(e)
αT
max w(e) < w(s) with αT =
e∈E(T )
e∈E(T )
min w(e)
e∈E(T )
≥ 1.
One will write for simplicity that HT (s) is verified.
Definition 3.10 (Weak homogeneity condition of a Cluster). Let us consider a graph G = (V, E, w) with
∗
K clusters C1∗ , . . . , CK
. A given cluster Ci∗ , i ∈ [K], Ci∗ is weakly homogeneous if: for all T an MST
of G, and ∀j ∈ [K], j 6= i, s.t. e(ij) ∈ CutG (T ), HT|C ∗ (e(ij) ) is verified. For simplicity, one denote
i
αi =
max αT|C ∗
i
T MST of G
¯
Definition 3.11 (Strong homogeneity condition of a Cluster). Let us consider a graph G = (V, E, w) with
∗
K clusters C1∗ , . . . , CK
. A given cluster Ci∗ , i ∈ [K], Ci∗ is strongly homogeneous if: for all T a spanning
tree (ST) of G, and ∀j ∈ [K], j 6= i, s.t. e(ij) ∈ CutG (T ), HT|C ∗ (e(ij) ) is verified. For simplicity, one
i
denote ᾱi = max αT|C ∗
T ST of G
i
We show that the weak homogeneity condition is implied by the strong homogeneity condition.
∗
Proposition 3.1. Let us consider a graph G = (V, E, w) with K clusters C1∗ , . . . , CK
. If a given cluster
∗
Ci , i ∈ [K] is strongly homogeneous, then, it is weakly homogeneous.
Proof. If T a spanning tree of G, and ∀j ∈ [K], j 6= i, s.t. e(ij) ∈ CutG (T ), HT|C ∗ (e(ij) ) is verified, then
i
in particular, it is true for any MST.
Strong homogeneity condition appears to be naturally more constraining on the edge weights than
the weak one. The accuracy of DBMSTClu is proved under the weak homogeneity condition, while the
accuracy of its differentially-private version is only given under the the strong homogeneity condition.
7
∗
Theorem 7. Let us consider a graph G = (V, E, w) with K homogeneous clusters C1∗ , . . . , CK
and T an
MST of G. Let now assume that at step k < K − 1, DBMSTClu built k + 1 subtrees C1 , . . . , Ck+1 by cutting
e1 , e2 , . . . , ek ∈ E.
Then, Cutk := CutG (T ) \ {e1 , e2 , . . . , ek } =
6 ∅ =⇒ DBCVIk+1 ≥ DBCV Ik , i.e. if there are still
edges in Cutk , the algorithm will continue to perform some cut.
Proof. See supplementary material.
∗
Theorem 8. Let us consider a graph G = (V, E, w) with K homogeneous clusters C1∗ , . . . , CK
and T an
MST of G.
Assume now that at step k < K − 1, DBMSTClu built k + 1 subtrees C1 , . . . , Ck+1 by cutting
e1 , e2 , . . . , ek ∈ E. We still denote Cutk := CutG (T )\{e1 , e2 , . . . , ek }.
If Cutk 6= ∅ then
argmax
DBCV Ik+1 (e) ⊂ Cutk i.e. the cut edge at step k + 1 is in Cutk .
e∈T \{e1 , e2 , ..., ek }
Proof. See supplementary material.
∗
Theorem 9. Let us consider a graph G = (V, E, w) with K weakly homogeneous clusters C1∗ , . . . , CK
and
T an MST of G. Let now assume that at step K − 1, DBMSTClu built K subtrees C1 , . . . , CK by cutting
e1 , e2 , . . . , eK−1 ∈ E. We still denote CutK−1 := CutG (T )\{e1 , e2 , . . . , eK−1 }.
Then, for all e ∈ T \{e1 , e2 , . . . , eK−1 }, DBCV IK (e) < DBCV IK−1 i.e. the algorithm stops: no
edge gets cut during step K.
Proof. See supplementary material.
∗
Corollary 3.1. Let us consider a graph G = (V, E, w) with K weakly homogeneous clusters C1∗ , . . . , CK
and T an MST of G. DBM ST Clu(T ) stops after K − 1 iterations and the K subtrees produced match
exactly the clusters i.e. under homogeneity condition, the algorithm finds automatically the underlying
clustering partition.
Proof. Th. 7 and 9 ensure that under homogeneity condition on all clusters, the algorithm performs
the K − 1 distinct cuts within CutG (T ) and stops afterwards. By definition of CutG (T ), it means the
DBMSTClu correctly builds the K clusters.
3.3
Private MST-based clustering
This section presents our new node clustering algorithm PTClust for weight differential privacy. It relies
on a mixed adaptation of PAMST algorithm [Pinot, 2018] for recovering a differentially-private MST of
a graph and DBMSTClu.
3.3.1
PAMST algorithm
Given a simple-undirected-weighted graph G = (V, E, w), PAMST outputs an almost minimal weight
spanning tree topology under differential privacy constraints. It relies on a Prim-like MST algorithm,
and an iterative use of the graph based Exponential mechanism. PAMST takes as an input a weighted
graph, and a utility function. It outputs the topology of a spanning tree which weight is almost minimal.
Algorithm 3 presents this new method, using the following utility function:
uG
: WE × R
(w, r)
→
7→
R
−|w(r) − min
w(r0 )|.
0
r ∈R
PAMST starts by choosing an arbitrary node to construct iteratively the tree topology. At every iteration,
it uses the Exponential mechanism to find the next edge to be added to the current tree topology while
keeping the weights private. This algorithm is the state of the art to find a spanning tree topology under
differential privacy. For readability, let us introduce some additional notations. Let S be a set of nodes
8
from G, and RS the set of edges that are incident to one and only one node in S (also denoted xorincident). For any edge r in such a set, the incident node to r that is not in S is denoted r→ . Finally, the
restriction of the weight function to an edge set R is denoted w|R .
Algorithm 2 PAMST(G, uG , w, )
1:
2:
3:
4:
5:
6:
7:
8:
9:
Input: G = (V, E, w) a weighted graph (separately the topology G and the weight function w), a
degree of privacy and uG utility function.
Pick v ∈ V at random
SV ← {v}
SE ← ∅
while SV 6= V do
)
r = MExp (G, w, uG , RSV , |V |−1
SV ← SV ∪ {r→ }
SE ← SE ∪ {r}
return SE
Theorem 10 states that using PAMST to get an almost minimal spanning tree topology preserves
weight-differential privacy.
Theorem 10. Let G = (V, E) be the topology of a simple-undirected graph, then ∀ > 0, PAMST(G, uG , •, )
is - differentially private on G.
3.3.2
Differentially private clustering
The overall goal of this Section is to show that one can obtain a differentially private clustering algorithm
by combining PAMST and DBMSTClu algorithms. However, PAMST does not output a weighted
tree which is inappropriate for clustering purposes. To overcome this, one cloud rely on a sanitizing
mechanism such as the Laplace mechanism. Moreover, since DBMSTClu only takes weights from (0,1],
two normalizing parameters τ and p are introduced, respectively to ensure lower and upper bounds to
the weights that fit within DBMSTClu needs. This sanitizing mechanism is called the Weight-Release
mechanism. Coupled with PAMST, it will allows us to produce a weighted spanning tree with differential
privacy, that will be exploited in our private graph clustering.
Definition 3.12 (Weight-Release mechanism). Let G = (G, w) be a weighted graph, > 0 a privacy
parameter, s a scaling parameter, τ ≥ 0, and p ≥ 1 two normalization parameters. The Weight-Release
mechanism is defined as
w + (Y1 , ..., Y|E| ) + τ
Mw.r (G, w, s, τ, p) = G, w0 =
p
where Yi are i.i.d. random variables drawn from Lap (0, s). With w + (Y1 , ..., Y|E| ) meaning that if one
gives an arbitrary order to the edges E = (ei )i∈[|E|] , one has ∀i ∈ [|E|], w0 (ei ) = w(ei ) + Yi .
The following theorem presents the privacy guarantees of the Weight-Release mechanism.
Theorem 11. Let G = (V, E) be the topology of a simple-undirected graph, τ ≥ 0, p ≥ 1, then ∀ > 0,
Mw.r G, •, µ , τ, p is - differentially private on G.
Proof. Given τ ≥ 0, p ≥ 1, and > 0, the Weight release mechanism scaled to µ can be break down into
a Laplace mechanism and a post-processing consisting in adding τ to every edge and dividing them by p.
Using Theorems 1 and 5, one gets the expected result.
So far we have presented DBMSTClu and PAMST algorithms, and the Weight-Release mechanism.
Let us now introduce how to compose those blocks to obtain a Private node clustering in a graph, called
9
Algorithm 3 PTClust(G, w, uG , , τ, p)
Input: G = (V, E, w) a weighted graph (separately the topology G and the weight function w), a
degree of privacy and uG utility function.
2: T = PAMST(G, w, uG , /2)
2µ
3: T 0 = Mw.r (T, w|E(T ) , , τ, p)
4: return DBMSTClu(T 0 )
1:
PTClust. The algorithm takes as an input a weighted graph (dissociated topology and weight function),
a utility function, a privacy degree and two normalization parameters. It outputs a clustering partition.
To do so, a spanning tree topology is produced using PAMST. Afterward a randomized and normalized
version of the associated weight function is released using the Weight-release mechanism. Finally the
obtained weighted tree is given as an input to DBMSTClu that performs a clustering partition. The
following theorem ensures that our method preserves -differential privacy.
Theorem 12. Let G = (V, E) be the topology of a simple-undirected graph, τ ≥ 0, and p ≥ 1, then ∀ > 0,
PTClust(G, •, uG , , τ, p) is -differentially private on G.
Proof. Using Theorem 10 one has that T is produced with /2-differential privacy, and using Theorem 11
one has that w0 is obtained with /2-differential privacy as well. Therefore using Theorem 4, T 0 is
released with -differential privacy. Using the post-processing property (Theorem 5) one gets the expected
result.
3.4
Differential privacy trade-off of clustering
The results stated in this section present the security/accuracy trade-off of our new method in the
differentially-private framework. PTClust relies on two differentially private mechanisms, namely PASMT
and the Weight-Release mechanism. Evaluating the accuracy of this method amounts to check whether using these methods for ensuring privacy does not deteriorate the final clustering partition. The accuracy is
preserved if PAMST outputs the same topology as the MST-based clustering, and if the Weight-Release
mechanism preserves enough the weight function. According to Def. 3.8, if a tree has a partitioning
topology, then it fits the tree-based clustering. The following theorem states that with high probability
PAMST outputs a tree with a partitioning topology.
∗
Theorem 13. Let us consider a graph G = (V, E, w) with K strongly homogeneous clusters C1∗ , . . . , CK
and T = PAMST(G, uG , w, ), > 0. T has a partitioning topology with probability at least
1−
K
X
i=1
(|Ci∗ |
− 1) exp −
A
2∆uG (|V | − 1)
with A = ᾱi max(w(e))
− min (w(e))
+ ln |E|.
e∈E G|C ∗
i
e∈E G|C ∗
i
Proof. See supplementary material.
The following theorem states that given a tree T under the strong homogeneity condition, if the subtree
associated to a cluster respects Def. 3.9, then it still holds after applying the Weight-Release mechanism
to this tree.
∗
Theorem 14. Let us consider a graph G = (V, E, w) with K strongly homogeneous clusters C1∗ , . . . , CK
0
and T = P AM ST (G, uG , w, ), T = (T, w|T ) and T = Mw.r (T, w|T , s, τ, p) with s << p, τ . Given some
10
(a) Homogeneous graph (b) DBMSTClu
(c) PTClust, = 1.0 (d) PTClust, = 0.7 (e) PTClust, = 0.5
Figure 1: Circles experiments for n = 100. PTClust parameters: wmin = 0.1, wmax = 0.3, µ = 0.1.
cluster Ci∗ , and j 6= i s.t e(ij) ∈ CutG (T ), if HT|C ∗ (e(ij) ) is verified, then HT 0 |C ∗ (e(ij) ) is verified with
i
i
probability at least
V(ϕ)
1−
V(ϕ) + E(ϕ)2
with the following notations :
• ϕ = (max Yj )2 − min Zj × X out
j∈[|Ci∗ |−1]
• Yj ∼ Lap
j∈[|Ci∗ |−1]
max w(e)+τ
e∈E(T )
p
iid
• Zj ∼ Lap
min w(e)+τ
e∈E(T )
p
iid
• X out ∼ Lap
, ps
,
w(e(ij) +τ s
,p
p
s
p
,
Proof. See supplementary material.
Note that Theorem 14 is stated in a simplified version. A more complete version (specifying an analytic
version of V(ϕ) and E(ϕ)) is given in the supplementary material.
4
Experiments
So far we have exhibited the trade-off between clustering accuracy and privacy and we experimentally
illustrate it with some qualitative results. Let us discuss hereafter the quantitative performances of our
algorithm. We have performed experiments on two classical synthetic graph datasets for clustering with
nonconvex shapes: two concentric circles and two moons, both in their noisy versions. For the sake
of readability and for visualization purposes, both graph datasets are embedded into a two dimensional
Euclidean space. Each dataset contains 100 data nodes that are represented by a point of two coordinates.
Both graphs have been built with respect to the strong homogeneity condition: edge weights within
clusters are between wmin = 0.1 and wmax = 0.3 while edges between clusters have a weight strictly
2
above wmax
/wmin = 0.9. In practice, the complete graph has trimmed from its irrelevant edges (i.e.
not respecting the strong homogeneity condition). Hence, those graphs are not necessarily Euclidean
11
76
69 74
62
60
66
67
65
64
7068 71
73
63
61
58
59
49
55
4
50
54
6
5
9 78
51
52
40
1112
38
3942
37 36
34
33
16 10
31
35
32
3029
20
26
24
28
27
25
23
22
19 17
1
76
69 74
83
81
8486
85
88 91
90
87
89 92
62
60
66
67
65
64
7068 71
73
78
83
81
82
80
8486
85
88 91
90
87
49
97
45
44
55
41
43
4
50
54
6
5
9 78
51
52
99
40
34
14
18
21
1112
38
3942
37 36
13
15
33
31
35
32
3029
16 10
20
26
24
28
27
25
23
92
93
3
53
56
98
89
0
2
57
47
48
46
96
95
94
77
79
72
75
63
61
58
59
93
3
53
56
41
43
78
82
80
0
2
57
47
48
46
45
44
77
79
72
75
22
19 17
1
96
95
94
97
98
99
13
15
14
18
21
(a) Homogeneous graph (b) DBMSTClu
(c) PTClust, = 1.0 (d) PTClust, = 0.7 (e) PTClust, = 0.5
Figure 2: Moons experiments for n = 100. PTClust parameters: wmin = 0.1, wmax = 0.3, µ = 0.1.
since close nodes in the visual representation may not be connected in the graph. Finally, weights are
normalized between 0 and 1.
Figures 1 and 2 (best viewed in color) show for each dataset (a) the original homogeneous graph G
built by respecting the homogeneity condition, (b) the clustering partition1 of DBMSTClu with the used
underlying MST, the clustering partitions for PTClust with µ = 0.1 obtained respectively with different
privacy degrees2 : = 0.5 (c), = 0.7 (c) and = 1.0 (e). The utility function uG corresponds to the
graph weight. Each experiment is carried out independently and the tree topology obtained by PAMST
will eventually be different. This explains why the edge between clusters may not be the same when the
experiment is repeated with a different level of privacy. However, this will marginally affect the overall
quality of the clustering.
As expected, DBMSTClu recovers automatically the right partition and the results are shown here
for comparison with PTClust. For PTClust, the true MST is replaced with a private approximate
MST obtained for suitable τ and p ensuring final weights between 0 and 1.
When the privacy degree is moderate ( ∈ {1.0, 0.7}), it appears that the clustering result is slightly
affected. More precisely, in Figures 1c and 1d the two main clusters are recovered while one point is
isolated as a singleton. This is due to the randomization involved in determining the edge weights for the
topology returned by PAMST. In Figure 2c, the clustering is identical to the one from DBMSTClu in
Figure 2b. In Figure 1d, the clustering is very similar to the DBMSTClu one, with the exception of an
isolated singleton. However, as expected from our theoretical results, when is decreasing, the clustering
quality deteriorates, as DBMSTClu is sensitive to severe changes in the MST (cf. Figure 1e, 2e).
5
Conclusion
In this paper, we introduced PTClust, a novel graph clustering algorithm able to recover arbitrarilyshaped clusters while preserving differential privacy on the weights of the graph. It is based on the release
of a private approximate minimum spanning tree of the graph of the dataset, by performing suitable
cuts to reveal the clusters. To the best of our knowledge, this is the first differential private graphbased clustering algorithm adapted to nonconvex clusters. The theoretical analysis exhibited a trade-off
between the degree of privacy and the accuracy of the clustering result. This work suits to applications
where privacy is a critical issue and it could pave the way to metagenomics and genes classification using
1 For the sake of clarity, the edges in those Figures are represented based on the original weights and not on the privately
released weights.
2 Note that, although the range of is in R? , it is usually chosen in practice in (0, 1] [Dwork and Roth, 2013, Chap 1&2].
+
12
individual gene maps while protecting patient privacy. Future work will be devoted to deeply investigate
these applications.
References
T. Asano, B. Bhattacharya, M. Keil, and F. Yao. Clustering algorithms based on minimum and maximum
spanning trees. In Proceedings of the Fourth Annual Symposium on Computational Geometry, SCG ’88,
pages 252–257, New York, NY, USA, 1988. ACM.
A. Blum, K. Ligett, and A. Roth. A learning theory approach to non-interactive database privacy. In
Proceedings of the Fortieth Annual ACM Symposium on Theory of Computing, STOC ’08, pages 609–
618, New York, NY, USA, 2008. ACM.
L. Chen, T. Yu, and R. Chirkova. Wavecluster with differential privacy. In Proceedings of the 24th ACM
International on Conference on Information and Knowledge Management, CIKM ’15, pages 1011–1020,
New York, NY, USA, 2015. ACM.
C. Dwork. A firm foundation for private data analysis. Commun. ACM, 54(1):86–95, Jan. 2011.
C. Dwork and A. Roth. The algorithmic foundations of differential privacy. Foundations and Trends R in
Theoretical Computer Science, 9(3-4):211–407, 2013.
C. Dwork, K. Kenthapadi, F. McSherry, I. Mironov, and M. Naor. Our data, ourselves: Privacy via
distributed noise generation. In Eurocrypt, volume 4004, pages 486–503. Springer, 2006a.
C. Dwork, F. McSherry, K. Nissim, and A.Smith. Calibrating noise to sensitivity in private data analysis.
In Theory of Cryptography, pages 265–284. Springer Berlin Heidelberg, 2006b.
O. Grygorash, Y. Zhou, and Z. Jorgensen. Minimum spanning tree based clustering algorithms. In 2006
18th IEEE International Conference on Tools with Artificial Intelligence (ICTAI’06), pages 73–81, Nov
2006.
M. Hay, C. Li, G. Miklau, and D. Jensen. Accurate estimation of the degree distribution of private
networks. In 2009 Ninth IEEE International Conference on Data Mining, pages 169–178, Dec 2009.
S.-S. Ho and S. Ruan. Preserving privacy for interesting location pattern mining from trajectory data.
Trans. Data Privacy, 6(1):87–106, Apr. 2013.
S. P. Kasiviswanathan, K. Nissim, S. Raskhodnikova, and A. Smith. Analyzing graphs with node differential privacy. In Proceedings of the 10th Theory of Cryptography Conference on Theory of Cryptography,
TCC’13, pages 457–476, Berlin, Heidelberg, 2013. Springer-Verlag.
F. McSherry. Privacy integrated queries. In Proceedings of the 2009 ACM SIGMOD International Conference on Management of Data (SIGMOD). Association for Computing Machinery, Inc., June 2009.
F. McSherry and K. Talwar. Mechanism design via differential privacy. In Annual IEEE Symposium on
Foundations of Computer Science (FOCS), Providence, RI, October 2007. IEEE.
A. Morvan, K. Choromanski, C. Gouy-Pailler, and J. Atif. Graph sketching-based massive data clustering.
SIAM Data Mining 2018 (to appear), 2017.
Y. Mülle, C. Clifton, and K. Böhm. Privacy-integrated graph clustering through differential privacy. In
EDBT/ICDT Workshops, 2015.
H. H. Nguyen, A. Imine, and M. Rusinowitch. Detecting communities under differential privacy. In
Proceedings of the 2016 ACM on Workshop on Privacy in the Electronic Society, WPES ’16, pages
83–93, New York, NY, USA, 2016. ACM.
13
K. Nissim, S. Raskhodnikova, and A. Smith. Smooth sensitivity and sampling in private data analysis. In
Proceedings of the thirty-ninth annual ACM symposium on Theory of computing - STOC. ACM Press,
2007.
R. Pinot. Minimum spanning tree release under differential privacy constraints. ArXiv e-prints, Jan. 2018.
S. E. Schaeffer. Graph clustering. Computer Science Review, 1(1):27 – 64, 2007.
A. Sealfon. Shortest paths and distances with differential privacy. In Proceedings of the 35th ACM
SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems - PODS. ACM Press, 2016.
C. T. Zahn. Graph-theoretical methods for detecting and describing gestalt clusters. IEEE Trans. Comput., 20(1):68–86, Jan. 1971.
14
SUPPLEMENTARY MATERIAL
6
Proof regarding the accuracy of DBMSTClu
6.1
Proof of Theorem 7
This theorem relies on the following lemma:
∗
Lemma 1. Let us consider a graph G = (V, E, w) with K clusters C1∗ , . . . , CK
and T an MST of G. If
∗
for all i ∈ [K], Ci is weakly homogeneous, then argmax w(e) ⊂ CutG (T ) i.e. the heaviest edges in T are
e∈T
in CutG (T ).
Proof. Let us consider Ci∗ a cluster of G. As Ci∗ is weakly homogeneous, ∀j ∈ [K] s.t. e(ij) ∈ CutG (T ),
max w(e) < w(e(ij) ). Hence, argmax w(e) ⊂ CutG (T ).
e∈T|C ∗
e∈E(T )
i
∗
Theorem. 7 Let us consider a graph G = (V, E, w) with K homogeneous clusters C1∗ , . . . , CK
and T an
MST of G. Let now assume that at step k < K − 1, DBMSTClu built k + 1 subtrees C1 , . . . , Ck+1 by
cutting e1 , e2 , . . . , ek ∈ E.
Then, Cutk := CutG (T ) \ {e1 , e2 , . . . , ek } =
6 ∅ =⇒ DBCVIk+1 ≥ DBCV Ik , i.e. if there are still
edges in Cutk , the algorithm will continue to perform some cut.
Pk+1
Proof. Let note DBCVI at step k, DBCV Ik = i=1 |CNi | VC (Ci ). Let assume that Cutk 6= ∅. Therefore,
there is e∗ ∈ Cutk and i ∈ {1, . . . , k + 1} s.t. e∗ ∈ E(Ci ). Since e∗ ∈ CutG (T ), using Lem. 1, one can
always take e∗ ∈ argmax w(e). Then, if we denote Ci1 , Ci2 the two subtrees of Ci induced by the cut of e∗
e∈E(Ci )
(see Fig. 3 for an illustration) and DBCV Ik+1 (e∗ ) the associated DBCVI value,
∆ = DBCV Ik+1 (e∗ ) − DBCV Ik
|Ci2 |
|Ci |
|Ci1 |
SEP(Ci1 ) − DISP(Ci1 )
SEP(Ci2 ) − DISP(Ci2 )
SEP(Ci ) − DISP(Ci )
+
−
=
.
N
max(SEP(Ci1 ), DISP(Ci1 ))
N
max(SEP(Ci2 ), DISP(Ci2 ))
N
max(SEP(Ci ), DISP(Ci ))
|
{z
}
{z
}
{z
}
|
|
VC (Ci1 )
VC (Ci2 )
VC (Ci )
There are two possible cases:
1. VC (Ci ) ≤ 0, then SEP(Ci ) ≤ DISP(Ci ) = w(e∗ ). As for l ∈ {1, 2}, SEP(Cil ) ≥ SEP(Ci ) and
DISP(Cil ) ≤ DISP(Ci ) because e∗ ∈ argmax w(e), then, for l ∈ {1, 2},
e∈E(Ci )
SEP(Cl ) − DISP(Ci )
SEP(Ci )
SEP(Cil ) − DISP(Cil )
≥
=
−1
max(SEP(Ci ), DISP(Ci ))
w(e)
max(SEP(Cil ), DISP(Cil ))
and ∆ ≥ 0.
2. VC (Ci ) ≥ 0, then SEP(Ci ) ≥ DISP(Ci ) = w(e∗ ) i.e. max(SEP(Ci ), DISP(Ci )) = SEP(Ci ), for l ∈
{1, 2}, DISP(Cil ) ≤ DISP(Ci ) i.e. DISP(Cil ) ≤ w(e∗ ), SEP(Cil ) = w(e∗ ) hence SEP(Cil ) ≥ DISP(Cil ).
Thus, VC (Ci ) = 1 −
VC (Cil )
DISP(Ci )
SEP(Ci )
≥ VC (Ci ) and ∆ ≥ 0.
and for l ∈ {1, 2}, VC (Cil ) = 1 −
DISP(Cil )
.
SEP(Cil )
Then, for l ∈ {1, 2},
For both cases, ∆ = DBCV Ik+1 (e∗ ) − DBCV Ik ≥ 0. Hence, at least the cut of e∗ improves the current
DBCVI, so the algorithm will perform a cut at this stage.
15
Ci
Ci2
e∗
Ci1
Figure 3: Illustration for Th. 7’s proof.
6.2
Proof of Theorem 8
∗
and T an
Theorem. 8 Let us consider a graph G = (V, E, w) with K homogeneous clusters C1∗ , . . . , CK
MST of G.
Let now assume that at step k < K − 1, DBMSTClu built k + 1 subtrees C1 , . . . , Ck+1 by cutting
e1 , e2 , . . . , ek ∈ E. We still denote Cutk := CutG (T )\{e1 , e2 , . . . , ek }.
Then, Cutk 6= ∅ =⇒
argmax
DBCV Ik+1 (e) ⊂ Cutk i.e. the edge that the algorithm cuts at
step k + 1 is in Cutk .
e∈T \{e1 , e2 , ..., ek }
Proof. It is sufficient to show that, at step k, if there exists an edge e∗ whose cut builds two clusters, then
e∗ maximizes DBCVI among all possible cuts in the union of itself and both resulting clusters. Indeed,
showing this for two clusters, one can easily generalize to the whole graph as a combination of couples of
clusters (see Fig. 5 for an illustration): if for each couple, the best local solution is in Cutk , then the best
general solution is necessary in Cutk .
Let us consider at step k of the algorithm two clusters C1∗ and C2∗ such that e∗ the edge separating
them in T is in Cutk (see Fig. 4 for an illustration). For readability we denote T|C1∗ = C1∗ and T|C2∗ = C2∗
Let us proof that for all ẽ ∈ T|C1∗ ∪C2∗ , one has: DBCV Ik+1 (e∗ ) > DBCV Ik+1 (ẽ). W.l.o.g. let assume
∗
∗
ẽ ∈ C1∗ and let denote C1,1
and C1,2
the resulting subtrees from the cut of ẽ. We still denote DBCV Ik+1 (e)
the value of the DBCVI at step k + 1 for the cut of e.
∆ := DBCV Ik+1 (e∗ ) − DBCV Ik+1 (ẽ)
|C ∗ |
|C2∗ |
SEP(C1∗ ) − DISP(C1∗ )
SEP(C2∗ ) − DISP(C2∗ )
+
= 1
N
max(SEP(C1∗ ), DISP(C1∗ ))
N
max(SEP(C2∗ ), DISP(C2∗ ))
|
{z
}
A
!
!!
∗
∗
∗
∗
∗
∗
|
) − DISP(C1,1
)
|
) − DISP(C1,2
)
|C1,1
SEP(C1,1
|C1,2
SEP(C1,2
−
+
∗ ), DISP(C ∗ ))
∗ ), DISP(C ∗ ))
N
max(SEP(C1,1
N
max(SEP(C1,2
1,1
1,2
|
{z
}
B
By weak homogeneity of C1∗ and C2∗ , A =
|C1∗ |
N
1−
DISP(C1∗ )
SEP(C1∗ )
16
+
|C2∗ |
N
1−
DISP(C2∗ )
SEP(C2∗ )
>0
∗
|C1,1
|
B=
N
|
!
∗
∗
∗
|C1,2
|
SEP(C1,1
) − DISP(C1,1
)
+
∗
∗
max(SEP(C1,1 ), DISP(C1,1 ))
N
{z
} |
By Lem. 1, e∗ ∈
B1
!
∗
∗
SEP(C1,2
) − DISP(C1,2
)
∗ ), DISP(C ∗ ))
max(SEP(C1,2
1,2
{z
}
B2
∗
w(e) so DISP(C1,2
) = w(e∗ ).
argmax
e∈E(T|C ∗ ∪C ∗ )
∗
1
2
∗
Since e ∈ CutG (T ), one has w(e∗ ) ≥ max(SEP(C1∗ ), SEP(C2∗ )). Moreover, as C2∗ is a subtree of C1,2
,
∗
∗
|C
|
SEP(C
)
1,2
1,2
∗
∗
∗
∗
then SEP(C1,2 ) ≤ SEP(C2 ). Thus, w(e ) ≥ SEP(C1,2 ). Finally, B2 = N
DISP(C ∗ ) − 1 ≤ 0.
1,2
∗
∗
Besides, w(ẽ) ≤ SEP(C1∗ ) =⇒ SEP(C1,1
) = w(ẽ) ≤ max∗ w(e) and DISP(C1,1
)=
e∈E(C1 )
max
∗ )
e∈E(C1,1
w(e) ≥
min w(e). Then, two possibilities hold:
e∈E(C1∗ )
1. B1 < 0 =⇒ B < 0 < A.
2. B1 ≥ 0, thus one has B1 =
condition, there is:
∗
|C1,1
|
N
DISP(C1∗ )
SEP(C1∗ )
1−
∗
DISP(C1,1
)
∗ )
SEP(C1,1
≤
∗
|C1,1
|
N
min w(e)
1−
∗
e∈C1
max w(e)
!
. Under weak homogeneity
∗
e∈C1
min w(e)
<
∗
e∈C1
max∗ w(e) .
Thus,
e∈C1
∗
|C1,1
|
DISP(C1∗ )
1−
N
SEP(C1∗ )
∗
|C |
DISP(C1∗ )
∗
< 1 1−
because C1,1
is a subtree ofC1∗
N
SEP(C1∗ )
B1 <
<A
So, B1 + B2 = B < A = DBCV Ik+1 (e∗ ).
Since B < A, ∆ > 0 and e∗ maximizes DBCVI among all possible cuts in the union of itself and both
resulting clusters. Q.E.D.
C1∗
∗
C1,1
ẽ
e∗
C2∗
∗
C1,2
Figure 4: Illustration for Th. 8’s proof.
6.3
Proof of Theorem 9
∗
Theorem. 9 Let us consider a graph G = (V, E, w) with K weakly homogeneous clusters C1∗ , . . . , CK
and T an MST of G. Let now assume that at step K − 1, DBMSTClu built K subtrees C1 , . . . , CK by
cutting e1 , e2 , . . . , eK−1 ∈ E. We still denote CutK−1 := CutG (T )\{e1 , e2 , . . . , eK−1 }.
17
Figure 5: Illustration for Th. 8’s proof. Each circle corresponds to a cluster. The six clusters are handled
within five couples of clusters.
Then, for all e ∈ T \{e1 , e2 , . . . , eK−1 }, DBCV IK (e) < DBCV IK−1 i.e. the algorithm stops: no
edge gets cut during step K.
Proof. According to Th. 7 and Th. 8, for all k < K, if Cutk 6= ∅, the algorithm performs some cut from
CutG (T ). We still denote for all j ∈ [K] Cj∗ = T|Cj∗ .Since |CutG (T )| = K − 1, the K − 1 first steps produce
P |Cj∗ |
∗
K − 1 cuts from CutG (T ). Therefore, DBCVIK−1 =
N VC (Cj ).
j∈[K−1]
∗
∗
Let be e the (expected) edge cut at step K, splitting the tree Ci∗ into Ci,1
and Ci,2
.
∆ = DBCVIK−1 − DBCVIK
∗
∗
|Ci,1
|
|Ci,2
|
|Ci∗ |
∗
∗
VC (Ci∗ ) −
VC (Ci,1
)−
VC (Ci,2
)
N
N
N
∗
∗
∗
∗
∗
∗
) − DISP(Ci,1
)
) − DISP(Ci,2
)
|Ci,1
| SEP(Ci,1
|Ci,2
| SEP(Ci,2
|C ∗ | SEP(Ci∗ ) − DISP(Ci∗ )
= i
−
−
∗
∗
∗
∗
∗
∗ ))
N max(SEP(Ci ), DISP(Ci ))
N max(SEP(Ci,1 ), DISP(Ci,1 ))
N max(SEP(Ci,2 ), DISP(Ci,2
=
Since Ci∗ is a weakly homogeneous cluster, therefore SEP(Ci∗ ) ≥ DISP(Ci∗ ). Then, minimal value of
∗
∗
∗
∗
∗
∗
∆, ∆min is reached when SEP(Ci,1
) ≥ DISP(Ci,1
), SEP(Ci,2
) ≥ DISP(Ci,2
), SEP(Ci,1
) = SEP(Ci,2
) =
0
∗
∗
0
min
w(e
),
DISP(C
)
=
DISP(C
)
=
max
w(e
).
Then,
i,1
i,2
∗
∗
0
0
e ∈E(Ci )
N × ∆min
e ∈E(Ci )
!
!
∗
∗
∗
DISP(C
)
DISP(C
)
DISP(C
)
i,1
i,2
i
∗
∗
− |Ci,1
| 1−
− |Ci,2
| 1−
= |Ci∗ | 1 −
∗ )
∗ )
SEP(Ci∗ )
SEP(Ci,1
SEP(Ci,2
max w(e0 )
max w(e0 )
∗
0 ∈E(C ∗ )
0 ∈E(C ∗ )
e
e
DISP(C
)
i
i
i
∗
∗
− |Ci,2
= |Ci∗ | 1 −
− |Ci,1
| 1−
| 1−
0)
SEP(Ci∗ )
min
min
w(e
w(e0 )
∗
∗
0
0
= |Ci∗ | −
DISP(Ci∗ )
+
SEP(Ci∗ )
By weak homogeneity condition on
e ∈E(Ci )
max∗ w(e0 )
0
min ∗ w(e0 )
0
e ∈E(Ci )
e ∈E(Ci )
e ∈E(Ci )
DISP(C ∗ )
Ci∗ , SEP(C ∗i)
i
<
min
w(e0 ))
max
w(e0 )
e0 ∈E(C ∗ )
i
e0 ∈E(C ∗ )
i
and ∆ > 0.
18
max
≤
w(e0 ))
e0 ∈E(C ∗ )
i
)
e0 ∈E(C ∗
i
minw(e0 ) .
Therefore, ∆min > 0
7
7.1
Proofs regarding the accuracy of PTClust
Proof of Theorem 13
∗
Theorem. 13 Let us consider a graph G = (V, E, w) with K strongly homogeneous clusters C1∗ , . . . , CK
and T = PAMST(G, uG , w, ), > 0. T has a partitioning topology with probability at least
− 2∆u (|V |−1) (ᾱi max(w(e))−min (w(e)))+ln(|E|)
K
X
G
e∈E(G|C ∗ )
e∈E(G|C ∗ )
i
i
1−
(|Ci∗ | − 1)e
i=1
Proof. Let T = PAMST(G, uG , w, ), {R1 , ..., R|V |−1 } denotes the ranges used in the successive calls of
the Exponential mechanism in PAMST(G, uG , w, ), rk = MExp (G, w, uG , Rk ,
), and Steps(Ci∗ )
|V | − 1
| {z }
0
the set of steps k of the algorithm were Rk contains at least one edges from G|Ci∗ . Finally for readability
we denote uk = uG (w, rk )
P[T has a partitioning topology]
=P[∀i, j ∈ [K], i 6= j, |{(u, v) ∈ E(T ), u ∈ Ci∗ , v ∈ Cj∗ }| = 1] = 1 − P[¬A]
|
{z
}
A
If we denote B = “∀i ∈ [K], ∀k > 1 ∈ Steps(Ci∗ ), if rk−1 ∈ E(G|Ci∗ ) then rk ∈ E(G|Ci∗ )” One easily
has: B =⇒ A, therefore P[¬A] ≤ P(¬B). Moreover, by using the privacy/accuracy trade-off of the
exponential mechanism, one has
2∆uG
≤ exp(−t).
∀t ∈ R, ∀i ∈ [K], ∀k ∈ Steps(Ci∗ ) P
u
≤
−
(t
+
ln
|R
|)
k
k
0 {z
|
}
Ak (t)
Moreover one can major P[¬B] as follows
P ∃i ∈ [K], ∃k ∈ Steps(Ci∗ ) s.t rk−1 ∈ E(G|Ci∗ ) and rk ∈
/ E(G|Ci∗ )
By using the union bound, one gets
≤
X
i∈[K]
P ∃k ∈ Steps(Ci∗ ) s.t rk−1 ∈ E(G|Ci∗ ) and rk ∈
/ E(G|Ci∗ )
Using the strong homogeneity of the clusters, one has
X
=
P ∃k ∈ Steps(Ci∗ ) s.t uk ≤ −|ᾱi max w(e) − min w(r)|
e∈E(G|C ∗ )
i∈[K]
i
r∈Rk
≤
X
i∈[K]
By setting tk,i =
P ∃k ∈ Steps(Ci∗ ) s.t uk ≤ −|ᾱi max w(e) − min w(e) |
e∈E(G|C ∗ )
i
0
2∆uG (ᾱi max(w(e))
e∈E(G|C ∗ )
i
=
X
i∈[K]
e∈E(G|C ∗ )
− min (w(e))) + ln(|Rk |) one gets
e∈E(G|C ∗ )
i
P [∃k ∈ Steps(Ci∗ ) s.t Ak (tk,i )]
19
i
Since for all i ∈ [K], and k ∈ Steps(Ci∗ ), |Rk | ≤ |E|, and using a union bound, one gets
≤
X
X
P
i∈[K] k∈Steps(Ci∗ )
[Ak (tk,i )] ≤
K
X
≤
(|Ci∗ | − 1) exp −
i=1
7.2
X
X
exp (−ti,k )
i∈[K] k∈Steps(Ci∗ )
ᾱi max w(e) − min w(e) + ln(|E|)
2∆uG (|V | − 1)
e∈E(G|C ∗ )
e∈E(G|C ∗ )
i
i
Proof of Theorem 14
Let recall the theorem from S. Kotz on the Laplace distribution and generalizations (2001):
Theorem 15. Let n ∈ N, (Xi )i∈[n] ∼ Lap(θ, s), denoting Xr:n the order statistic of rank r one has for
iid
all k ∈ N,
h
i
k
E (Xr:n − θ) = sk
n−r
r−1
X
X
n!Γ(k + 1)
(−1)k
aj,r,k +
bj,r,k
(r − 1)!(n − r)!
j=0
j=0
{z
}
|
α(n,k)
∗
Theorem. 14 Let us consider a graph G = (V, E, w) with K strongly homogeneous clusters C1∗ , . . . , CK
0
∗
and T = PAMST(G, uG , w, ), and T = Mw.r (T, w|T , s, τ, p) with s << p, τ . Given some cluster Ci ,
and j 6= i s.t e(ij) ∈ CutG (T ), if HT|C ∗ (e(ij) ) is verified, then HT 0 |C ∗ (e(ij) ) is verified with probability at
i
i
least
2
2
2
Λ1 + (θ(ij) + δ)Λ2 − (Λ3 + θ(ij) Λ24 )
1−
2
Λ1 + (θ(ij)
+ δ)Λ2 + 2Λ3 Λ4
with the following notations:
• δ = ps , θmin =
min w(e)+τ
e∈E(T )
p
max w(e)+τ
• θmax =
e∈E(T )
p
,θ(ij) =
w(e(ij) +τ
p
2
3
4
• Λ1 = 24δ 4 nα(n, 4) + 12θmax δ 3 nα(n, 3) + 12θmax
δ 2 nα(n, 2) + 4θmax
δnα(n, 1) + θmax
2
• Λ2 = 2δ 2 nα(1, 2) + 2θmin δnα(1, 1) + θmin
2
• Λ3 = 2δ 2 nα(n, 2) + 2θmax δnα(n, 1) + θmax
• Λ4 = δnα(1, 1) + θmin
Proof. Let τ > 0 and p > 1, according to the weight-release mechanism, all the randomized edge
weights w0 (e) with e ∈ E(T 0 ) are sampled from independents Laplace distributions Lap( w(e)+τ
, ps ).
p
Given some cluster Ci∗ , and j 6= i s.t e(ij) ∈ CutG (T ), HT|C ∗ (e(ij) ) is verified. Finding the probabili
2
ity that HT 0 |C ∗ (e
i
(ij)
) is verified is equivalent to find the probability P
, ps ) and X out ∼ Lap( w(e
Lap( w(e)+τ
p
(ij)
p
)+τ
(max Xe )
e∈E (C ∗ )
i
min Xe
e∈E (C ∗ )
i
< X out with Xe ∼
indep
, ps ). Denoting with Yi ∼ Lap (θmax , δ) , Zi ∼ Lap (θmin , δ) and X out ∼
iid
20
iid
Lap(θ(ij) , δ), one can lower bounded this probability by P
min Zi
i∈[|Ci∗ |−1]
(max Yi )2
i∈[|C ∗ |−1]
i
min Zi
i∈[|C ∗ |−1]
i
< X out . Choosing τ big enough s.t
< 0 is negligible, one has
P
(max Yi )2
i∈[|Ci∗ |−1]
min Zi
< X out
i∈[|Ci∗ |−1]
out
2
.
×
X
(max
Y
)
−
min
Z
=P
<
0
i
i
i∈[|Ci∗ |−1] i∈[|Ci∗ |−1]
|
{z
}
ϕ
Moreover since τ , p >> s, one has E(ϕ) ≤ 0. Therefore,
P [ϕ < 0] =P ϕ − E(ϕ) < −E(ϕ)
| {z }
≥0
=1 − P [ϕ − E(ϕ) > −E(ϕ)]
Using the one-sided Chebytchev inequality, one gets
≥1 −
V(ϕ)
V(ϕ)
=1−
V(ϕ) + E(ϕ)2
E(ϕ2 )
By giving an analytic form to E(ϕ) and V(ϕ) by using Theorem 15 one gets the expected result.
21
| 8cs.DS
|
Polynomial Transformation Method for Non-Gaussian
Noise Environment
Jugalkishore K. Banoth, Pradip Sircar*
Department of Electrical Engineering
Indian Institute of Technology Kanpur
Kanpur 208016, India
*Corresponding author; Email address: [email protected]
Abstract - Signal processing in non-Gaussian noise
environment is addressed in this paper. For many real-life
situations, the additive noise process present in the system is
found to be dominantly non-Gaussian. The problem of
detection and estimation of signals corrupted with nonGaussian noise is difficult to track mathematically. In this
paper, we present a novel approach for optimal detection and
estimation of signals in non-Gaussian noise. It is
demonstrated that preprocessing of data by the orthogonal
polynomial approximation together with the minimum errorvariance criterion converts an additive non-Gaussian noise
process into an approximation-error process which is close to
Gaussian. The Monte Carlo simulations are presented to test
the Gaussian hypothesis based on the bicoherence of a
sequence. The histogram test and the kurtosis test are carried
out to verify the Gaussian hypothesis.
Keywords: Orthogonal polynomial approximation, Signal
detection and estimation, Non-Gaussian noise
1
Introduction
In the signal detection and estimation problems, we often
assume that the additive random noise process is Gaussian
distributed because this distribution is simple and
mathematically tractable, and the assumption makes analytical
results possible. However for many real-life situations, the
additive noise process is found to be dominantly nonGaussian. Some examples are the ocean acoustic noise and the
urban radio-frequency (RF) noise [1]. The RF receivers
designed to perform in white Gaussian noise can not perform
satisfactorily when the electromagnetic environment
encountered by the receiver system is non-Gaussian in nature
[2]. For detection and estimation of radar signals in high
clutter environments and similar processing of sonar signals in
presence of high reverberation, we need to deal with nonGaussian noise [1, 2].
There are two existing approaches for solving the problems of
detection and estimation of signals in non-Gaussian noise
environment. The first approach is to use the robust statistics
in lieu of the classical mathematical statistics, and to look for
procedures which are consistent or in other words, insensitive
to deviations of the noise distribution from the idealized
model, i.e., the Gaussian distribution [3]. An optimally robust
procedure minimizes the maximum degradation of
performance for a preset deviation of the noise distribution.
The robust techniques, however, can not provide consistent
performance for a noise process with an arbitrary probability
density function (PDF).
The second approach to deal with a non-Gaussian noise
environment is to use a noise model which is general enough
to depict an arbitrary PDF, yet the model retains the desirable
simplicity of manipulation as that of a Gaussian PDF.
Accordingly, the Gaussian-mixture PDF, the generalized
Gaussian PDF, the Middleton class A PDF, and some such
PDFs are employed to model non-Gaussian noise [4].
Incidentally, as the noise model is required to be more
accurate, the ease of analysis as that of a Gaussian PDF
disappears.
In this paper, we present a third approach to deal with a nonGaussian noise environment, by employing the polynomial
transformation method. Preprocessing of data by the
orthogonal polynomial approximation (OPA) together with
the minimum error-variance criterion (MEC) has an excellent
noise-rejection capability [5, 6]. The OPA based
transformation was originally proposed to convert nonuniformly sampled data into uniformly sampled data [5].
However, since the transformation provides significant signal
enhancement by rejecting the high frequency interference,
preprocessing of data may be useful in detection and
estimation problems for better accuracy even for uniformly
sampled data [6]. Perhaps the most desirable feature of
preprocessing the signal samples by the OPA based method is
that the statistical distribution of the approximation-error
process in the preprocessed data becomes close to Gaussian
when the noise process is not necessarily Gaussian distributed
[5]. Based on this argument, the maximum likelihood
estimator (MLE) can be designed to estimate parameters of a
signal corrupted with non-Gaussian noise [7].
In the present work, we take a closer look of the preprocessing
of data by the OPA based method, and we test the hypothesis
that the approximation-error process in the preprocessed data
is Gaussian distributed even when the noise process
corrupting the sampled data is non-Gaussian. Several types of
tests are applied for testing the hypothesis. We plot the
histogram of a given sequence and look for the proverbial bell
shape as a simple test for its Gaussian distribution [8]. We
compute the kurtosis [9] and apply the Hinich test [10, 11] for
validation of the Gaussian hypothesis. We consider the
following noise processes for the Monte Carlo simulation: (i)
Gaussian, (ii) Laplacian, (iii) Uniform, and (iv) Gamma
distributed.
2
Orthogonal Polynomial
The real-valued discrete-time signal g[n] is to be detected/
estimated utilizing the sampled sequence x[n] g[n] w[n] ,
where w[n] is the noise sequence which may not be Gaussian
x[n] are preprocessed by the
distributed. The sampled data
orthogonal polynomial transformation
transformed data y[n] as follows [6],
to
obtain
the
(1)
nm
we can compute the autocorrelation functions (ACFs) of the
error process [6], provided the ACFs of the noise process are
known. Furthermore, by invoking the central limit theorem
when the random variables w[n] are independent with zero
mean and identical variance, and the coefficients nm are
bounded [5, 13], we can argue that the error process will be
close to Gaussian even when the distribution of the noise
process is non-Gaussian.
3
Transformations
y PQ1PT x
where
Gaussian Hypothesis Testing
The third order cumulant of the noise/ error process u[n] is
given by
C3u [k1 , k2 ] E u[n]u[n k1 ]u[n k2 ]
(3)
where E is the expectation operator, and the third order
spectrum, commonly known as the bispectrum, is defined as
the two-dimensional Fourier transform of the third order
cumulant,
S3u (1, 2 )
where
x x[0], x[1], , x[ N 1] ,
p [n] p [m]
j
j
N 1
,
2
j 0
p [i]
i 0 j
J 1
C
k1 k2
3u
[k1, k2 ]exp{ j (1k1 2 k2 )}
T
y y[0], y[1], , y[ N 1] ,
The squared bicoherence B3u 1 , 2
P ij p j [i]; i 0,1,, N 1; j 0,1,, J 1;
follows
T
N 1
N 1
N 1 2
Q P P Diag
p0 [m],
p12 [m], ,
pJ2 1[m]
m 0
m 0
m 0
T
Bu 1 , 2
2
The orthogonal polynomials p j [n] are computed by the
recurrence relation given in [5, 12], and the order of
approximation J is chosen such that the error-variance is
minimum.
The
transformed
sequence
y[n]
is
given
by
y[n] g[n] e[n] , where e[n] is the approximation error.
By utilizing the relation between the error sequence e[n] and
the noise sequence w[n] ,
e[n]
N 1
m 0
nm w[m]
(2)
S3u 1 , 2
2
is determined as
2
S2u 1 S2u 2 S2u 1 2
(5)
where S2u ( ) is the power spectrum.
The Hinich test is based upon the squared bicoherence at a
bifrequency 1 , 2 being zero for Gaussianity of the
underlying sequence. The B
2
value is averaged over the
principal domain [10, 11], and the resulting statistics is central
2 distributed under the null hypothesis: S3u 1 , 2 0 .
Hence, it is easy to devise a statistical test to determine
whether the observed squared bicoherence is consistent with a
2
central distribution by computing a probability of false
alarm (PFA) value. If the null hypothesis of the bispectrum
being zero is not rejected, we then compute the average
kurtosis K u given by [9]
(4)
Ku
E u 4
E u 2
2
3
5
(6)
where the value is averaged over each element of the
sequence u[n] . The kurtosis test is based on the null
hypothesis: Ku 0 for a Gaussian distribution of the
underlying process.
4
Concluding Remarks
In this paper, we present a new technique for optimal
detection and estimation of signals corrupted with nonGaussian noise. We preprocess the sampled data by the
polynomial transformation method which converts the noise
process into an approximation-error process which is
Gaussian distributed.
Simulation Results
The real part of the complex-exponential transient discretetime signal
3
g[n] bi exp( ji )exp( si nT )
(7)
i 1
b1 1.0, s1 0.2 j 2.0, 1 0;
where
b2 0.5, s2 0.1 j 4.0, 3 4;
b3 0.5, s3 0.3 j1.0, 3 6;
Figure 1(a): The Bispectrum of the input noise (Gaussian) and
the output error process
corrupted with non-Gaussian noise setting the signal-to-noise
ratio (SNR) at 10 dB, is sampled at 60 uniformly spaced
points with time interval T 0.15 .
We present the three cases of Laplacian, Uniform, and
Gamma distributed noise environments in this work, beside
the Gaussian noise case. In each case, after applying the
polynomial transformation we obtain the transformed data,
and then, subtracting the transient signal from the transformed
data, the error process is separated. The input noise and the
output error processes are tested for Gaussianity. Figs. 1− 4
show the bispectrum and the histogram plots.
For the Gaussian noise case, we calculate the bicoherence of
the output error and check whether the squared bicoherence is
2
consistent with a central distribution by computing the
PFA value. The PFA is computed to be 0.9479, which is high,
and we cannot reject the null hypothesis. The average kurtosis
value for the output error is computed to be −0.1526, whereas
the kurtosis value for the input noise is computed to be
−0.0857 (theoretical value zero). For the Laplacian case, the
PFA for the input noise is 0.396, and the PFA for the output
error is 0.9975. Since the PFA of the output error is high, we
cannot reject the null hypothesis. The kurtosis values are
2.9359 for the input noise and 0.0148 for the output error. For
the Uniform noise case, we compute the PFA for the input
noise to be 0.6973 and the PFA for the output error to be
0.9649. The average kurtosis values are computed to be
−1.2408 for the input noise and −0.1346 for the output error.
For the Gamma distributed noise environment, the PFA for
the input noise is 0.7379, and the PFA for the output error is
0.9847. The kurtosis values are 0.6962 for the input noise and
0.1025 for the output error. In all cases, we find that the
average kurtosis value of the output error process is near zero,
confirming that the error process is close to Gaussian.
Figure 1(b): The Histogram of the input noise (Gaussian) and
the output error process
Figure 2(a): The Bispectrum of the input noise (Laplacian)
and the output error process
Figure 2(b): The Histogram of the input noise (Laplacian) and
the output error process
[3] P.J. Huber and E.M. Ronchetti, Robust Statistics, John
Wiley, Hoboken, N.J., 2009.
[4] S. A. Kassam, Signal Detection in Non-Gaussian Noise,
Springer-Verlag, New York, 1988.
Figure 3(a): The Bispectrum of the input noise (Uniform) and
the output error
[5] P. Sircar and A.C. Ranade, “Nonuniform sampling and
study of transient systems response”, IEE Proceedings F Radar and Signal Processing, vol. 139 , no. 1, pp. 49–55,
1992.
[6] M. Ravi Shankar and P. Sircar, “Nonuniform sampling
and polynomial transformation method”, IEEE Int’l Conf. on
Commun. ICC 2002, vol. 3, pp. 1721–1725, 2002.
[7] P. Sircar and S. Mukhopadhyay, “Parameter estimation of
transient signal in non-Gaussian noise”, IEEE Region 10 Int’l
Conf. on EC3 – Energy, Computer, Commun. & Cont.
Systems TENCON '91, vol. 3, pp. 236–240, 1991.
Figure 3(b): The Histogram of the input noise (Uniform) and
the output error
[8] D.C. Montgomery and G. C. Runger, Applied Statistics
and Probability for Engineers, John Wiley, New York, 1999.
[9] J.A. Cadzow, “Blind deconvolution via cumulant
extrema”, IEEE Signal Process. Mag., vol. 13, no. 5, pp. 24–
42, 1996.
[10] M.J. Hinich, “Testing for Gaussianity and linearity of a
stationary time series”, J. Time Series Analysis, vol. 3, no. 3,
pp. 169–176, 1982.
Figure 4(a): The Bispectrum of the input noise (Gamma
distributed) and the output error
[11] J.K. Tugnait, “Validation, testing, and noise modeling”,
In: V.K. Madisetti and D.B. Williams (Eds.), The Digital
Signal Processing Handbook, pp. 16.1–16.12, CRC Press,
Boca Raton, 1998.
[12] A. Ralston and P. Rabinowitz, A First Course in
Numerical Analysis, McGraw-Hill, Singapore, 1988.
[13] W. Feller, An Introduction to Probability Theory and Its
Applications, Vol. I & II, John Wiley, New York, 1965.
.
Figure 4(b): The Histogram of the input noise (Gamma
distributed) and the output error
6
References
[1] E.J. Wegman and J.G. Smith (Eds.), Statistical Signal
Processing, Marcel Dekker, New York, 1984.
[2] E.J. Wegman, S.C. Schwartz and J.B. Thomas (Eds.),
Topics in Non-Gaussian Signal Processing, Springer-Verlag,
New York, 1989.
| 5cs.CE
|
A natural-inspired optimization machine based on the
annual migration of salmons in nature
Ahmad Mozaffari1,*, Alireza Fathi2
1: Systems Design Engineering Department, University of Waterloo, ON, Canada
2: Department of Mechanical Engineering, Babol University of Technology, Babol, Iran
* Corresponding Author, E-mail: [email protected]
Abstract
Bio inspiration is a branch of artificial simulation science that shows pervasive contributions to variety of engineering
fields such as automate pattern recognition, systematic fault detection and applied optimization. In this paper, a new
metaheuristic optimizing algorithm that is the simulation of “The Great Salmon Run” (TGSR) is developed. The
obtained results imply on the acceptable performance of implemented method in optimization of complex nonconvex, multi-dimensional and multi-modal problems. To prove the superiority of TGSR in both robustness and
quality, it is also compared with most of the well-known proposed optimizing techniques such as Simulated
Annealing (SA), Parallel Migrating Genetic Algorithm (PMGA), Differential Evolutionary Algorithm (DEA), Particle
Swarm Optimization (PSO), Bee Algorithm (BA), Artificial Bee Colony (ABC), Firefly Algorithm (FA) and Cuckoo
Search (CS). The obtained results confirm the acceptable performance of the proposed method in both robustness and
quality for different bench-mark optimizing problems and also prove the author’s claim.
I.
INTRODUCTION
Metaheuristic algorithms are population based methods that work with a set of feasible solutions and try to
improve them gradually. These algorithms can be divided into two main categories: Evolutionary Algorithms
(EA) which attempt to simulate the phenomenon of natural evolution and swarm-intelligence-based
algorithms. There are many different variants of evolutionary algorithms. The common ideas behind these
techniques are the same: defining a population of individuals and survival of the fittest according to the
theory of evolution. Another branch of population-based algorithms which is known as swarm intelligence
focuses on the collective behavior of self-organized systems in order to develop metaheuristics procedures.
The interactive behavior between individuals with one another or with their environment contributes to the
collective intelligence of the society and often leads to convergence of global behavior. There is a wide
variety of swarm-based algorithms which mimic the natural behavior of insects and animals such as ants,
fishes, birds, bees, fireflies, penguins, frogs and many others [15, 16]. Particle swarm optimization algorithm
(PSO) [4], Mutable Smart Bee Algorithm (MSBA) [17], Bee Algorithm (BA) [5], Artificial Bee Colony
(ABC) [3], Firefly Algorithm (FA) [9] and Cuckoo Search (CS) [10] are some of the most well-known
swarm base algorithms.
In this paper, a new algorithm called TGSR [11-14] is proposed that is inspired by natural migration of
salmons and the dangers which are laid in their migration. In the next section, this phenomenon will be
described briefly. Thereafter, all steps for implementing a simulated form of this major natural event will be
scrutinized. Then, the features and controlling parameters of the proposed inspired algorithm will be
described precisely. At last, it will be compared with cited optimization techniques to elaborate on the
strength and weakness of the proposed technique.
II.
THE GREAT SALMON RUN IN NATURE
The salmon run phenomena is one of the great annual natural events happening in the North America
where millions of salmons migrate through mountain streams for spawning. Since these creatures provide
one of the major food sources for living organisms, their passage upstream is fraught with some grave
dangers. Among them, hungry Grizzly bears, human fishers and waterfalls are most crucial dangers that they
should face. The hungry Grizzly bears congregate in forested valleys where they forage for whatever food
source they can find. However, they can hardly find food and also they are in danger by hungry wolves.
Salmons are the most important food sources for these hungry bears. Bears communicate with each other to
find a passage with higher amount of chubby salmons. In fact, they follow the swarm intelligence rules for
hunting salmons with higher qualities. Humans are one of the other major hunters of salmons. These fishers
often congregate in Alaska where there is an appropriate condition for hunting plenty of salmons. Humans
often mimic some diverse heuristic methodologies to find a region that possesses salmons with higher quality
and quantity. They employ scout ships for investigating the complete passage region. The rest of the fishers
are integrated in areas with higher salmon’s intensity. Fig. 1 illustrates a scheme of great salmon run
phenomena.
Figure 1. A scheme of great salmon run : grizzly bear and commercial fishers
The experiments prove that the humans’ hunting technique is really effective since they first separate into
some subparts to investigate the whole area and then integrate in intense areas. There are still many other
elements that menace salmon’s life during their migration. Waterfalls are another major danger since they
can be found in all pathways. At the beginning of the salmon’s migration, they divide into subgroups using
their instinct and some stochastic interprets. Each of these groups follows different pathways to their
destination. Some of them choose forested passages which are full of different dangerous hunters such as
Grizzly bears and salmon sharks while others move towards oceans, lakes and ponds. Released annual
official reports about the incremental quality and quantity of hunted salmons, suggests a high potential of this
great natural phenomenon to be simulated in the form of a parallel evolutionary strategy. Here, this concept
has been utilized to create a novel intelligent optimization algorithm.
In the next section, all of the abovementioned stages that occur during the natural salmon’s migration will
be applied in the simulation of TGSR algorithm. Thereafter, the resulted algorithm will be compared with
other intelligent optimization techniques.
III.
THE GREAT SALMON RUN (TGSR) ALGORITHM
The proposed TGSR contains two independent intelligent evolutionary operators [18]. Each of these two
operators delegates an independent salmon’s migration pathway. The first one belongs to the salmons that
move through forested regions and mountain’s canyons. The other one belongs to the salmons which are
passed through oceans, lakes and ponds. As it was mentioned in the former section, salmons choose their
passage based on their instinct and without a meaningful inference. However, the experimental evidence
shows that these creatures prefer to cross from ponds and lakes rather than canyons and forested passages [7].
Each of these ways is incorporated with their own natural menaces. Commercial fishers are concentrating on
the ponds and ocean pathways while Grizzly bears hunt the salmons that pass through mountain’s canyons
and forested regions [8]. Each of these two main hunters utilizes different techniques for hunting salmons
with higher qualities. TGSR utilized all above steps to handle an optimization problem. Here, the main steps
of the algorithm are given.
A.
Initialization
In the proposed algorithm, each potential solution delegates the salmon intensity in a region (amount of
salmons in a sub group). In other words, region with higher salmon intensity yields a solution with higher
fitness. The solutions are initialized stochastically spanning to the passage dominance (between lower bound
and upper bounds). (1) represents a procedure which is used to initialize random solutions with respect to the
solution space.
= + ∗
− 1
where lb and ub are the lower and upper bounds, respectively and rand is a random number spanning to 0
and 1 with a uniform distribution.
After initializing the solutions, optimization procedure is started. At the beginning of the optimization
process, all of these initialized solutions (salmons sub groups) are prepared for their migration (iterative
movements). It is obvious that each iteration cycle is equivalent to a natural migration phenomenon.
B.
Choosing Pathways for Migration
Before migration, salmons choose their pathway based on their instinct. This suggests a stochastic
shuffling control parameter for thrusting the salmon groups (initial solutions) in both pathways (evolutionary
operators). (2) formulates a mathematical form of this process.
ℎ : !
"#$ = % & ∗ '( )
"#* = '( − "#$
+
2
where N.$ is the number of salmon groups passing through ocean and ponds, N.* is the number of salmon
group passing through forested regions and mountain canyons, P0 is the number of all salmon groups which
participate in the migration and μ is a sharing factor that represents the salmon’s instinct.. As seen, the
proposed formulation is a strategy for shuffling the solutions stochastically. The results will confirm the
effectiveness of this shuffling in the diversification of the solutions.
After exerting the sharing process, these subgroups are entered in their pathway (evolutionary operator).
They face different dangers while crossing these pathways. In the following, the details of the passage
traversing are given.
C.
Crossing Lakes and Ponds
In the first operator, the human hunting is simulated. Humans hire scout ships to investigate the passage
dominance (solution space). These scout ships applies some arithmetic graphical search (an intelligent
diversification methodology) to explore the passage as best as they can. This exploration has been
mathematically modeled in (3).
23 = 24 + 56, − 24 8
+ 3
!
2
8
23 = 24 + 56, 4 −
where t represents the current iteration number, X; represents a new detected region (new solution) and X<
shows the former region of the scout ship (former solution). δx, y is calculated using (4).
?
5?, @ = @ ∗ ∗ 1 − B
A
4
where T is the number of the maximum iteration, b is a random number larger than 1 and is a random
number spanning to 0 and 1 with a uniform distribution.
The remaining ships (known as commercial fishers) communicate with both scout ships and other active
commercial fishers to find better areas (with higher salmon intensity) for hunting salmons. Then they
congregate in the areas with higher intensity of salmons. Fisher groups often consist of two main hunter
ships and one recruited ship. First, the main hunters find regions with an acceptable salmon intensity
(solution fitness). After that, they inform the recruited agent to exploit nearby regions to find more intense
areas (solution with higher fitness). This exploitation has been mathematically modeled in (5).
2D = E ∗ 2FG − 2FH + 2FG 5
where β is a random number spanning to 0 and 1 with uniform distribution, XR represents the new detected
solution by the recruited agent, XLG is the solution obtained by the first main hunter and XLH is the solution
obtained by the second one.
D.
Crossing Mountain Canyons and Forested Regions
The second operator simulates the Grizzly bears hunting methodology. Similar to other animals, Grizzly
bears communicate with each other to find a region with higher salmon intensity. Their hunting method is
really simple. They always inform each other if they find an acceptable region. Then, the entire Grizzly bear
groups approach the best region and search nearby areas. If they find an area with higher salmon intensity,
they inform other bears. Otherwise, they leave the region and continue the local search. One of the main
disadvantages of the bears hunting procedure is the lack of an independent diverse exploration. Bears hunting
methodology is mathematically expressed in (6).
2M = N O ∗ PD − QD + PD 6
where XS represents a new detected region, BU is the best reported region by the hunting team, LU is the
current region for which the bears have decided to perform a local exploitation and φ is an arbitrary angle
spanning to 0 and 360 degrees. cos φ directs the bears to their destination. Fig. 2 shows the schematic of
Grizzly bear’s movements around the best solution. It is obvious that these animals perform an exploitation
search with different radii and angle of attacks.
E. Regrouping for Spawning
At the end of the migration, the survived salmons congregate in their destination for spawning. In TGSR,
this natural event is simulated through a collection container. After salmons pass through their pathways
(operator’s performance), the salmon subgroups (solutions) are collected in a unique container. In other
words, the solutions are extracted from both operators and make a unique population. At this state, the
algorithm has reached the end of the first iteration.
The change in climate and urge for spawning are two main motivations which force the remaining
salmons to begin another migration. Continuity of these permanent migrations turns the TGSR to a powerful
iterative optimization algorithm. Fig. 3 illustrates the flowchart of the proposed TGSR algorithm.
Figure 2. Schematic of Grizzly bear’s movemen
Figure 3. The flowchart of the TGSR algorithm
In the following section, different numerical benchmark optimization problems are applied to elaborate on
the high robustness and acceptable quality of the TGSR algorithm.
IV.
A.
RESULTS AND DISCUSSION
Definition of the Control Parameters and Comparison Metrics
For validating the performance of the proposed optimization algorithm, the obtained results have been
compared with most of the well-known optimization techniques. The comparison metrics are measure of the
quality and robustness of the algorithm. For all benchmark optimization problems, the maximum iteration
number is the stopping criterion. Table 1 indicates the control parameters which are considered in this case
study.
In the next section, the performance of TGSR will be compared to other optimization algorithms.
TGSR for Benchmark Probleems
B.
In this section, the performance of TGSR will be tested for optimizing some well-known unconstraint
problems (functions). These functions are extracted from Belkadi et al. [2]. All algorithms were executed in
30 different runs with different initial random seeds. Table 2 compares the quality (mean) and standard
deviation (robustness) of the tested algorithms in optimization problems with 30 independent dimensions.
It seems that TGSR is an optimization method with high robustness and quality in optimizing the
stochastic test functions with high dimensionality. It is also evident that the proposed algorithm dominates
all other techniques in both robustness and quality. In the first case, DEA, CS and TGSR reached the global
solution for the Schaffer function. It can also be seen that BA and SA show weaker results comparing to the
other optimization algorithms. The robustness and quality of ABC and PSO are acceptable in this case. The
results show that FA and SA possessed 95% and 90% success respectively which are acceptable in this case.
Again, for Sphere function, all of the optimization algorithms reach the local optimum condition. However,
it can be seen that TGSR and PMGA reach better optimum solutions compared to the other optimization
techniques. For Griewank function, PMGA, DEA and TGSR find the global optimum solution with 100%
success. It shows the high reliability of these techniques in optimizing unimodal problems. ABC, PSO and
SA find an acceptable local optimum solution and FA and CS show weaker results in both quality and
robustness. In optimizing the Rastrigin function; it is only TGSR that is completely successful in finding the
global optimum solution. CS shows an acceptable quality in finding a nearly optimum solution. It can also
be seen that BA, PSO, ABC and DEA show weak results in optimization of the Rastrigin function. In the
last case, TGSR shows an explicit predominance (100% success) in optimization of the Rosenbrock
problem versus other algorithms. After above numerical investigations, it can be concluded that the TGSR
algorithm is highly reliable in optimizing unconstraint problems.
V.
CONCLUSIONS
In this paper, a novel bio inspired metaheuristic method called TGSR has been proposed that is
implemented based on the natural phenomenon of salmon’s migration and the dangers that are laid behind
their passages. To confirm the high robustness and efficient performance of proposed method, it has been
compared with most of the famous optimization techniques. The obtained results were really promising.
After that the algorithm has been applied for optimal designing of spring elements. The gained results show
the high potential of TGSR for dealing with serious engineering problems. In the next work, the effects of
controlling parameters on the convergence speed and diversification of proposed algorithm will be verified
more closely. Besides, a multi objective form of TGSR will be introduced for optimizing some real life
engineering problems.
REFERENCES
[1] B. Akay, D. Karaboga, A Modified Artificial Bee Colony Algorithm for Real-Parameter Optimization,
Information Science , doi: 10.1016/j.ins.2010.07.015.
[2] K. Belkadi, M. Gourgand, M. Benyettou, Parallel Genetic Algorithm with Migration for the hybrid
flowshop scheduling problem, Journal of Applied Mathematics and Decision Sciences 2006 1-17.
[3] D. Karaboga, B. Basturk, On the performance of artificial bee colony (ABC) algorithm, Applied Soft
Computing 8(2007) 687-697.
[4] J. Kennedy, R.C. Eberhart,Particle swarm optimization, In: IEEE International Conference on Neural
Networks, January, Vol.7, No. 6, 1995, pp. 1942-1948.
[5] D.T. Pham, A. Ghanbarzadeh, E. Koc, S. Otri, S. Rahim, M. Zaidi, The Bees Algorithm, Technical
Note, Manufacturing Engineering Centre, Cardiff University, UK, 2004.
[6] R. Storn, K. Price, Differential Evolution: A Simple and Efficient Heuristic for Global Optimization
over Continuous Spaces, International Journal of Global Optimization 11(1997) 341–359.
[7] L.F. Sundstom, E. Petersson, J.I. Johnsson, Heart rate responses to predation risk in Salmo trutta are
affected by the rearing environment, International Journal of Fish Biology 67(2005) 1280–1286.
[8] E.B. Taylor, A review of local adaptation in Salmonidae, with particular reference to Atlantic and
Pacific salmon, International Journal of Aquaculture 98(1991) 185–207.
[9] X.S. Yang, Firefly algorithms for multimodal optimization, In: H. Broersma (Editor), Proceedings of
the Stochastic Algorithms: Foundations and Applications, Sapporo, Springer, 2009, 5792, pp. 178-178.
[10] X.S. Yang, S. Deb, Cuckoo search via L´evy flights, In: F. Rothlauf (Editor), Proceeding of World
Congress on Nature & Biologically Inspired Computing, India: IEEE CS Press, 2009, pp. 210-214.
[11] A. Mozaffari, A. Fathi, S. Behzadipour, The Great Salmon Run: A novel bio-inspired algorithm for
artificial system design and optimization, International Journal of Bio-Inspired Computation 4 (2012)
286-301.
[12] A. Fathi, A. Mozaffari, Modeling a shape memory alloy actuator using an evolvable recursive black-box
and hybrid heuristic algorithms inspired based on the annual migration of salmons in nature, Applied
Soft Computing 14 (2014) 229-251.
[13] A. Mozaffari, A. Fathi, Identifying the behavior of laser solid freeform fabrication system using
aggregated neural network and the great salmon run optimization algorithm, International Journal of
Bio-Inspired Computation 4 (2012) 330-343.
[14] A. Mozaffari, A.M. Goudarzi, A. Fathi, P. Samadian, Bio-inspired methods for fast and robust
arrangement of thermoelectric modulus, International Journal of Bio-Inspired Computation 5 (2013) 19434.
[15] A. Mozaffari, M. Gorji-Bandpy, T.B. Gorji, Optimal design of constraint engineering systems:
application of mutable smart bee algorithm, International Journal of Bio-Inspired Computation 4 (2012)
167-180.
[16] A. Fathi, A. Mozaffari, Stochastic social particle swarm optimizer: A simple and efficient tool for global
optimization, Journal of Bioinformatics and Intelligent Control 2 (2013) 253-264.
[17] A. Mozaffari, A. Ramiar, A. Fathi, Optimal design of classic Atkinson engine with dynamic specific
heat using adaptive neuro-fuzzy inference system and mutable smart bee algorithm, Swarm and
Evolutionary Computation 12 (2013) 74-91.
[18] A. Fathi, A. Mozaffari, TGSR: The great salmon run optimization algorithm, International Journal of
Computer Applications in Technology, In-Press (Available Online).
TABLE I.
CONTROL PARAMETERS OF THE BENCHMARK OPTIMIZATION ALGORITHMS
Algorithm
TGSR
PMGA
[2]
Parameters
&= 0.75
'\]^((^_`] =
0.8
'( = 40
[=10
'( =100
[=100
[=100
=1.6
WFP=0.1
'Fab =0.05
MigNum=3
ABC
[3]
Limit= 10
NP=200
PSO
[4]
nSW=100
[=100
FA, Ref.
[9]
FN=50
[=100
c=0.5
CS
[10]
Nests=40
[=100
'=0.25
c=0.9
SA
[1]
Material=1
DEA
[6]
'( =50
F=1.25
CR=0.3
BA
[5]
n=100
nep=25
[=100
[=100
[=100
nsp=10
Prob=0.1
c1= 2
IW=0.72
c2=2
e=0.95
E=0.2
d=1.0
e=8
ngh=
rang/10
m=20
TABLE II.
PERFORMANCE OF THE OPTIMIZATION ALGORITHMS FOR UNCONSTRAINT PROBLEMS WITH A
DIMENSION OF 30
Problems (Dimensions=30)
Algorithms
ABC
[1]
PMGA
BA
[9]
PSO
[10]
SA
DEA
[1]
CS
[10]
FA
[9]
TGSR
Quality
(Success %)
Schaffer
11.43E+000
(100%)
Sphere
356.06E+000
(100%)
Griewank
1.27E-001
(100%)
Rastrigin
989.44E+000
(100%)
Rosenbrock
123.72E+000
(75%)
Robustness
(Success %)
Quality
(Success %)
44E+000
(100%)
12.63E+000
(100%)
151.03E+000
(100%)
34.43E+001
(100%)
14.44 E-002
(100%)
0
(100%)
165.29E+000
(100%)
120.34E+000
(90%)
209.87E+000
(75%)
101.11E+000
(90%)
Robustness
(Success %)
Quality
(Success %)
23.55E-019
(100%)
23.52E+000
(66%)
4.92E+000
(100%)
3425.34E+000
(78%)
0
(100%)
23.84E+000
(100%)
23.05E+000
(90%)
6823.32E+000
(70%)
342.72E+000
(90%)
1142.45E+001
(77%)
Robustness
(Success %)
Quality
(Success %)
29.76E+000
(66%)
0.53E-007
(88%)
2131.02E+000
(78%)
3110.12E-008
(74%)
21.96E+000
(100%)
11.33E+000
(67%)
5644.32E+000
(70%)
1322.11E+000
(80%)
4591.47E+000
(77%)
4545.23E+000
(58%)
Robustness
(Success %)
Quality
(Success %)
9.00E-008
(88%)
27.53E+000
(90%)
2314.65E+000
(74%)
8554.34E+000
(88%)
78.95E+000
(67%)
19.88E+000
(%90)
121.44E+000
(80%)
7752.98E+000
(66%)
2242.12E+000
(58%)
29.65E+000
(63%)
Robustness
(Success %)
Quality
(Success %)
18.00E+000
(90%)
0
(100%)
1111.02E+000
(88%)
21.21E-004
(68%)
12.46E+000
(90%)
0
(100%)
1112.76E-001
(66%)
235.31E+000
(83%)
16.61E+000
(63%)
143.67E+000
(100%)
Robustness
(Success %)
Quality
(Success %)
0
(100%)
0
(100%)
34.00E+000
(68%)
260.78E-001
(90%)
0
(100%)
210.78E+000
(100%)
27.44E+001
(83%)
10.34E+000
(72%)
20.23E+001
(100%)
2.12E+002
(100%)
Robustness
(Success %)
Quality
(Success %)
0
(100%)
6.63E-015
(95%)
311.13E-001
(90%)
19.57E+000
(80%)
110.02E+000
(100%)
211.32E+000
(100%)
4.10E+000
(72%)
1323.39E+000
(90%)
5.06E+001
(100%)
19.76E+000
(97%)
Robustness
(Success %)
Quality
(Success %)
12.31
(95%)
144.73E+000
(80%)
145.45E+000
(100%)
1231.07E+000
(90%)
25.69E+000
(97%)
0
(100%)
5.33E-017
(100%)
0
(100%)
1.23E-002
(97%)
1.04E-002
(100%)
Robustness
(Success %)
0
(100%)
11.09E-023
(100%)
0
(100%)
0.34E-014
(97%)
2.96E-009
(100%)
| 9cs.NE
|
Container Relocation Problem:
Approximation, Asymptotic, and Incomplete Information
arXiv:1505.04229v2 [cs.DS] 7 Oct 2015
Setareh Borjian
∗
Virgile Galle
†
Vahideh H. Manshadi
Patrick Jaillet ¶
‡
Cynthia Barnhart
Abstract
The Container Relocation Problem (CRP) is concerned with finding a sequence of moves of
containers that minimizes the number of relocations needed to retrieve all containers respecting
a given order of retrieval. While the problem is known to be NP-hard, there is much evidence
that certain algorithms (such as the A∗ search [20]) and heuristics perform reasonably well on
many instances of the problem.
In this paper, we first focus on the A∗ search algorithm, and analyze lower and upper bounds
that are easy to compute and can be used to prune nodes and also to determine the gap
between the solution found and the optimum. Our analysis sheds light on which bounds result
in fast computation within a given approximation gap. We also present extensive simulation
results that improve upon our theoretical analysis, and further show that our method finds the
optimum solution on most instances of medium-size bays. On the “hard” instances, our method
finds an approximate solution with a small gap and within a time frame that is fast for practical
applications. We also study the average-case asymptotic behavior of the CRP where the number
of columns grows. We calculate the expected number of relocations in the limit, and show that
the optimum number of relocations converges to a simple and intuitive lower-bound. This gives
strong evidence that the CRP is “easier” in large instances, and heuristics such as [4] finds near
optimal solution.
We further study the CRP with incomplete information by relaxing the assumption that the
order of retrieval of all containers are initially known. This assumption is particularly unrealistic
in ports without an appointment system. In our model, we assume that the retrieval order of
a subset of containers is known initially and the retrieval order of the remaining containers is
observed later at a given specific time. Before this time, we assume a probabilistic distribution
on the retrieval order of unknown containers. We extend the A∗ algorithm and combine it with
sampling technique to solve this two-stage stochastic optimization problem. We show that our
algorithm is fast and the error due to sampling and pruning the nodes is reasonably small. Using
∗
Oracle, Burlington. Email: [email protected].
ORC, MIT. Email: [email protected].
‡
School of Management, Yale. Email: [email protected].
§
CEE, MIT. Email: [email protected].
¶
EECS and ORC, MIT. Email: [email protected].
†
1
§
our framework, we study the value of information and its timing effect on the average number
of relocations.
1
Introduction
With the growth in international container shipping in maritime ports, there has been an increasing
interest in improving the operations efficiency in container terminals on the sea side and land side.
The operations on the sea side include loading export containers on the vessels or discharging
import containers from vessels and loading them onto internal trucks. The import containers are
then transferred to the land side and are stacked in the storage area. These containers are later
retrieved and delivered to external trucks to be distributed in the city.
Due to limited space in the storage area, containers are stacked in tiers on top of each other. As
shown in Figure 1, several columns of containers in a row create a bay of containers. If a container
that needs to be retrieved (target container) is not located at a top most tier and is covered by
other containers, the blocking containers must be relocated to another slot. As a result, during the
retrieval process, one or more relocation moves are performed by the yard cranes. Such relocations
are costly for the port operators and result in delay in the retrieval process. Thus, reducing the
number of relocations is one of the main challenges faced by port operators. Finding the sequence
of moves that minimizes the number of relocations while retrieving containers from a bay in a predefined order is referred to as the Container Relocation Problem (CRP) or the Block Relocation
Problem (BRP). A common assumption of the CRP is that only the containers that are blocking
the target container can be relocated. We refer to the CRP with this setting as the restricted CRP.
Figure 1: Illustration of a bay of containers in storage area
The CRP has been widely studied in the literature and has been shown to be NP-hard [4]. A
few papers have developed mathematical models for the CRP and have solved small-sized instances
of the problem exactly ([4], [14], [1]). In other papers, different heuristics are designed and tested on
real-sized instances ([9], [3], [11], [5]). Many of these heuristics are fast and shown to be empirically
close to optimal for small-sized instances. However, none of them have provable guarantees and their
performance on large-sized instances cannot be well assessed. Among the optimization methods
2
that provide exact solution for the CRP, the A∗ algorithm (a branch and bound framework) has
shown to be very effective and able to solve large instances of the problem in a time frame that is
suitable for practical application [20], [22], [16].
In this paper, we apply the A∗ algorithm ([20], [22],[16]) to the CRP, and analyze various upper
and lower bounds. First, we study the properties of a counting lower bound that simply counts the
number of blocking containers in the initial configuration (introduced in [9]). We show that this
bound is non-decreasing on each path of the A∗ tree, implying that the guaranteed optimality gap
given by the A∗ algorithm may only be improved as more levels of the tree are constructed (see
Proposition 2.2). Moreover, we analyze a generalized family of look-ahead lower bounds that also
counts the future relocation moves due to unavoidable blocking (introduced in [22]), and we prove
that this family provide tighter lower-bounds that are still monotone in each path (see Proposition
2.4 and 2.5 and Figure 10). Our result implies that using look-ahead lower bounds (with a slightly
higher computation time compared to the counting bound) can decrease the number of branching
needed to achieve a desired optimality gap.
We also study several properties of the H heuristic presented in [4] which indicates that this
heuristic serves well as an upper bound. In particular, we prove that the A∗ algorithm terminates
after a certain level when the H heuristic is used combined with the appropriate lower bound (see
Proposition 2.8, 2.7, and 2.9). Our numerical simulations show that, the algorithm typically finds
the optimal solution much faster than what is suggested by the theoretical results (see Figures 8
and 9). We also use the idea of branching to design a heuristic called the Tree-Heuristic (TH-L,
see Algorithm 2). We use randomly generated instances to benchmark several existing heuristics
and the new Tree Heuristic against the optimal solution given by the A∗ algorithm. We show that
TH-L has the best performance (smallest optimality gap) on average (see Table 2).
Building upon the previous results, we perform an average case analysis of the CRP in a bay
with finite height. First we compute the expected value of the counting lower bound ([4], see
Proposition 3.1) for an evenly distributed bay. Then we provide an asymptotic result and prove
that as the number of columns grows, in expectation, the optimum number of relocations converges
to the counting lower bound (see Theorem 3.3). The intuition behind this result is that as the
number of columns grows, the probability of finding a good column for a blocking container (i.e.,
a column where the blocking container will never be relocated again) approaches 1. In that case,
the optimal number of relocations is essentially the number of blocking containers in the initial
bay. Furthermore, we strengthen this result by showing experimentally that the expected number
of relocations performed by heuristic H ([4]) also converges to the counting lower bound (see Figure
6a). This further highlights the effectiveness of both heuristics H and TH-L in the case of large
instances.
A critical assumption of the CRP is that the retrieval order of all containers in the bay is
considered to be known in advance. However, in practice, such information is not available far in
advance and the exact departure order of containers becomes known only as their retrieval times
3
approach. Here we relax this assumption in the following way: We consider a 2-stage setting,
where the retrieval order of a subset of containers is known initially and the retrieval order of
the remaining containers is revealed all at once at a later time. We refer to this problem as the
CRP with incomplete information. Before the entire information becomes available, we assume
a probabilistic distribution on the unknown retrieval orders. We develop a 2-stage approximate
stochastic optimization algorithm, called ASA*, and use it to solve medium-sized instances of the
CRP with incomplete information. This algorithm is an extension of the A∗ algorithm combined
with sampling and pruning techniques based on the bounds defined earlier. We give theoretical
bounds on the approximation error incurred by this algorithm (see Propositions 4.1, 4.2, and 4.3).
Furthermore, we introduce a myopic heuristic, which is fast and performs well on average compared
to the ASA* algorithm (see Figure 12). We use the heuristic to solve large instances of the problem
and show through experiments that the ratio of expected number of relocations in the incomplete
information setting to that of complete information converges to a constant, as the bay gets larger
(see Figure 13). Moreover, we use the ASA* algorithm and the myopic heuristic to assess the value
of information (see Figure 12). Our experiments show that when the retrieval order of 50%-75% of
containers is known at the beginning, the loss from the missing information is negligible and the
average number of relocations is very close to that of the CRP with full information.
Our algorithms can serve as a decision support tool for port operators as well as an analytical
tool to draw insights on managerial decisions such as determining the size/capacity of the bays
and assessing the value of information. This issue of lack of information is particularly critical as
it captures port operations more realistically. Understanding the value of information could help
port operators design efficient appointment systems for trucks.
1.1
Literature Review
Stahlbock and Voß[15] provide a general survey on port operations literature. For a recent general
review and classification survey of the existing literature on the CRP as well as the stacking problem,
we refer the reader to Lehnfeld and Knust [12].
Many heuristics have been developed for the CRP. Kim and Hong suggest a decision rule that
uses an estimate of expected number of additional relocations for a column [9]. Caserta et al. [3]
propose a binary description of configurations and use it in a four steps heuristic. Lee and Lee [11]
consider an extended objective incorporating working time of cranes in a three phases heuristic.
Caserta et al. [5] derive a meta-heuristic based on the corridor method, which again optimizes
over a restricted set of moves and Caserta et al. in [4] present a heuristic that is defined later in
this paper. Ünlüyurt and Aydın [17] propose a branch and bound approach with several heuristics
based on this idea. Foster and Bortfeld [7] use a tree search procedure to improve a greedy initial
solution. Finally, Petering and Hussein [14] describe a look-ahead heuristic that does not make the
common assumption of the restricted CRP. In our experimental result, we use randomly generated
4
instances to compare the performance of these heuristics (and our new TH-L heuristic) to that of
the optimal solution given by the A* algorithm. We show that on average our TH-L outperforms
the existing heuristics.
In the mathematical programming formulations that insure optimality on the entire feasible set
of moves, the first binary formulation was suggested by Kim and Hong [9]. Petering and Hussein [14]
propose another formulation which is more tractable but cannot solve real-sized instances efficiently.
Caserta et al. [4] derive a different Integer Program (IP). Borjian et al. [2] develop an IP formulation
which allows for incorporating a time component into the model. This model is used to jointly
optimize the number of relocations and total service delay. In all these IP formulations, due to the
combinatorial nature of the problem, the number of variables and constraints dramatically increases
as the size of the bay grows, and the IP cannot be solved for large instances. A way to bypass
this problem has been to look at a new type of algorithm called A∗ . It was first introduced to this
problem by Zhang et al. [20], studied by Zhu et al. [22], and Tanaka and Takii [16] proposed a new
lower bound for the algorithm. Our paper contributes to this line of research (applying A∗ to CRP)
by theoretically analyzing several lower and upper bounds, also by computing (both theoretically
and via simulation) how the optimality gap behaves as the number of branches increases.
Average case analysis of CRP is fairly new. The only other paper in this direction is the recent
paper by Olsen and Gross [13]. They also provide a probabilistic analysis of the asymptotic
CRP when both the number of columns and tiers grow to infinity. They show that there exists a
polynomial time algorithm that solves this problem close to optimality with high probability. Our
model departs from theirs in two (related) aspects: (i) We keep the maximum height (number of
tiers) a constant whereas in [13] the height also grows. Our assumption is motivated by the fact
that the maximum height is limited by the crane height, and it cannot grow arbitrarily. (ii) We
assume the ratio of the number of containers initially in the bay to the bay size stays constant (i.e.,
the bay is almost full at the beginning). On the other hand, in [13], the ratio of the number of
containers initially in the bay to the bay size decreases (and it approaches zero) as the number of
columns grows. In other words, in the model of [13], in large bays, the bay is under-utilized.
Finally, we mention that another major challenge in port optimization is uncertainty (together
with the value of information) which has not been considered in the literature until very recently.
Zhao and Goodchild [21] model uncertainty in truck arrivals at the terminal. They assume that
trucks arrive in groups and that the information is revealed for the whole group at the same time.
Using this assumption of rolling information, they use a myopic heuristic to get insights on the
trade-off between information and optimality. Ku [10] consider a model where the trucks arrive
in groups, and the retrieval order within in each group is unknown in advance. They propose a
stochastic dynamic programming for this problem. They note that the size of decision tree grows
exponentially with the bay size, making the computation time prohibitive. Thus they develop
a heuristic called Expected Reshuffling Index (ERI) to approximately solve the problem. This
heuristic, moves each blocking container to the column that minimizes the expected number of
5
future relocations. In our work, we model uncertainty as a 2-stage stochastic optimization problem,
and we use a generalization of the A∗ algorithm combined with sampling and pruning techniques
to find the near optimal solution of this stochastic optimization problem.
This paper is structured as follows: Section 2 presents the A∗ Algorithm in the full information
case. It describes the method, introduces lower and upper bounds and derives theoretical results for
these bounds. Section 3 presents the asymptotic analysis of the CRP. Section 4 provides a stochastic
scheme for the case of incomplete information and studies the value of information about container
departure times. In Section 5, we present several experimental results for the average case analysis
of the CRP in both the complete and incomplete information cases. Finally, Section 6 concludes
and suggests future research topics. The proofs for all theoretical statements are given in the
Appendix.
2
Optimal solution for the CRP with full information
In this section and for the sake of completeness, we present the A∗ algorithm applied to the CRP,
as previously introduced by [22]. We also discuss bounds developed in the literature ([4] and [22])
and derive new theoretical results on them.
2.1
Description of the algorithm
Before describing the A∗ algorithm, we review the Container Relocation Problem (CRP) and introduce a few notations used throughout the paper: We are given a bay B with C columns and P
tiers, where C ≥ P ≥ 3. Initially N containers are stored in the bay. We label the containers based
on their departure order, i.e., container 1 is the first one to be retrieved. The CRP corresponds to
finding a sequence of moves to retrieve containers 1, 2, . . . , N (respecting the order) with a minimum
number of relocations. For bay B, we denote the minimum number of relocation by zopt (B).
To find a solution, we use the A∗ algorithm, which is basically a move-based approach using a
special decision tree. This approach takes 4 entries: B is the bay with N containers to be solved
optimally. R(D) and S(D) are functions giving respectively upper and lower bounds on zopt (D) for
any configuration D. In this paper, upper bound functions that are considered also give us feasible
solutions. Finally, N is the maximum number of nodes in the decision tree.
The algorithm returns two results: The best incumbent (in objective function) found by A∗ ,
denoted zA , to which we associate the sequence of moves needed to find this incumbent called σA
and the gap guaranteed between zA and zopt (B) denoted by γ. Notice that if the algorithm did not
reach N nodes, then the algorithm found an solution (i.e. γ = 0 and zA = zopt (B)).
Let us define the following notations: l is the level of the tree. A node is at level l, if l relocations
have been done from the initial bay to get to this node. This is the reason why this method is called
a move-based approach. Below we describe the algorithm in detail and introduce the notation used
in the description of Algorithm 1.
6
Algorithm 1 A∗ Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
procedure [zA ,γ]=A∗ (B, R, S, N )
Pre-processing:
while target container n is on top do
retrieve target container n of B and n ← n + 1
end while
Initialize:
zA ← ∞, l ← 0, Bl ← B, m ← 0 (number of nodes)
while there exists a non-visited node do
for all nodes Bl at level l do
Mark node Bl as visited
U(Bl ) = R(Bl ) + l and L(Bl ) = S(Bl ) + l
Updating the incumbent:
if U(Bl ) < zA then
zA ← U(Bl )
end if
Pruning:
if U(Bl ) ≤ L(Bl ) or L(Bl ) ≥ zA then
Prune node Bl (i.e. stop branching from this node)
else
Branching:
for Every “Child” of Bl do
if m ≥ N then
Stop
else
add a non-visited “Child” to the tree at level l + 1 and m = m + 1
end if
end for
end if
end for
l← l+1
end while
Gap:
Lmin =
min
(L(D))
D non-visited leaf
γ ← zA − Lmin
35: end procedure
34:
7
The tree represents the sequence of relocation moves. Suppose n is the target container. If
n is not blocked, it can be retrieved since a retrieval is optimal. In that case, no decision has to
be made, therefore no new node is created. If n is blocked, the topmost blocking container needs
to be relocated to another column. This column has to be chosen in order to minimize all future
relocations. In order to do so, we generate all possible feasible moves for the blocking container by
creating children of the corresponding node in a decision tree procedure. Considering all possible
nodes gives the optimal solution.
Building the whole tree of feasible moves would not be tractable since the number of nodes
increases exponentially with the number of relocations. Consequently, we use certain decision rules
to prune the tree with the guarantee of keeping the optimal path in the tree.
Let Bl be any configuration that appears at level l in the tree. By definition,
S(Bl ) ≤ zopt (Bl ) ≤ R(Bl )
(1)
Let us define the cumulative upper and lower bounds as the sum of the bound evaluated at the
current state and the number of relocations already performed (i.e., l relocations have been done
at level l of the tree) U(Bl ) = R(Bl ) + l and L(Bl ) = S(Bl) + l (line 11 of Algorithm 1). Notice that in
this case
L(Bl ) ≤ zopt (B) ≤ U(Bl ), for all Bl in the A∗ tree
(2)
At any level l, the algorithm prunes paths using the following two rules (line 17 of Algorithm 1):
(i) If L(Bl ) = U(Bl ), then zopt (Bl ) = U(Bl ) and we can stop branching on this path. From this
point, we can simply follow the feasible solution given by the upper bound.
(ii) If L(Bl ) ≥ zA (i.e. the lower bound L(Bl ) is larger than the best solution found so far zA ) then
this node can be pruned because the optimal solution for this node is going to be greater
than its lower bound, hence greater than zA . The best incumbent is updated if a Bl such that
U(Bl ) < zA is found (lines 13, 14 of Algorithm 1).
By construction, the above rules maintain the optimal path in the A∗ tree.
6
1
5
4
2
3
Figure 2: Bay with 3 tiers, 3 columns and 6 containers
8
U=5
6
1
5
4
2
U=4
1
5
4
2
5
4
2
3
5
6
2
3
3
L=3
6
U=5 , L=4
6
4
U=5 , L=4
3
U=4
L=2
5
1
5
6
4
2
3
4
6
2
U=4 , L=4
3
5
6
L=3
4
2
U=5 , L=4
3
Figure 3: Example of A∗ algorithm using H as upper bound and S0 as lower bound (defined in
subsection 2.2.1)
Example 1 Figure 3 presents the A∗ where the initial bay is shown in Figure 2. It uses upper
and lower bounds that are presented in Section 2.2. None of the nodes at level 1 can be pruned.
Now let us consider the 4 nodes at level 2 (the 4 rightmost configurations in Figure 3). The third
one uses rule (i) to stop branching from this node since lower and upper bounds are both equal to
4. At this level zA = 4. Therefore we can also prune the first, second and fourth nodes using rule
(ii). Therefore, the tree is complete at level 2. We follow the path until the third node at level 2,
and then follow the feasible solution given by the upper bound.
We know from [4] that the restricted CRP is NP hard. Therefore, there exist some instances for
which the number of nodes needed to find the optimal solution with this method grows exponentially
with the problem size. However to limit the computation time, we set a maximum number of nodes
N allowed in the A∗ tree. If the number of nodes reaches N , we return our best upper-bound as a
feasible solution. In this case, we cannot insure optimality, but we can provide a guaranteed gap
with the optimum solution. Let Lmin be the minimum of the cumulative lower bounds of every leaf
of the tree that is not pruned by A∗ . Notice that zopt (B) ≥ Lmin . Thus zA − zopt (B) ≤ zA − Lmin = γ
(line 24 of Algorithm 1).
We show later that given certain properties on the lower bounds, γ is nonincreasing in N . In
any case, there exists N ∗ such that if N > N ∗ then γ = 0. In our numerical average-case analysis
9
(Section 5), we show how N and γ are related.
The size of the A∗ tree depends on the local lower and upper bounds used. The closer they
are to the optimal solution at each node, the faster the tree is pruned, hence the faster it solves.
In this paper, we define some of the possible candidates, explain these choices and provide details
about the efficiency of the chosen bounds. Note that the algorithm can also be implemented using
depth first search instead of breadth first search; in this paper, we implement the latter as we find
experimentally that it is the most efficient one on average.
2.2
2.2.1
The Candidate Bounds
The Lower Bounds
The counting lower bound. This bound was introduced by Kim and Hong in [9] and it is based
on the following simple observation. In the initial configuration, if a container is blocking, then it
must be relocated at least once. Thus we count the number of blocking containers in B and we
denote it S0 (B). Note that if a container blocks more than one container (for instance, container 6
blocks containers 1 and 4 in Figure 2), we count this container only once.
Observation 2.1. For any configuration B,
S0 (B) ≤ zopt (B)
We define L0 as the cumulative bound computed with S0 .
As mentioned in section 2.1, we want our algorithm to have the following property: As we increase
N , γ should not increase. One way to enforce this is to insure that given any path of the tree,
the cumulative lower bound is nondecreasing on this path. We will indeed show that L0 follows
this property. Before proceeding let us define a few notations: Let ci be the ith column of the
configuration and min(ci ) be the index of the container with the smallest retrieval time in ci . Now
suppose we want to move the blocking container r. If we relocate r to column ci and r < min(ci ),
then we call such a move a “good” move and column ci is a “good” column for r. On the other
hand, if r > min(ci ), we call this move a “bad” move and ci a “bad” column for r.
Proposition 2.2. For any configuration B, any level l ≥ 0, any configuration Bl in level l of A∗
and any child Bl+1 of Bl , we have L0 (Bl+1 ) ≥ L0 (Bl ).
The look-ahead lower bounds. Note that the counting lower bound (S0 ) is only taking into
account the initial configuration of the bay. By definition, a “bad” move implies a future relocation
for the relocated container. Therefore we can construct lower bounds that anticipate “bad” moves
and hence are closer to the optimal solution. This idea has been used by Zhu et al ([22]) to define
10
the following family of lower bounds. For the sake of completeness, we redefine the bounds formally. In this paper, we prove in Proposition 2.5 that these lower bounds have the non-decreasing
property as we branch in the A∗ tree.
The basic idea is the following: We want to count unavoidable “bad” moves in any feasible
sequence of moves. In order to do so, we need to keep track of the “maximum of minimums of
each column of B” denoted by MM(B) = max (min(ci )). Suppose r is the first blocking container;
i∈{1,...,C}
a “bad” move for r is unavoidable if MM(B) ≤ r. As we empty the bay, the maximum of minimums
is changing. More precisely it depends on the sequence of moves that we use. We can, however,
compute an upper bound on this value for every feasible sequence of moves. One way is to assume
that every container that is relocated is discarded from the bay. At any time, the maximum of
minimums in the “discarded” bay will be not smaller than if we computed it given any feasible
sequence of moves. Before introducing formal notations, we give an example.
6
1
5
4
2
5
3
4
(a) B = B1
2
3
4
(b) B2
3
(c) B3
Figure 4: Example of discarded bays with 3 tiers, 3 columns and 6 initial containers
Example 2 Consider the leftmost bay B in Figure 4. First, one can see that S0 (B) = 2 since 6
is blocking 1, and 5 is blocking 2. Now in B1 , 6 is blocking 1 and MM(B1 ) is 3; thus 6 has to be
relocated at least twice. Therefore the optimum is at least 1 more than S0 . Now consider bay B2
where 1 and 6 have been discarded. 5 is blocking 2 and MM(B2 ) is 4; thus 5 has to be relocated at
least twice. Therefore the optimum is at least 2 more than S0 .
Now we formalize this idea: Let n1 be the smallest container in the bay. Let k be a container
(k ∈ {n1 , . . . , N}) and let Rk (B) be the set of containers blocking k and not blocking any container k′
such that k′ < k.
Let Bn1 = B and let Bn1 +1 be the bay where container n1 and containers in Rn1 (B) have all been
discarded from Bn1 . By recursion, we can define a sequence of bays Bk for k ∈ {n1 + 1, . . . , N}.
For p ∈ N, we now define the pth look-ahead lower bound (denoted by Sp (B)):
Sp (B) = S0 (B) +
min(p+n1 −1,N )
X
X
k=n1
r∈Rk (B)
χ(r > MM(Bk )),
(3)
where χ(·) is the indicator function. The corresponding cumulative lower bound is denoted by Lp (B).
11
Observation 2.3. For every configuration B and p ∈ 1, . . . , N,
Sp (B) ≤ zopt (B)
Proposition 2.4. For p ∈ 1, . . . , N and any bay B, we have Lp (B) ≥ Lp−1 (B).
This result shows that as we increase p, Sp gets closer to the optimal solution. In Section 5,
we numerically study the effect of the choice of lower bounds in the A∗ algorithm in terms of the
number of nodes needed to solve a random set of instances. We observe that on average, most of
the gain comes from the first look ahead.
We remark that SN (B) = SN−C (B) simply because when we reach N-C, there is always at least
one empty column. Note that if k is the smallest integer such that Bk has an empty column, then
the lower bound does not change, i.e., SN (B) = Sk−1 (B). Thus, in practice, the process of computing
the lower bounds terminates much before N − C.
In the next proposition, we show that similar to L0 , the cumulative upper bound Lp has the
desirable monotone property, i.e., they do not decrease as we branch in the A∗ tree.
Proposition 2.5. For any p ∈ 1, . . . , N, configuration B, level l ≥ 0, configuration Bl in level l and
any child Bl+1 of Bl in the A∗ tree, we have Lp (Bl+1 ) ≥ Lp (Bl ).
In general it is hard to find the rate at which the lower bound increases on a given path. In
Section 5, we show how the lower bound SN increases on the optimal path of random instances.
2.2.2
The Upper Bound
Any feasible solution can be used as an upper bound. Thus, we use heuristics to construct feasible
solutions that serve as upper bounds.
Ideally we would want a heuristic that is close to the optimal solution and at the same time
easy to compute. Note that we need to compute an upper bound at every node. The heuristic
proposed by Casserta et al. ([4]) seems to meet those criteria. For completeness, we redefine this
heuristic. In this article we prove several properties of the upper bound that we will use in later
parts (see Propositions 2.7, 2.8 and 2.9).
The Heuristic H ([4]) Suppose n is the target container located in column c, and r is the
topmost blocking container in c. We use the following rule to determine c∗ , the column where r
should be relocated to. Recall that min(ci ) is the minimum of column ci . We set min(ci ) = N + 1 if
ci is empty. We have:
argmin {min(ci ) : min(ci ) > r}
if ∃ ci such that min(ci ) > r
ci ∈{1,...,C}\c
∗
c =
{min(ci )}
otherwise
cargmax
∈{1,...,C}\c
i
12
The rule says: if there is a column where min(ci ) is greater than r (r can do a “good” move),
then choose such a column where min(ci ) is minimized, since columns with larger minimums can
be useful for larger blocking containers. If there is no column satisfying min(ci ) > r (r can only
do “bad” moves), then choose the column where min(ci ) is maximized in order to delay the next
unavoidable relocation of r as much as possible. We will refer to this heuristic as heuristic H and
denote its number of relocations by zH (B).
Example 2 Consider the bay B in Figure 2. Using heuristic H, the sequence is going to be:
relocation of 6 to column c3 , retrieval of 1, relocation of 5 to column c1 , retrieval of 2, relocation of
6 to column c2 , retrieval of 3, relocation of 5 to column c2 , retrievals of 4,5 and 6. Thus zH (B) = 4.
Notice that we had S2 (B) = 4, so heuristic H is optimal for this initial configuration.
By definition, H is a feasible solution, so it is an upper bound on the optimal solution.
Observation 2.6. For any configuration B, we have zopt (B) ≤ zH (B).
The heuristic H has a certain number of useful properties stated below.
Proposition 2.7. In a bay with C columns, for any configuration B with at most C containers, we
have
S0 (B) = zopt (B) = zH (B).
(4)
Proposition 2.8. In a bay with C columns, for any configuration B with at most C + 1 containers,
we have S1 (B) = zopt (B) = zH (B).
This result implies that if we use heuristic H as an upper bound and the S1 lower bound in the
approach, the tree will stop at most after N − C − 1 retrievals with the guarantee of an optimal
solution.
We provide two similar bounding results on H.
A∗
Proposition 2.9. In a bay with C columns, for any configuration B with at most C + k containers,
we have:
• zH (B) ≤ zopt (B) + 2, if k = 2
• zH (B) ≤ zopt (B) +
k(k+1)
2 ,
if 3 ≤ k ≤ C.
Remark 2.10. The following example shows that this upper bound can increase as we branch in the
A∗ tree. In Figure 5, bay (a) shows the initial configuration and bay (b) is one of its two “children”.
It is easy to check that heuristic H needs 7 relocations for bay (a) and 8 for bay (b).
13
5
5
9
7
9
7
1
2
4
2
4
3
6
8
6
8
3
(a)
(b)
Figure 5: Heuristic H can increase on a path of A∗ tree
3
An asymptotic analysis of CRP
In this section, we study CRP for random large bays and we show that as the number of columns
in the bay grows, the problem gets “easier” in the sense that the gap between the optimal solution
and our simplest lower bound (S0 ), does not increase on average, and in fact it is bounded by a
constant. The basic intuition is that, as the number of columns grows, for any blocking container,
we can find a “good” column with high probability. This implies that each blocking container is
only relocated once with high probability.
Throughout this section, we assume that initial configurations have P tiers, C columns, N containers and that there are exactly h containers in each column, where h ≤ P − 1. For more clarity
we denote by BC such a bay with C columns.We assume that the initial configuration is a uniform
sample among all possible such configurations. Notice that in that case, when C grows to infinity,
N = h × C also grows to infinity.
First let us explain how a uniformly random bay is generated. We view a bay as an array of
P × C columns. The slots are numbered from bottom to top, and left to right from 1 to PC. For
example, let P = 4 and C = 7, the second topmost slot of the third column is 11. The goal is to
get a bay B with uniform probability, meaning each container is equally likely to be anywhere in
the configuration, with the restriction that there are h containers per column. We first generate
uniformly at random a permutation of {1, . . . , N} called π. Then we assign a slot for each container
with the following relation: B(j, i) = π(h × (i − 1) + j) for j ≤ h and B(j, i) = 0 for j ≥ h + 1. One
1
can see that each bay is generated with probability N!
. There is a one to one mapping between
configurations with C columns and permutations of {1, . . . , hC}, denoted by ShC . Finally, we denote
the expectation of random variable X over this uniform distribution by EC [X].
Now we compute the expected counting lower bound.
Proposition 3.1. Let S0 be the counting lower bound (defined in Section 2), we have
EC [S0 (BC )] = αh × C,
14
(5)
where αh is the expected number of blocking containers in one column and can be computed as
h−1
X
αh =
k=1
k × pk,h
(6)
and pk,h = P[there are k blocking containers in a column with h containers] can be computed by recrusion as
k
∀h ≥ 0, p0,h =
X1
1
and ∀k ≥ 1, pk,h =
pk−j+1,h−j
h!
h
j=1
Remark 3.2. Note that αh only depends on the distribution of the relative order of the h containers
in one column (and not on the actual labels). Thus the expected number of blocking containers in
one column only depends on its height.
The major result of this part is the following. In the asymptotic case where the number of
columns increases to infinity, the expected optimal number of relocations is asymptotically proportional to the expected number of blocking containers (the counting lower bound).
Theorem 3.3. Let S0 be the counting lower bound (defined in Section 2) and zopt be the optimal
number of relocations. Then for C ≥ h + 1, we have
h
i
EC zopt (BC )
≤ f (C)
(7)
1≤
EC [S0 (BC )]
where
f (C) = 1 +
K
C
→ 1
C→ ∞
(8)
where K is a constant defined by equation (19).
The proof of Theorem 3.3 is given in the Appendix. Here we just give an intuition of the
h
i
proof. We show that as C grows, with high probability EC+1 [zopt (BC+1 )] − EC zopt (BC ) is exactly
h
i
αh . Therefore, for C large enough, EC zopt (BC ) essentially behaves like αh × C, which is equal to
EC [S0 (BC )] (according to Proposition 3.1).
In the next corollary, we show that the optimal solution of the unrestricted CRP has a similar
asymptotic behavior. We remind that the unrestricted CRP refers to the problem where we can
also relocate non-blocking containers.
Corollary 3.4. Let zgen (BC ) be the optimal number of relocations for the unrestricted CRP. For
C ≥ h + 1, we have
h
i
EC zgen (BC )
1≤
≤ f (C)
(9)
αh C
where f is the function defined in Theorem 3.3.
15
The above theorem gives insights on how the expected optimal solution of the CRP behaves
asymptotically on random bays. To conclude this section, we show experimentally that the same
result holds for heuristic H, i.e., the ratio of E [zH (BC )] and E [S0 (BC )] converges to 1 as C goes to
infinity. For each size C, we compute both expectations over a million instances, take their ratio
E[zopt (B )]
E[z (B )]
and plot the result in Figure 6a. Notice that we have 1 ≤ E[S0 (BCC)] ≤ E[SH0 (BCC)] , so Figure 6a also
shows experimentally that Theorem 3.3 holds.
1.4
1.3
1.2
1.25
E[H-S0]
E[H]/E[S0]
1
1.2
1.15
0.6
1.1
0.4
1.05
1
0.8
0.2
0
20
40
60
80
100
(a) Convergence of the ratio
0
200
400
600
800
1000
C: Number of columns
C: Number of columns
E[zH (BC )]
(b) Convergence of the difference E [zH (BC )] −
E [S0 (BC )]
E[S0 (BC )]
Figure 6: Simulations for the Asymptotic Theorem
First, note that Figure 6a implies that the relative gap between heuristic H and S0 shrinks to
E[zH (BC )]−E[zopt (BC )]
E[z (BC )]−E[S0 (BC )]
0 as C increases. Moreover we have
and thus the relative gap
≤ H E[S
0 (BC )]
E[zopt (BC )]
of H with optimality also converges to 0 as C grows to infinity.
h
i
In the proof of Theorem 3.3, we also study the function g(C) = EC zopt (BC ) − EC [S0 (BC )].
Note that g(C) ≤ EC [zH (BC )] − EC [S0 (BC )] where the right-hand side of the inequality is the
function plotted in Figure 6b. The plot shows that g(C) ≤ 1.25 for all C, meaning that g(C) is
bounded as we proved in Theorem 3.3. Moreover, the plot implies that heuristic H is on average
at most 1.25 away from the optimal solution, so heuristic H is relatively more efficient in the case
of large bays. Intuitively, the probability of having a good column converges to 1, as we increase
the number of columns; hence the problem tends to become easier as C grows.
Finally, in the proof, we note that the rate of convergence for the optimal solution to S0 is at
least C1 . Interestingly, we can infer from Figure 6a. that the rate of convergence of the ratio for
heuristic H is also proportional to C1 .
4
CRP with Incomplete Information
The A∗ algorithm as explained in Section 2, relies on the assumption that the departure order of
containers is known in advance. However, in practice, we often only know the departure order of
16
some of the containers (those that are going to depart in the near future).
In this section, we explain how the A∗ algorithm can be adapted in a Dynamic Programming
(DP) framework for solving the CRP in the case that some of the decisions should be made with
incomplete information. We use the A∗ algorithm to obtain an approximate solution for this
problem. In what follows, we describe the setting of the problem, introduce some notations, and
explain the algorithm for the CRP with incomplete information. Moreover, we show how to bound
the approximation error. In Section 5, we present the results of computational experiments.
The CRP with incomplete information involves retrieving N containers from a bay with C
columns and P tiers, where partial information about the departure order of the containers is
initially available. To model the CRP with incomplete information, we discretize time into time
steps of equal length and assume that each move (a relocation or a retrieval) takes exactly one
time step. Further we assume that the label (or index) of a container indicates the earliest time
that it can be retrieved (i.e., container n can be retrieved at time n or later). Usually, very little
information is available about the containers that are going to be retrieved far in the future. Thus,
it is reasonable to assume that at any given time step t, we only know the departure order of a
certain number of containers in the bay (the containers that are going to be retrieved within a
short time horizon after t). We refer to such containers in the bay as known containers. Similarly,
we refer to the remaining containers in the bay as unknown containers. By definition, all unknown
containers have larger indices (i.e., later departure times) than the known containers.
As time passes, some of the known containers are retrieved and as more information becomes
available, some of the unknown containers become known. In general, the information might be
updated multiple times during the retrieval process (every time that the arrival time of a truck is
provided by the truck driver, the information gets updated and some unknown containers become
known). In the most general case, the information can be updated every time step. Alternatively,
we can consolidate several small pieces of information into one or a few pieces, and assume that
the information is revealed at Γ different times during the retrieval process (i.e., Γ is the number
of time that information is updated). In this case, we have a multi-stage problem and Γ sets of
decisions need to be made.
In this paper, we focus on a 2-stage setting; we assume that a subset of containers is initially
known and that the departure order of all the remaining ones becomes known at time t∗ . We denote
the set of known containers by K and assume that containers {1, 2, . . . , |K|} are those known at time
zero. Similarly, the set of unknown containers is denoted by U, and containers {|K| + 1, . . . , N}
become known at time step t∗ > 0. We refer to this setting as the 2-stage CRP since there are two
types of decisions that need to be made: first-stage decisions (retrievals and relocations before time
step t∗ ) and second-stage decisions (retrievals and relocations after time step t∗ ). We assume that
before t∗ , we have probabilistic information about the containers in U, meaning that we know the
probability of realization of each possible departure order. Such information can be obtained from
historical data or from an appointment system that provides some estimate of departure times of
17
the containers (for example a time window for retrieving each container). We denote the set of
possible departure orders of the containers in U (possible scenarios) by Q. From now on, we assume
1
that all scenarios are equally likely, i.e., the probability of each scenario is
, and the number of
|Q|
scenarios is (N − |K|)!. Note, however, that we could use the algorithm to solve the CRP with any
other probability distribution on the departure order of unknown containers.
We use a 2-stage stochastic optimization technique to solve this problem, where in the first-stage
we minimize E[z] as follows:
min
E[z] =
σ1 ,...,σt∗ −1
X 1
z(B(q)).
|Q|
(10)
q∈Q
where B(q) is the resulting bay when scenario q is realized, z(B(q)) is the total number of
relocations for B(q), and σ1 , . . . , σt∗ −1 are the first-stage decisions. The 2-stage problem can be
solved with the A∗ algorithm as follows:
(1). We build the tree with t∗ − 1 moves for time-steps 1, 2, . . . , t∗ − 1, in a similar way as
illustrated in Figure 3; we denote this tree by T[1,t∗ −1] .
(2). For each node at level t∗ − 1, we need to compute the expected number of remaining
relocations. We enumerate all possible scenarios and solve the CRP with complete information
corresponding to each scenario, using the A∗ algorithm.
(3). We find p∗ (the optimal path or sequence of moves) that minimizes the expected total
number of relocations over all paths.
(4). Once p∗ up to time-step t∗ − 1 is selected and we observe the information at t∗ , we use the
A∗ algorithm to solve a specific instance through the end.
Notice that to find the optimal path for time interval [1, t∗ − 1], we need to solve up to (C −
∗
1)(t −1) (N − |K|)! instances in Step (2) with the A∗ algorithm. Although A∗ is fast, the number of
scenarios is prohibitively large and enumerating all scenarios is not feasible due to limited resources
of memory and long computation time. We next explain how we use sampling and pruning to
overcome these issues. We also quantify the error incurred as a result of sampling and pruning. We
refer to the resulting algorithm as ASA∗ (Approximate Stochastic A∗ ).
Limiting the number of possible scenarios on each path. We overcome this issue by
sampling S scenarios on each path and computing the number of relocations for the sampled
departure orders rather than for all possible orders. For each path p, let z̄ be the number of
relocations averaged over samples. Also let E[z] be the true mean. To determine the number of
samples needed to get a good approximation, we use the following version of Hoeffding’s inequality:
!
−2Sδ2
P(|E[z] − z̄| > δ) ≤ 2 exp
,
(rmax − rmin )2
18
(11)
where δ is a pre-specified level of desired precision, and rmax /rmin are lower/upper bounds on
random variable z.
Note that we can set rmin = 0; for rmax , we do not have a tight bound, but we can use N(P − 1)
as an upper bound (since each of the N containers is blocked by at most P − 1 containers). Let us
denote the desired probability for bounding the error (i.e., RHS of (11)) by ǫ. For a given δ and ǫ,
the required number of samples can be computed as follows.
S≥
r2 max ln 2ǫ
−2δ2
.
(12)
By sampling from the possible scenarios on each path, we can significantly reduce the number
of scenarios (and thus the computation time). For example, for a bay with 7 columns, 4 tiers,
21 containers, and |K| = 6, the total number of possible scenarios on any of the paths at time t∗
is about 1011 . Using inequality (12), the total number of scenarios would be around 30,000 for
δ = 0.5, ǫ = 0.05, and rmax = 63. Note that by sampling, we incur an error and ASA∗ may choose
a suboptimal path pASA where E[zpASA ] > E[zp∗ ] (recall that p∗ is the optimal path that would
be chosen without sampling). In the next proposition, we show that such an error, denoted by
∆
e1 = E[zpASA ] − E[zp∗ ], is bounded in expectation (over a uniform distribution on the initial bay).
Proposition 4.1. Suppose for each path, we estimate the number of relocations using S independent
samples, where S is given in (12). Also suppose
ASA∗ chooses path pASA as the optimal path, and
s
π
e1 = E[zpASA ] − E[zp∗ ]. We have E[e1 ] ≤ 2δ
ǫ .
− ln( )
2
Pruning the paths of T[1,t∗ −1] . To address this issue, we use the upper bound and lower
bounds to prune the nodes of T[1,t∗ −1] , similar to the A∗ algorithm. However, since some of the
containers are unknown before t∗ , we have to compute the expectations (E[L] and E[U]). Again,
we use the idea of sampling and we estimate these values by computing L and U using S samples,
where S is obtained from Inequality (12). Because of the sampling error, we may prune an optimal
path by mistake, resulting in an error that is illustrated in Figure 7. This error is the difference
between E(Up̂ ) (the true upper bound of the path with the minimum estimated upper bound), and
E(Lp̃ ) (the true lower bound of the path selected by ASA∗ for pruning). Intuitively, this is because
E(Up̃ ) is the worst we would achieve if we prune p̃ (by mistake), and E(Lp̂ ) is the best we could
achieve if we do not prune p̃.
In the next two propositions, we show that the error due to pruning is bounded, whether we
prune some paths once at t∗ − 1 (resulting in error e2 ) or prune some paths at several time-steps
before t∗ − 1 (resulting in error e3 ).
Proposition 4.2. Suppose that for each path p at time-step t∗ − 1, we estimate the expected lower
and upper bounds (Lp and Up ) from S samples, where S is given in (12). Also, suppose ASA∗
19
Figure 7: Illustration of the error due to pruning: the thick line segment is the error as a result of
pruning the optimal path by mistake.
s
chooses to prune one or more paths. We have E[e2 ] ≤ 2δ
π
ǫ .
− ln( )
2
Proposition 4.3. Suppose we prune some paths at m time-steps t1 , t2 , ..., tm < t∗ − 1. At each
time of pruning and for each path p, we estimate the expected lower and upper bounds (Lp and Up )
from S samples where S is given in (12). Also, suppose that at each time, i, of pruning, ASA∗
chooses to prune one s
or more paths (denoted by p̃i ). The expected total loss, E[e3 ], is bounded
ǫ
2
2
s
)π
−ln(
d
d
min
min
max
max
ǫ 2
π
dmin
ǫ 2
2
δ
2δ
+ Up̂ , where Up̂ = max{Up̂1 , . . . , Up̂m },
( ) δ
+
by m ( )
ǫ
2
δ
2
2
− ln( )
2
di = Lp̃i − Up̂i , dmin = min{d1 , . . . , dm }.
In Propositions 4.1-4.3, we bound the loss that can be incurred by pruning some paths at time
− 1 (e1 and e2 ) or at t < t∗ − 1 (e3 ). Notice that for fixed δ and ǫ, the errors e1 and e2 are
independent of the bay size if we increase the number of samples S according to (12); thus if we
prune paths only at t∗ − 1, the loss remains unchanged and the relative loss (as a percentage of
total relocations) decreases as the bay gets larger. Therefore, for large bays and “hard-to-solve”
configurations, one can pick larger δ and ǫ that result in a smaller number of samples, and the
relative error would still be small.
Table 1 shows e1 , e2 and e3 . The losses are very small even for combinations of δ and ǫ that
result in reasonably small number of samples. For example, for δ ≤ 1 and ǫ < 0.1, E[e1 ] and E[e2 ]
e1 + e2
are no more than 2. Because e1 and e2 are independent of the bay size, the relative error (
)
E[z]
decrease as the bay gets larger (E[z] is the average number of relocations when full information is
available). For the e3 , this measure is almost constant for different bay sizes and can be controlled
by changing ǫ, δ, and m (number of times that we do pruning at t < t∗ − 1),
t∗
20
δ
ǫ
E[e1 ] and E[e2 ]
C
E[e3 ]
3]
% error( E[e
E[z] )
0.1
0.1
0.1
0.5
0.5
0.5
1
1
1
0.01
0.05
0.1
0.01
0.05
0.1
0.01
0.05
0.1
0.15
0.18
0.2
0.77
0.92
1.02
1.54
1.85
2.05
10
15
20
25
30
35
40
45
50
0.91
1.36
1.81
2.26
2.72
3.17
3.62
4.07
4.52
0.07
0.073
0.074
0.075
0.076
0.076
0.076
0.076
0.076
Table 1: Left table: Expected loss due to sampling and pruning at t∗ − 1, for C = 7, P = 4, and
N = 21; Right table: Expected loss due to pruning at t < t∗ − 1 for ǫ = 0.05, δ = 0.5, dmin = 1,
max
Up̂ = 2N, and m = 5
In Section 5, we use ASA∗ to solve the CRP with incomplete information for a bay with 7
columns and 4 tiers and with different amounts of information initially available. Moreover, we
introduce a heuristic for this problem and compare the results of ASA∗ with those of the heuristic.
5
Experimental Results
In this section, several experimental results are presented to further understand the effectiveness
of our algorithms in both complete and incomplete information settings.
In the complete information case, we study thoroughly the A∗ algorithm and the effect of
parameters on the efficiency and performance of the algorithm. First we show that using the Nth
look-ahead lower bound improves dramatically the computational tractability of the algorithm.
Moreover, our experiments shows that on the optimal path of the tree, the Nth look-ahead lower
bound reaches the optimum solution after a few levels. We also show the trade-off between N
(maximum number of nodes in the tree) and the number of instances solved optimally. Further we
introduce a new class of heuristics, the Tree Heuristic (TH-L), compare it with 3 existing heuristics
([9], [4], [14]) and show that it outperforms the existing heuristics.
In the incomplete information setting, we show through experiments that the ASA∗ algorithm
is fast and efficient and most of the instances are solvable within a reasonable time for mediumsized instances. We introduce a myopic heuristic which expands on the H heuristic in [4]. Using
the ASA∗ algorithm and the heuristic, we study the value of information. Our simulations results
show that while more information results in fewer relocations, the marginal value of information
decreases with the level of information. We also use the myopic heuristic to examine the effect of
21
the level of information on the number of relocations for different bay sizes. We show that for any
level of information, the ratio of relocations with incomplete information and complete information
converges to a constant as the bay gets larger.
5.1
Experimental results for complete information
In most of the experiments, we consider medium-sized instances of the CRP. We consider bays of
size 4 tiers and 7 columns with 21 containers (P = 4 = h + 1, C = 7 and N = 21 = hC with 3
containers per column). We generate randomly 100,000 instances using the uniform distribution
on bays described in Section 4.
The effect of lower bounds on the size of the A∗ Tree. In Section 2, we introduced a
family of lower bounds (S0 , S1 ,. . . ,SN ). Clearly SN is more computationally expensive to use than
S0 , but we proved that SN was a tighter lower bound. We show here that using SN improves the
performance of the A∗ algorithm significantly.
For each instance, we solve it using 4 types of lower bounds S0 , S1 , S2 and S21 (here N = 21).
For each of them we record the number of nodes needed to solve the instance optimally. Figure 8
presents the results. First, we show in the box-plots, the distribution of the number of nodes for
each of those lower bounds. Second, we give the average number of nodes. We give two main
insights from this experiment. First, the average number of nodes needed to solve with S21 is 1/3
of the number of nodes with S0 . Also, we point out that using S1 and S2 instead of S0 decreases
the average number of nodes by 1/3 and 1/2, respectively. The fact that the number of nodes is
decreasing as we use tighter lower bounds is not surprising and it is aligned with what we proved
in Section 2. The more surprising insight from the experiment is that introducing lower bounds
that look just one or two steps ahead (i.e., S1 and S2 ) makes a great improvement in terms of
performance of the A∗ approach, without affecting the computation time significantly.
Second, it can be seen that the box is concentrated around 0 for each lower bound, which implies
that most instances are solved within a few hundred nodes. What influences the average number
of nodes are the “hard-to-solve” cases where a large number of nodes is needed. Using S21 instead
of S0 makes the A∗ algorithm more efficient by decreasing the number of nodes by a factor of 3 for
those cases. From now on, we use S21 as the lower bound. Recall that actually we only need to
compute Sk for k ≤ N − C, as Sk+1 , . . . , SN are all equal to Sk . Taking advantage of this property
further improves computational tractability.
The convergence rate of the lower bound. After introducing the lower bounds, we have
shown in Propositions 2.2 and 2.5 that on any given path, the lower bound is non-decreasing. The
efficiency of the algorithm mainly depends on the rate of increase of the lower bound on each path.
If this rate is high, then the algorithm has a greater chance to find the optimal solution with fewer
22
Figure 8: The effect of the Lower Bound on A∗
Difference between L0 and optimum
nodes, as the tree is pruned faster. The lower bounds can either “collapse” with the upper bounds,
or exceeds the best incumbent available.
In this experiment and for the sake of clarity, we only focus on the rate of convergence of the
lower bound on the optimal path. For each instance, we find the optimal path, and at each level l,
we compute the difference between S21 (Bl ) and zopt (Bl ). We average this difference on all instances
and plot the results in Figure 9
0.25
0.2
0.15
0.1
0.05
0
0
2
4
6
8
10
12
14
level of the A* Tree
Figure 9: The convergence rate of the lower bound on the optimal path
First of all, we can see that even at level 0 (meaning at the root node), the average difference
is less than 0.25. This shows that on average the lower bound is close to optimal, therefore it does
23
not need to increase a lot to reach the optimal solution. Second, we can see that after level 10, the
average gap is less than 0.001, meaning that after 10 relocations, in most cases, the lower bound
has reached optimality on the optimal path. Therefore, this path will terminate at level 10 if the
upper bound also collapsed to the optimal. Finally, the trend of the curve shows that the gap
between optimal and the lower bound decreases faster at the beginning of the tree. A simple way
to explain this is that the lower bound increases when a future “bad” move can be predicted. But
such bad moves become harder to predict as the bay gets emptier (there are more empty columns
in the bay).
Percentage of instances solved optimally
The effect of N . By construction, the A∗ method is tunable. In this experiment, we study
the impact of N on the average guaranteed gap with optimality, and the percentage of instances
solved optimally. Notice that from Propositions 2.2 and 2.5, we know that the guaranteed gap is
a non-increasing function of N . We solve each instance with different N and we record the gap
with the optimal solution obtained by the A∗ algorithm. We show the results in Figure 10. On the
left, we give the average gap as a function of N and on the right, we show the percent of instances
solved optimally.
0.4
0.35
Average Gap
0.3
0.25
0.2
0.15
0.1
0.05
0
0
1
2
3
4
5
6
1
0.95
0.9
0.85
0.8
0.75
0.7
0
1
2
3
4
5
6
log(N)
log(N)
(a) Average Gap as a function of N
(b) Percentage of instances solved optimally as
a function of N
Figure 10: The Effect of N
First, we observe that the average guaranteed gap at the root node is very small (less than 0.4).
Also, in more than 70% of instances, the optimal solution is found at the root node. Second, as we
increase N , the gap sharply decreases and the number of instances solved to optimality increases.
This experiment is very useful from a practical a point of view in order to determine N . We
can set an average gap that we allow in our solutions and infer the minimum N that is needed to
achieve this performance on average. For example for bays of 4 by 7, if we allow for a gap of 0.05
on average, we can set our N to 100.
This experiment is related to the first experiment. As shown in Figure 8, most instances only
require few hundreds of nodes, but there are some that require hundreds of thousands. This explains
24
the “concavity” of the function in the left plot and the “convexity” of the function in the right
plot. In conclusion, this experiment highlights the trade-off between the quality of the solution
(represented by the average guaranteed gap) and its tractability (represented by N ).
Benchmark of a new heuristic and existing heuristics. Many heuristics have been developed for the CRP. From a practical point of view, heuristics are valuable since they are fast and
easy to implement in container terminals. Therefore, it is relevant to evaluate their performance
to advise operators. In order to benchmark existing heuristics, one can measure:
• The distribution of the difference between heuristics and the optimal solution; and
• The expected performance ratio of the heuristic defined as
PR(heuristic) =
zheuristic − zopt
zopt
(13)
,
where zheuristic is the number of relocations in the heuristic solution. PR shows the relative
gap with optimality for a given heuristic.
From Figure 8, if we set N to 400, 000, then all 100, 000 instances are solved optimally using the
algorithm. In addition to heuristic H, we study two other existing heuristics presented by Kim
and Hong (KH from [9] which uses the estimation of future relocation) and Petering and Hussein
(LA-5 from [14] that takes into account “repositioning moves”, i.e. repositioning containers that
are not in the same column as the target container).
A∗
Gap with
Heuristic H
Heuristic KH
Heuristic LA-5
Heuristic TH-2
Optimal
Distribution
Distribution
Distribution
Distribution
0
87.0%
31.8%
83.5%
95.7%
1
11.4%
21.4%
13.7%
3.98%
2
1.4%
21.3%
2.0%
0.3%
≥ 3
0.2%
11.4%
0.3%
0.02%
E [PR]
1.44%
16.0%
1.81%
0.44%
Table 2: Benchmarks of Heuristics on 100,000 instances
We use the same idea as in the A∗ algorithm and introduce a new class of heuristics that can
improve any existing heuristics; we refer to this class as Tree Heuristic (TH-L). The basic idea of
TH-L is to take the L best columns and branch on them to construct a decision tree. The L best
columns can be chosen using any of the existing heuristics that compute a score for each column.
Using this principle and considering several good candidates for each relocation, we are less likely
25
to make a mistake. Here, we implement the TH-L with the H heuristic. The algorithm is presented
in Algorithm 2. In our experiment, we set L = 2. Note that TH results in less or the same number
of relocations compared to H since the path of H is included in the tree of TH-L.
Algorithm 2 Tree Heuristic
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
procedure [ZTH ] = TreeHeuristic(B, L)
ZTH ← 0
while B is not empty do
Retrieval:
if target container n is on top of its column then
Retrieve n form B
Relocation:
else
r ← topmost blocking container
C1 ← sort arg increasingly {min(ci )| min(ci ) > r}
ci ∈{1,...,C}\c
11:
C2 ← sort arg decreasingly {min(ci )| min(ci ) < r}
12:
C3 ← [C1 , C2 ]
S ← C3 [1 : L]
ZTH = ZTH + min{TreeHeuristic (B where r moves to column s,L)}
13:
14:
end if
16:
end while
17: end procedure
15:
ci ∈{1,...,C}\c
s∈S
Results for the four heuristics are summarized in Table 2. First, H is optimal in most instances
(87%). This is one of the main reasons that heuristic H was chosen as an upper bound in the A∗
method. Second, TH-2 is indeed closer to optimality than H, KH and LA-5, in distribution and in
terms of the average performance ratio.
Finally, notice that the percentages for LA-5 do not sum to 100% since there are cases for which
LA-5 is better than A∗ , which is because LA-5 considers “repositioning moves”, i.e., it solves for
the unrestricted CRP. As a result, for 0.5% of the 100,000 instances, LA-5 solves the instance with
one relocation less than A∗ .
In our last experiment for complete information, we study how the parameter L affects the
performance of the TH-L heuristic. We consider the same 100,000 instances and solve them with
L varying from 1 to 6, and we record their performance ratio. Notice that TH with L=6 considers
all possibilities for the blocking container therefore it gives the same solution as the A∗ without the
use of bounds. Thus PR(TH-6) = 0. Further note that L = 1 gives heuristic H.
The main observation is that the marginal gain of branching is maximum when we increase L
from 1 to 2. By considering two promising columns instead of one, TH-2 finds better solutions
26
L
PR (TH-L)
1
2
3
4
5
6
1.44%
0.44%
0.27%
0.20%
0.16%
0%
Table 3: Effect of parameter L on the performance of heuristic TH-L
for most instances where H was not optimal. Note that increasing L (considering more candidate columns) will further improve the solution; however the gain from more branching is small
considering the exponentially increasing cost of computation.
5.2
Experimental results for the CRP with Incomplete Information
Number of nodes and computation time for ASA∗ . Stochastic optimization methods that
are based on enumerating scenarios are usually computationally expensive. The ASA∗ algorithm,
however, is fast and tractable due to the use of sampling and pruning that allow for suppressing
many nodes in the tree. More importantly, the ASA∗ is tunable in the sense that one can set ǫ and
δ to change the number of nodes and thereby solve an instance within a desired amount of time,
and yet ensure that the loss from using large ǫ or small δ is bounded (as shown in propositions
4.1-4.3).
Figure 11a shows the cumulative distribution of the number of nodes (after pruning) for the
tree up to t∗ =11, for a bay with 7 columns, 4 tiers, 3 containers per column, and |K| = 11 (Recall
that t∗ is the time step at which the set of unknown containers become known and |K| is the number
of containers that are initially known). We implement ASA∗ with δ = 0.5 and ǫ = 0.05. It can be
seen that for half of the instances, the tree up to t∗ has 100 or fewer nodes. Also, about 90% of
instances have 200 or fewer nodes. Note that the nodes of the trees that are constructed after time
t∗ for solving instances corresponding to different scenarios do not have too much of an effect on
computation time because the A∗ algorithm is very fast. Figure 11b shows the cumulative density
of the average computation time. It can be seen that half of the instances are solved in less than
two minutes and 90% of the instances are solved in 15 minutes or less.
Although the ASA∗ algorithm is fast for medium-sized bays, it might not be tractable for large
bays. We next introduce a myopic heuristic and use both ASA∗ and the heuristic to provide several
insights on the value of information.
Myopic Heuristic. This is an extension of the heuristic H that we explained in Section 2.
Similar to the setting that we explained in Section 4, the problem is to retrieve N containers from
a bay with C column and P tiers. In the incomplete information setting, the departure order of
containers {1, 2, . . . , |K|} are known at time zero; the departure order of containers {|K|, |K|+1, . . . , N}
become known at time step t∗ > 0.
Suppose n is the target container located in column c. If n is not blocked by any container, it
is retrieved without any relocation. Otherwise, let r be the topmost blocking container in column
27
(a) ASA∗ implemented with δ = 0.5 and ǫ = 0.05
(b) Experiments were done on desktop computer with
Intel i73770s processor (3.90 GHz) and 16Gb RAM.
Figure 11: Tractability of the ASA∗ algorithm
c, and c∗ be the column where r should be relocated. The myopic heuristic determines c∗ using the
same rules as the H heuristic explained in Section 2, except that we assign an index of N + 1 to all
unknown containers and we set min(ci ) = N + 2 if ci is empty. Ties are broken arbitrarily. In the
following experiments we study myopic heuristic and compare its performance to that of ASA*.
Value of information. To study the effect of the level of information initially available (|K|), we
take 1000 random instances of a bay with 7 columns, 4 tiers, N = 21 containers, and 3 containers
per column. We solve each instance with 6 levels of information: |K| = ⌈0.25N⌉, ⌈0.375N⌉, ⌈0.5N⌉,
⌈0.625N⌉, ⌈0.75N⌉, and ⌈0.9N⌉, using the ASA∗ algorithm and the myopic heuristic (⌈x⌉ is the
smallest integer larger than x). For all cases, we fix t∗ at ⌈0.25N⌉ + 1 to ensure that at every timestep before t∗ , at least one container is known. For each of the 12 cases, we compute the average
of relocations over the 1000 instances. We then compare the relative gap of each case with the
E[zASA∗ ]−E[zopt ]
E[zMH ]−E[zopt ]
average relocation for the CRP with full information:
and
, where zMH is
E[zopt ]
E[zopt ]
the number of relocations obtained by the myopic heuristic.
Figure 12a shows the relative gap in relocations for different levels of information. With the
ASA∗ algorithm (implemented with δ = 0.5 and ǫ = 0.05), the gap is about 8% when 25% of the
containers (6 containers) are known at time zero. The gap reduces to 3% when half of the containers
are initially known and is almost zero when 90% of containers are known at time zero. The same
behaviour can be observed for the heuristic. In both cases, the marginal value of information
becomes smaller when more information is available. This is more significant for the heuristic. For
example, when the level of information increases from 25% to 50%, there is a significant drop in
the gap; then the gap decreases more slowly and approaches zero at 100% information. Note that
28
with 100% information, the myopic heuristic is the same as the H heuristic.
To get an insight into the value of information for the myopic heuristic, recall that it behaves
similar to the H heuristic as long as min(ci ) for all columns are known. Since all unknown containers
have larger indices than the known containers, knowing at least one container in each column is
sufficient to obtain the same min(ci ) as in the heuristic H. Thus, after some point, having more
information does not have much of effect on the number of relocations when we use the myopic
heuristic.
Figure 12b shows how the gap between the myopic heuristic and ASA∗ shrinks as the level of
information increases. When 25% of the containers are initially known, using the myopic heuristic
results in 12% more relocations on average (compared to ASA∗ ). This gap drops to less than 2%
when all containers are known. Note that when more than 50% of the containers are initially
known, the solution provided by the myopic heuristic is reasonably close to the solution of ASA∗
(less than 5%). Therefore, the myopic heuristic can be used in practice as it is easy to implement
and efficient.
(a) Value of information: relative gap in relocations text (b) Comparing the myopic heuristic and the ASA∗ algowith the case of full information
rithm: relative gap in relocations for different levels of
information
Figure 12: Experiment results for 7×4 bay
It is interesting to see the effect of different levels of information on the number of relocations in
larger bays. Figure 13 shows the average relative gap in relocations for bays of different sizes. For
each bay size, the average relative gap is computed for 100,000 instances, using the myopic heuristic.
The most important observation is that the ratio of E[zMH ] and E[zH ] does not constantly increase
with bay size. And, for each information level, this ratio converges to a constant as the bay gets
larger. Two other observations can be made from Figure 13. First, the converging ratio for large
bays drops fast as the level of information increases. For example, when 25% of containers are
29
initially known, the ratio converges to 1.4. This number drops to 1.03 when the level of information
is 75%. Second, the rate of convergence is much faster when more information is available. For
example, when 75% or more of the containers are initially known, the ratio is almost constant
independent of the bay size. When 25% of the containers are known, the ratio converges to 1.4 for
bays with 40 or more columns. Last, note that the asymptotic behaviour of the myopic heuristic
in Figure 13, is as an upper bound on the performance of the ASA∗ algorithm.
What is shown in Figure 13 provides useful insights for port operators. Considering the gain
(fewer relocations) at each level of information, port operators can design appointment systems
that capture this gain by offering faster or discounted service to the customers who provide their
arrival time in advance.
Figure 13: Asymptotic behaviour of myopic heuristic:
larger.
E[zMH ]
E[zH ]
approaches a constant as the bay gets
Comparing the myopic heuristic with existing heuristic. To the best of our knowledge,
only a few recent papers studied CRP with incomplete information, and developed heuristic algorithms for this problem [10, 21]. Here we compare our myopic heuristic with the RDH (revised
difference heuristic of [21]) and show that myopic outperforms the RDH. Zhao et al. [21] show the
percentage savings of RDH over a nearest relocation strategy for different bay sizes and for different
amounts of information initially available. The greatest savings is about 50% and is realized for a
bay with 12 columns and 7 maximum height, with 50 or more containers initially known (see Figure
13 of [21]). The total number of containers is not indicated for this simulation, so we assume the
bay is full with 72 containers. (With 50 known containers, the level of information is about 70%).
For a bay with 12 columns and 3 containers per column, the maximum savings is about 40% and
is realized when 22 or more containers are known (i.e., when the information level is about 60%).
30
Figure 14: Percentage of saving of myopic heuristic over nearest relocation strategy.
In Figure 14, we show the percentage savings of the myopic heuristic over the nearest location
strategy for bays with different numbers of columns, with 4 tiers and 3 containers per column.
The savings are shown for 4 different levels of information. It can be seen that when level of
information is 50%, the saving is about 48%, which is greater than the saving for RDH when even
more information (about 70%) is available. When the level of information is 25%, the savings of
the myopic heuristic is about 35%, whereas for the same setting, RDH results in less than 25%
savings over the nearest relocation strategy.
6
Conclusion
Managing the relocation moves is one of the main challenges in the storage yard of container
terminals and has a direct effect on the costs and efficiency of yard operations. The Container
Relocation Problem (CRP) is notorious for its computational intractability and most research
studies have designed heuristics in order to solve the problem, particularly for large bays.
In this paper, we revisited a fast and efficient method, the A∗ algorithm that provides optimal
solution for the CRP. We proved several properties of the method as applied to the CRP and
assessed its performance and computational tractability through experiments. We also studied the
asymptotic behavior of optimal solutions of the CRP and showed that the ratio of the optimal
number of relocations and a simple counting lower bound converges to 1 as the bay gets very
large. This gives strong evidence that CRP is easier in large bays. Moreover, we showed through
experiments that the H heuristic (introduced in [4]) has the same behavior. An important result
of our experiments is that it is recommended to apply the heuristic H to large blocks of containers
(several bays) in order to minimize the loss from using a heuristic rather than optimization methods
31
(that are computationally expensive). We also presented an improvement over H heuristic, and show
through experiments that it outperforms existing heuristics.
Further, we extended our study to the CRP with incomplete information, which is an important
case in practice because the retrieval order of containers are usually not known far in advance.
We introduced a two-stage approximate stochastic optimization framework (ASA∗ ) that optimizes
the expected number of relocations given a probabilistic distribution on the departure order of
containers. We also presented a myopic heuristic that is very fast and efficient for the CRP with
incomplete information. We used ASA∗ and the heuristic to study the value of information for
different bay sizes and showed that the relative gap between ASA∗ and the heuristic shrinks as
the level of information increases. In fact, when the available information is more than 50%, the
myopic heuristic is very efficient and can be used in practice. We also showed that the ratio of the
number of relocations between heuristic H and the myopic heuristic converges to a constant fast
as bays get larger, which again implies that the loss from using the myopic heuristic is minimized
when applied to large bays.
Our paper opens directions to many further studies in both complete information setting and
the incomplete one. A very interesting problem in the complete information case is to generalize
our model to a setting where stacking and retrieving processes overlap in time. In such a setting,
the goal would be to jointly optimize the number of relocation for stacking and retrieving. We
believe that one of the major questions of this problem is how to prioritize between stacking and
retrieving at a given time. In the incomplete information case, one can easily generalize our twostage stochastic optimization problem to a multi-stage setting where ASA* could be extended.
However, as we increase the number of stages, the computation time grows rapidly. Additional
assumptions and/or other efficient heuristics would be needed to ensure a reasonable running time.
Finally, another interesting problem is to study the case where no information about any containers
is known in advance. Intuitively, it appears that the policy of placing the container on the emptiest
column should be the optimal solution; however this remains to be proven.
Other storage systems such as steel plate stacking and warehousing systems face the relocation
problem with complete and incomplete information (see Kim et al. [8], Zäpfel and Wasner [19] for
the former, and Chen et al. [6] for the latter). We believe that the frameworks proposed in this
paper could also be applicable to those problems.
References
[1] Akyüz, M. H. and Lee, C.-Y. (2014). A mathematical formulation and efficient heuristics for
the dynamic container relocation problem. Naval Research Logistics (NRL), 61(2):101–118.
[2] Borjian, S., Manshadi, V. H., Barnhart, C., and Jaillet, P. (2015). Managing relocation and
delay in container terminals with flexible service policies. ArXiv: 1503.01535.
32
[3] Caserta, M., Schwarze, S., and Voß, S. (2009). A new binary description of the blocks relocation
problem and benefits in a look ahead heuristic. Evolutionary Computation in Combinatorial
Optimization, 5482:37–48.
[4] Caserta, M., Schwarze, S., and Voß, S. (2012). A mathematical formulation and complexity
considerations for the blocks relocation problem. European Journal of Operational Research,
219(1):96–104.
[5] Caserta, M., Voß, S., and Sniedovich, M. (2011). Applying the corridor method to a blocks
relocation problem. OR Spectrum, 33(4):915–929.
[6] Chen, L., Langevin, A., and Riopel, D. (2011). A tabu search algorithm for the relocation
problem in a warehousing system. International Journal of Production Economics, 129(1):147–
156.
[7] Forster, F. and Bortfeldt, A. (2012). A tree search procedure for the container relocation
problem. Computers & Operations Research, 39(2):299–309.
[8] Kim, B.-I., Koo, J., and Sambhajirao, H. P. (2011). A simplified steel plate stacking problem.
International Journal of Production Research, 49(17):5133–5151.
[9] Kim, K. H. and Hong, G.-P. (2006). A heuristic rule for relocating blocks. Computers &
Operations Research, 33(4):940–954.
[10] Ku, D. (2014). Rehandling problem of pickup containers under truck appointment system.
In Computational Collective Intelligence. Technologies and Applications, volume 8733 of Lecture
Notes in Computer Science, pages 272–281. Springer International Publishing.
[11] Lee, Y. and Lee, Y.-J. (2010). A heuristic for retrieving containers from a yard. Computers &
Operations Research, 37(6):1139–1147.
[12] Lehnfeld, J. and Knust, S. (2014). Loading, unloading and premarshalling of stacks in storage
areas: Survey and classification. European Journal of Operational Research, 239(2):297–312.
[13] Olsen, M. and Gross, A. (2014). Average case analysis of blocks relocation heuristics. 5th
International Conference ICCL Proceedings, 8760:81–92.
[14] Petering, M. E. and Hussein, M. I. (2013). A new mixed integer program and extended lookahead heuristic algorithm for the block relocation problem. European Journal of Operational
Research, 231(1):120–130.
[15] Stahlbock, R. and Voß, S. (2008). Operations research at container terminals: a literature
update. OR Spectrum, 30(1):1–52.
33
[16] Tanaka, S. and Takii, K. (2014). A faster branch-and-bound algorithm for the block relocation
problem. 2014 IEEE International Conference on Automation Science and Engineering (CASE),
14:7–12.
[17] Ünlüyurt, T. and Aydın, C. (2012). Improved rehandling strategies for the container retrieval
process. Journal of Advanced Transportation, 46(4):378–393.
[18] Vondrák, J. M. J. (2001). The Probabilistic Method: Lecture Notes. (KAM-DIMATIA series.
Charles Univ).
[19] Zäpfel, G. and Wasner, M. (2006). Warehouse sequencing in the steel supply chain as a
generalized job shop model. International Journal of Production Economics, 104(2):482–501.
[20] Zhang, H., Guo, S., Zhu, W., Lim, A., and Cheang, B. (2010). An investigation of IDA*
algorithms for the container relocation problem. Proceedings of the 23rd International Conference
on Industrial Engineering and Other Applications of Applied Intelligent Systems (IEA/AIE), 3140.
[21] Zhao, W. and Goodchild, A. V. (2010). The impact of truck arrival information on container terminal rehandling. Transportation Research Part E: Logistics and Transportation Review, 46(3):327–343.
[22] Zhu, W., Qin, H., Lim, A., and Zhang, H. (2012). Iterative deepening A* algorithms for
the container relocation problem. IEEE Transactions on Automation Science and Engineering,
9(4):710–722.
34
Appendix
Proof of Proposition 2.2. Between level l and l + 1, one move has been operated (r moved to
column ci ). If r < min(ci ) (‘good” move) then r was a blocking container in Bl but not in Bl+1 , thus
S0 (Bl+1 ) = S0 (Bl ) − 1. If r > min(ci ), we have S0 (Bl+1 ) = S0 (Bl ). So S0 (Bl+1 ) = S0 (Bl ) − χ(r < min(ci ))
where χ is the indicator function. Therefore L0 (Bl+1 ) = S0 (Bl+1 ) + (l + 1) = S0 (Bl ) − χ(r < min(ci )) +
1 + l ≥ S0 (Bl ) + l = L0 (Bl ).
Proof of Proposition 2.4. We have Lp (B) − Lp−1 (B) =
negative function, the results follows.
P
r∈Rp (B) χ(r
> MM(Bp )). Since χ is a non
Proof of Proposition 2.5. Intuitively, this property comes from the fact that Sp (Bl+1 ) ≥ Sp (Bl ) −
1. Indeed, if this holds, we have
Lp (Bl+1 ) = Sp (Bl+1 ) + l + 1 ≥ Sp (Bl ) − 1 + l + 1 = Sp (B) + l = Lp (B)
So let us prove Sp (Bl+1 ) ≥ Sp (Bl ) − 1. Let r be the container that is relocated between Bl and Bl+1 .
We have two cases:
• If r can do a good move, r has only been counted once in Sp (Bl ) as it is a blocking container
in Bl and it might or might not contribute to Sp (Bl+1 ). Therefore Sp (Bl+1 ) ≥ Sp (Bl ) − 1.
• If r cannot do a good move, it has been counted twice in Sp (Bl ). But since it cannot do a
good move, it is necessarily a blocking container in Bl+1 so it counts at least once in Sp (Bl+1 ).
Thus Sp (Bl+1 ) ≥ Sp (Bl ) − 1.
This completes the proof.
Proof of Proposition 2.7. Using heuristic H, a container is relocated only if it is blocking. For
any blocking container, there exists at least one empty column which means there is a possible
good move. So each blocking container is relocated at most once, i.e, zH (B) = S0 (B). Since S0 (B) ≤
zopt (B) ≤ zH (B) = S0 (B) and we have equation (4).
Proof of Proposition 2.8. For simplicity of proofs we assume that n1 , the target container in
the bay is 1. If 1 is not blocked then it is retrieved. In that case, there are C containers left and
thus we use Proposition 2.7 (note that in that case S1 (B) = S0 (B)).
Then let us suppose 1 is blocked by some containers. If there is more than one blocking
container, all except the last one can be relocated to empty columns. So we come back to the case
where 1 is only blocked by container r, and there is no empty column.
Now we consider two cases: If r can do a good move and H relocates r to a good column and then
retrieves 1. We call B′ the new configuration. By Proposition 2.7, we have zH (B′ ) = S0 (B′ ) = S1 (B′ ).
Notice that zH (B) = zH (B′ ) + 1 and S1 (B) = S1 (B′ ) + 1 and the result follows.
35
Now suppose r cannot do a good move. This can only happen if r = C + 1 and all the other
columns have exactly one container. It is easy to check that in that case, zH (B) = S1 (B).
Proof of Proposition 2.9. First let us introduce some notations. Let B be the initial configuration and B′ be the configuration obtained after retrieving container 1 using heuristic H. We denote
R1 the set of containers blocking container 1 in B.
Now let us consider the case k = 2. Note that using Proposition 2.8 we have zH (B′ ) = S1 (B′ )
and S1 (B′ ) ≤ S0 (B′ ) + 1. There are four possibilities to consider:
• If |R1 | = 0, then it is easy to see that zH (B) = zH (B′ ) = zopt (B′ ) = zopt (B).
• If |R1 | = 1, we know that S0 (B′ ) ≤ S0 (B). Therefore zH (B) = 1 + zH (B′ ) = 1 + S1 (B′ ) ≤
1 + S0 (B′ ) + 1 ≤ S0 (B) + 2 ≤ zopt (B) + 2.
• If |R1 | > 2, then there are at least |R1 | − 2 columns that are empty in B. Therefore the |R1 | − 2
topmost containers blocking 1 find necessarily a ”good” column in B. Hence this case can be
reduced to the case where |R1 | = 2.
• If R1 = 2, then we claim that S1 (B′ ) ≤ S0 (B). If this is true, then zH (B) = zH (B′ ) + 2 =
S1 (B′ ) + 2 ≤ S0 (B) + 2 ≤ zopt (B) + 2. Now let us denote r1 and r2 the two containers blocking 1
in B.
– If r1 ≤ C or r2 ≤ C then S0 (B′ ) ≤ S0 (B) − 1 and therefore S1 (B′ ) ≤ S0 (B′ ) + 1 ≤ S0 (B).
– If r1 > C and r2 > C, then if one of them finds an empty column then again S0 (B′ ) ≤
S0 (B)− 1. Otherwise it means that all columns have one container and C + 1 and C + 2 are
on top of 1. In those two bays we can verify that zH (B) = 4 ≤ 2 + 2 = S0 (B) + 2 ≤ zopt + 2.
This concludes the case k = 2.
For 3 ≤ k ≤ C, the proof is simpler and works by induction. For k = 2, the previous paragraph
proves a stronger inequality. Let us suppose it is true for k − 1, and B has C + k containers. We
k(k−1)
have by induction hypothesis, zH (B′ ) ≤ S0 (B′ ) + 2 . We have three cases:
• If |R1 | = 0, then zH (B) = zH (B′ ) ≤ S0 (B′ ) +
k(k−1)
2
≤ S0 (B) +
k(k+1)
2
≤ zopt (B) +
k(k+1)
2 .
• If |R1 | > k, then there are at least (C − 1) − (C + k − (|R1 | + 1)) = |R1 | − k empty columns and
k(k−1)
therefore S0 (B′ ) ≤ S0 (B) + k − |R1 | which implies zH (B) = |R1 | + zH (B′ ) ≤ |R1 | + S0 (B′ ) + 2 ≤
k(k−1)
k(k+1)
k(k+1)
|R1 | + S0 (B) + k − |R1 | + 2 = S0 (B) + 2 ≤ zopt (B) + 2 .
• If 0 < |R1 | ≤ k, then since S0 (B′ ) ≤ S0 (B) we have zH (B) = |R1 | + zH (B′ ) ≤ |R1 | + S0 (B′ ) +
k(k−1)
k(k+1)
k(k+1)
k + S0 (B) + 2 = S0 (B) + 2 ≤ zopt (B) + 2 .
k(k−1)
2
≤
which concludes the proof.
36
Proof of Proposition 3.1. Let Si0 (BC ) be the number of blocking containers in column i. By the
linearity of expectation, we have
C
C
h
i
X
X
i
EC [S0 (BC )] = EC
S0 (BC ) =
E Si0 (BC ) = αh × C,
i=1
i=1
h
i
where αh = E S10 (BC ) . The last equality comes from the fact that each column is identically
distributed.
P
Now let us compute αh . By definition, αh = h−1
k=1 kpk,h . Let us start with p0,h . Given h randomly
chosen containers, the probability that they are placed in the first column in a descending order
(meaning there is no blocking container) is p0,h = h!1 . Now for 1 ≤ k ≤ h − 1, we compute pk,h by
recursion (conditioning on the position of the smallest container). Let n∗ be the smallest container
among these h containers. n∗ is located in jth topmost tier with probability 1h . Conditioned on that
event, it means that (j − 1) containers are blocking n∗ . So if j > k + 1, then the event of having
k blocking container is not possible. Therefore for j ≤ k + 1 there should be (k − (j − 1)) blocking
containers below the jth topmost tier and this happens with probability pk−( j−1),h−j (since there are
h − j tier below the jth topmost tier). Summing over all j (possible positions of n∗ ), we have
pk,h =
k
X
1
j=1
h
pk−j+1,h−j ,
which completes the proof
Proof of Theorem 3.3. Since for all configurations BC , zopt (BC ) ≥ S0 (BC ) then
Moreover, we have:
h
i
h
i
EC zopt (BC ) − EC [S0 (BC )]
EC zopt (BC )
=1+
EC [S0 (BC )]
EC [S0 (BC )]
i
1 h
=1+
EC zopt (BC ) − αh C
αh C
g(C)
=1+
αh C
EC [zopt (BC )]
EC [S0 (BC )]
≥ 1.
(14)
where
h
i
g(C) = EC zopt (BC ) − αh C.
(15)
h
i
In Lemma 6.1, we study how EC zopt (BC ) evolves and we show that it increases almost linearly in
α which shows that the function g(.) is essentially bounded.
Lemma 6.1. Let zopt (BC ) be the optimal number of relocations for configuration BC , then there
exists a constant θ (defined in equation (37)) such that:
h
i
h
i
EC+1 zopt (BC+1 ) ≤ EC zopt (BC ) + αh + h(P − 1)(C + 1)e−θ(C+1) , ∀C ≥ h + 1
(16)
37
Using Lemma 6.1, we have for all C ≥ h + 1:
h
i
h
i
EC+1 zopt (BC+1 ) ≤ EC zopt (BC ) + αh + h(P − 1)(C + 1)e−θ(C+1)
h
i
h
i
=⇒ EC+1 zopt (BC+1 ) − αh (C + 1) ≤ EC zopt (BC ) − αh C + h(P − 1)(C + 1)e−θ(C+1)
=⇒ g(C + 1) ≤ g(C) + h(P − 1)(C + 1)e−θ(C+1)
C
∞
X
X
−θi
=⇒ g(C) ≤ g(h + 1) + h(P − 1)
ie
≤ g(h + 1) + h(P − 1)
ie−θi
i=1
i=h+2
=⇒ g(C) ≤ g(h + 1) +
eθ h(P
(eθ
− 1)
− 1)
2
= K′ .
(17)
Therefore using equations (14) and (17), we have
h
i
EC zopt (BC )
K
≤ 1 + = f (C),
EC [S0 (BC )]
C
(18)
where
K=
K′
α
g(h + 1) +
=
α
which completes the proof of the theorem.
eθ h(P−1)
(eθ −1)
2
(19)
Proof of Lemma 6.1. Now we need to prove equation (16). Define a column to be “special” if
all its containers are not smaller than ω = (h − 1)(C + 1) + 1. Now let us consider the following event:
n
o
Ω = The bay with C + 1 columns has at least one “special” column .
(20)
The intuition is the following: the probability of having a “special” column grows quickly to 1
as a function of C implying that the event Ω happens with high probability. Now, conditioned on
Ω, we more easily express the difference between bays of size C + 1 and C in the following way. We
claim that
h
i
h
i
EC+1 zopt (BC+1 )|Ω ≤ EC zopt (BC ) + αh ,
(21)
Let BC+1 be a given bay of size C + 1 that verifies Ω. Since columns in bays can be interchanged,
we suppose that a “special” column is the first (leftmost) column of the bay. We also denote
n1 , n2 , . . . , nh the containers of the first column. We know that n1 , n2 , . . . , nh ≥ ω and n1 , n2 ,
. . . , nh . Finally let BC be the bay BC+1 without its first column (see Figure 15).
First we prove that
n1
(22)
zopt (BC+1 ) ≤ zopt (BC ) + S0 . . . .
nh
38
Figure 15: Bay decomposition of BC+1 (The part on the right has C columns)
To prove equation (22), we construct a feasible sequence σ for the bay of size C + 1 for which the
number of relocations is equal to the right side of equation (22). Let σopt (BC ) the optimal sequence
for BC , t′ be the first time step when the target container in σopt (BC ) is larger than min{n1 , n2 , . . . , nh }
and B′C be the bay obtained at t′ using σopt (BC ). Let the first t′ − 1 moves of σ be the first t′ − 1
moves of σopt (BC ). Note that B′C has at most C − h containers due to the choice of ω. By Proposition
2.7, the number of relocations performed by σopt (BC ) from t′ until the end is S0 (B′C ). Therefore
zopt (BC ) = # relocations up to t′ done by zopt (BC ) + S0 (B′C ).
(23)
n1
′
′
′
After t , we run heuristic H on BC+1 = . . . ∪ BC .
nh
We claim that zσ is exactly the right side of equation (22). There are at most C containers in
′
BC+1 , therefore using Proposition 2.7, we know that if we apply the heuristic H to this configuration,
n1
then the number of relocations done by H is S0 (B′C+1 ) = S0 (B′C ) + S0 . . . . Therefore
nh
n1
zopt (BC+1 ) ≤ zσ (BC+1 ) = # relocations up to t′ done by zopt (BC ) + S0 (B′C ) + S0 . . . ,
nh
and we have
n1
zopt (BC+1 ) ≤ zopt (BC ) + S0 . . . ,
nh
(24)
which proves equation (22).
Now we can take the expectation from both sides of equation (21) over a uniform distribution
of the rest of the hC containers that are not in the first column. We claim that the first term on
39
h
i
the right hand-side of equation (24) is exactly EC zopt (B) . For any configuration that appears in
BC , we can map it to a unique configuration where all containers are between 1 and hC. Thus,
n1
n1
h
i
(25)
EC zopt (BC+1 ) . . . ≤ EC zopt (BC ) + S0 . . . .
nh
nh
Next, we take the expectation of both sides of equation (25) over possible first columns, which
is a “special” column. Now notice that if BC+1 is generated uniformly in the sets of bays of size
C + 1, then conditioned on Ω, the probability of having a certain column [n1 , . . . , nh ]T is identical
for any n1 , . . . , nh ≥ ω and it is given by
n1
(C − h)!
P . . . Ω =
.
(26)
C!
nh
Moreover, if we withdraw this first column from BC+1 and look at the last C columns (after renumbering the containers so that the numbers lie between 1 and hC), then this bay is also uniformly
distributed in the sets of bays with C columns. Therefore we can write:
n
n
1
1
X
h
i
EC zopt (BC+1 ) . . . , Ω × P . . . Ω
(27)
EC+1 zopt (BC+1 )|Ω =
(n1 ,...,nh )
nh
nh
n, n
i
j
ni ≥ ω
X
E z (B )
=
C opt C+1
(n1 ,...,nh )
ni , n j
ni ≥ ω
≤ EC
h
n1
n1
. . . × P . . .
nh
nh
n1
n1
n1
X
i X
S0 . . . × P . . .
P . . . Ω +
zopt (BC )
(n1 ,...,nh )
(n1 ,...,nh ) n
nh
nh
h
n, n
n, n
i
i
j
≤ EC
Ω
j
n1
n1
X
i
S0 . . . × P . . .
zopt (BC ) +
(n1 ,...,nh )
nh
nh
n, n
i
(28)
ni ≥ ω
ni ≥ ω
h
Ω
Ω
(29)
j
ni ≥ ω
The equality between 27 and 28 comes from the fact that if we know that BC+1 has a “special”
col
n1
P
P . . . Ω =
umn, then we do not need to condition on Ω. Equation 29 uses the fact that
n1 , ..., nh
ni ≥ ω
nh
1.
40
Note that, given any (n1 , . . . , nh ) such that ni , n j , we have
n1
E S0 . . . = αh ,
nh
when the expectation is over a random order of (n1 , . . . , nh ). This is true regardless of the set
(n1 , . . . , nh ) that is drawn from (See Remark 3.2). This implies that the second term in the right
hand side of equation (29) is equal to αh ; Therefore, we get equation (21).
Lemma 6.2 states that the event Ω has a probability that increases exponentially fast to 1 as a
function of C.
Lemma 6.2. Let Ω be the event defined by equation (20), then there exists a constant θ > 0 such
that
P(Ω) ≤ e−θ(C+1)
(30)
where θ is given by equation (37).
h
i
Now we want to focus on the event Ω. We give an upper bound on EC+1 zopt (BC+1 )|Ω . For any
configuration, in order to retrieve one container, we need at most P − 1 relocations (since at most
P − 1 containers are blocking it), thus for any configuration, the optimal number of relocations is
at most P − 1 times the number of containers (h(C + 1)) which gives us h(P − 1)(C + 1) as an upper
bound on the optimal number of relocations. We use this universal bound to get
h
i
EC+1 zopt (BC+1 )|Ω ≤ h(P − 1)(C + 1).
(31)
Finally using Lemma 6.2, we have
h
h
h
i
i
i
EC+1 zopt (BC+1 ) = EC+1 zopt (BC+1 )|Ω P(Ω) + EC+1 zopt (BC+1 )|Ω P(Ω)
h
i
h
i
≤ EC+1 zopt (BC+1 )|Ω + EC+1 zopt (BC+1 )|Ω × e−θ(C+1)
h
i
≤ EC zopt (BC ) + αh + h(P − 1)(C + 1)e−θ(C+1) ,
which completes the proof of Lemma 6.1.
Proof of Lemma 6.2. Recall that
Ω=
n
The bay with C + 1 columns has at least one “special” column
o
.
We know that each bay of size C + 1 can be mapped to a permutation π of Sh(C+1) taken
uniformly at random. Let q(.) be the function from Sh(C+1) to R+ defined by
q : π 7−→ number of “special” columns in the resulting bay of π.
41
Note that
P Ω = P q(π) = 0 .
First we compute the expected value of q(.)
C+1
X
EC+1 [q] = EC+1
χ (ci is a “special” column)
i=1
EC+1 [q] = (C + 1) × P ({c1 is a “special” column}) ,
(32)
where we use linearity of expectation and the fact that columns are identically distributed.
A simple counting implies that 1 :
(C + 1)[(C + 1) − 1] . . . [(C + 1) − h + 1]
h(C + 1)[h(C + 1) − 1] . . . [h(C + 1) − h + 1]
!h
(C + 1) − h + 1
≥
h(C + 1)
!h
2
≥
,
h(h + 1)
P ({c1 is a “special” column}) =
(33)
where we use C + 1 ≥ h + 1 to show the last inequality.
Therefore we know that
!h
2
.
EC+1 [q] ≥ (C + 1) ×
h(h + 1)
(34)
We claim that q(.) is well concentrated around its mean. In order to do so, we prove that q(.)
is 1-Lipschitz. Define ρ the distance between two permutations π1 , π2 ∈ Sh(C+1) as ρ(π1 , π2 ) =
| {i ∈ [h(C + 1)] : π1 (i) , π2 (i)} |. We want to prove that
|q(π1 ) − q(π2 )| ≤ ρ(π1 , π2 ), ∀ (π1 , π2 ) ∈ Sh(C+1) .
(35)
Let π1 , π2 ∈ Sh(C+1) . Let us first consider the case where ρ(π1 , π2 ) = 2. (Notice that if ρ(π1 , π2 ) , 0
then ρ(π1 , π2 ) ≥ 2). In that case, we have i, j ∈ {1, . . . , n} such that π1 (i) = π2 (j) and π1 (j) = π2 (i).
Let B1 and B2 be the configurations generated by π1 and π2 . Having ρ(π1 , π2 ) = 2 corresponds to
the fact that if we swap 2 containers in B1 , we get B2 and we denote those containers a = π1 (i) and
b = π1 (j). We have three cases:
• a and b are both in “special” columns in B1 . In this case, swapping them will not change
anything since both their new columns in B2 will also be “special” and hence |q(π1 )−q(π2 )| = 0.
1
1
Notice that when C → ∞, the probability is equivalent to ( )h which would guarantee a faster convergence rate.
h
42
• a and b are both in columns that are not “special” columns in B1 . If a, b ≥ ω or a, b < ω then
we do not create any new special column in B2 . Now suppose that a ≥ ω and b < ω, then the
column of a in B2 might be a “special” column, but the column of b in B2 cannot be “special”.
Therefore in that case, |q(π1 ) − q(π2 )| ≤ 1.
• a is in a “special” column in B1 but b is not. Now we know that a ≥ ω. If b < ω then
the column of b in B2 cannot be “special” but the column of a might be and in that case
|q(π1 ) − q(π2 )| ≤ 1. If b ≥ ω, then the column of b in B2 is “special” and the column of a in B2
is not “special” which gives us |q(π1 ) − q(π2 )| = 0. Note that the proof is identical if b is in a
“special” column in B1 but a is not.
So far we have shown that
If ρ(π1 , π2 ) = 2, then |q(π1 ) − q(π2 )| ≤ 1.
(36)
Now we suppose that ρ(π1 , π2 ) = k where 2 ≤ k ≤ h(C + 1). Note that we can construct a
sequence of permutations (π′1 , π′2 , . . . , π′k ) such that π′1 = π1 , π′k = π2 and ρ(π′i , π′i+1 ) = 2.
Now using this fact and equation (36),
|q(π1 ) − q(π2 )| =
k−1
X
i=1
q(π′i ) −
q(π′i+1 )
≤
k−1
X
i=1
|q(π′i )
−
q(π′i+1 )|
≤
k−1
X
i=1
1 = k − 1 ≤ k = ρ(π1 , π2 ),
which proves that q(.) is 1-Lipschitz.
Now we use Theorem 8.3.3 of Matoušek and Vondrák [18] which states that P q ≤ EC+1 [q] − t ≤
t2
e− 8h(C+1) and apply it with t = EC+1 [q] and equation (34) to get
2
(EC+1 [q])
P q = 0 = P q ≤ EC+1 [q] − EC+1 [q] ≤ e− 8h(C+1) ≤ e−θ(C+1) ,
where
2
1
θ=
8h h(h + 1)
!2h
> 0,
(37)
which concludes the proof.
Proof of Corollary 9. Using that S0 (BC ) ≤ zgen (BC ) ≤ zopt (BC ), we have 1 ≤
EC [zopt (BC )]
EC [S0 (BC )] ≤ f (C). Finally we know that EC [S0 (BC )] = αh C.
EC [zgen (BC )]
EC [S0 (BC )]
≤
Proof of Proposition 4.1. We can write E[e1 ] as follows:
E[e1 ] = (E[E[zpASA ] − z̄pASA ]) − (E[E[zp∗ ] − z̄p∗ ]) + E[z̄pASA − z̄p∗ ] ≤ E[|E[zpASA ] − z̄pASA |] + E[|E[zp∗ ] − z̄p∗ |].(38)
43
The above inequality holds because ASA∗ chose pASA , and thus we have z̄pASA < z̄p∗ . Moreover,
we can compute E[|E[zpASA ]− z̄pASA |] and E[|E[zp∗ ]− z̄p∗ |] using their CDF. We denote |E[zpASA ]− z̄pASA |
by ∆. Note that 0 ≤ ∆ ≤ rmax and inequality (11) gives a bound on its CDF, when estimating z̄pASA
with S samples. For E[∆], we have:
Z
rmax
E[∆] =
Z0 rmax
1 − F∆ (x)dx
(39)
−2Sx2
)dx
r2 max
0
Z rmax
Z ∞
s
ǫ x22
π
ǫ x22
= 2
( ) δ dx ≤ 2
( ) δ dx = δ
ǫ .
2
2
0
0
− ln( )
2
≤
2 exp(
(40)
(41)
We can compute E[|E[zp∗ ] − z̄p∗ |] in a similar way.
The first equality in (41) is obtained by substituting S with the expression in (12). Using the
bound on E[∆] and inequality (38), we can compute E[e1 ] as follows:
s
E[e1 ] ≤ 2δ
π
ǫ .
− ln( )
2
(42)
Proof of Proposition 4.2. For an arbitrary initial bay B, let p̂ be the path with the minimum
estimated upper bound. Notice that pruning p̃ may be a mistake if the true mean of Lp̃ is less
than the true mean of Up̂ , i.e., E[Lp̃ ] < E[Up̂ ]; in such a case, e2 ≤ E[Up̂ ] − E[zp̃ ] ≤ E[Up̂ ] − E[Lp̃ ].
Otherwise e2 is zero; thus we have e2 ≤ (E[Up̂ ] − E[Lp̃ ])+ .
Let us denote Lp̃ − E[Lp̃ ] and E[Up̂ ] − Up̂ by xL and xU , respectively. Also, let d be Lp̃ − Up̂ .
The true and estimated values of the upper bound and lower bound, and the loss are illustrated in
Figure 7; e2 is shown by the thick line segment. For any bay with an arbitrary initial configuration,
we can bound e2 as follows:
+
e2 ≤ (E[Up̂ ] − Up̂ ) + (Up̂ − Lp̃ ) + (Lp̃ − E[Lp̃ ]) = (xU − d + xL )+ .
(43)
The expected loss, E[e2 ], can be bounded as shown in (46).
E[e2 ] = E[(xL + xU − d)+ ] ≤ E[(xL + xU )+ ] ≤ E[|xL + xU |]
≤ E[|xL | + |xU |] = E[|xL |] + E[|xU |]
s
π
≤ 2δ
ǫ .
− ln( )
2
44
(44)
(45)
(46)
The first inequality in 44 holds because d is always positive. Inequality 45 results from the
triangular inequality, and the last inequality is obtained by replacing E[|xL |] and E[|xU |] by the
expression in (41).
Proof of Proposition 4.3. For an arbitrary initial bay B, let p̂max be the path that maximizes
E[Up̂ ] among p̂1 , . . . , p̂m . We bound e3 (the expected loss due to prununig) as follows:
e3 ≤ P(mistake at t1 ∪ · · · ∪ mistake at tm )(E[zp̂max ] − E[zp∗ ])
(47)
≤ m P(mistake at t1 )(E[Up̂max ] − 0)
(48)
= m P(mistake at t1 )(E[Up̂max ] − Up̂max + Up̂max )
(49)
≤ m P(mistake at t1 )(|xU | +
(50)
max
Up̂ ).
Notice that inequality 48 is obtained by using the union bound. Moreover, we replace the
probability of mistakes at each step by the maximum probability. Without loss of generality we
assume that maximum probability is at t1 . Also note that we do not know which path is p̂max
because we have not observed E[Up̂1 ], . . . , E[Up̂m ]. Nevertheless, regardless of the path, we can
max
replace E[Up̂max ] − Up̂max with |xU | (by definition) to obtain (50). Also, we can bound Up̂max by Up̂ .
Now we compute P(mistake at t1 ). Let us denote |xL | and |xU | by x and y, respectively. Also,
let w = x + y and fW (w) be the PDF of w. Notice that the probability of making a mistake depends
on the value of d; the maximum probability corresponds to the stage with the smallest value of d
and can be computed as follows:
Z ∞
Z ∞Z w
P(mistake at t1 ) = P(x + y − dmin > 0) =
fW (w)dw =
fX (w − y) fY (y) dy dw.
(51)
dmin
dmin
0
fX (x) and fY (y) can be obtained from (11), and the above integral can be bounded as follows:
Z
∞
dmin
Z
0
w
2)
− ln( ǫ ) 2
ǫ (y2 +(w−y)
2
δ2
y(w
−
y)(
)
4
2
δ2
dmin 0
r
p
2
2
−ln( 2ǫ )π ǫ dmin
dmin −ln( 2ǫ )
ǫ dmin
d
min
( ) δ2 +
)
( ) 2δ2 Erf(
√
2
δ
2
2
2δ
r
2
−ln( 2ǫ )π ǫ d2min
ǫ dmin
d
min
2
( ) δ +
( ) 2δ2 .
2
δ
2
2
E[|xU |], it follows that:
Z
fX (w − y) fY (y) =
=
≤
From 50, 52, and expression 41 for
∞
Z
w
(52)
v
t
E[e3 ] ≤ m
2
ǫ
−ln( )π ǫ d2min s π
max
2
,
U
( ) 2δ2 δ
+
p̂
ǫ
2
2
− ln( )
2
= max{Up̂1 , . . . , Up̂m }, and dmin = min{d1 , . . . , dm }.
ǫ dmin dmin
( ) δ2 +
2
δ
max
where Up̂
45
(53)
46
| 8cs.DS
|
arXiv:1611.01415v2 [math.AC] 17 Nov 2016
Explicit Implicit Function Theorem for All Fields
YINING HU
CNRS, Institut de Mathématiques de Jussieu-PRG
Université Pierre et Marie Curie, Case 247
4 Place Jussieu
F-75252 Paris Cedex 05 (France)
[email protected]
Abstract
We give an explicit implicit function theorem for formal power series that is valid for
all fields, which implies in particular Lagrange inversion formula and and Flajolet-Soria
coefficient extraction formula known for fields of characteristic 0.
Theorem 1. Let K be an arbitrary field. If P (X, Y ) ∈ K[[X, Y ]] and f (X) ∈ K[[X]] are such
that f (0) = 0, P (X, f (X)) = f (X) and PY′ (0, 0) = 0. Then
X
[X n ]f =
[X n Y m−1 ](1 − PY′ (X, Y ))P m (X, Y ).
m≥1
If the characteristic of K is 0, we also have the following form
[X n ]f =
X 1
[X n Y m−1 ]P m (X, Y ).
m
m≥1
Remark 1. The conditions P (X, f (X)) = 0 and f (0) = 0 imply that P (0, 0) = 0. As PY′ (0, 0)
is also 0, the sums in both expressions of [X n ]f are finite.
Remark 2. When P (X, Y ) = Xφ(Y ), where φ(X) ∈ K[[X]] and φ(0) 6= 0, we obtain the
Lagrange inversion formula
[X n ]f = [Y n−1 ](φ(X)n − Y φ′ (X)φ(X)n−1 ).
If the characteristic of K is 0, we also have the following form
[X n ]f =
1 n−1
[Y
]φ(Y )n .
n
Remark 3. When P (X, Y ) is a polynomial in X and Y , we obtain a generalisation of FlajoletSoria coefficient extraction formula [2].
Definition 1. Let P (X, Y ) ∈ K[[X, Y ]],
P (X, Y ) =
∞
X
j=0
where aj (x) ∈ K[[X]]. We define
1
aj (x)Y j ,
[m]
P
∞
X
j
aj (x)Y j−m .
(X, Y ) =
m
j=m
for m ∈ N.
The motivation of the definition of P [m] is to avoid the factorials in the denominators in
the Taylor series, which do not make sense in positive characteristic. Once this obstacle is
circumvented, the Taylor formula works as expected.
Proposition 1. Let K be an arbitrary field. Let P (X, Y ) ∈ K[[X, Y ]] and f (X) ∈ K[[X]] with
f (0) = 0. Then
∞
X
P (X, Y ) =
(Y − f (X))m P [m] (X, f (x)).
(*)
m=0
Proof. Let P (X, Y ) =
∞
P
aj (X)Y j with aj (X) ∈ K[[X]] for j ∈ N. We prove that for all k ∈ N,
j=0
the coefficient of Y k in the left side and right side of (∗) is equal. Indeed, we have
[Y k ]
∞
X
(Y − f (X))m P [m] (X, f (x))
m=0
∞
∞
X
X
j
m
m−k
aj (x)f (X)j−m
=
(−f (X))
m
k
j=m
m=k
=
∞
X
aj (X)f (X)j−k
j=k
=
∞
X
j
X
j
m
(−1)m−k
m
k
m=k
aj (X)f (X)j−k
j=k
j
X
j
(−1)m−k
j − m, m − k, k
m=k
=ak (X).
j
P
We have the last equality because
m=k
j
j−m,m−k,k
(−1)m−k = 1 if j = k and 0 if j > k. This is
because we have the multinomial expansion
X
j
j
(a + b + c) =
aj−m bm−k ck
j − m, m − k, k
k≤m≤j
=
j X
j
X
k=0 m=k
j
aj−m bm−k ck
j − m, m − k, k
When we take a = 1, b = −1, the identity becomes
j
c =
j
X
k=0
j
X
m=k
!
j
m−k
ck .
(−1)
j − m, m − k, k
Seeing this as a polynomial identity in the variable c gives us the desired result.
2
The following corollary is immediate.
Corollary 1. Let K be an arbitrary field. Let Q(X, Y ) ∈ K[[X, Y ]] and f (X) ∈ K[[X]] be
such that f (0) = 0 and Q(X, f (X)) = 0. Then there exists R(X, Y ) ∈ K[[X, Y ]] such that
Q(X, Y ) = (Y − f (X))R(X, Y ).
Definition 2. For the formal power series in K((X1 , ..., Xm ))
X
nm
P (X1 , X2 , ..., Xm ) =
an1 n2 ...nm X1n1 X2n2 ...Xm
ni >−µ
its (principal) diagonal Df (t) is defined as the element in κ((T ))
X
DP (T ) =
ann...n T n .
The following proposition is a generalization of Proposition 2 from [1], the only difference in
the proof is in the first step where we use Corollary 1 to factorize Q(X, Y ).
Proposition 2. Let K be an arbitrary field. Let Q(X, Y ) ∈ K[[X, Y ]] and f (X) ∈ K[[X]] be
such that f (0) = 0, Q(X, f (X)) = 0 and Q′Y (0, 0) 6= 0, then
Q′ (XY, Y )
.
f (X) = D Y 2 Y
Q(XY, Y )
Proof. Using Corollary 1 we can write Q(X, Y ) = (Y −f (X))R(X, Y ) with R(X, Y ) ∈ K[[X, Y ]].
We have R(0, 0) 6= 0 because f (0) = 0 and Q′Y 6= 0. Then
1
R′ (X, Y )
1
Q′Y (X, Y ) =
+ Y
.
Q(X, Y )
Y − f (X)
R(X, Y )
Replacing X by XY and multiplying by Y 2 we get
Q′ (XY, Y )
R′ (XY, Y )
Y2
D Y2 Y
=D
+D Y2 Y
.
Q(XY, Y )
Y − f (XY )
R(XY, Y )
(†)
For the first term on the right side of (†) we have
Y2
D
Y − f (XY )
Y
=D
1 − Y −1 f (XY )
!
∞
X
−n+1
n
=D
Y
f (XY )
n=0
=D (f (XY ))
=f (X).
R′ (XY,Y )
Y
For the second term, as R(0, 0) 6= 0, R(XY,Y
) is a power series in XY and Y , so when we
2
multiply this by Y there is no diagonal term.
3
Proof of Theorem 1. Let the power series Q(X, Y ) be defined as Q(X, Y ) = P (X, Y ) − Y , then
Q′Y (0, 0) = PY′ (0, 0) − 1 6= 0, and Q(X, f (X)) = P (X, f (X)) − f (X) = 0. According to Proposition 2,
f = D{Y 2 Q′Y (XY, Y )/Q(XY, Y )}
= D{Y 2 (PY′ (XY, Y ) − 1)/(P (XY, Y ) − Y )}
P (XY, Y )
)}
Y
X P (XY, Y )
)m )}.
= D{Y (1 − PY′ (XY, Y ))(1 +
(
Y
= D{Y (1 − PY′ (XY, Y ))/(1 −
m≥1
)
has no constant term, and
We have the last equality due to the fact that PY′ (0, 0) = 0, P (XY,Y
Y
P
P (XY,Y )
P (XY,Y ) m
therefore 1/(1 −
) = 1 + m≥1 ( Y
) .
Y
As in each term of Y (1−PY′ (XY, Y )) the power of Y is larger than that of X, it cannot contribute
to the diagonal. Therefore,
fn = [X n Y n ]Y (1 − PY′ (XY, Y ))(1 +
X P (XY, Y )
(
)m )
Y
m≥1
X P (XY, Y )
= [X n Y n ]Y (1 − PY′ (XY, Y ))(
)m )
(
Y
m≥1
X
n m−1
′
=
[X Y
](1 − PY (X, Y ))P (X, Y )m .
m≥1
References
[1] H. Furstenberg, “Algebraic functions over finite fields”, J. Algebra 7, 271–277 (1967).
[2] M. Soria, Thèse d’habilitation (1990), LRI, Orsay.
4
| 0math.AC
|
Nonparametric Regression for Spatial Data with Wavelets ∗
arXiv:1609.06744v3 [math.ST] 11 Feb 2018
Johannes T. N. Krebs†
February 13, 2018
Abstract
We study nonparametric regression estimates for random fields. The data satisfies certain strong mixing conditions
and is defined on the regular N -dimensional lattice structure. We show consistency and obtain rates of convergence.
The rates are optimal modulo a logarithmic factor in some cases. As an application, we estimate the regression function
with multidimensional wavelets which are not necessarily isotropic. We simulate random fields on planar graphs with
the concept of concliques (cf. Kaiser et al. (2012)) in numerical examples of the estimation procedure.
Keywords: Asymptotic statistics; Concliques; Graphical networks; Multidimensional wavelets; Nonparametric regression; Rates of convergence; Random fields; Sieve estimation; Spatial lattice processes; Strong spatial mixing conditions
MSC 2010: Primary 62G08, 62H11, 65T60; Secondary: 65C40, 60G60
1 Introduction
In this article we consider a nonparametric regression model with random design for data which is observed on a spatial
structure such as a regular N -dimensional lattice or a finite and undirected graph G = (V, E) with a set of nodes V and
a set of edges E. Consider the random field (X, Y ) = {(X(s), Y (s)) : s ∈ ZN } ⊆ Rd × R. We assume that (X, Y )
has equal marginal distributions, e.g., (X, Y ) is stationary. Denote the probability distribution of the X(s) by µX . The
process satisfies the regression model
Y (s) = m(X(s)) + ς(X(s)) ε(s),
s ∈ ZN ,
(1.1)
where m and ς are two elements of the function space L2 (µX ). The collection of error terms ε = {ε(s) : s ∈ ZN }
is independent of X. The ε(s) have mean zero and unit variance. There is a vast literature on nonparametric regression models, e.g., the books of Härdle (1990), Györfi et al. (2002) and Györfi et al. (2013). A particular choice for
the estimation of (1.1) are sieve estimators, see Grenander (1981). One class of sieve estimators are neural networks:
Hornik (1991) investigates approximation properties of multilayer feedforward networks. Rates of L2 -convergence for
sigmoidal neural networks have been studied by Barron (1994) and McCaffrey and Gallant (1994). Franke and Diagne
(2006) use neural networks for modeling financial time series. Kirch and Tadjuidje Kamgaing (2014) model autoregressive processes by a feedforward neural network.
Another popular choice for the construction of the sieve are wavelets, see Härdle et al. (2012) and Fan and Gijbels
(1996). While the kernel method is efficient if the probability distribution of the X(s) has unbounded support, it can
have disadvantages if the distribution is compactly supported because of the boundary bias. Moreover, the kernel method
requires a smooth regression function, e.g., two-times continuous differentiability. In situations where these requirements
are not satisfied, the wavelet method is an alternative which performs relatively well because it adapts automatically to the
∗ This research was supported by the German Research Foundation (DFG), Grant Number KR-4977/1 and by the Fraunhofer ITWM, 67663
Kaiserslautern, Germany which is part of the Fraunhofer Gesellschaft zur Förderung der angewandten Forschung e.V.
† Department of Statistics, University of California, Davis, CA, 95616, USA, email: [email protected]
1
regularity of the underlying regression function. The wavelet estimator does not depend on the smoothness parameters
of the curve of interest, however, it behaves as if the true curve is known in advance.
In this article, we consider the sieve estimator as defined in Györfi et al. (2002) and we construct the sieve in applications with general multidimensional wavelets. The wavelet method has already been studied both in the classical
i.i.d. case and for dependent data in various ways: Donoho et al. (1996) and Donoho and Johnstone (1998) use wavelets
for univariate density estimation with i.i.d. data. Cai (1999) studies block thresholding of the wavelet estimator in the
regression model with fixed design. Kerkyacharian and Picard (2004) construct warped wavelets for the random design regression model which admit an orthonormal basis w.r.t. the design distribution. Kulik and Raimondo (2009) use
warped wavelets in the regression model with dependent data and heteroscedastic error terms. Brown et al. (2010) study
the wavelet method in the context of nonparametric regression estimators for exponential families.
Recently, the analysis of spatial data has gained importance in many applications, e.g., in astronomy, image analysis,
environmental sciences or more general in GIS applications. Spatial data, which is often referred to as random fields,
is mostly indexed by the discrete set ZN , N ∈ N+ . The monographs of Cressie (1993) and Kindermann and Snell
(1980) offer a detailed introduction to this topic. Consequently, nonparametric regression models (with random design) for dependent data have become a major tool in spatial statistics. We only mention a few related references:
Koul (1977), Roussas and Tran (1992), Baraud et al. (2001), Guessoum and Saı̈d (2010), Yahia and Benatia (2012),
Li and Xiao (2017).
So far, the kernel method has been a popular when considering regression models for spatial data, see e.g., Carbon et al.
(2007) or Hallin et al. (2004). The wavelet method however has received little attention: Li (2016) studies a wavelet
estimator for the nonparametric regression model in the context of spatial data under the assumption that the design
distribution of the X(s) is known. In this article, we continue with these ideas but we remove the assumption that the
design distribution is known. We transfer the nonparametric regression model of Györfi et al. (2002) for i.i.d. data to
spatially dependent data. The model of Györfi et al. (2002) has three important features. Firstly, the regression function
m can be any function in L2 (µX ). It is not required that m belongs to a certain range of function classes. E.g., other
papers in the wavelet context often assume that the regression function belongs to the class of Besov spaces. Secondly,
the function classes we construct the estimator from can take a very general form; we could use neural networks instead
of multidimensional wavelets. Thirdly, the predicted variables Y (s) are not necessarily bounded and neither the design
distribution of the X(s) nor the distribution of the error terms ε(s) needs to admit a density w.r.t. the Lebesgue measure.
Furthermore, in this paper, we enrich the model with the following novelties. The data is not necessarily i.i.d. distributed
any more. We prove consistency and derive rates of convergence of the least-squares estimator under strong mixing
conditions. We relax the assumptions on the marginal distributions of the random field (X, Y ): the design distribution
does not have to be known and does not have to admit a density w.r.t. the d-dimensional Lebesgue measure. The latter
condition is assumed for instance in Hallin et al. (2004). In applications we choose d-dimensional wavelets to construct
the sieve. These wavelets can take a very general form and do not have to be isotropic.
Moreover, we remove the usual assumption of stationarity: we show that our estimator is consistent if the random
field has equal marginal distributions. This is useful in applications to (Markov) random fields defined on irregular
graphical networks which do not satisfy the usual definitions of stationarity. A Gaussian random field defined on a finite
graph G = (V, E) is such an example. There, the dependency structure of the data is determined by the adjacency matrix
of G and is supposed to vanish with an increasing graph-distance. Particular applications we have in mind are data like
traffic intensity or road roughness indices on road networks, which may be represented on graphs.
The simulation examples in the present manuscript are constructed with the algorithm of Kaiser et al. (2012) and
use the concept of concliques. This approach puts us in position to consider our simulation as iterations of an ergodic
Markov chain and we achieve a fast convergence of the simulated random field when compared to the Gibbs sampler.
We give two simulation examples where we consider one bivariate and one univariate nonparametric linear regression
problem on real graphical structures. The results give encouraging prospects in the handling of random fields on graphs.
Altogether, on the one hand, the main contribution of the paper is the generalization of the theory of distributionfree nonparametric regression of Györfi et al. (2002) to spatial data. On the other hand, we demonstrate how practical
inference on irregular graphs can be performed with the studied estimation technique.
The remainder of this manuscript is organized as follows: we introduce the basic notation in detail in Section 2.
Besides, we present two general theorems on the consistency and the rate of convergence of the truncated nonparametric
2
linear least-squares estimator. In Section 3 we use general d-dimensional wavelets to construct a consistent estimator of
the regression function. Additionally, we derive rates of convergence for this estimator in examples where the regression
function satisfies certain smoothness conditions. Section 4 is devoted to numerical applications: we present simulation
concepts for random fields on graphical structures and discuss the developed theory in two examples. Section 5 contains
the proofs of the presented theorems. Appendix A consists of useful exponential inequalities for dependent sums.
Appendix B contains a deferred verification of an example.
2 Linear Regression on Spatial Data
In this section, we present two main results of this article: we prove the consistency of the nonparametric estimator and
derive its rate of convergence under very general conditions. The section is divided into three subsections. We start
with the necessary notation and definitions in the first subsection and explain the estimation procedure in the second
subsection. The last subsection contains the results.
2.1 Notation and Definitions
Since we focus on random variables defined on a spatial structure, we introduce some notation usually used in this
context. We work on a probability space (Ω, A, P) that is equipped with a generic random field Z; in our application Z
will often be the random field (X, Y ). So Z is a collection of random variables {Z(s) : s ∈ ZN }, where N is the lattice
dimension. Each Z(s) maps from Ω to S, where (S, S) is a measurable space.
The random field is called (strictly) stationary if for each k ∈ N+ , for all points s1 , . . . , sk ∈ ZN and for each
translation t ∈ ZN , it is true that the joint distribution of {Z(s1 + t), . . . , Z(sk + t)} coincides with the joint distribution
of {Z(s1 ), . . . , Z(sk )}.
If U is a random variable on (Ω, A, P) with values in [−∞, ∞], we write kU kP,p for the p-norm of U w.r.t. P for
p ∈ [1, ∞]. Similarly, if ν is a measure on (Rd , B(Rd )) and f a real-valued function on Rd , we write kf kLp (ν) for the
Lp -norm of f w.r.t. ν for p ∈ [1, ∞].
We define the 2-norm of a square matrix A = (ai,j )1≤i,j≤d ∈ Rd×d as kAk2 = supx:kxk2 =1 kAxk2 and write λ
(resp. λd ) for the one-dimensional (resp. d-dimensional) Lebesgue measure on (R, B(R)) (resp. (Rd , B(Rd ))). Denote
the space of square integrable Borel functions on Rd w.r.t. the d-dimensional Lebesgue measure by L2 Rd , B(Rd ), λd .
We sometimes abbreviate it by L2 (λd ).
Furthermore, we write k · k∞ for the maximum norm on RN and d∞ for the corresponding metric which is extended
to subsets I, J of RN via d∞ (I, J) := inf{d∞ (s, t) : s ∈ I, t ∈ J}. Additionally, we write s ≤ t for s, t ∈ RN if and
only if the single coordinates satisfy si ≤ ti for each 1 ≤ i ≤ N . The N -dimensional vector (1, . . . , 1) is abbreviated
by eN .
Let I be a subset ZN . Then the σ-algebra which is generated by the Z(s) with s in I is denoted by F(I). The
α-mixing coefficient was introduced by Rosenblatt (1956), we define it for a random field as
α(k) :=
sup
Z
sup |P(A ∩ B) − P(A)P(B)| ,
I,J⊆ N , A∈F(I),
d∞ (I,J)≥k B∈F(J)
k ∈ N.
A random field is strongly spatial mixing if α(k) → 0 as k → ∞. The β-mixing coefficient was introduced by
Kolmogorov and Rozanov (1960), it is defined for two sub-σ-algebras F, G of A as
X X
1
P(Ui ∩ Vj ) − P(Ui )P(Vj ) : (Ui )i∈I ⊆ F, (Vj )j∈J ⊆ G are finite partitions of Ω .
β(F, G) := sup
2
i∈I j∈J
We define the β-mixing coefficient of the random field Z as
β(k) :=
sup
β(F(I), F(J)).
Z
I,J⊆ N ,
d∞ (I,J)≥k
3
We have the following relation between the two mixing coefficients 2α(k) ≤ β(k) ≤ 1, see Bradley (2005).
N
In the following we associate to each n ∈ NN
+ an index sets In := {s ∈ Z : eN ≤ s ≤ n}. As we study regression
estimates based on data defined on this index set, we need to precise the asymptotics of the vector n. Consider a sequence
(n(k) : k ∈ N) ⊆ NN such that
min(ni (k) : i = 1, . . . , N ) ≥ C ′ max(ni (k) : i = 1, . . . , N ) for some C ′ ∈ R+
and max(ni (k + 1) : i = 1, . . . , N ) > max(ni (k) : i = 1, . . . , N ).
(2.1)
We say that such a sequence converges to infinity and write n → ∞. Moreover, if (An(k) : k ∈ N) is sequence which
is indexed by the sequence (n(k) : k ∈ N), we also write An for this sequence. In particular, we characterize limits for
real-valued sequences An in this notation, i.e., we agree to write limn→∞ An for limk→∞ An(k) .
Equation (2.1) allows us to proceed at different speeds in each direction, as long as the ratio between the minimum
and the maximum does not fall below a certain level. The amendment that the running maximum is strictly increasing ensures that we select sufficiently many data points in the sampling process and guarantees a strongly universally
consistent estimator.
We need two regularity conditions to prove the consistency of the sieve estimator. The first condition concerns both
the index set on which the data is defined and the distribution of the data. We consider two models (α) and (β):
Condition 2.1. Z := {Z(s) : s ∈ ZN } is an Rd -valued random field for N, d ∈
distributions, i.e., LZ(s) = LZ(t) for all s, t ∈ ZN . Furthermore,
N+ which has equal marginal
(α) the α-mixing coefficients decrease exponentially, i.e., α(k) ≤ c0 exp(−c1 k), k ∈ N and for certain c0 , c1 ∈ R+ .
(β) for each pair (s, t) ∈ ZN × ZN the joint distribution of (Z(s), Z(t)) is absolutely continuous w.r.t. the product
measure PZ(eN ) ⊗ PZ(eN ) such that the corresponding Radon-Nikodým derivatives are uniformly bounded in that
sup
s,t∈
Z
N
dP(Z(s),Z(t))
d PZ(eN ) ⊗ PZ(eN )
P,∞
< ∞.
Moreover, the β-mixing coefficients of Z decrease exponentially, i.e., β(k) ≤ c0 exp(−c1 k), k ∈
certain c0 , c1 ∈ R+ .
(2.2)
N and for
Condition 2.1 (α) is a very weak condition if the regression estimator is expected to be consistent. A usual assumption in this context is stationarity, as in Hallin et al. (2004) or Li (2016). However, since we want to cover irregular
networks, we need this relaxed assumption because there is no definition of stationarity for random fields on a general
(finite) graphical structure. Clearly, the dependence within the data has to vanish with increasing distance on the lattice.
The decay of the α-mixing coefficients is not unusual. In particular, one can show that for time series exponentially
decreasing α-mixing coefficients are guaranteed under mild conditions (Davydov (1973), Withers (1981)).
Condition 2.1 (β) implies the first condition. The assumption on the Radon-Nikodým derivatives is reasonable as
the dependence between the observations vanishes with increasing distance. Note that this condition does not imply
that the marginal laws of the Z(s) have to admit a density w.r.t. the Lebesgue measure. Assuming this condition,
we obtain optimal rates of convergence. The technical reason why β-mixing ensures optimal rates is that the data
(X(s), Y (s)) can be coupled with a sample (X ∗ (s), Y ∗ (s)) which is sufficiently independent, we give more details
below. Chen and Christensen (2013) also obtain optimal rates of convergence for regression estimates of time series
under a β-mixing condition.
At this point, it is important to remember the relation of Condition 2.1 (β) to m-dependence. Bradley (1989) shows
that for random fields β-mixing and stationarity imply m-dependence if N is at least two. Thus, each strong mixing
condition which is stricter than α-mixing implies m-dependence if the random field is stationary. In the following, we
will work only in a few cases with a combination of Condition 2.1 (β) and stationarity. Especially, we do not need
the assumption of stationarity when establishing rates of convergence under β-mixing. So our results apply indeed to
β-mixing data and not only to m-dependent data.
4
In order to study sieve estimators, we have to quantify the approximability of function classes by a finite collection
of functions. For that reason let ε > 0 and let Rd , B(Rd ) be endowed with a probability measure ν. Consider a class
of real-valued Borel functions G on Rd . Every finite collection g1 , . . . , gM of Borel functions on Rd is called an ε-cover
of size M of G w.r.t. the Lp -norm k · kLp (ν) if for each g ∈ G there is a j, 1 ≤ j ≤ M , such that kg − gj kLp (ν) =
1/p
R
p
< ε. The ε-covering number of G w.r.t. k · kLp (ν) is defined as
Rd |g − gj | dν
n
o
N ε, G, k · kLp (ν) := inf M ∈ N : ∃ ε − cover of G w.r.t. k · kLp (ν) of size M .
(2.3)
N is monotone, i.e., N ε2 , G, k · kLp (ν) ≤ N ε1 , G, k · kLp (ν) if ε1 ≤ ε2 . The covering number can be bounded
uniformly over all probability measures under mild regularity conditions, compare the theorem of Haussler (1992)
which is stated as Proposition A.1 in the appendix. Thus, many function classes G satisfy the following condition
Condition 2.2. G is a class of uniformly bounded, measurable functions f : Rd → R, i.e., there is a B ∈ R+ such that
kf k∞ = sup{|f (x)| : x ∈ Rd } ≤ B for all f ∈ G. Additionally, for all ε > 0 and all M ≥ 1:
for any choice z1 , . . . , zM ∈ Rd the ε-covering number of G w.r.t. the L1 -norm of the discrete measure
with point masses M −1 in z1 , . . . , zM is bounded
by a deterministic
function depending only on ε and G.
P
We denote this function by HG (ε), i.e., N ε, G, k · kL1 (ν) ≤ HG (ε), where ν = M −1 M
k=1 δzk .
The key requirement of Condition 2.2 is that the covering number (which can be stochastic) admits a deterministic
bound which only depends on the function class itself and on the parameter ε. A comparison with Proposition A.1 shows
that almost every deterministic class of uniformly bounded functions admits such a bound. In particular, Condition 2.2
is valid for classes of uniformly bounded Sobolev functions or Hölder continuous functions.
2.2 The Estimation Procedure
In this subsection, we describe the estimation procedure in detail. It coincides largely with the framework of Györfi et al.
(2002). We assume that the random field (X, Y ) satisfies Condition 2.1 (α) or (β). The Y (s) are real-valued, and each
pair (X(s), Y (s)) satisfies the relation (1.1) for each observation location s ∈ ZN . The error terms ε(s) are independent
of X, have mean zero and unit variance. Note that we do not require any specific distribution of the error terms, e.g., a
Gaussian distribution is not necessary.
In addition, let Fn(k) ⊆ L2 (µX ) be deterministic increasing function classes the union of which is dense in L2 (µX ).
The function classes are indexed by the sequence from (2.1). In this context, increasing means that Fn(k) ⊆ Fn(k+1) for
k ∈ N+ . Preferably, one chooses function classes which satisfy the universal approximation property, see e.g. Hornik
(1991): as the distribution µX is unknown in practice, the union of the function spaces Fn should be dense in L2 (µ) for
any probability measure µ on (Rd , B(Rd )). We come back to this in more detail in Section 3.
The least-squares estimator is defined for a function class Fn and a sample In as
X
2
mn := arg min |In |−1
Y (s) − f (X(s)) .
(2.4)
f ∈Fn
s∈In
Later, we will choose finite-dimensional linear spaces spanned by real-valued functions fj : Rd → R for the Fn , i.e.,
Fn =
(
Kn
X
j=1
)
aj fj : aj ∈ R, j = 1, . . . , Kn .
(2.5)
Note that the basis functions fj are ordered, so that Fn(k) ⊆ Fn(k+1) if Kn(k) ≤ Kn(k+1) . Using linear spaces as
function classes has the computational advantage that the minimization is an unrestricted ordinary least-squares problem
on the domain of the parameters without an additional penalizing term, i.e., the minimizing function in Equation (2.4)
P
PKn
can be determined by the parameters (a1 , . . . , aKn ) which minimize |In |−1 s∈In (Y (s) − i=1
ai fi (X(s)))2 .
In examples of application, the parameters ai are estimated with principal component regression and singular value
decomposition. Nevertheless, the subsequent results are derived for more general function classes Fn . These merely
5
have to satisfy a technical condition on the measurability of the random variables (X(s), Y (s)) mapping from the
probability space (Ω, A, P) to Rd × R; we indicate this by the writing Ω ∋ ω 7→ (X(s, ω), Y (s, ω)).
In what follows, let (ρn(k) : k ∈ N+ ) be a real-valued and positive sequence which tends to infinity. Denote the
truncation operator by TL y := max(min(y, L), −L) (L > 0). Then we define the truncated function classes of Fn by
Tρn Fn := {Tρn f : f ∈ Fn }. The function classes Fn , resp. Tρn Fn , must not be too complex in the sense that taking
the supremum preserves measurability: to be more precise, we need that the map
Ω ∋ ω 7→
sup
f ∈Tρn Fn
|In |−1
X
s∈In
h
i
2
2
|f (X(s, ω)) − TL Y (s, ω)| − E |f (X(eN )) − TL Y (eN )|
(2.6)
is A-B(R)-measurable for all n(k) and for all L > 0. This is necessary to apply exponential inequalities to (2.6). Finitedimensional linear spaces satisfy (2.6), so this condition is satisfied in our applications. In order to obtain a consistent
estimator in regions of Rd with sparse data, we consider the truncated least-squares estimator
m̂n := Tρn mn .
(2.7)
Consequently, the properties of the truncated least-squares estimator are determined by the sample In , by the sequence ρn and by the function classes Fn . In the case of linear spaces the latter are defined in terms of the number of
basis functions Kn . In the following results the interplay between ρn , In and the Fn determines the consistency of the
estimator and its rate of convergence.
2.3 Consistency and Rate of Convergence
This subsection contains the main results of Section 2. We start with a result on the consistency of the truncated leastsquares estimator m̂n from Equation (2.7):
Theorem 2.3. Let the random field (X, Y ) satisfy (1.1). Let the Fn be increasing function classes the union of which is
dense in L2 (µX ) and fulfills (2.6). Assume that Condition 2.2 is satisfied for the truncated function classes Tρn Fn and
define κn (ε, ρn ) := log HTρn Fn (ε/(4ρn )) . Assume that κn (ε, ρn ) → ∞ as n → ∞ in NN for each ε > 0.
Let Condition 2.1 (α) be satisfied. If for each ε > 0
κn (ε, ρn )ρ4n (log |In |)2 /|In |1/N → 0 as n → ∞,
then the estimate m̂n is weakly universally consistent, i.e., limn→∞ E
tionally (X, Y ) is stationary and if additionally
R
R
d
|m̂n − m|2 dµX
(2.8)
= 0. Moreover, if addi-
4
ρ4n (log |In |) /|In |1/N → 0 as n → ∞,
then m̂n is strongly universally consistent, i.e., limn→∞
Let Condition 2.1 (β) be satisfied. If for each ε > 0
R
R
d
|m̂n − m|2 dµX = 0 a.s.
κn (ε, ρn )ρ4n (log |In |)N /|In | → 0 as n → ∞,
then the estimate m̂n is weakly universally consistent. Moreover, if (X, Y ) is stationary and if additionally
ρ4n (log |In |)N +2 /|In | → 0 as n → ∞,
then m̂n is strongly universally consistent.
The growth rates of the truncation sequence and of the covering number are upper bounds which guarantee a consistent estimator. We see that the conditions in the case of α-mixing data are more restrictive than in the case of β-mixing
data. In the first case the growth in ρ4n times the logarithm of the covering number, κn (ε, ρn ), have to be overcompensated by the sample size |In |1/N modulo a logarithmic factor for a weakly universally consistent estimator. In the second
6
case of β-mixing data, the sample size |In | is not corrected by the exponent 1/N . This last result corresponds to the
classical case of i.i.d. data, see Györfi et al. (2002). Below we will see again this analogy between i.i.d. and β-mixing
data. Moreover in both dependence settings, we need an additional growth condition which ensures a strongly universally consistent estimate. Note that in the case where Condition 2.1 (β) is satisfied and (X, Y ) is stationary, we have that
(X, Y ) is indeed m-dependent if N > 1 with the result of Bradley (1989). Hence, the last statement in Theorem 2.3 is
actually achieved under m-dependence if N > 1.
In the next corollary, we give an application to the linear spaces from Equation (2.5). For linear function spaces we
can compute an upper bound of the covering number with Proposition A.1. This corollary is also a generalized result of
Györfi et al. (2002) Theorem 10.3:
Corollary 2.4. Let Fn be the linear span of continuous and linear independent functions f1 , . . . , fKn as in (2.5) such
that ∪k∈N+ Fn(k) is dense in L2 (µX ).
Assume Condition 2.1 (α). m̂n is weakly universally consistent if Kn ρ4n log ρn (log |In |)2 /|In |1/N → 0 as n → ∞.
m̂n is strongly universally consistent if additionally (X, Y ) is stationary and if additionally ρ4n (log |In |)4 /|In |1/N →
0 as n → ∞.
Assume Condition 2.1 (β). m̂n is weakly universally consistent if Kn ρ4n log ρn (log |In |)N /|In | → 0 as n → ∞.
m̂n is strongly universally consistent if additionally (X, Y ) is stationary and if additionally ρ4n (log |In |)N +2 /|In | →
0 as n → ∞.
One usually chooses a truncation sequence ρn growing at a rate of O(log |In |) which is negligible, e.g., see Kohler
(2003) who considers piecewise polynomials as basis functions in the case of i.i.d. data.
The next result gives the rate of convergence of the truncated least-squares estimator m̂n in both dependence scenarios. This rate can be divided into an empirical error which depends on the realization ω ∈ Ω and an approximation error
which relates the regression function m to its projection onto the function classes Fn .
However, in order to derive a rate of convergence result, we need an additional requirement on the error terms
because we do not rule out conditional dependence between two distinct observations Y (s) and Y (t). Thus, we need a
condition on the conditional covariance matrix of the observations Y (s) given the observations X(s). We denote this
matrix by Cov(Y (In ) | X(In )). Note that in the special case with uncorrelated error terms ε(s), Cov(Y (In ) | X(In )) is
a diagonal matrix and it is sufficient to impose a restriction on the conditional variances.
Theorem 2.5. Assume that the regression function and the conditional variance function are essentially bounded, i.e.,
kmk∞ , kςk∞ ≤ L. If the error terms ε(s) are correlated, assume that E |ε(eN )|2+γ < ∞ for some γ > 0. The
function classes Fn are linear spaces as in Equation (2.5).
If Condition 2.1 (α) is satisfied, assume Kn log(|In |)/|In |1/(2N ) → 0 as n → ∞. Then there is a C ∈ R+ such that
Z
Z
Kn log |In |
.
E
|m̂n − m|2 dµX ≤ 8 inf
|f − m|2 dµX + C
f ∈Fn Rd
|In |1/(2N )
Rd
If Condition 2.1 (β) is satisfied, assume Kn (log |In |)N +2 /|In | → 0 as n → ∞. Then there is a C ∈ R+ such that
E
Z
R
2
d
|m̂n − m| dµX
≤ 16 inf
f ∈Fn
Z
R
d
2
|f − m| dµX + C
Kn (log |In |)N +2
.
|In |
The first term appearing on the right-hand side of both inequalities is a multiple of the the approximation error which
depends on the function class Fn . The second term is the estimation error. The number of basis functions Kn has a
linear influence on this error. This influence is negative because if Kn increases, there are more parameters which need
to be estimated. Conversely, a growing sample size reduces this error. In the case of β-mixing data, an increasing sample
size reduces the error more than in the case of α-mixing data.
The boundedness of the functions m and ς and the assumption that we know this bound are essential to derive rates
of convergence, see Györfi et al. (2002) for more details. However, note that we do not assume the error terms to be
bounded. We only require a moment condition, i.e., E |ε(eN )|2+γ < ∞ for some γ > 0. This is not unexpected if we
want to bound the summed conditional covariances in our model from (1.1) which has the multiplicative heteroscedastic
structure.
7
In the case of linear function spaces, Györfi et al. (2002) find that the estimation error can be bounded by Kk (log k)/k
times a constant under similar assumptions for the case of an i.i.d. sample of size k. This guarantees an optimal rate of
convergence in terms of Stone (1982) up to a logarithmic factor. We see that for β-mixing data our result guarantees the
same rate up to a logarithmic factor. We discuss this in detail in Section 3 below.
3 Linear Wavelet Regression on Spatial Data
In this section we consider an adaptive wavelet estimate of the regression function m. The section is divided into two
subsections. We give the main definitions in the first subsection, the results follow in the second subsection.
3.1 Preliminaries
A detailed introduction to the properties of wavelets, in particular the construction of wavelets with compact support,
can be found in Meyer (1995) and Daubechies (1992). Since we consider d-dimensional data, we give a short review
on important concepts of wavelets in d dimensions indexed by the lattice Zd . The definitions are taken from Strichartz
(1993). In the following, M ∈ Rd×d is a matrix which preserves the lattice, i.e., M Zd ⊆ Zd . Moreover, M is strictly
expanding in that all eigenvalues ζ of M satisfy |ζ| > 1. Denote the absolute value of the determinant of M by |M |.
A multiresolution analysis (MRA) of L2 λd with a scaling function Φ : Rd → R is an increasing sequence of
subspaces . . . ⊆ U−1 ⊆ U0 ⊆ U1 ⊆ . . . such that the following four conditions are satisfied
S
(1) (Denseness) j∈Z Uj is dense in L2 λd ,
T
(2) (Separation) j∈Z Uj = {0},
(3) (Scaling) f ∈ Uj if and only if f (M −j · ) ∈ U0 ,
(4) (Orthonormality) {Φ( · − γ) : γ ∈ Zd } is an orthonormal basis of U0 .
The relationship between an MRA and an orthonormal basis of L2 (λd ) is summarized in the next theorem:
Theorem 3.1 (Strichartz (1993)). Suppose Φ generates a multiresolution analysis and the ak (γ) satisfy for all 0 ≤
j, k ≤ |M | − 1 and γ ∈ Zd the equations
X
X
aj (γ ′ ) ak (M γ + γ ′ ) = |M | δ(j, k) δ(γ, 0) and
a0 (γ) = |M |.
γ′∈
Z
d
γ∈
Z
d
P
Furthermore, define the functions Ψk := γ∈Zd ak (γ) Φ(M · −γ) for k = 1, ..., |M | − 1. Then the set of functions
{|M |j/2 Ψk (M j · −γ) : j ∈ Z, k = 1, . . . , |M | − 1, γ ∈ Zd } forms an orthonormal basis of L2 (λd ):
M
M
L2 (λd ) = U0 ⊕
Wj =
Wj , where Wj := h |M |j/2 Ψk (M j · −γ) : k = 1, . . . , |M | − 1, γ ∈ Zd i.
j∈
N
j∈
Z
The scaling function Φ is also called the father wavelet and also denoted by Ψ0 . The Ψk are the mother wavelets
for k = 1, . . . , |M | − 1. We sketch in a short example how to construct a d-dimensional MRA provided that one has a
father and a mother wavelet on the real line.
Example 3.2 (Isotropic d-dimensional MRA from one-dimensional MRA via tensor products). Let d ∈ N+ and let ϕ
be a father wavelet on the real line R together with a mother wavelet ψ, so that ϕ and ψ are related by the identities
√ X
√ X
ϕ≡ 2
hγ ϕ(2 · −γ) and ψ ≡ 2
gγ ϕ(2 · −γ),
γ∈
Z
γ∈
Z
for real sequences (hγ : γ ∈ Z) and (gγ : γ ∈ Z). Let ϕ generate an MRA of L2 (λ) with the corresponding spaces Uj′ ,
j ∈ Z. The d-dimensional wavelets are derived as follows. Define M by 2Id , where Id is the identity matrix in Rd×d .
8
Denote the mother wavelets as pure tensors by Ψk := ξk1 ⊗ . . . ⊗ ξkd for k ∈ {0, 1}d \ 0, where ξ0 := ϕ and ξ1 := ψ.
The scaling function is given as Φ := Ψ0 := ⊗di=1 ϕ.
In Appendix B we demonstrate that Φ and the linear spaces Uj := ⊗di=1 Uj′ form an MRA of L2 (λd ) and that the
functions Ψk , generate an orthonormal basis in the sense that
D
E
M
M
Wj =
Wj , where Wj = |M |j/2 Ψk M j · −γ : γ ∈ Zd , k ∈ {0, 1}d \ 0 .
L2 (λd ) = U0 ⊕
j∈
N
j∈
Z
3.2 Consistency and Rate of Convergence
In the sequel, we bridge the gap between nonparametric regression and wavelet theory. As indicated in Theorem 2.3 the
S
function spaces k∈N+ Fn(k) are preferably dense in L2 (µ) for any probability measure µ. The next theorem states that
wavelets satisfy this universal approximation property.
Theorem 3.3. Consider an isotropic MRA on Rd with corresponding scaling function Φ constructed as in Example 3.2
from a compactly supported real scaling function ϕ. Let µ be a probability measure on B(Rd ) and let 1 ≤ p < ∞.
S
Then j∈Z Uj is dense in Lp (µ).
In what follows we assume that Φ is a compactly supported scaling function and that M is a diagonalizable matrix,
i.e., M = S −1 DS for a diagonal matrix D which contains the eigenvalues of M . Denote the maximum of the absolute
values of the eigenvalues by ζmax := max{|ζi | : i = 1, . . . , d}. Set
Φj,γ := |M |j/2 Φ(M j · −γ), where γ ∈ Zd and j ∈ Z.
Let (wn(k) : k ∈ N) ⊆ Z and (j(n(k)) : k ∈ N) ⊆ Z be two increasing sequences with limn→∞ wn = ∞ and
limn→∞ j(n) = ∞. We set Kn := {γ ∈ Zd : kγk∞ ≤ wn } ⊆ Zd . Then, we define the linear function space by
Fn :=
X
γ∈Kn
aγ Φj(n),γ : aγ ∈ R
⊆ Uj(n) .
(3.1)
So the j(n) scale Φ, whereas the wn control which translations are used for the construction of the function space Fn .
Based on the results from the previous section, the following statements are true for the linear wavelet estimate.
Theorem 3.4. Assume that the wavelet basis is dense in L2 (µX ). Set ρn := c log |In | for some constant c ∈ R+ . Define
the wavelet estimator m̂n by Equations (2.7) and (3.1). Let limn→∞ (ζmax )j(n) /wn = 0.
Assume that Condition 2.1 (α) is satisfied, then m̂n is weakly universally consistent if
lim wnd (log |In |)6 log log |In | |In |1/N = 0.
(3.2)
n→∞
m̂n is strongly universally consistent if (X, Y ) is stationary, if (3.2) holds and if limn→∞ (log |In |)8 |In |1/N = 0.
Assume that Condition 2.1 (β) is satisfied, if limn→∞ wnd (log |In |)N +4 log log |In | |In | = 0, then m̂n is weakly
universally consistent. m̂n is strongly universally consistent if additionally (X, Y ) is stationary and if additionally
limn→∞ (log |In |)N +6 |In | = 0.
Theorem 3.5. Let Condition 2.1 (β) and the assumptions of Theorem 2.5 be satisfied, then there is a constant C independent of n such that
Z
Z
2
|m̂n − m| dµX ≤ C wnd (log |In |)N +2 |In | + 16 inf
|f − m|2 dµX .
E
R
f ∈Fn
d
R
d
We give a short application in the case where the wavelet basis is generated by isotropic Haar wavelets in d dimensions and where the regression function m is (A, r)-Hölder continuous on a compact subset of Rd . This means that
r
|m(x) − m(y)| ≤ A kx − yk∞ for all x, y in the domain of m, for an A ∈ R+ and for an r ∈ (0, 1].
9
Corollary 3.6. Let the conditions of Theorem 3.5 be satisfied such that the data satisfies Condition 2.1 (β). Let the
conditional mean function m be (A, r)-Hölder continuous. Define the level j as a function of n by 2j ≃ |In |1/(d+2r) .
Then the mean integrated squared error satisfies
Z
E
(3.3)
|m̂n − m|2 dµ = O (log |In |)N +2 |In |−2r/(d+2r) .
R
d
Proof. Note that by construction it suffices to choose wn proportional to 2j because the domain of the function m
is bounded, we can cover it with 2jd wavelets from the j-th scale. This means that the estimation error behaves as
O(2jd (log |In |)N +2 /|In |).
It remains to compute the approximation error: there is a function f ∈ Fn piecewise constant on dyadic ddimensional cubes of edge length 2−j with values
f (x) = m (γ1 , . . . , γd )/2j for x ∈ (γ1 , . . . , γd )/2j , ((γ1 , . . . , γd ) + eN )/2j ,
where γi ∈ Z for i = 1, . . . , d such that γ = (γ1 , . . . , γd ) is an admissible element from Kn . Hence for this f
Z
|f − m|2 dµX ≤ sup |f − m|2 ≤ A2 2−2rj(n) .
R
d
dom m
The choice of j as 2j ≃ |In |1/(d+2r) approximately equates the estimation and the approximation error.
The interpretation of the two parameters d and r in the rate of convergence is well known: on the one hand, an
increase in d deteriorates the rate (the curse of dimensionality). On the other hand, an increase in r towards 1 increases
the rate of convergence because the regression function becomes smoother and can be better approximated by finite
linear combinations of functions.
We compare the above result to the results for the classical case of i.i.d. data: if the regression function is Hölder
continuous, the rate of convergence is in O k −2r/(d+2r) up to a logarithmic factor, where the sample size is k, see
Kohler (2003) or Györfi et al. (2002). This is nearly optimal when compared to Stone (1982). The additional log-loss is
due to the increasingly complex sieves.
Hence, our rate of O (log |In |)N +2 |In |−2r/(d+2r) is the same modulo a logarithmic factor. Note that this result is
independent of the lattice dimension N on which the data is defined.
Chen and Christensen (2013) also consider regression function estimates with wavelets for β-mixing time series.
Their results can be compared to the present findings in the special case where the lattice dimension N equals 1. They
also obtain a nearly optimal rate w.r.t. the sup-norm.
Li (2016) considers a wavelet based regression estimator for spatial data similar to our model from (1.1). He also
obtains a nearly optimal rate. However, some of the regularity conditions are more restrictive than those in Condition 2.1:
the design distribution of the regressors X(s) has to admit a density and, additionally, this density must be known. This
knowledge is used in the estimation process of the regression function. Furthermore, the response variables Y (s) have
to be bounded. Our results are derived without these additional restrictions.
4 Examples of Application
We begin this section with some well-known results on Gaussian Markov random fields necessary for the following
applications. The monograph of Kindermann and Snell (1980) gives an introduction to Markov random fields. Let
G = (V, E) be a finite graph. We write Ne(s) for the neighbors of a node s w.r.t. the graph G and −s for the set V \ {s}.
Assume that {Y (s) : s ∈ V } is multivariate normally distributed with expectation α ∈ R|V | and covariance matrix
d
1
Σ ∈ R|V |×|V | . Then Y has the density fY (y) = (2π)− 2 det(Σ)− 2 exp − (y − α)T Σ−1 (y − α)/2 . In particular, if we
write P for the precision matrix Σ−1 , the conditional distribution of Y (s) given the remaining observations Y (−s) is
X
Y (s) | Y (−s) ∼ N α(s) − (P (s, s))−1
P (s, t) y(t) − α(t) , P (s, s)−1 .
t6=s
10
Since P = Σ−1 is symmetric and since we can assume that P (s, s)−1 > 0, Y is a Markov random field if and only if
P (s, t) 6= 0 for all t ∈ Ne(s) and P (s, t) = 0 for all t ∈ V \ Ne(s), for all nodes s ∈ V .
Cressie (1993) investigates the conditional specification
X
c(s, t) Y (t) − α(t) , τ 2 (s) ,
(4.1)
Y (s) | Y (−s) ∼ N α(s) +
t∈Ne(s)
where C = c(s, t) 1≤s,t≤|V | is a |V | × |V | matrix and T = diag(τ 2 (s) : s ∈ V ) is a |V | × |V | diagonal matrix such
that the coefficients satisfy the condition τ 2 (s)c(t, s) = τ 2 (t)c(s, t) for s 6= t and c(s, s) = 0 as well as c(s, t) = 0 =
c(t, s) if s, t are no neighbors. This means P (s, t) = −c(s, t)P (s, s), i.e., Σ−1 = P = T −1 (I − C). If I − C is
invertible and if (I − C)−1 T is symmetric and positive definite, then the entire random field is multivariate normal with
Y ∼ N α, (I − C)−1 T .
It is plausible to use equal weights c(s, t) in many applications, see Cressie (1993). Thus, we can write the matrix C
as C = ηH, where H is the adjacency matrix of G, i.e., H(s, t) is 1 if s and t are neighbors, otherwise it is 0. Denote
the maximal (resp. minimal) eigenvalue of H by hm (resp. h0 ). Assume that h0 < 0 < hm which is often satisfied
in applications. We know from the properties of the Neumann series that in this case the matrix I − C is invertible if
(h0 )−1 < η < (hm )−1 .
This insight allows us to simulate a Gaussian Markov random field with an MCMC-algorithm using concliques
with a full conditional distribution. Here we refer to Kaiser et al. (2012) for a general introduction to the concept of
concliques. An overview of the exact simulation procedure can be found in Krebs (2018a). In the present simulation
examples, we run 15k iterations of the MCMC-algorithm. These suffice to ensure a nearly stationary distribution of the
Gaussian random field.
We sketch the simulation procedure: let G = (V, E) be a finite graph with nodes s1 , . . . , s|V | . We simulate
a d-dimensional random field Z on G such that each component Zi takes values in R|V | , for i = 1, . . . , d. We
use copulas to simulate some of the components Zi as dependent. Each random field Zi has a specification Zi ∼
N α (1, . . . , 1)′ , σ 2 Σ , where α, σ ∈ R and σ > 0. Furthermore, Σ is a correlation matrix which satisfies the relation
(I − ηH)−1 T = σ 2 Σ.
(4.2)
The parameter η is chosen such that I − ηH is invertible and T is a diagonal matrix T = diag τ 2 (s1 ), . . . , τ 2 (s|V | ) . A
large absolute value of η indicates a strong dependence within the random variables of one component, whereas η = 0
indicates independence within the component. The marginal distributions within a component equal each other, i.e.,
Zi (s) ∼ N(α, σ 2 ) for s ∈ V . However, the conditional variances τ 2 ( · ) within a component Zi may differ.
In the next step, we use some of the components to construct the random field {X(s) : s ∈ V } and use another
independent component to construct the error terms {ε(s) : s ∈ V }. We specify this below. Then we simulate the
random field Y as in Equation (1.1) for a choice of m and a constant ς. So the conditional heteroscedastic part is
constant. However, depending on the underlying graph G, the error terms ε(s) can have complex dependence patterns
among each other. We estimate m with the truncated least-squares estimator from Equation (2.7). In the situation where
the regression function m is known, the L2 -error can serve as a criterion for the goodness-of-fit of m̂: we split the whole
sample into a learning sample VL and a testing sample VT . Here both VL and VT should be two connected sets w.r.t.
the underlying graph if this is possible. We estimate m̂ from the learning sample and compute the approximate L2 -error
with Monte Carlo integration over the testing sample, i.e.,
Z
X
|m̂ − m|2 dµX ≈ |VT |−1
|m̂(X(s)) − m(X(s))|2 .
(4.3)
R
d
s∈VT
We run this entire simulation procedure 1000 times. Afterwards, we compute the mean and the standard deviation of the
(approximate) L2 -error from (4.3) based on these simulations.
Example 4.1 (Bivariate nonparametric regression). We simulate a random field on a planar graph G = (V, E) which
represents the administrative divisions in the Sydney bay area on the statistical area level 1. For further reference see the
11
website of the Australian bureau of statistics, www.abs.gov.au. It comprises 7,713 nodes and approximately 47k edges
in total. Hence, G is highly connected compared to the standard four-nearest neighborhood lattice. Figure 1(a) illustrates
the graph.
On this graph we model a three-dimensional Gaussian Markov random field Z = (Z1 , Z2 , Z3 ). Every Zi has a specification as in Equation (4.1) such that the marginals Zi (s) within each component are standard normally distributed. The
parameter space of η is derived from the adjacency matrix of the graph G and contains the interval (−0.2221, 0.1312).
Note that the range for the lattice with a four-nearest-neighborhood structure is (−0.25, 0.25).
Then we adjust the marginal conditional variance of the variable Zi (s) given by τi2 (s) such that the entire random
vector Zi has a covariance structure of the type Σi as in (4.2) for a correlation matrix Σi for i = 1, 2, 3.
j
1
2
3
4
Estimates on the graph
D4 wavelet Haar wavelet
0.264
0.413
(0.006)
(0.008)
0.122
0.258
(0.009)
(0.008)
0.163
0.198
(0.036)
(0.010)
0.422
0.259
(0.075)
(0.012)
Independent reference estimates
D4 wavelet
Haar wavelet
0.260
0.406
(0.006)
(0.007)
0.119
0.254
(0.009)
(0.007)
0.170
0.196
(0.044)
(0.010)
0.435
0.257
(0.077)
(0.012)
Table 1: L2 -error of the bivariate regression problem: the estimated mean and in parentheses the estimated standard
deviation for a level j = 1, . . . , 4. The first two columns give the results for the random field, the last two columns those
of the independent reference sample.
In order to obtain dependent components Z1 and Z2 , we draw the error terms from a two-dimensional Gaussian
copula in each iteration. The exact simulation parameters are given by µZi = 0, σi = 1 for i = 1, 2, 3, η1 = 0.12,
η2 = −0.18 and η3 = 0.12. The covariance between the first two components is 0.7. The third component Z3 is
simulated as independent.
n
o−1
(s) for i = 1, 2, 3.
The vectors τi2 ∈ R|V | are computed with the formula τi2 (s) = diag∼ inv I − ηi H
Here we denote the inverse of a matrix by inv, the operator that maps the diagonal of a matrix to a vector by diag∼ and
the elementwise inversion of a vector by {·}−1 . Afterwards, we transform the first two components Z1 and Z2 with the
distribution function of a two-dimensional standard normal distribution onto the unit square and obtain the random field
(X1 , X2 ). We specify the mean function in this example as
m:
R2 → R, (x1 , x2 ) 7→ (2 − 3x22 + 4x42 ) exp − (2x1 − 1)2 .
Figure 1(b) shows the function plot of m. We simulate Y (s) = m(X1 (s), X2 (s)) + Z3 (s) and we use two different
wavelet scaling functions for the estimation of m: we perform the first regression with the Haar scaling function ϕ =
1[0,1) and the second with Daubechies 4-scaling function D4 (which is also known as db2). Figure 2 displays the
results: Figure 2(a) depicts the estimate with Daubechies 4-scaling function, Figure 2(b) the one with the Haar scaling
function. Table 1 gives the L2 -error statistics. Note that the L2 -error minimizing j for the Haar wavelet differs from
the j minimizing the error for Daubechies 4-scaling function. The latter outperforms the Haar scaling function in this
example. Moreover, Table 1 shows the L2 -error statistics with independent sample data of the same size. Note that this
estimator with independent data is slightly better than the estimate from the random field for both wavelet types.
Example 4.2 (Univariate nonparametric regression). In this example we consider a one-dimensional spatial regression
problem based on a graph which represents Australia divided into administrative divisions on the statistical area level
3. The graph consists of 330 nodes and 1600 edges, cf. Figure 3(a). Again, this graph is highly connected in certain
regions relatively to the four nearest neighborhood structure on a lattice.
12
(a) The Sydney bay area (on statistical area level 1-scale)
3
2.5
z
2
1.5
1
0.5
0
0
1
0.2
0.8
0.4
0.6
0.6
0.4
0.8
x
0.2
1
0
y
(b) Function plot of m
Figure 1: Input graph and regression function for the bivariate regression problem
13
3
2.5
z
2
1.5
1
0.5
0
0
1
0.2
0.8
0.4
0.6
0.6
0.4
0.8
0.2
1
x
0
y
(a) Estimate of m with the D4 scaling function for j = 2
3
2.5
z
2
1.5
1
0.5
0
0
1
0.2
0.8
0.4
0.6
0.6
0.4
0.8
x
0.2
1
0
y
(b) Estimate of m with the Haar scaling function for j = 3
Figure 2: Estimated regression functions for a bivariate regression problem
14
10° S
15° S
20° S
25° S
30° S
35° S
40° S
115° E
120° E
125° E
130° E
135° E
140° E
145° E
150° E
155° E
(a) Administrative divisions of mainland Australia
6
5
4
3
2
1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
(b) A realization of X and the mean function m
Figure 3: Graph and true regression function.
15
0.9
1
4.5
4.5
4
4
3.5
3.5
3
3
4.5
4
3.5
3
2.5
2.5
2.5
2
2
2
1.5
1.5
1.5
0
0.5
1
1
0
0.5
1
0
0.5
1
(a) Estimate of m with the D4 scaling function for j = 3, 4, 5.
4.5
4.5
4.5
4
4
4
3.5
3.5
3.5
3
3
3
2.5
2.5
2.5
2
2
2
1.5
1.5
0
0.5
1
1.5
0
0.5
1
0
0.5
(b) Estimate of m with the Haar scaling function for j = 3, 4, 5.
Figure 4: The estimates for the univariate regression problem.
16
1
We simulate two Gaussian random fields Z1 and Z2 on G with marginal means 0 and marginal variances 1 with the
Markov chain method as in Example 4.1. The parameter space for η contains the interval (−0.3060, 0.1615). We choose
η for both components equal to 0.15 and run 15k iterations of the MCMC-algorithm. Then we use the inverse of the
standard normal distribution to retransform the component Z1 onto the unit interval and obtain the random field X with
marginals uniformly distributed on [0, 1]. The conditional mean function is defined as the discontinuous function
p
m : [0, 1] → R, x 7→ 2 + 8x2 − (1.7x)4 1{x≤0.7} + 2 4(x − 0.7) + 1 1{0.7<x} .
We specify Y as Y (s) = m(X(s)) + Z2 (s)/2. Figure 3(b) depicts the simulated random field. Figure 4(a) shows
the estimation with the Daubechies 4-scaling function, while Figure 4(b) depicts the case for the Haar wavelet. Table 2
shows that in all cases the L2 -error is minimized for the level j = 4. Note that in this example the Daubechies wavelet
always outperforms the Haar wavelet when measured by the L2 -error. Furthermore, we see that the L2 -error of the
independent reference estimates are minimal for the same level j = 4.
j
2
3
4
5
6
Estimates on the graph
D4 wavelet Haar wavelet
0.326
0.405
(0.031)
(0.059)
0.241
0.344
(0.033)
(0.064)
0.224
0.284
(0.077)
(0.073)
0.319
0.349
(0.172)
(0.117)
0.772
0.753
(0.437)
(0.213)
Independent reference estimates
D4 wavelet
Haar wavelet
0.321
0.401
(0.029)
(0.061)
0.233
0.341
(0.035)
(0.067)
0.213
0.280
(0.062)
(0.078)
0.299
0.333
(0.134)
(0.093)
0.712
0.727
(0.380)
(0.212)
Table 2: L2 -error of the univariate regression problem: the estimated mean and in parentheses the estimated standard
deviation for a level j = 2, . . . , 6. The first two columns give the results for the random field, the last two columns those
of an independent reference sample of the same size.
5 Proofs of the Results in Section 2 and Section 3
The first lemma is a consequence of the coupling lemma of Berbee (1979). In the case of β-mixing, we construct another
sample (X ∗ , Y ∗ ) which has good properties.
Lemma 5.1. Let (X, Y ) be a random field on ZN . For each n ∈ NN
+ and q ∈ N+ such that 2q < min{ni : i =
1, . . . , N }, there is a partition of In which is denoted by {I(l, u) : l = 1, . . . , 2N , u = 1, . . . , R} and collection
N
of random variables Z ∗ (l, u) = ((X ∗ (s), Y ∗ (s)) : s ∈ I(l, u)) ∈ R(d+1)q such that for each l the collection
Z ∗ (l, 1), . . . , Z ∗ (l, R) is independent and P(Z ∗ (l, u) 6= Z(l, u)) = β(q) where Z(l, u) = ((X(s), Y (s)) : s ∈ I(l, u))
and β is the β-mixing coefficient of (X, Y ). Moreover, Z ∗ (l, u) is independent of Z(l, 1), . . . , Z(l, u − 1) for each
u = 1, . . . , R and for each l = 1, . . . , 2N .
Proof. The proof follows as in Carbon et al. (1997) where a similar coupling result is established under the α-mixing
condition. We only sketch the main parts. Firstly, we give the construction of the partition. We choose R1 , . . . , RN such
that
2q(Ri − 1) < ni ≤ 2qRi =: n∗i for each k = 1, . . . , N.
For the k-th coordinate direction, we partition the summation index set {1, . . . , n∗i } ⊇ {1, . . . , ni } into Ri subsets each
consisting of two disjoint intervals of length q. So, we have a union of 2Ri intervals of length q.
17
Combining the partitions in all N coordinate directions, we get a partition of the N -dimensional rectangle In∗ =
{s ∈ ZN ; eN ≤ s ≤ n∗ } ⊇ In into R = R1 ·. . .·RN blocks containing (2q)N points of the N -dimensional integer lattice
each. Within each block, there are 2N smaller subsets, which are N -dimensional rectangles with all edges of length q.
Write I(l, u) for the l-th subset in the u-th block, l = 1, . . . , 2N and u = 1, . . . , R. Its cardinality is q N . Moreover,
using the requirement on n, we have that Ri ≥ 2 for each k = 1, . . . , 2N . Thus, q N R ≤ |In | and |In |/2N ≤ q N R. The
subcubes I(l, u) have the property that for fixed l the distance between I(l, u) and I(l, u′ ) is at least q.
Secondly, we apply recursively (as in Carbon et al. (1997)) the lemma of Berbee (1979) to the collection of random
N
variables Z(l, 1), . . . , Z(l, R) for each l. We obtain random variables Z ∗ (l, 1), . . . , Z ∗ (l, R) ∈ R(d+1)q with the desired properties. The claim follows now when defining the X ∗ (s) and the Y ∗ (s) such that Z ∗ (l, u) = ((X ∗ (s), Y ∗ (s)) :
s ∈ I(l, u)) a.s. for each u = 1, . . . , R and for each l = 1, . . . , 2N .
The next proposition is a well-known result of Györfi et al. (2002) which states sufficient conditions for a consistent
estimator. It holds as well for dependent data because in the proof of the proposition those terms which are related to the
dependence structure of the data converge to zero by assumption. So it is our task to verify these assumptions later. More
precisely, it is assumed that the function classes can approximate the regression function m arbitrarily exactly both in
part (a) and in part (b) of this proposition. In our case this assumption does not depend on the data. However, the second
requirement in both parts of the proposition is affected by the dependence structure of the data: here it is assumed that a
certain empirical mean uniformly converges to the corresponding true mean for each possible function in the sieve. This
requirement crucially depends on the data and we can verify this assumption later.
Proposition 5.2 (Modified version of Györfi et al. (2002) Theorem 10.2). Let (Ω, A, P) be a probability space endowed
with the random field (X, Y ) which satisfies the model (1.1) and Condition 2.1 such that each X(s) is Rd -valued
and each Y (s) is R-valued. Let Fn(k) ⊆ L2 (µX ) be a class of functions f : Rd → R for each k ∈ N+ . Let
(ρn(k) : k ∈ N) ⊆ R+ be an sequence which increases to infinity. Denote the truncated least-squares estimate of m
from Equation (2.7) by m̂n . In addition, let the map from (2.6) be A-B(R)-measurable.
(a) If for all L > 0 both
inf
kf − mkL2 (µX ) = 0 and
lim E
n→∞
f ∈Fn ,||f ||∞ ≤ρn
"
#
2
2
X
−1
TL Y (s) − f (X(s)) − E TL Y (eN ) − f (X(eN )
sup |In |
= 0,
lim E
n→∞
f ∈Tρn Fn
s∈In
i
hR
2
then m̂n is weakly universally consistent, i.e., limn→∞ E Rd (m̂n − m) dµX = 0.
P
(b) If furthermore, limn→∞ |In |−1 s∈In |Y (s) − TL Y (s)|2 = E |Y (eN ) − TL Y (eN )|2 a.s. and if both
lim
inf
n→∞ f ∈Fn ,||f ||∞ ≤ρn
lim
sup
n→∞ f ∈Tρ Fn
n
|In |
kf − mkL2 (µX ) = 0 a.s. and
−1
2
2
X
=0
TL Y (s) − f (X(s)) − E TL Y (eN ) − f (X(eN ))
a.s.
n∈In
for all L > 0, then m̂n is strongly universally consistent, i.e., limn→∞
We proceed with the proof of the first main theorem of Section 2.
2
R
R
d
(m̂n − m) dµX = 0
a.s.
Proof of Theorem 2.3. We verify that in both dependence scenarios the sufficient criteria of Proposition 5.2 are satisfied
for the given choices of the parameters. The structure of the proof is quite similar to the one of Theorem 10.3 in
Györfi et al. (2002). Therefore we sketch those parts which differ because of the dependence in the data and the assumed
covering condition (Condition 2.2). The approximation property of the function classes is satisfied by assumption.
Moreover, we can assume w.l.o.g. that L < ρn in both cases because ρn tends to infinity. We have to consider the
function classes
Hn := h : Rd × R → R, h(x, y) = |f (x) − TL (y)|2 for all (x, y) ∈ Rd × R, for some f ∈ Tρn Fn .
18
We begin with the case of α-mixing data. From Condition 2.2 we obtain a uniform bound on the ε-covering number
N ε, Hn , k · kL1 (ν) which we denote by HHn (ε), here ν is an arbitrary probability measure with equal mass concentrated at certain points z1 , . . . , zu ∈ R, u ∈ N+ . Provided that L ≤ ρn , we have for the covering number of this class
Hn
ε
ε
ε
HH n
≤ HTρn Fn
= HTρn Fn
= exp κk (ε/32, ρn ).
32
32(4ρn )
128ρn
For details on the first inequality, see the proof of Theorem 10.3 in Györfi et al. (2002). Note that the functions in Hn are
bounded by 4ρ2n if L ≤ ρn . By assumption, ρ4n κn (ε/32, ρn ) (log |In |)2 |In |1/N → 0 as n → ∞. We use Theorem
A.4 to give an upper bound on the following probability (note that we only need to consider the exponential term in
(A.2) which decays at a slower rate)
!
i
h
1 X
2
2
|f (X(s)) − TL Y (s)| − E |f (X(eN )) − TL Y (eN )|
>ε
(5.1)
sup
P
f ∈Tρn Fn |In | s∈I
n
(
)
A2 |In |1/N ε2
≤ A1 exp {κn (ε/32, ρn)} exp −
2
16ρ4n + 4ρ2n ε (log |In |)
!)
(
(5.2)
2
κn (ε/32, ρn)[16ρ4n + 4ρ2n ε (log |In |) ]
ε2 |In |1/N
,
A2 −
= A1 exp −
2
ε2 |In |1/N
16ρ4n + 4ρ2n ε (log |In |)
for suitable constants A1 and A2 . The weak consistency follows from (5.2): let ε > 0 be arbitrary but fixed, then
#
"
i
h
1 X
2
2
sup
|f (X(s)) − TL Y (s)| − E |f (X(eN )) − TL Y (eN )|
E
f ∈Tρn Fn |In | s∈I
n
(
)
Z ∞
A2 |In |1/N t2
≤ ε + A1 exp {κn (ε, ρn )}
exp −
dt → ε,
16ρ4n + 4ρ2n t (log |In |)2
ε
as n → ∞. Concerning the a.s. convergence of the estimate, we find that under the condition of α-mixing and station2
arity the random variables {|Y (s) − TL Y (s)| : s ∈ ZN } are ergodic, see Theorem B.4 in Krebs (2018a). This implies
that
X
a.s.
lim |In |−1
|Y (s) − TL Y (s)|2 = E |Y (eN ) − TL Y (eN )|2
n→∞
s∈In
for all L > 0. Furthermore, if additionally
4
ρ4n (log |In |) /|In |1/N → 0 as n → ∞,
Equation (5.2) remains summable for a sequence of index sets In(k) which satisfies the condition in (2.1). Thus, an
application of the Borel-Cantelli Lemma to the same equation yields that the estimator is strongly universally consistent.
This finishes the case for α-mixing data.
It follow the considerations for the case of β-mixing data. Again, we assume that ρn > L. Therefore we use the
partition of In which is provided by Lemma 5.1 for the choice q = ⌈2 log |In |/c1 ⌉. As in Tran (1990) we assume that
Ri q = ni for each i = 1, . . . , N . We use the coupled random field (X ∗ , Y ∗ ) to obtain the estimator m̂∗n of the regression
function m. We split the integrated error as follows
Z
Z
Z
2
∗
2
|m̂n − m| dµX ≤ 2
|m̂n − m| dµX + 2
|m̂∗n − m̂n |2 dµX
(5.3)
R
d
R
R
d
d
Exploiting the properties of (X ∗ , Y ∗ ), we find that the second term is at most
Z
|m̂∗n − m̂n |2 dµX ≤ 4ρ2n 1{(X ∗ (s), Y ∗ (s)) 6= (X(s), Y (s)) for one s ∈ In } .
R
d
19
Using that β(q) ∈ O(|In |−2 ), we have the following bound for the expectation
Z
∗
2
|m̂n − m̂n | dµX ≤ 4ρ2n β(q) ≤ Cρ2n /|In |2 → 0,
E
R
d
n → ∞,
where we use that by assumption ρ4n /|In | vanishes. Moreover, we have
∞
X
k=1
P
Z
R
d
|m̂∗n(k) − m̂n(k) |2 dµX > ε
≤ 4ε−1
∞
X
k=1
ρ2n(k) β(q(n(k))) ≤ C
∞
X
k=1
|In(k) |−1.5 < ∞.
Hence, Rd |m̂∗n − m̂n |2 dµX both vanishes in the mean and a.s. Consequently, the first integral in (5.3) remains and we
need to study the probability in (5.1) in this scenario, it equals
2N X
R
i
h
X
X
1
2
2
P sup
|f (X ∗ (s)) − TL Y ∗ (s)| − E |f (X ∗ (eN )) − TL Y ∗ (eN )|
> ε
f ∈Tρn Fn |In | l=1 u=1
s∈I(l,u)
2N
R
i
h
X
X
X
|I
|ε
1
n
2
2
P sup
|f (X ∗ (s)) − TL Y ∗ (s)| − E |f (X ∗ (eN )) − TL Y ∗ (eN )| > N .
≤
2 R
f ∈Tρn Fn R u=1
R
l=1
s∈I(l,u)
(5.4)
Using the properties of the coupled process and the stationarity of (X, Y ), we see that the summands over the index
sets I(l, 1), . . . , I(l, R) are i.i.d. for each l = 1, . . . , 2N . Consequently, we can apply Theorem 9.1 in Györfi et al.
(2002). Note that in the proof of this theorem it is only necessary that the data is independent but not that it is identically
distributed. Hence, we obtain for Equation (5.4) the bound
ε2 |In |
ε2 |In |
εq N
εq N
N +3
HTρn Fn
exp − 2N +13 N 4
exp − 2N +13 N 4 ≤ 2
2
HH n
2N +3
2
q ρn
2N +3 (4ρ2n )
2
q ρn
N
2
εq
ε |In |
= 2N +3 exp κn
, ρn − 4N +13
2N +3
2
(c1 )−N (log |In |)N ρ4n
N +3
(5.5)
using the definition of κn and of the block size q. Note that the factor q N inside κn can be neglected as it only decreases
κn marginally if ε > 0 is fixed. Now, the same computations as in the case of α-mixing data yield the result. We do not
go into the details.
The proof of Corollary 2.4 requires the concept of the Vapnik-Chervonenkis-dimension (VC-dimension). The definition of the VC-dimension is rather technical and can be found in the book of Györfi et al. (2002), Definition 9.6.
Proof of Corollary 2.4. It remains to consider some technical issues. Clearly, the map
RKn × Ω ∋ (a, ω) 7→
Kn
X
i=1
ai fi (X(s, ω)) is B(RKn ) ⊗ A-measurable.
The desired measurability of the map in Equation (2.6) follows from the fact that for any measurable function g on a
product space (S × T, S ⊗ T) the set
t ∈ T : sup g(s, t) > c = t ∈ T ∃s ∈ S : g(s, t) > c
s∈S
= πTS×T {(s, t) ∈ S × T : g(s, t) > c} ∈ T,
where πTS×T is the projection from S × T onto T .
Furthermore, the Vapnik-Chervonenkis-dimension is at least 2 if Kn ≥ 2. Indeed, choose functions f1 and f2 .
Without loss of generality, there is an x̄ in Rd and an a in R such that af1 (x̄) = f2 (x̄) > 0. Since f1 and f2 are
20
linearly independent, exactly one of the following three cases occurs: (1) either there are x1 and x2 in a neighborhood
of x̄ such that af1 (x1 ) > f2 (x1 ) and f2 (x2 ) > af1 (x2 ), (2) or af1 = f2 on U and af1 > f2 on Rd \ U , where
U ⊂ Rd contains x̄, (3) or f2 = af1 on U and f2 > af1 on Rd \ U . In the last two cases we can modify a such that
we achieve the first case, by linear independence. Thus, the two points pi := (xi , ti ) (for i = 1, 2) with the property that
af1 (x1 ) > t1 > f2 (x1 ) and f2 (x2 ) > t2 > af1 (x2 ) are shattered by the set of all subgraphs of the linear space hf1 , f2 i,
hence, Vhf1 ,...,fn i+ ≥ Vhf1 ,f2 i+ ≥ 2. Consequently, the conditions of Theorem A.1 are satisfied. We have
κn (ε, ρn ) = log HTρn Fn
ε
4ρn
!
V
24eρ2n (Tρn Fn )+
16eρ2n
≤ log 3
log
ε
ε
e 2
ρ4n = O(Kn log ρn ).
≤ log 3 + (Kn + 1) log (24)2
ε
The statement follows now from Theorem 2.3.
We need another proposition and a piece of notation to prove the rate of convergence of the regression estimator.
Notation 5.3. Let f be a real-valued function on Rd and let the distribution of the X(s) be given by µX . Let X ′ =
{X ′ (s) : s ∈ ZN } be an i.i.d. ghost sample with the same marginals as X. Moreover, X ∗ is constructed for each
†
∗
n ∈ NN
+ as in Lemma 5.1 and the random field X is an independent copy of X . Define the following empirical
L2 -norms
kf k|In | :=
|In |−1
and kf k∼
|In | :=
as well as
X
f (X(s))2
s∈In
(2|In |)−1
∗
kf k|In |
:=
X
!1/2
,
′
′
f (X(s))2 + f (X (s))2
s∈In
|In |−1
X
s∈In
|In |−1
kf k|In | :=
f (X ∗ (s))2
!1/2
!1/2
and
X
′
f (X (s))2
s∈In
†
kf k|In |
:=
|In |−1
!1/2
X
s∈In
f (X † (s))2
!1/2
.
Consider the random point measure ν with equal masses which is induced by the sample of the random field and of
P
the ghost sample (X(In ), X ′ (In ), i.e., ν = (2|In |)−1 s∈In δX(s) + δX ′ (s) . We abbreviate the ε-covering number of
a function class G w.r.t. 2-norm of ν by
′
N2 ε, G, (X(In ), X (In )) := N ε, G, k · kL2 (ν) .
The next two statements prepare the second main theorem of Section 2 which is Theorem 2.5. The first is intended
for α-mixing data, the second for β-mixing data.
Proposition 5.4. Assume that the random field X satisfies Condition 2.1 (α). Let G be a class of R-valued functions on
Rd which are all bounded by a universal constant B. Then for all ε > 0
!
!
√
′
2ε
, G, (X(In ), X (In ))
P sup kf k − 2 kf kIn > ε ≤ A1 N2
32
f ∈G
P,∞
(5.6)
!
!
4
1/N
A3 ε2 |In |
A2 ε |In |
+ exp −
· exp −
B2
B 4 + B 2 ε2 (log |In |)2
for constants 0 < A1 , A2 , A3 < ∞ which neither depend on the bound B, nor on ε, nor on the index set In .
Provided that the Vapnik-Chervonenkis dimension VG+ is at least 2 and that ε is sufficiently small, the bound from
21
Proposition 5.4 is non-trivial: we have with Proposition A.1
3 2
′
24 · 162 eB 2
16 eB
ε
√ , G, (X(In ), X (In ))
.
· log
≤ log 3 + VG+ log
log N2
ε2
ε2
16 2
P,∞
Proof of Proposition 5.4. Let {X(s) : s ∈ In } be a subset of the strongly mixing and stationary random field X and let
′
{X (s) : s ∈ In } be the corresponding ghost sample. We use the relation
!
!
!
ε
ε
′
′
P sup kf k − 2 kf kIn > ε ≤ P sup kf k − 2 kf kIn >
+ P sup kf kIn − kf kIn >
.
2
4
f ∈G
f ∈G
f ∈G
We only consider the second probability on the right-hand-side of the last inequality, bounds on the first probability
are given by the second term
Let
√ in the second line of (5.6) and are derived in2 Theorem 11.2 of Györfi et al. (2002).
′
of
G
with
respect
to
the
empirical
L
-norm
of
the
sample
X(I
),
X
(I
)
with
U1 , . . . , UH ∗ be an ε/(16 2)-covering
n
n
√
√
′
the definition H ∗ := N2 ε/(16 2), G, X(In ), X (In ) and Uk := {f ∈ G : kf − gk k∼
In < ε/(16 2)}, where the
′
covering functions are g1 , . . . gH ∗ . Note that H ∗ and the Uk are random and that both k · kIn and k · kIn are bounded by
√
∼
2 k · kIn . Then,
||H ∗ ||P,∞
X
′
′
ε
ε
P ∃f ∈ G : kf kIn − kf kIn >
P ∃f ∈ Uk : kf kIn − kf kIn > .
≤
4
4
(5.7)
k=1
Now, we use that kf kIn ≤
√
∼
2 kf kIn to obtain for f ∈ Uk the inequality
′
′
′
′
kf kIn − kf kIn = kf kIn − kgk kIn + kgk kIn − kgk kIn + kgk kIn − kf kIn
√
′
′
≤ kf − gk kIn + kgk kIn − kgk kIn + kf − gk kIn ≤ 2 2
′
ε
√ + kgk kIn − kgk kIn .
16 2
′
′
Hence, ∃f ∈ Uk : kf kIn − kf kIn > 4ε is a subset of kgk kIn − kgkIn > 8ε . Since the inequality a − b > c implies
a2 − b2 > c2 for a, b, c ≥ 0, we get for the probabilities on the right-hand-side of Equation (5.7) the following bounds
2
2
′
′
ε
ε2
kgk kIn − kgk kIn >
P kgk kIn − kgk kIn >
≤P
8
64
!
io
h
′
′
1 X
1 Xn
ε2
2
2
2
2
−
gk (X (s)) − E gk (X (eN ))
gk (X(s)) − E gk (X(eN ))
>
≤P
|In |
|In |
64
s∈In
s∈In
!
i
h
′
′
ε2
1 X
≤P
gk (X (s))2 − E gk (X (eN ))2 >
|In |
128
s∈In
!
(5.8)
ε2
1 X
2
2
.
+P
>
gk (X(s)) − E gk (X(eN ))
|In |
128
s∈In
The first term from (5.8) can be bounded by Hoeffding’s inequality, we have
!
i
h
′
′
|In |
ε2
1 X
2
2
≤ 2 exp −Cε4 4 .
>
gk (X (s)) − E gk (X (eN ))
P
|In |
128
B
(5.9)
s∈In
We apply Proposition A.3 to the second term and obtain that
P
ε2
1 X
gk (X(s))2 − E gk (X(s))2 >
|In |
128
s∈In
22
!
≤ exp −
Cε4 |In |1/N
B 4 + B 2 ε2 (log |In |)
2
!
.
(5.10)
Obviously, the bound in (5.10) dominates the bound in (5.9). This finishes the proof.
The next proposition is a generalization of Theorem 11.2 of Györfi et al. (2002) for β-mixing data.
Proposition 5.5. Assume that the random field X satisfies Condition 2.1 (β). Let G be a class of R-valued functions on
Rd which are all bounded by B ∈ R+ . Let n be sufficiently large such that both 8/c1 log |In | < min{ni : i = 1, . . . , N }
and C ∗ ≤ 23N −1 (c1 )−N (log |In |)N where the constant C ∗ is defined in (5.13). Then for all ε > 0
!
!
√
′
2ε
∗
N
P sup kf k − 2 kf kIn ≥ ε ≤ 3 · 2 N2
, G, (X(In ), X (In ))
32
f ∈G
P,∞
(5.11)
ε2 |In |
· exp −
N
25N +5 B 2 c−N
1 (log |In |)
Proof of Proposition 5.5. Set q = ⌈2/c1 log |In |⌉ and apply Lemma 5.1. We obtain a partition of In given by {I(l, u) :
l = 1, . . . , 2N , u = 1, . . . , R} such that we can write
N
N
|In |−1
X
f (X ∗ (s))2 = R−1
R
2 X
X
Z ∗ (l, u)2
and
l=1 u=1
s∈In
|In |−1
X
f (X † (s))2 = R−1
R
2 X
X
Z † (l, u)2 ,
l=1 u=1
s∈In
1/2
1/2
P
P
. Note that
and Z † (l, u) = R|In |−1 s∈I(l,u) f (X † (s))2
where Z ∗ (l, u) = R|In |−1 s∈I(l,u) f (X ∗ (s))2
0 ≤ Z ∗ (l, u), Z † (l, u) ≤ B.
In the following, let f˜ be a function in G such that f˜ − 2 f˜
In
≥ ε if there is such a function. Otherwise, f˜ is
any other function. We write P for the conditional probability measure and E∗ for the conditional expectation given
the data X ∗ (In ).
The remaining proof is a modification of Theorem 11.2 in Györfi et al. (2002) and is split in three steps. In the first
step, we show that
!
!
ε
3
∗
∗
†
P sup kf k − 2 kf kIn ≥ ε ≤ P sup kf kIn − kf kIn ≥
,
(5.12)
2
4
f ∈G
f ∈G
∗
if B 2 /ε2 ≤ |In |/(22N +6 C ∗ ) where
C ∗ :=
√ p
2 1 + CP (1 + CN β̄∞ ).
(5.13)
Here CP is a uniform
p bound of the essential suprema of the Radon-Nikodým derivatives in (2.2) and the factor β̄∞
P∞
equals k=0 k N −1 β(k) < ∞; additionally, the constant CN depends on the lattice dimension N and is given below.
For this result, we need that
2 !!
2
2
†
2
†
ε
ε
∗
∗
P 2 f˜ + ≥ f˜ ≥ 1 − P 3 f˜ + ≤ 4 f˜ − f˜
(5.14)
2
4
In
In
Indeed, this follows with some calculations (see the proof of Theorem 11.2 Györfi et al. (2002)). Furthermore, we need
a result, which follows using the β-mixing property and a lemma in Krebs (2018b),
h
i √ p
h
i X
X
β(ks − tk∞ )1/2
E∗ f˜(X † (s))2 f˜(X † (t))2 ≤ 2 1 + CP E∗ f˜(X † (eN ))4
s,t∈I(l,u)
s,t∈I(l,u)
√ p
≤ 2 1 + CP (1 + CN β̄∞ )B 2 f˜
for a certain constant CN which depends on the lattice dimension N .
23
2
qN ,
Moreover, using that for a fixed l the blocked random variables {X † (I(l, u)) : u = 1, . . . , R} are independent, the
probability on the right-hand-side of (5.14) is at most
2
P
P
∗
˜(X † (s))2
N
∗ P
†
2
E
f
˜
2
R
s∈I(l,u)
l,u
s∈I(l,u) f (X (s))
2N +4 X X
16 Var
≤
2
2
2
2
2
2
|In |
|In |
2
2
l=1 u=1
3 f˜ + ε4
3 f˜ + ε4
2
N
∗ 2 ˜
N
22N +6 C ∗ B 2
2N +4 2 RC B f q
≤
.
≤
2
2
|In |2
3|In |ε2
2
3 f˜ + ε
(5.15)
4
This last term is at most 1/3 if |In | ≥ 22N +6 C ∗ B 2 /ε2 . In particular, the right-hand-side of (5.14) is then at least 2/3.
Using once more a result of Györfi et al. (2002), we have that
!
†
∗
ε
ε
†
∗
+ ≥ f˜
≥ E 1 f˜ − 2 f˜
≥ ε P∗ 2 f˜
P sup kf k|In | − kf k|In | ≥
4
2
|In |
|In |
f ∈G
Consequently, (5.12) follows from this last inequality
if |In | ≥ 22N +6 C ∗ B 2 /ε2 .
√
In the second step, consider an ε/(16 2)-covering of G with respect to the empirical L2 -norm of the sample
∗
X (In ), X † (In ) . It follows as in the proof of Proposition 5.4 that
P sup
f ∈G
where ||H ∗ ||P,∞ ≤ N2
†
kf kIn
√
−
∗
kf kIn
ε
≥
4
!
′
2ε
32 , G, (X(In ), X (In ))
≤
||H ∗ ||P,∞
X
k=1
P,∞
P kgk k†In − kgk k∗In ≥
ε
,
8
(5.16)
.
Consequently, it remains to bound the last probability in (5.16). This is done in the third step. Consider a function f
such that |f | ≤ B, then
1/2
1/2
X
X
ε
ε
P kf k†In − kf k∗In ≥
= P R−1
− R−1
Z † (l, u)2
≥
Z ∗ (l, u)2
8
8
l,u
l,u
P
†
2
∗
2
−1
R
ε
l,u Z (l, u) − Z (l, u)
= P
1/2 ≥ 8
1/2
P
P
+ R−1 l,u Z ∗ (l, u)2
R−1 l,u Z † (l, u)2
PR
2N
−1
†
2
∗
2
X
R
ε
u=1 Z (l, u) − Z (l, u)
≤
P
1/2 ≥ 2N +3
1/2
P
P
R
R
l=1
+ R−1 u=1 Z ∗ (l, u)2
R−1 u=1 Z † (l, u)2
!
2N
R
R
1/2
1/2
X
X
X
ε
−1
†
2
−1
∗
2
=
P R
(5.17)
Z (l, u)
− R
Z (l, u)
≥ N +3 .
2
u=1
u=1
l=1
Next, we use a trick which induces additional randomness and which can be applied to the last probabilities. W.l.o.g.
we consider the case l = 1. Then, choose i.i.d. random variables V (1), . . . , V (R) which are uniformly distributed on
{−1, 1} and define
†
U (u) :=
(
Z † (1, u)
∗
Z (1, u)
if V (u) = 1
if V (u) = −1
and
24
∗
U (u) :=
(
Z ∗ (1, u)
†
Z (1, u)
if V (u) = 1
if V (u) = −1.
As the Z ∗ (1, u) and Z † (1, u) are independent and have for each u identical distributions, we can replace their distribution with the distribution of the U ∗ (u) and U † (u). Now, write P∗ for the probability measure conditioned on
σ(Z † (1, u), Z ∗ (1, u), u = 1, . . . R). Then if l = 1, the probability in (5.17) equals
P
R
R
1/2
1/2
X
X
R−1
− R−1
U † (u)2
≥
U ∗ (u)2
=E
P∗
!
2N +3
)#
(
R
R
R
1/2
1/2
X
X
X
ε
−1
†
2
∗
2
−1
−1
†
2
∗
2
R
V (u)(Z (1, u) − Z (1, u) ) ≥ N +3
.
R
+ R
Z (1, u)
Z (1, u)
2
u=1
u=1
u=1
u=1
"
ε
u=1
Due to the independence between the V (u) and the (Z † (1, u), Z ∗ (1, u)), we can bound the inner conditional probability
with Hoeffding’s inequality and obtain the bound
!
PR
†
2
∗
2
Z
(1,
u)
+
Z
(1,
u)
Rε2
ε2 |In |
u=1
2 exp − 2N +5 PR
≤ 2 exp − 3N +5 2 N
†
2
∗
2 2
2
2
B q
u=1 |Z (1, u) − Z (1, u) |
ε2 |In |
.
(5.18)
≤ 2 exp −
N
25N +5 B 2 c−N
1 (log |In |)
N
We use for the last inequality the three relations R ≥ |In |/(2N q N ) and q N ≤ 22N /cN
1 (log |In |) as well as
|Z † (1, u)2 − Z ∗ (1, u)2 |2 ≤ Z † (1, u)4 + Z ∗ (1, u)4 ≤ B 2 (Z † (1, u)2 + Z ∗ (1, u)2 ).
Combining Equations (5.12) to (5.18) yields the result given in (5.11) if |In | ≥ 22N +6 C ∗ B 2 /ε2 . Otherwise in the case
that |In | < 22N +6 C ∗ B 2 /ε2 , the exponential in (5.18) is at least e−1 if C ∗ ≤ 23N −1 (c1 )−N (log |In |)N , hence, the
right-hand-side of (5.11) is greater than one; so the inequality is also true in this case.
Proof of Theorem 2.5. We begin with the case of α-mixing data and use the decomposition
Z
2
2
|m̂n − m|2 dµX = km̂n − mk = km̂n − mk − 2 km̂n − mkIn + 2 km̂n − mkIn
R
d
≤ 2 max km̂n − mk − 2 km̂n − mkIn , 0
2
+ 8 km̂n − mkIn
2
(5.19)
The exponentially decreasing mixing rates ensure that the norm of the conditional covariance matrix remains bounded
and that we can use Theorem 11.1 of Györfi et al. (2002) even in the case where the error terms ε(s) are not uncorrelated.
There is a constant C1 such that kCov(Y (In ) | X(In ))k2 ≤ C1 for all k ∈ N. Indeed, consider the operator norms for
matrices which are defined for a matrix A ∈ Ru1 ×u2 and p ∈ [1, ∞] by the corresponding
p-norm on Ru1 (resp. on Ru2 )
p
as kAkp = maxx∈Ru2 :kxkp =1 kAxkp . We have the norm inequality kAk2 ≤ kAk1 kAk∞ . As the covariance matrix
is symmetric, the ∞- and the 1-norm are equal. We consider a line (resp. a column) of the covariance matrix that contains
the conditional covariances of the Y (s). By assumption, the error terms satisfy E |ε(s)|2+γ < ∞ for some γ > 0.
We use Davydov’s inequality from Appendix A.2 and the bound on the mixing coefficients, α(k) ≤ c0 exp(−c1 k) for
certain c0 , c1 ∈ R+ . We obtain
X
| Cov(Y (s), Y (t) | X(In ))|
t∈In
2
≤ kςk∞
X
t∈In
2/(2+γ) X
2
α(ks − tk∞ )γ/(2+γ)
| Cov(ε(s), ε(t))| ≤ 10 kςk∞ E |ε(s)|2+γ
2/(2+γ)
2
≤ 10 kςk∞ c0 E |ε(s)|2+γ
t∈In
max1≤i≤N ni
X
u=0
(2u + 1)N − (2u − 1)N exp(−c1 γ/(2 + γ)u) ≤ C1 ,
for a universal constant C1 < ∞ and for all s ∈ In . Hence, kCov(Y (In ) | X(In ))k2 ≤ C1 . Thus, we find with Theorem
25
11.1 of Györfi et al. (2002), which is applicable to dependent data too, that
Z
i
h
Kn
2
+ inf
(f − m)2 dµX .
E km̂n − mkIn ≤ C1
|In | f ∈Fn Rd
(5.20)
Next, consider the expectation of the first term in Equation (5.19), it admits the upper bound
h
i
E max km̂n − mk − 2 km̂n − mkIn , 0 2
Z ∞
P max km̂n − mk − 2 km̂n − mkIn , 0 2 > u du
≤v+
Zv ∞
√
P ∃f ∈ TL Fn : kf − mk − 2 kf − mkIn > u du,
≤v+
(5.21)
v
for each v > 0 and if |In | is large enough.
We apply Proposition 5.4; note that the second exponential term in (5.6)
so we only consider the first
is negligible,
2(Kn +1)
2
term here. We find with Proposition A.1 that the covering number is in O L /v
provided that v < 162 L2 ;
w.l.o.g. this is the case. Hence, (5.21) can be bounded by
!
2 2(Kn +1) Z ∞
A2 u2 |In |1/N
L
exp −
v + A1
du.
(5.22)
2
v
L4 + L2 u (log |In |)
v
Define v := Kn log(|In |)/|In |1/(2N ) which converges to zero by assumption. One finds that (5.22) is in O(v). Combining this result with Equations (5.19) and (5.20) implies the assertion for α-mixing data.
Next, we consider the case of β-mixing data. For that reason we need the coupled process (X ∗ , Y ∗ ) obtained from
Lemma 5.1 for q = ⌈2/c1 |In |⌉. We also compute the truncated least-squares estimate for the coupled regression problem
and denote it by m̂∗n . Then the following upper bound of Equation (5.19) is true in terms of the estimate m̂∗n
4 max km̂∗n − mk − 2 km̂∗n − mkIn , 0
2
2
2
2
+ 16 km̂n − mkIn + 16 km̂n − m̂∗n kIn + 2 km̂n − m̂∗n k .
(5.23)
Consider the expectation of the last two terms: we have
i
h
h
i
E km̂n − m̂∗n k2In + E km̂n − m̂∗n k2 ≤ 2N +3 L2Rβ(q) = o(|In |−1 ),
this follows from the coupling property. Consequently, these two terms are negligible. A bound on the expectation of
the second term in (5.23) has already been established in (5.20). We can bound the expectation of the first term in (5.23)
similar as in the case of α-mixing data in (5.21) but this time using Proposition 5.5. Thus, instead of (5.22) and if |In | is
large enough, we obtain for the expectation of this term the bound
v + A1
L2
v
2(Kn +1) Z
v
∞
exp −
u |In |
5N
+5
2
N
2
L c−N
1 (log |In |)
du,
(5.24)
where v is positive and again A1 is a positive constant. One finds that for the choice v = Kn (log |In |)N +2 /|In | both
terms in (5.24) are in O(v). This proofs the result in the case of β-mixing data.
We come to the proofs of the theorems in Section 3.
R
S
Proof of Theorem 3.3. If j∈Z Uj is not dense in Lp (µ), there is a 0 6= g ∈ Lq (µ) which satisfies Rd f g dµ = 0 for all
S
f ∈ j∈Z Uj where q is Hölder conjugate to p. We show that the Fourier transform of g is zero which contradicts the
S
assumption that g 6= 0. This proves in particular that j∈Z Uj is dense. Consider the Fourier transform of this element
26
g which we define here for reasons of simplicity as
Fg :
R → C, ξ 7→
d
Z
R
g(x) eihx,ξi µ(dx),
d
where h·, ·i is the Euclidean inner product on Rd .
Since the scaling function Φ is of the form Φ = ⊗di=1 ϕ and ϕ is a compactly supported one-dimensional scaling
function, we can assume that the support of Φ is contained in the cube [0, A]d for some A ∈ N+ . Choose 1 > ε > 0
arbitrary, there is an n ∈ N such that we have for Q := [−An, An]d
µ(Rd \ Q)1/p < ε
3 · 2d−1 max(kgkLq (µ) , 1)).
Consider ξ ∈ Rd arbitrary, then we get with the assumed properties of g that
|Fg(ξ)| ≤
Z
R
d
(coshx, ξi − F1 (x))g(x) µ(dx) +
Z
R
d
(sinhx, ξi − F2 (x))g(x) µ(dx)
(5.25)
S
S
for all F1 , F2 ∈ j∈Z Uj . We show that the first term in Equation (5.25) is smaller than ε for suitable F ∈ j∈Z Uj ;
the second term can be treated in the same way. Therefore, we use several times the trigonometric identities sin =
− cos · + π2 , as well as, cos(α + β) = cos α cos β − sin α sin β: we can split cosh · , ξi in 2d−1 terms as coshx, ξi =
P2d−1
i=1 bi cos(ξ1 x1 + ai,1 ) · . . . · cos(ξd xd + ai,d ), where the bi are in {−1, 1}. First, we prove that each of the functions
cos(ξk · +ai,k ) can be uniformly approximated on finite intervals. Indeed, define the kernel
K:
R2 → R, (x, y) 7→
X
k∈
Z
ϕ(x − k) ϕ(y − k)
and the associated linear wavelet projection operator Kj for j ∈ Z by
Kj : L2 (λ) → Uj ,
f 7→
XD
k∈
Z
E
f, 2j/2 ϕ(2j · −k) 2j/2 ϕ(2j · −k).
Then, K satisfies the moment condition M (N ) from Härdle et al. (2012) for N = 0: since ϕ is a scaling function, we
R
P
have R K( · , y)dy = k∈Z ϕ( · − k) ≡ 1. Furthermore,
|K(x, y)| =
X
k∈
Z
2
ϕ(x − k) ϕ(y − k) ≤ (A + 1) kϕk∞ 1{|x−y|≤A} =: F (x − y),
where we assume w.l.o.g. that supp ϕ ⊆ [0, A]. Thus, F is integrable w.r.t. the Lebesgue measure λ and K satisfies
the moment condition M (0). Next, let I(i, k) ⊇ [−An, An] be a finite interval such that cos(ξk · +ai,k ) is zero at the
boundary of I(i, k). Then by Theorem 8.1 and Remark 8.4 in Härdle et al. (2012) the uniformly continuous restriction
cos(ξk · +ai,k ) 1I(i,k) can be approximated in L∞ (λ) with elements from some Uj , i.e.,
cos(ξk · +ai,k ) 1I(i,k) − Kj cos(ξk · +ai,k ) 1I(i,k)
L∞ (λ)
→ 0.
Thus, if ε̃ > 0 is arbitrary but fixed, we can choose for each factor cos(ξk · +ai,k ) 1I(i,k) an approximation fi,k in some
Uj such that cos(ξk · +ai,k )1I(i,k) − fi,k L∞ (λ) ≤ ε̃. This implies that for each of the i = 1, . . . , 2d−1 products we
have
cos(ξ1 x1 + ai,1 )1I(i,1) · . . . · cos(ξd xd + ai,d )1I(i,d) − fi,1 ⊗ . . . ⊗ fi,d
≤ (1 + ε̃)d − 1 ≤ dε̃edε̃ ≤ ded ε̃.
L∞ (λ)
This means that the d-dimensional approximation follows from the one-dimensional approximations.
27
(5.26)
bi fi,1 ⊗ . . . ⊗ fi,d and ε̃ := ε/ 3 · 2d−1 ded kgkLq (µ) , then we arrive at
P2d−1
Set now F1 :=
i=1
Z
≤
(cos hx, ξi − F1 (x)) g(x) µ(dx)
RZ
d
| cos hx, ξi − F1 (x)| |g(x)|µ(dx) +
Q
Z
R
d \Q
| cos hx, ξi − F1 (x)| |g(x)|µ(dx)
(5.27)
We consider the terms in (5.27) separately. We can estimate the first term by
Z
Q
| cos hx, ξi − F1 (x)| |g(x)|µ(dx) ≤
d−1
2X
i=1
Z
Q
ε
ded ε̃ |g(x)| µ(dx) ≤ 2d−1 ded kgkLq (µ) ε̃ = .
3
(5.28)
Likewise, for the second term we infer that
Z
| cos hx, ξi − F1 (x)| |g(x)|µ(dx)
R
≤
d \B
d−1
2X
i=1
Z
...+
d
Y
R
d \B
d−1
2X
i=1
k=1
Z
R
d \B
!
cos(ξk xk + ai,k ) 1×dk=1 I(i,k) −
d
Y
k=1
d
Y
k=1
!
fi,k (xk ) |g(x)| µ(dx) + . . .
cos(ξk xk + ai,k ) 1Rd \×dk=1 I(i,k) |g(x)| µ(dx)
1
≤ 2d−1 ded ε̃ kgkLq (µ) µ Rd \ B p + 2d−1 kgkLq (µ) µ
ε
ε
ε
+ .
= ·
3 3 · 2d−1 max(kgkLq (µ) , 1) 3
Rd \ B
p1
(5.29)
All in all, we have when combining Equations (5.28) and (5.29) that (5.27) is less than ε as desired.
R
2
Proof of Theorem 3.4 and of Theorem 3.5. We prove that inf f ∈Fn ,kf k∞ ≤ρn Rd |f − m| dµX → 0. Let ε > 0. Since
S
U is dense in L2 (µX ), there is a function f and a j0 ∈ N such that for all j ≥ j0 , we have f ∈ Uj and
R j∈N j 2
Rd |f − m| dµX < ε/4. We can write for each level j(n)
f=
X
aj,γ Ψj,γ +
γ∈Kn
for coefficients aj,γ ∈ R. Set gn :=
P
γ ∈K
/ n
X
aj,γ Ψj,γ
γ ∈K
/ n
aj,γ Ψj,γ . The support of the gn decreases monotonically to zero:
{gn 6= 0} ⊆ x ∈ Rd : M j x − γ ∈ [0, L]d , kγk∞ > wn
⊆ x ∈ Rd : M j x ∞ ≥ kγk∞ − L, kγk∞ > wn ⊆ x ∈ Rd : M j x
⊆ x ∈ Rd : S −1 2 (ζmax )j kSk2 kxk2 ≥ wn − L ↓ ∅ (n → ∞),
2
≥ wn − L
by the assumption that (ζmax )j /wn → 0 as n → ∞. Furthermore, there is a k1 ∈ N such that we have the estimate
R
d
d
2
Rd f 1 R \ [−k1 , k1 ] dµX < ε/4 for all k ≥ k1 . Hence, there is a k2 ∈ N such that both
[−k1 , k1 ]d ⊆
[
supp Ψj,γ and f
γ∈Kn
1 [−k1 , k1 ]d
∞
≤ ρn
for all k ≥ k2 . In particular, the function f 1 [−k1 , k1 ]d is admissible in the sense that it is in Tρn Fn and that
R
d
2
Rd |m − f 1 [−k1 , k1 ] | dµX < ε as desired.
For the second part, it remains to compute κn (ε, ρn ) = log HTρn Fn (ε/(4ρn )). We use the bound which is given in
28
Proposition A.1, we have
HTρn Fn (ε/(4ρn )) ≤ 3 exp 2((2wn + 1)d + 1) log(384e ρ2n/ε) , i.e., κn (ε, ρn ) = O wnd log(ρn )
for ε > 0 which is arbitrary but fixed. The statement concerning the consistency properties follows now from Theorem 2.3. The statement which concerns the rate of convergence follows from Theorem 2.5.
A
Exponential Inequalities for Dependent Sums
In this section, we give a short review on important concepts which we use throughout the article. We begin with a result
concerning the covering number of a function class G of real-valued functions on Rd . Denote the class of all subgraphs
of this class G by G+ := (z, t) ∈ Rd × R : t ≤ g(z) : g ∈ G and the Vapnik-Chervonenkis-dimension of G+ by
VG+ . In this case Condition 2.2 is satisfied if ε is sufficiently small and if the Vapnik-Chervonenkis dimension of G+ is
at least two. More precisely, we have the following statement
Proposition A.1 (Haussler (1992)). Let [a, b] ⊆ R. Let G be a class of uniformly bounded real valued functions
g : Rd → [a, b] such that VG+ ≥ 2. Let 0 < ε < (b − a)/4. Then for any probability measure ν on (Rd , B(Rd ))
3e(b − a)p
2e(b − a)p
.
log
log N ε, G, k · kLp (ν) ≤ log 3 + VG+ log
εp
εp
In particular, VG+ ≤ r + 1 in the case where G is an r-dimensional linear space.
Davydov’s inequality relates the covariance of two random variables to the α-mixing coefficient:
Proposition A.2 (Davydov (1968)). Let (Ω, A, P) be a probability space and let G, H be sub-σ-algebras of A. Set
α := sup{|P(A ∩ B) − P(A)P(B)| : A ∈ G, B ∈ H}. Let p, q, r ≥ 1 be Hölder conjugate, i.e., p−1 + q −1 +
r−1 = 1. Let ξ (resp. η) be in Lp (P) and G-measurable (resp. in Lq (P) and H-measurable). Then |Cov(ξ, η)| ≤
10 α1/r kξkLp (P) kηkLq (P) .
In the remaining part of this appendix we derive upper bounds on the probability of events of the type
(
)
X
−1
sup |In |
g(Z(s)) − E [ g(Z(eN )) ] > ε ,
g∈G
(A.1)
s∈In
where G is a class of functions and Z is an α-mixing random field on ZN . (Here we assume that G is sufficiently regular,
so that this set is indeed measurable.)
Proposition A.3. Let the real valued random field Z satisfy Condition 2.1 (α). The Z(s) have expectation zero and are
′
bounded by B. Assume that n ∈ NN
+ satisfies min{ni : i = 1, . . . , N }/ max{ni : i = 1, . . . , N } ≥ C , for a constant
C ′ > 0. Then there is a constant A ∈ R+ which depends on the lattice dimension N , the constant C ′ and the bound on
the mixing coefficients but neither on n ∈ NN
+ , nor on ε, nor on B such that for all ε > 0
P |In |−1
X
s∈In
Z(s) ≥ ε ≤ exp −
A|In |1/N ε2
2
B + Bε(log |In |)(log log |In |)
.
Proof. One can apply the exponential inequality of Merlevède et al. (2009) for strongly mixing time series to the random
field Z as follows: consider a fixed n ∈ NN and set j ∗ = min{1 ≤ j ≤ N : nj = min{ni : i = 1, . . . , N }}. Then
define a time series Y by
X
Yk =
Zs , k = 1, . . . , nj ∗ .
s∈In ,sj∗ =k
We have that the Yk are bounded by B(C nj ∗ )N −1 . Moreover, this time series is strongly mixing with exponentially
decreasing mixing coefficients. The result follows now from Theorem 1 of Merlevède et al. (2009).
′
29
We can prove with the previous proposition an important statement
Theorem A.4. Assume that the conditions of Proposition A.3 are satisfied. Let G be a set of measurable functions
g : Rd → [0, B] for B ∈ [1, ∞) which satisfies Condition 2.2 and assume that (A.1) is measurable. Then there is a
constant A which is independent of ε, n and B such that for all ε > 0
!
X
−1
P sup |In |
g(Z(s)) − E [ g(Z(eN )) ] ≥ ε
g∈G
s∈In
(A.2)
ε
|In |ε2
A|In |1/N ε2
≤ 10HG
exp −
+ exp − 2
.
32
512B 2
B + Bε(log |In |)2
Proof of Theorem A.4. We assume the probability space to be endowed with the i.i.d. random variables Z ′ (s) for s ∈ In
which have the same marginal laws as the Z(s). We write for shorthand
Sn (g) :=
1 X
1 X
g(Z(s)) and Sn′ (g) :=
g(Z ′ (s)).
|In |
|In |
s∈In
s∈In
We can decompose the probability with these definitions as follows
!
P sup |Sn (g) − E [ g(Z(eN )) ]| ≥ ε
g∈G
≤P
ε
sup |Sn (g) − Sn′ (g)| ≥
2
g∈G
!
ε
sup |Sn′ (g) − E [ g(Z ′ (eN )) ]| ≥
2
g∈G
+P
!
.
Next, we apply Theorem 9.1 from Györfi et al. (2002) to second term on the right-hand side of (A.3) and obtain
!
ε
ε
|In |ε2
′
′
P sup |Sn (g) − E [ g(Z (eN )) ]| ≥
≤ 8HG
exp −
.
2
16
512B 2
g∈G
(A.3)
(A.4)
To get a bound on the first term of the right-hand side of (A.3), we use Condition 2.2 to construct an ε/32-covering.
ε
Write H ∗ := HG 32
for the upper bound on the covering number. Let gk∗ for k = 1, . . . , H ∗ be as in Condition 2.2.
Define
)
(
ε
1 X
.
g(Z(s)) − gk∗ (Z(s)) + g(Z ′ (s)) − gk∗ (Z ′ (s)) ≤
Uk := g ∈ G :
2|In |
32
s∈In
Then
P sup |Sn (g) −
g∈G
Sn′ (g)|
ε
≥
2
!
∗
≤
H
X
k=1
P sup |Sn (g) − Sn′ (g)| ≥
g∈Uk
ε
2
.
(A.5)
Thus, using the approximating property of the functions gk∗ , we get for each probability in (A.5)
7ε
∗
′
∗
P sup |Sn (g) −
≤ P |Sn (gk ) − Sn (gk )| ≥
16
g∈Uk
7ε
7ε
∗
′
∗
∗
∗
′
+ P |Sn (gk ) − E [ gk (Z (eN )) ]| ≥
.
≤ P |Sn (gk ) − E [ gk (Z(eN )) ]| ≥
32
32
Sn′ (g)|
ε
≥
2
The second term on the right-hand side of (A.6) can be estimated using Hoeffding’s inequality, we have
98 |In | ε2
7ε
≤ 2 exp −
.
P |Sn′ (gk∗ ) − E [ gk∗ (Z ′ (eN )) ]| ≥
32
322 B 2
30
(A.6)
We apply Proposition A.3 to the first term of Equation (A.6). Finally, we use that HG
ε
16
≤ HG
ε
32
.
B Details on Example 3.2
We show how to derive an isotropic MRA in d dimensions from a one-dimensional MRA. It is straightforward to show
that for a multiresolution analysis with corresponding scaling function Φ there is a sequence (a0 (γ) : γ ∈ Γ) ⊆ R such
R
P
that Φ ≡ γ∈Γ a0 (γ) Φ(M · −γ) and the coefficients a0 (γ) satisfy the equations a0 (γ) = |M | Rd Φ(x) Φ(M x−γ) dx
P
P
as well as γ∈Γ |a0 (γ)|2 = |M | = γ∈Γ a0 (γ).
S
In the first step, we show that the conditions for an MRA are satisfied. The spaces j∈Z Uj are dense: we have by
the definition
d
O
Uj′ = f1 ⊗ . . . ⊗ fd : fi ∈ Uj′ ∀i = 1, . . . , d .
Uj =
i=1
Note that the set of pure tensors g1 ⊗ . . . ⊗ gd : gi ∈ L2 (λ) is dense in L2 (λd ). Hence, it only remains to show
that we can approximate any pure tensor g1 ⊗ . . . ⊗ gd by a sequence (Fj ∈ Uj : j ∈ N+ ). Let ε > 0 and let
2
g1 ⊗ . . . ⊗ gd ∈ L2 (λd ) be a a pure tensor.
n Choose a sequence of pure tensors (fi,j :oj ∈ N+ ) converging to gi in L (λ)
for i = 1, . . . , d. Denote by L := sup kfi,j kL2 (λ) , kgi kL2 (λ) : j ∈ Z, i = 1, . . . , d < ∞. Then
kg1 ⊗ . . . ⊗ gd − f1,j ⊗ . . . ⊗ fd,j k2L2 (λd ) ≤ d2 L2(d−1) max kgi − fi,j k2L2 (λ) → 0 as j → ∞.
1≤i≤d
T
Pn
Furthermore, j∈Z Uj = {0}: Let f = i=1 ai fi,1 ⊗ . . . ⊗ fi,d be an element of each Uj . Then each fi,k is an element
of each Uj′ for all j and, hence, zero. The scaling property is immediate, too. Indeed,
f ∈ Uj ⇔ f =
n
X
i=1
⇔f =
ai fi,1 ⊗ . . . ⊗ fi,d and fi,k ∈ Uj′ ,
n
X
i=1
k = 1, . . . , d
ai fi,1 ⊗ . . . ⊗ fi,d and fi,k (2−j · ) ∈ U0′ ⇔ f (M −j · ) ∈ U0 .
The functions {Φ( · − γ) : γ ∈ Γ} form an orthonormal basis of U0 . We have for γ, γ ′ ∈ Zd
Z
Z
Φ(x − γ) Φ(x − γ ′ ) dx =
⊗dk=1 ϕ(xk − γk ) · ⊗dk=1 ϕ(xk − γk′ ) dx
R
=
R
d
d
d
Y
k=1
Z
R
ϕ(xk − γk )ϕ(xk − γk′ ) dxk = δγ,γ ′
Pn
Pn
and for each f ∈ U0 by definition f = i=1 ai ϕ( · − γ1i ) · . . . · ϕ( · − γdi ) = i=1 ai Φ( · − γ i ) for γ 1 , . . . , γ n ∈ Zd .
This proves that Φ together with the linear spaces Uj generates an MRA of L2 (λd ). It remains to prove that the wavelets
generate an orthonormal basis of L2 (λd ).
√
√
d
For an index k ∈ i=1 {0, 1}, define akl i by 2hl if ki = 0 and 2gl if ki = 1 for i = 1, . . . , d. Furthermore, set
ak (γ) := akγ11 · . . . · akγdd . Then, the scaling function and the wavelet generators satisfy
×
Ψk =
X
γ1 ,...,γd
akγ11 · . . . · akγdd ϕ(2 · −γ1 ) ⊗ . . . ⊗ ϕ(2 · −γd ) =
X
γ
ak (γ)Φ(M · −γ).
Since ϕ is a scaling function, the coefficients a0 (γ) of the scaling function Φ satisfy the relation
X
γ
a0 (γ) = 2
d/2
X
γ1 ,...,γd
h γ1 · . . . · h γd = 2
31
d/2
X
γ1
h γ1
!d
= 2d .
Furthermore, we have for j, k ∈ {0, 1}d and γ ∈ Γ,
X
X
X
ajγd′ ak2γdd +γ ′ = 2d δj,k δγ,0 .
ajγ1′ ak2γ1 1 +γ ′ · . . . ·
aj (γ ′ )ak (M γ + γ ′ ) =
1
1
d
d
′
′
′
γ
γ1
γd
Indeed, for s = 1, . . . , d and z := γs
X
ajγss′ ak2γs s +γs′
γs′
P
2 l hl g2z+l
2 P h h
l l 2z+l
=
P
2
l gl h2z+l
P
2 l gl g2z+l
if js = 0 and ks = 1,
if js = ks = 0,
if js = 1 and ks = 0,
if js = ks = 1.
Since the ϕ( · − z) form an ONB of U0′ , we have
Z
X
X
δz,0 =
ϕ(x − z) ϕ(x) dx =
hl hm δ2z+l,m =
hl h2z+l .
R
l,m
l
In the same way,
δz,0 =
Z
R
ψ(x − z)ψ(x) dx =
X
gl gm δ2z+l,m =
l,m
X
gl g2z+l .
l
In addition, as U1′ = U0′ ⊗ W0′ , we get
Z
X
X
X
ψ(x − z) ϕ(x) dx =
gl hm δ2z+l,m =
gl h2z+l =
gl−2z hl ,
0=
R
l,m
l
l
for all z ∈ Z. Hence, the conditions of Theorem 3.1 (Theorem 1.7 in Strichartz (1993)) are satisfied and the family of
L
functions {|M |j/2 Ψk (M j · −γ) : γ ∈ Γ, k = 1, . . . , |M | − 1} forms an ONB of Wj and L2 (λd ) = j∈Z Wj .
References
Y. Baraud, F. Comte, and G. Viennet. Adaptive estimation in autoregression or β-mixing regression via model selection.
The Annals of Statistics, pages 839–875, 2001.
A. R. Barron. Approximation and estimation bounds for artificial neural networks. Machine Learning, 14(1):115–133,
1994.
H. C. Berbee. Random walks with stationary increments and renewal theory. MC Tracts, 1979.
R. C. Bradley. A caution on mixing conditions for random fields. Statistics & Probability Letters, 8(5):489–491, 1989.
R. C. Bradley. Basic properties of strong mixing conditions. a survey and some open questions. Probability surveys, 2
(2):107–144, 2005.
L. D. Brown, T. T. Cai, and H. H. Zhou. Nonparametric regression in exponential families. The Annals of Statistics, 38
(4):2005–2046, 2010.
T. T. Cai. Adaptive wavelet estimation: a block thresholding and oracle inequality approach. The Annals of Statistics,
pages 898–924, 1999.
M. Carbon, L. T. Tran, and B. Wu. Kernel density estimation for random fields (density estimation for random fields).
Statistics & Probability Letters, 36(2):115–125, 1997.
32
M. Carbon, C. Francq, and L. T. Tran. Kernel regression estimation for random fields. Journal of Statistical Planning
and Inference, 137(3):778–798, 2007.
X. Chen and T. Christensen. Optimal uniform convergence rates for sieve nonparametric instrumental variables regression. Cowles Foundation Discussion Paper No. 1923, 2013.
N. Cressie. Statistics for spatial data. Wiley series in probability and mathematical statistics: Applied probability and
statistics. J. Wiley, 1993.
I. Daubechies. Ten lectures on wavelets. SIAM, 1992.
Y. A. Davydov. Convergence of distributions generated by stationary stochastic processes. Theory of Probability & Its
Applications, 13(4):691–696, 1968.
Y. A. Davydov. Mixing conditions for Markov chains. Teoriya Veroyatnostei i ee Primeneniya, 18(2):321–338, 1973.
D. L. Donoho and I. M. Johnstone. Minimax estimation via wavelet shrinkage. The Annals of Statistics, 26(3):879–921,
1998.
D. L. Donoho, I. M. Johnstone, G. Kerkyacharian, and D. Picard. Density estimation by wavelet thresholding. The
Annals of Statistics, pages 508–539, 1996.
J. Fan and I. Gijbels. Local polynomial modelling and its applications: monographs on statistics and applied probability
66, volume 66. CRC Press, 1996.
J. Franke and M. Diagne. Estimating market risk with neural networks. Statistics and Decisions, 24:1001–1021, 2006.
U. Grenander. Abstract inference. John Wiley & Sons, 1981.
Z. Guessoum and E. O. Saı̈d. Kernel regression uniform rate estimation for censored data under α-mixing condition.
Electronic Journal of Statistics, 4:117–132, 2010.
L. Györfi, M. Kohler, A. Krzyżak, and H. Walk. A distribution-free theory of nonparametric regression. Springer Berlin,
New York, Heidelberg, 2002.
L. Györfi, W. Härdle, P. Sarda, and P. Vieu. Nonparametric curve estimation from time series, volume 60. Springer,
2013.
M. Hallin, Z. Lu, and L. T. Tran. Local linear spatial regression. The Annals of Statistics, 32(6):2469–2500, 2004.
W. Härdle. Applied nonparametric regression. Cambridge University Press, 1990.
W. Härdle, G. Kerkyacharian, D. Picard, and A. Tsybakov. Wavelets, approximation, and statistical applications. Lecture
Notes in Statistics. Springer New York, 2012.
D. Haussler. Decision theoretic generalizations of the pac model for neural net and other learning applications. Information and Computation, 100(1):78–150, 1992.
K. Hornik. Approximation capabilities of multilayer feedforward networks. Neural networks, 4(2):251–257, 1991.
M. S. Kaiser, S. N. Lahiri, and D. J. Nordman. Goodness of fit tests for a class of markov random field models. The
Annals of Statistics, pages 104–130, 2012.
G. Kerkyacharian and D. Picard. Regression in random design and warped wavelets. Bernoulli, 10(6):1053–1105, 2004.
R. Kindermann and L. Snell. Markov random fields and their applications. The American Mathematical Society, 1980.
C. Kirch and J. Tadjuidje Kamgaing. A uniform central limit theorem for neural network-based autoregressive processes
with applications to change-point analysis. Statistics, 48(6):1187–1201, 2014.
33
M. Kohler. Nonlinear orthogonal series estimates for random design regression. Journal of Statistical Planning and
Inference, 115(2):491–520, 2003.
A. N. Kolmogorov and Y. A. Rozanov. On strong mixing conditions for stationary gaussian processes. Theory of
Probability & Its Applications, 5(2):204–208, 1960.
H. L. Koul. Behavior of robust estimators in the regression model with dependent errors. The Annals of Statistics, pages
681–699, 1977.
J. T. N. Krebs. Orthogonal series estimates on strong spatial mixing data. Journal of Statistical Planning and Inference,
193:15–41, 2018a.
J. T. N. Krebs. A large deviation inequality for β-mixing time series and its applications to the functional kernel
regression model. Statistics & Probability Letters, 133:50–58, 2018b.
R. Kulik and M. Raimondo. Wavelet regression in random design with heteroscedastic dependent errors. The Annals of
Statistics, 37(6A):3396–3430, 2009.
L. Li. Nonparametric regression on random fields with random design using wavelet method. Statistical Inference for
Stochastic Processes, 19(1):51–69, 2016.
L. Li and Y. Xiao. Wavelet-based estimation of regression function with strong mixing errors under fixed design.
Communications in Statistics-Theory and Methods, 46(10):4824–4842, 2017.
D. F. McCaffrey and A. R. Gallant. Convergence rates for single hidden layer feedforward networks. Neural Networks,
7(1):147–158, 1994.
F. Merlevède, M. Peligrad, E. Rio, et al. Bernstein inequality and moderate deviations under strong mixing conditions.
In High dimensional probability V: the Luminy volume, pages 273–292. Institute of Mathematical Statistics, 2009.
Y. Meyer. Wavelets and operators, volume 1. Cambridge university press, 1995.
M. Rosenblatt. A central limit theorem and a strong mixing condition. Proceedings of the National Academy of Sciences,
42(1):43–47, 1956.
G. G. Roussas and L. T. Tran. Asymptotic normality of the recursive kernel regression estimate under dependence
conditions. The Annals of Statistics, pages 98–120, 1992.
C. J. Stone. Optimal global rates of convergence for nonparametric regression. The Annals of Statistics, pages 1040–
1053, 1982.
R. S. Strichartz. Construction of orthonormal wavelets. In J. J. Benedetto, editor, Wavelets: mathematics and applications, pages 23–50. CRC press, 1993.
L. T. Tran. Kernel density estimation on random fields. Journal of Multivariate Analysis, 34(1):37–53, 1990.
C. S. Withers. Conditions for linear processes to be strong-mixing. Zeitschrift für Wahrscheinlichkeitstheorie und
Verwandte Gebiete, 57(4):477–480, 1981.
D. Yahia and F. Benatia. Nonlinear wavelet regression function estimator for censored dependent data. Afrika Statistika,
7(1):391–411, 2012.
34
| 10math.ST
|
Integrability conditions for Compound Random Measures
arXiv:1707.06768v2 [stat.ME] 12 Nov 2017
Alan Riva Palacio and Fabrizio Leisen
School of Mathematics, Statistics and Actuarial Sciences, University of Kent
Sibson Building, Canterbury, Kent CT2 7FS
Abstract
Compound random measures (CoRM’s) are a flexible and tractable framework for
vectors of completely random measure. In this paper, we provide conditions to guarantee the existence of a CoRM. Furthermore, we prove some interesting properties of
CoRM’s when exponential scores and regularly varying Lévy intensities are considered.
1 Introduction
Recently, a growing literature in Bayesian non-parametrics (BNP) proposed new priors which can take into account different features of the data, such as partial exchangeability, see De Finetti (1938). In this case, one would like to consider different densities for different groups instead of a single common density for all the data. After the seminal paper of MacEachern (1999), the problem of modeling a finite number of dependent densities has become an active area of research in Bayesian nonparametrics. A common approach is to construct BNP priors based on functions
of Completely Random Measures (CRM’s), see Lijoi and Prünster (2010). For example, special attention has been given to the normalization of CRM’s starting with
the work of Regazzini, Lijoi and Prünster (2003). Roughly speaking, a CRM is a
generalization of a subordinator, that is a process with independent increments and
almost surely increasing paths; for a full account of CRM’s see Kingman (1993).
This property is very helpful to derive the Laplace functional transform which is
the basis to derive some analytical quantities of interest such as posterior and predictive distributions or the Exchangeable Partition Probability Function (EPPF), see
James, Lijoi and Prünster (2009). To build more flexible priors in possibly higher dimensional spaces, vectors of dependent CRM’s are constructed for example in Leisen and Lijoi
(2011), Leisen, Lijoi and Spano (2013) and Zhu and Leisen (2015) where respectively
they build vectors of Poisson-Dirichlet and Dirichlet processes. These papers deal
with the Lévy-Copula approach introduced in Cont and Tankov (2004) to induce dependence among the components of the vector. In a similar fashion, Lijoi, Nipoti and Prünster
1
(2014a) introduce a vector of random probability measures where the dependence
arises by virtue of a suitable construction of the Poisson random measures underlying the CRM’s; furthermore, in the framework of survival analysis, Lijoi and Nipoti
(2014) introduce a new class of vectors of random hazard rate functions that are expressed as kernel mixtures of dependent CRM’s. Camerlenghi, Lijoi and Prünster
(2017) focus on partial exchangeable models which arise from hierarchical specifications of CRM’s.
Compound random measures (CoRM’s), introduced by Griffin and Leisen (2017), are
a flexible and tractable framework for many dependent random measures including
many of the superposition and Lévy copula approaches. They have recently been applied to modeling graphs for overlapping communities by Todeschini, Miscouridou and Caron
(2017). Griffin and Leisen (2017) and Griffin and Leisen (2018) described posterior
sampling methods for a particular class of normalized compound random measure
mixtures which exploits a representation of the Laplace transform of the CoRM through
a univariate integral of a moment generating function.
In this paper we aim to provide explicit existence conditions for CoRM’s in order
to guarantee the existence of the marginal Lévy intensities. On the other hand, we
prove that the resulting CoRM is well posed in the sense that it satisfies the usual
integrability condition for multivariate Lévy processes. Furthermore, this paper provides an interesting result for CoRM’s when regularly varying Lévy intensities are
considered. The paper closes highlighting a representation on the multivariate Lévy
intensity of a CoRM when the score distribution is the result of marginal independent
and identically distributed exponential scores.
The outline of the paper is as follows. Section 2 will set the scene by introducing
the basic definitions which are required in the CoRM setting. Section 3 is devoted
to prove our main results. Section 4 deals with CoRM’s built with regularly varying
Lévy intensities and exponential scores. Section 5 concludes.
2 Preliminaries
Let (Ω, F , P ) be a probability space and X a Polish space with corresponding Borel
σ-algebra X . We denote by MX the space of boundedly finite measures on the measurable space (X, X ) and by MX the associated Borel σ-algebra, see Appendix 2 in
Daley and Vere-Jones (2003) for technical details.
Definition 1. A random measure µ on X is called a completely random measure (CRM) if
for any n > 1 and disjoint sets A1 , . . . , An ∈ X the random variables µ( A1 ), . . . , µ( An ) are
mutually independent.
In the following we consider only CRM’s without fixed jumps, namely CRM’s of the
form µ = ∑i∞=1 wi δui for collections of random variables {wi }i∞=1 in R + and {ui }i∞=1 in
X. Such CRM’s can be characterized by their Laplace transform
h
i
R
− s f ( x ))ν̃ (ds,dx )
−µ( f )
E e
= e − R + × X (1− e
2
R
where µ( f ) = X f ( x )µ(dx ), f : X → R + is such that µ( f ) < ∞ and ν̃(ds, dx ) is a
measure in (R + × X, B(R + ) ⊗ X ) such that
Z
R+ ×X
min{1, s}ν̃ (ds, dx ) < ∞
(1)
for any bounded set X ∈ X . A measure ν̃ satisfying the condition displayed in
equation (12) is called the Lévy intensity of µ. We say that ν̃ is homogeneous when
ν̃(ds, dx ) = ρ(ds)α(dx ) with ρ a measure in (R + , B(R + )) and α a measure in (X, X ).
The notion of a completely random measure can be generalized to higher dimensions
in a similar fashion to Definition 1, see for instance Griffin and Leisen (2017). As a
result, we have a representation in terms of a Laplace functional transform. Precisely,
for a vector of completely random measures µ = (µ1 , . . . , µd ) on X we have that
R
h
i
− (R + ) d ×X (1− e−s1 f1 ( x )−···−s d f d ( x ) ) ν̃d (dss ,dx )
− µ1 ( f 1 )−···− µ d ( f d )
E e
=e
with f j : X → R + , j ∈ {1, . . . , d} such that µ j ( f j ) < ∞, where for g : X → R + we
R
have µ j ( g) = X g( x )µ j (dx ). The measure ν̃d in (R + )d × X, B((R + )d ⊗ X must be
such that
Z
(R + ) d × X
min{1, ks k}ν̃d (dss , dx ) < ∞
(2)
for any bounded set X ∈ X ; we call ν̃d a multivariate Lévy intensity. We set the
notation
νj ( A, X ) =
Z
(R + ) d −1
ν̃d (ds1 , . . . , ds j−1, A, ds j+1, . . . , dsd , X )
with j ∈ {1, . . . , d} and A ∈ B (R + ). We call νj the j-th marginal of the d-variate
Lévy intensity ν̃d ; it follows that for each j ∈ {1, . . . , d}, µ j has Lévy intensity νj . In
this framework we can define the concept of Compound Random Measure (CoRM). The
following definition differs from the one in Griffin and Leisen (2017) since it takes
into account the inhomogeneous case, where the locations and associated weights in
the CRM are not independent.
Definition 2. A Compound Random Measure (CoRM) is a vector of CRM’s whose Lévy
intensity is given by
Z
s
s
(3)
ν̃d (dss , dx ) =
z−d h 1 , . . . , d dss ν⋆ (dz, dx )
z
z
R+
where h, the score distribution, is a d-variate probability density function and, ν⋆ , the directing
Lévy measure, is a Lévy intensity.
By performing a simple change of variable we note that
Z
s
s
1
, . . . , d dss = 1.
z− d h
z
z
(R + ) d
3
Therefore, z−d h
sd
s1
,
.
.
.
,
z
z
can be seen as the density of a distribution function
ds1
dsd
H
.
,...,
z
z
This allows to write the multivariate Lévy intensity in equation (3) as
Z
dsd
ds1
ν⋆ (dz, dx ).
,...,
ν̃d (dss , dx ) =
H
z
z
R+
(4)
To write the Lévy intensity of a CoRM in terms of distribution functions rather than
densities will be convenient for the results proved in the next section.
3 Integrability conditions
The specification of a CoRM needs the initial choice of a score distribution and a directing Lévy measure. Although this sounds straightforward, it is necessary to check
that theses choices lead to a well defined CoRM. Otherwise, the risk is to perform a
Bayesian statistical analysis based on an ill-posed prior. In this section we look at two
important aspects of Definition 2:
1. we provide conditions on the score distribution and the directing Lévy measure
for the existence of the marginal Lévy intensities of a CoRM, see Theorem 1 and
Corollary 1,
2. we provide conditions on the score distribution and the directing Lévy measure
for the existence of the multivariate Lévy intensity of a CoRM, see Theorem 2.
Summing up, Theorem 1 and Corollary 1 are focusing on the marginal existence of a
CoRM. On the other hand, Theorem 2 focuses on the global existence of a CoRM. The
proofs of the theorems can be found in the supplementary material.
Theorem 1. Let H be a d-variate score distribution and ν⋆ a directing Lévy measure defining
a measure ν̃d as in (4) with corresponding marginals νj for j ∈ {1, . . . , d}. Let X be a bounded
set in X , then the measure νj satisfies the integrability condition (12) if and only if
Z
1 ⋆
P Sj ≥
ν (dz, dx ) < ∞
(5)
z
(0,1)× X
and
Z
1
zν⋆ (dz, dx ) < ∞.
P Sj ≤
z
[1,∞)× X
Furthermore if the marginal score H j satisfies that
1
≤ z ∀ z ∈ (0, ǫ) for some ǫ > 0
1 − Hj
z
4
(6)
(7)
and
1
<∞
lim zH j
z→∞
z
(8)
then conditions (15), (16) are satisfied with an arbitrary choice of the directing Lévy measure
ν⋆ .
As set in Definition 2, we usually work with CoRM’s given by a score with a probability density; in such case the following corollary to Theorem 1 follows.
Corollary 1. If S j has probability density function h j then conditions (17), (18) reduce to
lim
hj
1
z
2
z
<1
(9)
lim h j (ǫ) < ∞.
(10)
z →0
and
ǫ →0
The previous results concerned conditions for the marginals of a CoRM to be well
defined, now we focus on such a result for the CoRM. For a score density function
h and directing Lévy measure ν⋆ to properly define a CoRM we need to check the
condition (2) which takes the form
Z
R+ ×X
Z
(R + ) d
min{1, ks k}h
s
1
z
,...,
sd dss ⋆
ν (dz, dx ) < ∞
z zd
(11)
for bounded set X ∈ X . As stated at the beginning of this section, in the next theorem we provide conditions on the score distribution and the directing Lévy measure
for the existence of the multivariate Lévy intensity of a CoRM. This is equivalent to
provide conditions such that the latter inequality holds true.
Theorem 2. Consider a CoRM which satisfies conditions (15) and (16) for each marginal νj ,
j ∈ {1, . . . , d}, then the integrability condition (11) is satisfied.
We conclude this section by providing two examples of the use of the previous
results when considering Gamma and Beta distributed scored distributions.
Example 1: Gamma scores
We consider the marginal gamma score case. Let h be the d-variate probability density
of the score distribution; for j ∈ {1, . . . , d} we denote the j-th marginal density h j and
let it correspond to a Gamma distribution with shape and rate parameters α j , β j , i.e.
αj
h j (s ) =
β j s α j −1 e − β j s
Γ(α j )
5
1{s∈(0,∞)} .
We check the constraints (17), (18) by making use of Corollary 1 as we have probability
densities. To check (24) we see that
βj
αj
h j 1s
β j e− s
=0
= lim
lim
s →0 Γ ( α j ) s α j + 1
s →0
s2
and constraint (25) is satisfied for arbitrary Lévy directing measure ν⋆ whenever α j ≥
1, as in the examples presented in Griffin and Leisen (2017). However for α j < 1 the
associated CoRM will be well posed depending on the choice of ν⋆ . If for example we
take the directing Lévy measure to be the σ-stable, i.e.
ν⋆ (dz, dx ) =
σ
z1+ σ
dzdx
then constraint (15) in Theorem 1 can be reduced to
Z 1Z ∞
0
1
z
h j (s )
σ
z1+ σ
Z ∞Z 1
dsdz =
1
s
1
h j (s )
σ
zσ+1
dzds =
Z ∞
1
h j (s)(sσ − 1)ds < ∞,
which is always satisfied since h j is a Gamma density. On the other hand, condition
(16) in Theorem 1 becomes
Z ∞Z
1
1
z
0
h j (s )
σz
z1+ σ
Z 1Z
dsdz =
0
1
s
1
σ
h j (s) σ dzds =
z
Z 1
0
h j (s)σ(sσ−1 − 1)ds < ∞
which is not satisfied when α j + σ < 1.
Example 2: Beta scores
In the setting as above, if the marginal scores are Beta distributed, i.e.
h j (s ) =
s α j − 1 (1 − s ) β j − 1
1{s∈(0,1)}
B(α j , β j )
then constraint (17) becomes
lim
hj
s →0
1
s
2
s
( s − 1) β j −1
= 0,
s →0 s α j + β j B ( α j , β j )
= lim
so it is always satisfied; and condition (18) is satisfied whenever α j ≥ 1. We consider
again a σ-stable Lévy intensity for ν⋆ when α j < 1. Proceeding as in the previous
example, constraint (15) becomes
Z ∞Z 1
1
s
1
h j (s )
σ
zσ+1
dzds = 0 < ∞
so it always holds; and constraint (16) becomes
Z 1
0
h j (s)σ(sσ−1 − 1)ds < ∞,
which does not hold if α j + σ < 1.
6
4 Other interesting properties
The aim of this section is to investigate two interesting properties of CoRM’s. First, we
focus on CoRM’s which arise from regurlarly varying directing Lévy measures. This
result is motivated by the recent papers of Caron and Fox (2017) and Todeschini, Miscouridou and Caro
(2017) which made use of regularly varying Lévy measures to construct sparse random graphs. Second, we provide an explicit expression of the multivariate Lévy
intensity of a CoRM with independent exponential scores. This result is interesting when compared with Theorem 3.2 in Zhu and Leisen (2015) and Corollary 2 in
Griffin and Leisen (2017) which provide, respectively, the Lévy copula representation and the Laplace exponent of CoRM’s with independent exponential scores. The
proofs of the theorems can be found in the supplementary material.
Both results deal with a d-variate CoRM given by an homogeneous directing Lévy
intensity ν⋆ (dz, dx ) = ρ⋆ (dz)α(dx ). Therefore, the corresponding marginals can be
written as νj (ds, dx ) = ρ j (ds)α(dx ) with ρ j a measure in (R + , B(R + )).
4.1 Regularly varying directing Lévy measure
In this section we focus on CoRM’s given by a directing Lévy measure that is regularly
varying. We recall that a real valued function L is slowly varying if limt→∞ L(at)/L(t) =
1 ∀ a > 0.
Definition 3. An homogeneous LévyR measure ρ⋆ (dz)α(dx ) in R + × X is said to be regularly
∞
varying if the tail integral U ⋆ (y) = y ρ⋆ (ds) is a regularly varying function, i.e. it satisfies
1 1
U (y) = L
y yσ
⋆
for some σ ∈ [0, 1) which we call the index and L a slowly varying function.
The following Theorem highlights an interesting link between the directing Lévy
measure and the marginal Lévy intensities in terms of the regularly varying property.
Theorem 3. Consider a CoRM with an homogeneous directing Lévy measure ρ⋆ (ds)α(dx )
such that the conditions of Theorem 2 are satisfied. If ρ⋆ is regularly varying with tail integral
U then the marginals ρ j , j = 1, . . . , d, are regularly varying.
Example 3: σ-stable directing Lévy measure
Consider a σ-stable directing Lévy measure
ν⋆ (ds, dx ) =
σ
dsdx.
Γ (1 − σ ) s σ + 1
The related tail integral is
U ⋆ (y) =
1
Γ (1 − σ ) y σ
7
which is a regularly varying function with index σ and slowly varying function L(y) =
1
. We see that the regularly varying tail integrals related to the CoRM construcΓ (1− σ )
h i
tion arise as a factor of U, namely U j (y) = E Sσj U ⋆ (y) and from Theorem 3 the
associated marginal tail integrals are regularly varying.
4.2 Independent Exponential scores
Consider a d-variate CoRM given by an homogeneous directing Lévy measure ν⋆ (dz, dx ) =
ρ⋆ (dz)α(dx ) and a score distribution corresponding to d independent standard exponential distributions, i.e.
d
h ( s1 , . . . , s d ) = ∏ e − s i .
i =1
We observe that each associated marginal takes the form νj (ds, dx ) = f (s)dsα(dx ),
R∞
s
where f (s) = 0 z−1 e− z ρ⋆ (dz). The following Theorem provides a characterization
for this class of CoRM’s.
Theorem 4. Consider a CoRM as described above; the corresponding d-variate Lévy intensity
ν̃d (dss , dx ) = ρ̃d (s )dss α(dx ) is such that
ρ̃d (s ) = (−1)d−1
∂ d−1
f (s )
∂sd−1
.
s= s1 + ...+ sd
5 Conclusions
In this paper, we proved some integrability condition for Compound Random Measures. The new findings can be useful to Statisticians which aim to use vectors of dependent completely random measures which arise from CoRM’s with directing Lévy
measure and score distribution that have not been considered so far. Furthermore,
in the homogeneous case, we proved that the marginal Lévy intensities are regularly
varying whenever a regularly varying directing Lévy measure is considered. Finally,
we provide a representation of the homogeneous CoRM’s when the score distribution
is the result of independent and identically distributed exponential distributions.
Acknowledgments
Fabrizio Leisen was supported by the European Community’s Seventh Framework
Programme [FP7/2007-2013] under grant agreement no: 630677.
8
References
References
Camerlenghi, F., Lijoi, A. and Prünster, I. (2017). ’Bayesian prediction with multiplesamples information’, Journal of Multivariate Analysis, 156, 18-28.
Caron, F. and Fox, E. (2017). ’Sparse graphs using exchangeable random measures’,
forthcoming in the Journal of the Royal Statistical Society-Series B.
Cont, R., and Tankov, P., (2004). Financial modelling with jump processes, Boca Raton, FL:
Chapman & Hall/CRC.
Daley, D.J., and Vere-Jones, D., (2003), An introduction to the theory of point processes.
Vol. 1., New York: Springer.
De Finetti, B. (1938). ’Sur la condition de ”equivalence partielle”’, Colloque consacré
à la théorie des probabilités, Vol. VI, Université de Genève, Hermann et C. ie, Paris.
Griffin J. E. and Leisen F. (2017). ’Compound random measures and their use in
Bayesian nonparametrics’, Journal of the Royal Statistical Society - Series B, 79, 525545.
Griffin J. E. and Leisen F. (2018). ’Modelling and computation using NCoRM mixtures
for density regression ’, forthcoming in Bayesian Analysis.
James, L. F., Lijoi, A., and Prünster, I. (2009). ’Posterior analysis for normalized random measures with independent increments’, Scandinavian Journal of Statistics,
36(1), 76-97.
Kingman, J.F.C., (1993), Poisson processes, Oxford: Oxford University Press.
Leisen, F., and Lijoi, A., (2011). ‘Vectors of Poisson-Dirichlet processes’, J. Multivariate
Anal., 102, 482–495.
Leisen, F., Lijoi, A., and Spano, D., (2013). ‘A Vector of Dirichlet processes’, Electronic
Journal of Statistics, 7, 62–90.
Lijoi, A. and Nipoti, B. (2014). ‘A class of hazard rate mixtures for combining survival
data from different experiments’, Journal of the American Statistical Association, 109,
802-814.
Lijoi, A., Nipoti, B., and Prünster, I. (2014). ‘Bayesian inference with dependent normalized completely random measures’, Bernoulli, 20, 1260-1291.
A. Lijoi and I. Prünster (2010). Beyond the Dirichlet process. Bayesian Nonparametrics
(Holmes, C.C., Hjort, N.L., Müller, P. and Walker, S.G., Eds.), 80–136, Cambridge
University Press, Cambridge.
9
MacEachern, S.N., (1999). ‘Dependent nonparametric processes’. ASA Proceedings of
the Section on Bayesian Statistical Science, 50–55.
E. Regazzini, A. Lijoi, I. Prünster (2003). Distributional results for means of normalized random measures with independent increments. Ann. Statist. 31, 560–585.
Todeschini A., Miscouridou X. and Caron F. (2017). ’Exchangeable random measures for sparse and modular graphs with overlapping communities’, arXiv:
1602.02114v2 .
Zhu W. and Leisen F. (2015). ’A multivariate extension of a vector of Poisson-Dirichlet
processes’, Journal of Nonparametric Statistics, 27, 89–105.
6 Supplemetary Material: Proofs
First, we recall the integrability condition for a Lévy measure:
Z
R+ ×X
min{1, s}ν̃ (ds, dx ) < ∞
(12)
see Section 2 of the main document.
Let H and ν⋆ be, respectively, a score distribution and a directing Lévy measure
which define a CoRM. We denote with H j , j ∈ {1, . . . , d}, the j-th marginal of a
d-dimensional score distribution H. A simple change of variable leads to the j-th
marginal of a CoRM, namely
Z
Z
Z
Z
⋆
⋆ dz
νj ( A, X ) =
, dx H j (ds).
(13)
H j (ds)ν (dz, dx ) =
ν
s
R + × X A/z
R + A× X
We can see the formula above as a mean. Let S j be a random variable with distribution
H j , then
"
!#
A
⋆
νj ( A, X ) = E ν
,X
(14)
Sj
for A ∈ B (R + ). We use the last identity to give conditions for the marginal intensity
νj to be a proper Lévy intensity, i.e. a measure that satisfies the condition displayed in
equation (12).
Proof of Theorem 1
We recall the conditions stated in the first part of the theorem:
Z
1 ⋆
P Sj ≥
ν (dz, dx ) < ∞
z
(0,1)× X
10
(15)
and
Z
1
zν⋆ (dz, dx ) < ∞,
P Sj ≤
z
[1,∞)× X
and the conditions stated in the second part of the Theorem:
1
≤ z ∀ z ∈ (0, ǫ) for some ǫ > 0
1 − Hj
z
(16)
(17)
and
1
lim zH j
< ∞.
z→∞
z
(18)
Proof. We recall that ν⋆ satisfies (12) since it is a Lévy intensity. Using (14), condition
(12) for νj becomes
!#
"Z
dz
,X
E
min{1, z}ν⋆
Sj
R+ ×X
= E
Z
⋆
0, S1
j
×X
zν (dz, dx ) + E
Therefore, νj satisfies (12) if and only if
and
E
Z
E
Z
0, S1
j
1
S j ,∞
×X
×X
Z
1
S j ,∞
×X
ν⋆ (dz, dx ) < ∞.
(19)
(20)
(21)
zν⋆ (dz, dx ) < ∞
ν⋆ (dz, dx ) < ∞.
The former can be decomposed using the Fubini-Tonelli theorem in
Z
Z
1
⋆
zν⋆ (dz, dx )
zν (dz, dx ) =
P Sj ≤
E
1
+
z
0, S × X
R ×X
j
Z
Z
1
1
⋆
=
P Sj ≤
zν (dz, dx ) +
zν⋆ (dz, dx ).
P Sj ≤
z
z
(0,1)× X
[1,∞)× X
Condition (16) ensures that the second term of the above equation is finite. It is easy
to see that the first term is finite as well. Indeed,
Z
Z
1
⋆
P Sj ≤
zν⋆ (dz, dx ) < ∞.
zν (dz, dx ) ≤
z
(0,1)× X
(0,1)× X
11
On the other hand, the second term in (19) can be decomposed in
Z
Z
1
⋆
≤ S j ν⋆ (dz, dx )
ν (dz, dx ) =
E
P
1 ,∞ × X
+ ×X
z
R
Sj
Z
Z
1 ⋆
1 ⋆
=
P Sj ≥
ν (dz, dx ) +
ν (dz, dx ).
P Sj ≥
z
z
(0,1)× X
[1,∞)× X
Condition (15) ensures that the first term of the above equation is finite. It is easy to
see that the second term is finite as well. Indeed,
Z
Z
1 ⋆
P Sj ≥
ν (dz, dx ) ≤
ν⋆ (dz, dx ) < ∞.
z
[1,∞)× X
[1,∞)× X
Therefore, the first part of the theorem follows from (19), (20) and (21).
For the remaining part of the Theorem we use that (12) is attained when considering
the directing Lévy measure ν⋆ . Indeed, if
1
<∞
(22)
lim zP S j ≤
z→∞
z
then as ν⋆ is a Lévy intensity
1
P Sj ≤
zν⋆ (dz, dx ) < ∞.
(23)
z
[1,∞)× X
so (20) holds. And if there exists ǫ > 0 such that 1 − H j 1z ≤ z ∀ z ∈ (0, ǫ) then
Z
Z
Z
1
⋆
P
≤ S j ν (dz, dx ) <
zν⋆ (dz, dx ) < ∞,
z
(0,1)× X
(0,1)× X
so (21) also holds. From the first part of the theorem the CoRM marginal νj satisfies
the integrability conditions for arbitrary ν⋆ .
Proof of Corollary 1
We recall the conditions stated in Corollary 1:
lim
hj
1
z
2
z
<1
(24)
lim h j (ǫ) < ∞.
(25)
z →0
and
ǫ →0
12
) and observe that f (0+ ) = 0 so the existence
of f ′ (0+ ) > 0 implies (17). As S j has a probability density we get that f ′ (0+ ) exists
and (17) is equivalent to f ′ (0+ ) > 0 which we write as
h j 1z
lim
< 1.
z →0
z2
Proof. We define f (z) = z − (1 − H j
1
z
Using the fundamental theorem of calculus we see that (18) reduces to
1
= lim h j (ǫ) < ∞
lim zP S j ≤
z→∞
ǫ →0
z
which is satisfied when h j is continuous at zero.
6.1 Proof of Theorem 2
Proof. Denote Pj = {s ∈ (R + )d : max{s1 , . . . , sd } = s j } for j ∈ {1, . . . , d}; then, by
using (13) and the fact that each νj is a Lévy intensity we get that for any bounded set
X in X
Z
Z
s
s dss ⋆
min{1, ks k}h 1 , . . . , d
ν (dz, dx )
z
z zd
R + × X (R + ) d
Z
s
d Z
s dss ⋆
ν (dz, dx )
=∑
min{1, ks k}h 1 , . . . , d
d
+ ×X P
z
z
z
R
j
j =1
≤
≤
=
d Z
Z
d Z
Z
d Z
Z
∑
+
j =1 R × X
∑
+
j =1 R × X
∑
j =1
=
≤
=
R+ ×X
Pj
min{1,
(R + ) d
R+
√
ds j }h
min{1,
min{1,
√
√
√
sd dss ⋆
ν (dz, dx )
,...,
z
z zd
s
1
ds j }h
ds}h j
s
1
z
s ds
z
d Z
min{1,
d Z
√ √
min{ d, ds}νj (ds, dx )
∑
+
j =1 R × X
∑
+
j =1 R × X
√
d Z
d∑
+
j =1 R × X
,...,
z
ν⋆ (dz, dx )
ds}νj (ds, dx )
min{1, s}νj (ds, dx ) < ∞.
13
sd dss ⋆
ν (dz, dx )
z zd
Proof of Theorem 3
We recall that for the case at hand
1 1
U (y) = L
y yσ
⋆
(26)
is a tail integral.
Proof. We note that equation (14) implies that
"
ρ j ( A) = E ρ⋆
A
Sj
!#
It follows that the marginals of the CoRM are given by
U j (y) = ρ j ((y, ∞))
!#
"
y
= E U⋆
Sj
σ
Sj
Sj
=E L
y
y
Sj
1
=E L
Sσj
.
y
yσ
i
h
Hence, it is enough to check if the function l (z) = E L(S j z)Sσj is slowly varying for
L a slowly varying function. Let a > 0, we need to check
i
h
σ
E
L
(
atS
)
S
j j
l (at)
i = 1.
= lim h
lim
t→∞
t→∞ l (t)
E L(tS )Sσ
j
j
For a fixed ǫ > 0 we can choose t0 such that ∀u > t0
ǫ
| L(au)/L(u) − 1| < ,
2
14
since L is slowly varying. Then for t > t0
i
h
h
i
E L(atS j )Sσj
E Sσj L(atS j ) − L(tS j )
h
i −1 =
h
i
σ
σ
E L(tS j )S j
E L(tS j )S j
σ
σ
n
n
o
o
E 1
S j L(atS j ) − L(tS j )
S j L(atS j ) − L(tS j )
E 1
t
t
S j > t0
S j ≤ t0
h
i
h
i
+
≤
E L(tS j )Sσj
E L(tS j )Sσj
ǫ
σ
σ
n
n
o
o
E 1
S j 2 L(tS j )
S j L(atS j ) − L(tS j )
E 1
t
t
S j > t0
S j ≤ t0
h
i
h
i
<
+
σ
σ
E L(tS j )S j
E L(tS j )S j
σ
o
n
S j L(atS j ) − L(tS j )
E 1
t
S j ≤ t0
ǫ
h
i
< +
2
E L(tS j )Sσj
σ
σ
o
n
S j t L(atS j ) − L(tS j )
E 1
t
S j ≤ t0
ǫ
= +
2
U j ( 1t )
R
sσ tσ | L(ats) − L(ts)| H j (ds)
t
ǫ
(0, t0 ]
= +
2
U j ( 1t )
R
du
σ | L( au) − L(u)| H
u
j
t
ǫ
(0,t0 ]
= +
1
2
Uj ( t )
(27)
We observe that limx→0 U j ( x ) = ∞. since U j is a tail integral. From (26) it follows
that limx→0 x σ ( L(ax ) − L( x )) = 0. Hence, the function g( x ) = x σ ( L(ax ) − L( x )) is
bounded in [0, t0 ] by a constant K1,t0 . Finally we observe that for t > t0
Z
Z
du
Hj
<
H j (du) ≤ 1.
t
(0,t0 ]
(0,1]
We set t1 > t0 such that for u > t1
2K1,t0
< U j (1/u).
ǫ
Choosing t > t1 we get
R
(0,t0 ]
uσ
| L(au) − L(u)| H j
U j ( 1t )
15
du
t
<
K1,t0
U j ( 1t )
<
ǫ
2
It follows from (27) that
i
h
E L(atS j )Sσj
h
i − 1 < ǫ.
E L(tS j )Sσj
Consequently, l defined above is slowly varying, implying that the marginal tail integral U j is regularly varying.
6.2 Proof of Theorem 4
Proof. It is straightforward to see that
ρ̃d (s ) =
Z ∞
0
z
−d
h
s
sd ⋆
ρ (dz) =
,··· ,
z
z
1
Z ∞
0
z− d e−
s1 + ...+ s d
z
ρ⋆ (dz).
From Example 1 in the main document, we know that, for arbitrary ρ⋆ and d ∈ N \
{0}, the previous integral is finite. Therefore for s 6= 0
Z ∞
∂ j −1 − s ⋆
z e z ρ (dz) = ρ̃ j+1 (s, 0, . . . , 0) < ∞
j
0
∂s
and this concludes the proof since using the Dominated Convergence Theorem we
can take the derivative under the integral sign as follows
(−1)d−1
∂ d−1
f (s )
∂sd−1
= (−1)d−1
s= s1 + ...+ sd
=
Z ∞
0
Z ∞ d−1
∂
−1 − zs
ρ⋆ (dz)
z
e
d−1
z− d e
16
0
∂s
s + ...+ s
− 1 z d
ρ⋆ (dz).
s= s1 + ...+ sd
| 10math.ST
|
WEAKLY TRIPOTENT RINGS
arXiv:1704.01303v1 [math.RA] 5 Apr 2017
S. BREAZ, A. CÎMPEAN
Abstract. We study the class of rings R with the property that for x ∈ R at least one
of the elements x and 1 + x are tripotent.
1. Introduction
An element x of a ring R is called tripotent if x3 = x, and a ring R is tripotent if all
its elements are tripotent. Hirano and Tominaga proved in [4, Theorem 1] that a ring R is
tripotent if and only if every element of R is a sum of two commuting idempotents. The
class of these rings was extended in [7, Section 4] to the class of rings R such that every
element is a sum or a difference of two commuting idempotents. This is a natural approach
since similar methods were applied for other classes of rings. For instance, the classes of
clean rings or nil-clean rings were extended in [1] and [2] to the classes of weakly clean,
respectively weakly nil-clean in the following ways: a ring R is weakly (nil-)clean if and
only if every element of R is a sum or a difference of a unit (nilpotent) element and an
idempotent.
Apparently the natural question which asks if we can properly extend in a similar way
the classes of rings which satisfy identities which involve tripotents instead of idempotents
has a negative answer, since −u is tripotent whenever u is tripotent. However, we want to
propose such an extension, starting by the following remarks:
(1) every element of a ring R is a sum or a difference of two (commuting) idempotents
if and only if for every element x ∈ R at least one of the elements x or 1 + x is a
sum of two (commuting) idempotents;
(2) a ring R is weakly clean if and only if for every element x ∈ R at least one of the
elements x or 1 + x is a sum of a unit and an idempotent.
(3) a ring R is weakly nil-clean if and only if for every element x ∈ R at least one of
the elements x or 1 + x is a sum of a nilpotent and an idempotent.
Therefore, we will say that a ring R is weakly tripotent if for every element x ∈ R at least
one of the elements x or 1 + x is tripotent. Two main differences between the properties
of tripotent rings and weakly tripotent rings are the following: while tripotent rings are
always commutative and the class of tripotent rings is closed under direct products, these
propoerties are not valid for weakly tripotent rings. One of the main result of the present
paper is Theorem 15 where we provide a characterization of commutative weakly tripotent
rings as subrings of rings of the form R0 × R1 × R2 , where R0 is a weakly tripotent ring
2010 Mathematics Subject Classification. 16R50, 16A32, 16U99.
1
2
S. BREAZ, A. CÎMPEAN
without nontrivial idempotents (these rings are described in Corollary 10), R1 is a Boolean
ring, and R2 is a tripotent ring of characteristic 3. Moreover, we will obtain connections
between the class of weakly tripotent rings and the rings studied in [7]. In particular, it
is proved in the end of the paper that the class of rings such that every element is a sum
or a difference of two commuting idempotents is strictly contained in the class of weakly
tripotent rings (Corollary 19).
2. Weakly tripotent rings
Before we will study the connections between these classes, we state some some basic
properties of weakly tripotent rings.
Lemma 1. Let R be a weakly tripotent ring. Then
(1) every subring of R is weakly tripotent;
(2) every homomorphic image of R is weakly tripotent;
(3) 24 = 0, hence R has a decomposition R = R1 × R2 such that R1 and R2 are weakly
tripotent and 8R1 = 3R2 = 0;
For the the case R is of characteristic 3 weakly tripotent rings are in fact tripotent, we
can restrict our study to rings on characteristic 2k , k ∈ {1, 2, 3}. We refer to [3] for other
extensions of other generalizations of tripotent rings of characteristic 3.
Lemma 2. If R is a ring that is weakly tripotent and has characteristic 3 then R is
tripotent, hence it is isomorphic to a subdirect product of a direct product of Z3 .
Proof. The identity (1 + x)3 = 1 + x is equivalent to 1 + 3x + 3x2 + x3 = 1 + x and since
3 = 0 this is equivalent to x3 = x.
We start with the
Proposition 3. The following are equivalent for a ring R:
(a) R is weakly tripotent;
(b) for every element x ∈ R at least one of the elements x or 1 − x is tripotent.
Proof. (a)⇒(b) Let x ∈ R such that x3 6= x. It follows that (−x)3 6= −x, hence (1 − x)3 =
1 − x.
(b)⇒(a) is proved in a similar way.
Remark 4. Let us remark that a similar equivalence is not valid neither for Boolean rings
nor for nil-clean conditions. More precisely:
(a) a ring is Boolean (clean, respectively nil-clean) if and only if for every element x ∈ R
at least one of the elements x or 1 − x is idempotent (clean, respectively nil-clean);
(b) in the ring Z3 for every element x ∈ R at least one of the elements x or 1 + x is
idempotent (resp. nil-clean), but Z3 is not Boolean or nil-clean, while Z(15) (i.e. the
ring of rational numbers with denominator coprime with 15) is a weakly clean ring
which is not clean, [1].
WEAKLY TRIPOTENT RINGS
3
In the following result we describe a connection between the class of weakly tripotent
rings and one of the classes studied in [7]. A second connection will be established in
Corollary 19.
Corollary 5. The class of all weakly tripotent ring is strictly contained in the class of
rings such that all elements are sums an idempotent and a tripotent that commute.
Proof. The first statement is obvious since for every element x of a weakly tripotent ring
such that x3 6= x we have that x − 1 is tripotent. To produce an example, we observe that
in the ring Z4 × Z4 every element is a sum of a idempotent and a tripotent, but x = (1, 2)
and (1, 1) + x are not tripotent elements.
It is easy to see that the class of weakly tripotent rings is contained in the class studied in
[7, Section 2] of those rings such that every element is a sum of two commuting tripotents.
The proof of the following proposition is easy.
Proposition 6. If a ring R is weakly tripotent then it can be embedded as a subring of a
direct product of subdirectly irreducible weakly tripotents rings.
However, in the case of characteristic 2k a direct product of weakly tripotent ring is not
necessarily weakly tripotent. The proof of the following proposition is standard. We refer
to the similar property stated for weakly nil-clean rings in [1, Theorem 1.7]
Q
Proposition 7. A direct product i∈I Ri of weakly tripotent rings is weakly tripotent if
and only if there exists i0 ∈ I such that Ri0 is weakly tripotent and for all i ∈ I \ {i0 } the
rings Ri are tripotent.
The following result was proved in [7, Theorem 3.6] by using the fact that if every
element of a ring R is a sum of an idempotent and a tripotent that commute then R is
strongly nil-clean, hence R/J(R) is Boolean and J(R) is nil, [5, Theorem 5.6]. We present
for reader’s convenience a proof which uses the general procedure described in [6, Section
12].
Theorem 8. [7, Theorem 3.6] Let R be a ring of characteristic 2k . The following are
equivalent:
(1) every element of R is a sum of an idempotent and a tripotent that commute;
(2) x4 = x6 for every x ∈ R;
(3) R/J(R) is Boolean and U (R) is a group of exponent 2;
(4) R/J(R) is Boolean, and for every x ∈ J(R) we have x2 = 2x.
Consequently, if R satisfies the above conditions then J(R) coincides to the set of all
nilpotent elements of R.
Proof. (1)⇒(2) This is the implication (2)⇒(3) from [7, Theorem 3.6].
(2)⇒(3) First, we observe that for every ring R which satisfies (2) and every x ∈ U (R)
we have x2 = 1.
The class of rings which satisfies (2) is closed with respect to subrings and factor rings.
More, precisely, we embed R/J(R) as a subring of a direct product of left primitive rings
4
S. BREAZ, A. CÎMPEAN
which are factor rings of R, [6, Theorem 12.5]. We will prove that every left primitive ring
which satisfies (2) is isomorphic to a division ring.
In order to prove this, we first observe that ifK is a division ring
of characteristic 2
1 1 0 ... 0
1 0 0 ... 0
then for every n ≥ 2, for the n × n matrix A = 0 0 0 . . . 0 with coefficients in
.. .. ..
.
. . . · · · ..
0 0 0 ... 0
4
6
K we have A 6= A . Then for every n ≥ 2 and every division ring K of characteristic 2,
the rings Mn (K) of n × n matrices over K do not satisfy (2).
Suppose that R is left primitive and satisfies (2). If R is not left artinian then we use
[6, Theorem 11.19] to conclude that there exists a subring S of R and a division ring (of
characteristic 2) such that there exists an onto ring homomorphism from S to the ring of
M2 (K), which is not possible. Therefore there exists a positive integer n and a division
ring K such that R ∼
= Mn (K). This is possible only if n = 1, hence K is a division ring
of characteristic 2. Moreover, for every x ∈ K we have x2 = 1, which is possible only if
K∼
= Z2 (if we have a nonzero element x 6= 1 in K then the subfield generated by x has 4
elements, and it follows that x2 6= 1).
Coming back to the general case, it follows that for every ring R which satisfies (2) then
ring R/J(R) is Boolean, since it is a subring of a direct product of copies of Z2 .
Corollary 9. If every element of R is a sum of an idempotent and a tripotent that commute
then
(i) the ideal J(R) is nil and it contains all nilpotent elements of R;
(ii) for every x ∈ J(R) we have x2 = 2x, 4x = 0, and x3 = 0 .
(iii) xy = yx for all x, y ∈ J(R);
(iv) 2J(R)2 = 0.
Proof. (i) is obvious.
(ii) For every x ∈ J(R) the elements 1 ± x are invertible. Hence (1 ± x)2 = 1, and it
follows that x2 = 2x = −2x. Hence 0 = 4x = x3 .
(iii) Since 1 + x and 1 + y are invertible, and the group U (R) is commutative, we have
(1 + x)(1 + y) = (1 + y)(1 + x), hence xy = yx.
(iv) Observe that 2(x + y) = (x + y)2 = x2 + xy + yx + y 2 = 2x + 2y + xy + yx, hence
xy = −yx. Therefore, 2xy = 0.
In the following corollary we have a characterization of weakly tripotent rings without nontrivial idempotents. This will be useful in order to describe commutative weakly
tripotent rings, since in this case every subdirectly irreducible ring is without nontrivial
idempotents.
Corollary 10. The following are equivalent for a ring R without nontrivial idempotents
such that 3 ∈ U (R):
(1) every element of R is a sum of an idempotent and a tripotent that commute;
WEAKLY TRIPOTENT RINGS
5
(2) R is local such that R/J(R) ∼
= Z2 , x2 = 1 for all x ∈ U (R);
∼
(3) R/J(R) = Z2 , for every x ∈ J(R) we have x2 = 2x;
(4) R is weakly tripotent.
Consequently, every weakly tripotent ring without nontrivial idempotents is commutative.
Proof. (1)⇒(2) This follows from Theorem 8 since every idempotent of R/J(R) can be
lifted to an idempotent of R.
(2)⇒(3) This can be proved as in Corollary 9(ii).
(3)⇒(4) Let x ∈ R \ J(R). Then 1 + x ∈ J(R), and it follows that (1 + x)2 = 2(1 + x).
This implies that x2 = 1. If x ∈ J(R) then 1 + x ∈
/ J(R), hence (1 + x)2 = 1, hence 1 + x
is tripotent.
For the last statement, it is enough to observe that every element of R \ J(R) is of the
form 1 + x with x ∈ J(R), and to apply Corollary 9.
(4)⇒(1) See Corollary 5.
We recall that a ring is abelian if every idempotent of R is central.
Corollary 11. Every abelian ring such that every element of R is a sum of an idempotent
and a tripotent that commute is commutative. In particular, every abelian weakly tripotent
ring is commutative.
Proof. If we have an abelian weakly tripotent ring, we embed the ring as a subdirect
product of a family of subdirectly irreducible abelian weakly tripotent rings (Proposition
6). Since every subdirectly irreducible abelian ring has no nontrivial idempotents, the
conclusion follows from Corollary 10.
But, let us remark that there exists non-commutative subdirectly irreducible weakly
tripotent rings which are not local rings:
Example 12. The ring T2 (Z2 ) of upper triangular matrices over Z2 is subdirectly irreducible and weakly tripotent.
The above corollary allows us to provide nontrivial examples:
k
Example 13. (1) If N is an abelian group such that
2 N = 0, with k ≤ 3, then the
x n
idealization of N by Z2k , i.e. the ring af all matrices
, is weakly tripotent if and
0 x
only if 2N = 0.
(2) Let 2 ≤ k ≤ 3 be a positive integer. If N = Z4 then the multiplication defined on the
abelian group R = Z2k × N by (x, n) ∗ (y, m) = (xy, xm + ny + 2mn) induces a structure
of weakly tripotent ring on Z4 × Z4 .
(3) The ring R = Z4 [X, Y ]/(X 2 , Y 2 , XY − 2, 2X + 2Y ) is a subdirectly irreducible
weakly tripotent ring. In order to prove that R is subdirectly irreducible, we observe that
if I is an ideal in R such that X or Y is in I then 2 ∈ I. Moreover, if 2 + X ∈ I then
2X = X(2 + X) ∈ I and 2Y + 2 = Y (2 + X) ∈ I. Therefore, 2 = 2X + 2Y + 2 ∈ I. It
follows that the ideal {0, 2} is the smallest ideal of R.
6
S. BREAZ, A. CÎMPEAN
In order to provide a characterization for general weakly tripotent rings, we need the
following lemma.
Lemma 14. Let R be a commutative weakly tripotent ring of characteristic 2k (k ≤ 3),
and let N be the nil radical of R. Then for every idempotent e in R one of the following
properties is true:
(a) en = 0 for all n ∈ N ;
(b) en = n for all n ∈ N .
Proof. We will denote f = 1 − n. If n is a nilpotent element then (e + n)3 = e + n or
(f − n)3 = f − n.
Suppose that (e + n)3 = e + n. Using the identities 8n = 0, n2 = 2n and n3 = 0 it
follows that (e + n)3 = e + en, hence en = n. This also implies that f n = 0.
If (f − n)3 = f − n it follows in the same way that f n = n, hence en = 0.
Now suppose that there exists n1 and n2 nonzero nilpotents such that en1 = 0 and
f n2 = 0. Since n1 + n2 is nilpotent, we can suppose that e(n1 + n2 ) = 0. But this implies
that en2 = 0. Therefore n2 = 0, a contradiction, and the proof is complete.
We have the following characterization:
Theorem 15. A commutative ring R is weakly tripotent if and only if R = R′ × R′′ such
that:
(1) R′′ is a tripotent ring of characteristic 3 or R′′ = 0;
Q
(2) R′ = 0 or R′ can be embedded as a subring of a direct product R0 ×
i∈I Ri such that
R0 is a weakly tripotent ring without nontrivial idempotents, and all Ri are boolean
rings.
Proof. We only have to prove that if R is a weakly tripotent ring of characteristic 2k then
it can be embedded in a direct product as in (2).
Let J be the Jacobson radical of R, and recall that J coincides to the set of all nilpotent
elements of R. We fix an ideal L which is maximal with the property J ∩ L = 0. Then
L∼
= (L + J)/J can be embedded in R/J. Therefore all elements of L are idempotents and
2L = 0.
Let e be an idempotent in R/L. Then R/L ∼
= eR/L × f R/L, where f = 1 − e. Since
(J + L)/L ⊆ J(R/L) we can suppose by Lemma 14 that (J + L)/L ⊆ eR. Therefore, if K
is an ideal in R such that L ⊆ K and K/L ∼
= f R/L then K ∩ J = 0. By the maximality
of L it follows that K = L, hence R/L has no nontrivial idempotents.
In order to complete the proof, we proceed as in the proof of Birkhoff’s Theorem, [6,
Theorem 12.3]. For every 0 6= x ∈ L we choose an ideal Ix which is maximal with the
property x ∈
/ Ix . By Zorn’s Lemma we can choose Ix such that J ⊆ Ix . Therefore R/Ix
Q
is Boolean, and there exists an injective ring homomorphism R → R/L ×
06=x∈L R/Ix
which provides a subdirect decomposition for R.
We complete this theorem with an example of a weakly tripotent ring R as in Theorem
15(2) such that it has no a direct factor isomorphic to R0 .
WEAKLY TRIPOTENT RINGS
7
Example 16. Let R0 be a subdirectly irreducible weakly Q
tripotent ring. For every positive
integer k we fix a ring Rk ∼
= Z2 . In the direct product k≥0 Rk we consider the ring R
generated
by 1 = (1k )k≥0 and the ideal J(R0 ) ⊕ (⊕k≥1 Rk ) of all families of finite support
Q
in k≥0 Rk such that the 0-th component is in J(R0 ). It is easy to see that R has the
desired properties.
Remark 17. Let us remark that the ring R0 from Theorem 15 is not necessarily subdirectly
irreducible. In order to see this, we consider the ring R from Example 13(1) with N =
Z2 × Z2. This ring
is not subdirectly irreducible since for every subgroup S ≤ N the set
0 n
IS = {
∈ R | x ∈ S} is a minimal ideal of R. Therefore, if we view R as a
0 0
subdirect product of a family of subdirectly irreducible rings, at least two factors of this
product are not tripotent rings.
Using the same proof as in Theorem 15 we obtain:
Proposition 18. The following are equivalent for a ring R:
(a) every element of R is a sum or a difference of two commuting idempotents;
(b) R is isomorphic to a subring of a direct product R0 × R2 × R2 such that R0 is 0 or Z4 ,
R1 is Boolean, and R2 is a subdirect product of a family of copies of Z3 ;
Proof. (a)⇒(b) Let R be a ring such that every element is a sum or a difference of two
commuting idempotents. From [7, Lemma 4.2(5)] it follows that R is abelian, hence R is
commutative by Corollary 11.
By [7, Theorem 4.4] we have R = R1 × R2 such that R1 /J(R1 ) is Boolean such that
J(R1 ) = 0 or J(R1 ) = {0, 2} and R2 is a tripotent ring, so it is enough to assume that
R/J(R) is Boolean such that J(R) = {0, 2}. In this hypothesis, we observe that for every
idempotent e ∈ R we have 2e = 0 or 2(1 − e) = 0. Therefore, we can apply the same
technique as in the proof of Theorem 15 to conclude that R is isomorphic to a subring of
a direct product of a subdirectly irreducible factor ring of R and a Boolean ring. Since in
this case subdirectly irreducible factor rings of R are isomorphic to Z2 or Z4 , the proof is
complete.
(b)⇒(a) This is obvious.
Corollary 19. If every element of a ring R is a sum or a difference of two commuting
idempotents then R is weakly tripotent. The converse is not true.
Proof. The first statement follows from Proposition 18 and Theorem 15.
In order to see that the converse is not true, let us observe that Z8 is weakly tripotent,
but not all elements of R are sums or differences of two commuting idempotents.
References
[1] M.-S. Ahn, D.D. Anderson, Weakly clean rings and almost clean rings, Rocky Mt. J. Math. 36 (2006),
783–798.
[2] S. Breaz, P. Danchev, Y. Zhou, Rings in which every element is either a sum or a difference of a
nilpotent and an idempotent, J. Algebra Appl. 15, No. 8, Article ID 1650148, 11 p. (2016).
8
S. BREAZ, A. CÎMPEAN
[3] H. Chen and M. Sheibani, On Yaqub nil-clean ring, preprint arXiv:1704.00213 [math.RA].
[4] Y. Hirano, H. Tominaga, Rings in which every element is the sum of two idempotents, Bull. Aust.
Math. Soc. 37 (1988), 161–164.
[5] T. Koşan, Z. Wang, Y. Zhou, Nil-clean and strongly nil-clean rings, J. Pure Appl. Algebra 220 (2016),
633–646.
[6] T.Y. Lam, A first course in noncommutative rings, Graduate Texts in Mathematics, 131. New York:
Springer-Verlag, 1991.
[7] Z. Ying, T. Koşan, Y. Zhou, Rings in which every element is a sum of two tripotents, Can. Math. Bull.
59 (2016), 661–672.
Simion Breaz: ”Babeş-Bolyai” University, Faculty of Mathematics and Computer Science,
Str. Mihail Kogălniceanu 1, 400084, Cluj-Napoca, Romania
E-mail address: [email protected]
Simion Breaz: ”Babeş-Bolyai” University, Faculty of Mathematics and Computer Science,
Str. Mihail Kogălniceanu 1, 400084, Cluj-Napoca, Romania
E-mail address: cimpean [email protected]
| 0math.AC
|
LinkedCT: A Linked Data Space for Clinical Trials
Oktie Hassanzadeh, M.Sc.1, Anastasios Kementsietsidis, Ph.D. 2, Lipyeow Lim, Ph.D. 2,
Renée J. Miller, Ph.D.1, Min Wang, Ph.D.2
1
University of Toronto, Ontario, Canada;
2
IBM T.J. Watson Research Center, Hawthorne, NY, U.S.A.
Abstract
The Linked Clinical Trials (LinkedCT) project aims
at publishing the first open semantic web data source
for clinical trials data. The database exposed by
LinkedCT is generated by (1) transforming existing
data sources of clinical trials into RDF, and (2)
discovering semantic links between the records in the
trials data and several other data sources. In this
paper, we discuss several challenges involved in
these two steps and present the methodology used in
LinkedCT to overcome these challenges. Our
approach for semantic link discovery involves using
state-of-the-art approximate string matching
techniques combined with ontology-based semantic
matching of the records, all performed in a
declarative and easy-to-use framework. We present
an evaluation of the performance of our proposed
techniques in several link discovery scenarios in
LinkedCT.
Introduction
Clinical trials data are used extensively in drug
development and clinical research1. The data
collected in trials are widely used by a variety of
different individuals for different purposes. Creating a
single web data source of clinical trials data that is
highly interlinked with existing medical data sources
and that can be queried with semantically rich and
complex queries, is a challenging task. This is the
goal of the Linked Clinical Trials (LinkedCT)
project. Such a web data source could significantly
enhance the discovery of clinical trials in order to
match patients with trials, perform advanced studies
and enable tailored therapeutics2.
The LinkedCT data space (http://linkedct.org)
is published according to the principles of publishing
linked data3. These principles greatly enhance
adaptability and usability of data on the web, either
by humans or machines. Each entity in LinkedCT is
identified by a unique HTTP dereferenceable
Uniform Resource Identifier (URI). When the URI is
looked up, related RDF statements about the entity is
returned in HTML or RDF/XML based on the user’s
browser. Moreover, a SPARQL endpoint is provided
as the standard access method for RDF data.
Figure 1. Sample interlinked entities in LinkedCT*.
Another important aspect of linked data publication is
providing typed links to other relevant information
sources for each resource. Without establishing such
links, the online sources will resemble islands of data
(or data silos), where each island maintains only part
of the data necessary to satisfy the information needs
of the particular users of that data source. In such a
setting, apart from finding relevant data in a given
island, users are also responsible for finding which
other islands might hold relevant data, and what the
specific pieces of relevant data are within these
islands. To illustrate, consider the simple case of
matching a patient with a clinical trial, where a
clinician is trying to match a patient in “Westchester
Medical Center” who is using the drug “Campath”
with a related trial. For this, the clinician needs to
identify the sources of information about trials, and in
addition she first needs to identify all the locations
close to where the patient lives, possibly using a
source
of
geographical
information
(e.g.,
GeoNames†), and then find all the synonyms (generic
names and brand names) of the drug using perhaps a
drug data source (e.g., DrugBank‡). All of this is
*
Only the links to external entities are shown.
http://geonames.org
‡
http://drugbank.ca
†
needed in order to identify, for example, a trial at
“Columbia University” on “Alemtuzumab” (the
generic name of the drug “Campath”).
Previous research in finding semantic links between
data items has mainly focused on a more restricted
version of the problem, namely, entity resolution or
record linkage, i.e., the identification of entities in
disparate sources that represent the same real-world
entity4. Yet, the general problem investigated here
considers links between entities that are not
necessarily identical, although they are semantically
related. For example, in Figure 1, “Columbia
University” is linked with “New York” using a link of
type based_near and “Beta-Thalassemia” is linked
with
“Blood
Disorders”
with
link
type
is_a_type_of. In practice, the type of the links is
usually specified using existing vocabularies, for
example, owl:sameAs specifies that two entities
are the same, rdfs:seeAlso specifies that the
target resource may contain additional information
about the source, and foaf:based_near specifies
that the source entity is located in the target resource
(which is of type location). Discovering such typed
links is currently an area of active research.
relational views over the XML data, and to
materialize the views to enhance the performance.
Third, the relational data is mapped into RDF format.
We use the popular D2RQ tool for transforming the
relational data to RDF6 and the companion D2R
server for publishing the relational data as a linked
data source. Table 1 shows the overall statistics of the
data and Table 2 shows the statistics of some of the
entities in the data space.
Table 1. Overall Statistics.
7,011,000
822,824
Number of links to external
linked data sources
308,904
Table 2. Sample Entities.
Triplification of Trials Data
In this section, we outline the linked data generation
(or triplification) of trials data in LinkedCT. The core
of the LinkedCT dataset is derived from
ClinicalTrials.gov5, a registry of clinical trials
conducted in the United States and around the world.
As of November 2008, this registry contained
information about more than 60,000 trials conducted
in 158 countries. Each trial is associated with a brief
description of the trial, related conditions and
interventions, eligibility criteria, sponsors, locations,
and several other pieces of information. The data on
ClinicalTrials.gov is semi-structured and is available
in HTML and XML formats. Transforming the data
into RDF involves several steps. First, the XML
collection of the trials needs to be crawled and stored
locally. This results in a collection of over 60,000
XML files. Second, the XML data needs to be
mapped into relational format. The mapping is
obtained by manually inspecting the data and the
Document Type Definition (DTD) of the data. In
addition to all the entities and facts in the data, the
mapping also extracts the relationships between the
entities and their types. The relational schema needs
to be reasonably normalized to ensure performance
and avoid update anomalies when the data is
refreshed. For LinkedCT, we have used a hybrid
relational-XML DBMS (IBM DB2) to create
Total number of RDF triples
Number of Entities
Entity
Trials
Count
60,520
Condition
14,243
Intervention
67,271
Location
222,115
Collaborator Agency
7,071
Overall Official
52,660
Primary Outcomes
55,761
Reference
45,110
Criteria
73,688
Total
822,824
Interlinking LinkedCT with Other Data Sources
After the triplification of the trials data, we are faced
with the challenge of finding useful links between the
data in LinkedCT and other data sources. We have
interlinked LinkedCT with several linked data
sources of medical information: DBpedia§ and
YAGO (linked data sources derived from Wikipedia
articles, containing many useful pieces of information
about diseases and drugs), DailyMed** (published by
the National Library of Medicine, providing high
quality information about marketed drugs),
Diseasome†† (containing information about 4,300
disorders and disease genes linked by known
disorder–gene associations for exploring known
phenotype and disease gene associations and
indicating the common genetic origin of many
diseases), DrugBank (a repository of roughly 5,000
§
http://dbpedia.org
http://www4.wiwiss.fu-berlin.de/dailymed/
(derived from http://dailymed.nlm.nih.gov/)
††
http://www4.wiwiss.fu-berlin.de/diseasome/
(derived from http://www.ncbi.nlm.nih.gov/omim)
**
FDA-approved
drugs),
Bio2RDF’s
PubMed
(Bio2RDF‡‡ project’s linked data version of
PubMed), GeoNames (linked data source of
geographical information) and Lingvoj (information
about languages). Also, the trials, conditions,
interventions, locations and references on LinkedCT
have been linked to their related web pages on
ClinicalTrials.gov, Wikipedia, Freebase and Pubmed
using foaf:page links. Table 3 shows the
statistics of the external links in LinkedCT.
Table 3. Statistics of the links between LinkedCT and
several other data sources.
Source / Target
Link Type
Count
LinkedCT (intervention)
↔ DBpedia (drug)
owl:sameAs
11,527
LinkedCT (intervention)
↔ DrugBank (drug)
rdfs:seeAlso
23,493
LinkedCT (intervention)
↔ DailyMed (drug)
rdfs:seeAlso
39,396
LinkedCT (condition) ↔
DBpedia (disease)
owl:sameAs
342
LinkedCT (condition) ↔
Diseasome (disease)
owl:sameAs
830
LinkedCT
Geonames
foaf:based_ne
ar
129,177
LinkedCT (reference) →
Bio2RDF’s PubMed
owl:sameAs
42,219
LinkedCT
(trial)
ClinicalTrials.gov
foaf:page
61,920
(trial)
→
→
The difficulty in discovering links between different
data sources lies in the absence of globally unique
identifiers in many of the disparate data sources. The
problem is exacerbated by presence of errors in the
data. The consequence is that establishing links using
exact matching of data fields is not robust enough to
capture the amount of real linkages. This calls for
approximate string matching in combination with
semantic matching of the data fields. Approximate
string matching is required when there are different
string representations for the same entity. For
example, “Alzheimer's disease” in LinkedCT should
be matched with “Alzheimer_disease” in Diseasome.
String matching is also robust against misspelled
words and alternative spellings (eg. "Thalassemia"
and "Thalassaemia").
Semantic matching is particularly useful for matching
clinical terms, since many drugs and diseases have
multiple names. Drugs may have generic names and
‡‡
http://bio2rdf.org/
several brand names. Diseases may also have
different scientific and non-scientific names or may
be
abbreviated.
For
example,
"Acquired
Immunodeficiency Syndrome" is abbreviated as
"AIDS", or the terms "Klinefelter Syndrome" and
"Hypogonadotropic hypogonadism" are alternatively
used to refer to the same disease. Obviously,
approximate string matching is not sufficient for data
with such different representations.
String Matching for Semantic Link Discovery
String data is prone to several types of inconsistencies
and errors including typos, spelling mistakes, using
abbreviations or different conventions. Therefore,
finding similar strings, or approximate string
matching, is an important part of an (online) link
discovery process. Approximate string matching is
performed based on a similarity function sim() that
quantifies the amount of closeness (as opposed to
distance) between two strings. A similarity threshold
θ is set by the user to specify that there is a link from
the base record to the target record if their similarity
score, returned by function sim(), is above θ.
Similarity Function There exists a variety of
similarity functions for string data in the literature.
The performance of a similarity functions usually
depends on the characteristics of data, such as length
of the strings, along with the type of the errors and
inconsistencies present in the data. An increasingly
popular class of string similarity functions is based on
tokenization of the strings into q-grams, i.e.,
substrings of length q of the strings. Using q-gram
tokens makes it possible to treat strings as sets of
tokens and use a set similarity measure as the measure
of similarity between the two strings. Based on
comparison of several such measures, we use
weighted Jaccard similarity along with q-gram tokens
as the measure of our choice due to its relatively high
efficiency and accuracy compared with other
measures, especially in matching records in medical
domain. In particular, this similarity function has the
ability to take into account the importance of the
tokens, and therefore could match two records that
differ in many unimportant tokens. This allows
matching “HIV virus” with “HIV”, and “Alzheimer”
with “Alzheimer’s Disease”. Moreover, this function
captures
token
swaps,
therefore
matching
“Adenocarcinoma of the Colon” with “Colon
adenocarcinoma” for example. Many traditional
string similarity functions (such as edit similarity) fail
to capture such differences in the strings.
Jaccard similarity between two string records r1 and
r2 is the fraction of tokens in r1 and r2 that are present
in both. Weighted Jaccard similarity is the weighted
version of Jaccard similarity, i.e.,
where w(t,R) is a weight function that reflects the
commonality of the token t in the relation R, i.e., the
higher the weight of a q-gram, the more important the
q-gram is. We choose RSJ (Robertson/Sparck Jones)
weight for the tokens which are similar to (but more
effective than) the commonly-used Inverse Document
Frequency (IDF) weights:
where N is the number of tuples in the base relation R
and nt is the number of tuples in R containing the
token t. The similarity value returned is between 0
(for strings that do not share any q-grams) and 1 (for
equal strings).
Implementation We have implemented the weighted
Jaccard similarity predicate declaratively in standard
SQL and used it as a join predicate in a standard
RDBMS engine by extending our previous work on
approximate selection7. Alternatively, this predicate
can be used with some of the specialized, high
performance, state-of-the-art approximate join
algorithms4 that make the process scalable to very
large databases. In our experiments we used q-grams
of size 2 and a specific q-gram generation technique
to enhance the effectiveness of the predicate: strings
are first padded with whitespaces at the beginning and
the end, then all whitespace characters are replaced
with q − 1 special symbols (e.g., $).
Ontology-based Semantic Matching for Link
Discovery
In order to find meaningful semantic links between
values, it is often necessary to leverage some
semantic knowledge about the domain from which
these values originate. In many domains, including
health care, there are existing, commonly accepted,
semantic knowledge bases that can be used to this
end. A common type of such semantic knowledge is
an ontology. In the health care domain, well-known
ontologies such as the NCI thesaurus are widely used
that encapsulate a number of diverse relationship
types between their recorded medical terms,
including, synonymy, hyponymy/hypernymy, etc.
Such relationship types can be conveniently
represented in the relational model and (recursive)
SQL queries can be used to test whether two values
are associated with a relationship of a certain type.
Implementation We extend the existing methods for
ontology-based keyword search over relational data
for ontology-based link discovery. Assume that the
semantic knowledge is stored in a table thesaurus
with columns src and tgt. The column src
contains concept IDs of the terms, and the column
tgt contains the terms. This is a common approach
in storing semantic knowledge, used in sources such
as NCI thesaurus and Wordnet’s synsets. For
example, the terms “Acetaminophen”, “Tylenol”,
“Paracetamol” and “APA” all have the same concept
ID “Acetaminophen” in the NCI thesaurus. In
practice, this data could be stored in a table thesaurus
with an additional column rel that stores the type of
the relationship, or it could be stored in XML format
which is a common approach particularly in the
health care domain. For the data stored in XML,
thesaurus can be a view defined in a hybrid XML
relational DBMS such as DB2. Due to space
constraints, we limit our discussion in this paper to
semantic knowledge stored as relational data.
Assume that the trial intervention names are stored in
the table ctintervention(tid,name) and the
DBpedia drug titles are stored in the table
dbpediadrug(tid,title). The following SQL
query returns names and titles related through the
semantic knowledge stored in the table thesaurus:
SELECT DISTINCT c.*, d.*
FROM ctintervention AS c,
dbpediadrug AS d, thesaurus AS t
WHERE src IN
(SELECT src FROM thesaurus
WHERE tgt = c.name)
AND tgt = d.title
For more complex relationships such as hyponymy or
hypernymy, a recursive query can be used to traverse
the ontology tree to a specific depth. In our
experiments in this paper, we only use synonym
information from NCI thesaurus.
Experiments
We show the effectiveness of semantic link discovery
using string and semantic matching techniques
described above using five linkage scenarios in
LinkedCT. Table 4 shows the summary of the results.
The columns “Link #” show the number of links
discovered using exact matching, string matching,
semantic matching and the overall number of links.
Matching each entity in LinkedCT with the target
data source may result in several links. For example,
if there are 10 trials on condition “AIDS”, and
“AIDS” is matched with its corresponding resource
on DBpedia, then 10 links are added to the total
number of links. The columns “Linked Entity #”
show the number of distinct entities in LinkedCT that
Table 4. Statistics of the Links Discovered Using String and Semantic Matching.
Intervention ↔
DBpedia (drug)
Intervention
DrugBank
(drug)
Intervention
DailyMed
(drug)
Condition
DBpedia
(disease)
Condition
Diseasome
↔
↔
↔
↔
Exact Match
Link Linked
Entity #
#
Link
#
String Matching
Linked
Diff
Entity #
%
8,442
867
9,716
9,867
926
11,865 +20.2%
1,938 +109.3% 12,127 +22.9% 1,641
+77.2% 23,493 +138.1% 2,477 +167.5%
14,257
673
24,461 +71.6%
1,296
+92.6% 27,685 +94.2% 1,099
+63.3% 39,396 +176.3% 1,688 +150.8%
164
164
333
+103.0%
330
+101.2%
173
+5.5%
173
+5.5%
342
+108.5%
342
+108.5%
232
192
778
+235.3%
575
+199.5%
301
+29.7%
247
+28.6%
830
+257.8%
615
+220.3%
+15.1%
1,558
Diff
%
Link
#
Semantic Matching
Diff Linked
Diff
Entity #
%
%
+79.7% 10,630 25.9%
are linked. The “Diff %” columns show the
improvement over the exact matching of the strings.
Note that since the links are automatically discovered,
some may be inaccurate. These can best be filtered by
a user from the query results. The benefit of using
automatically discovered links is that a very large
number of links is achieved at the cost of some
precision. In our case, manually inspecting a subset of
the links showed above 98% precision on average in
all the scenarios.
The results in Table 4 clearly show the effectiveness
of both string and semantic matching in all these
linkage scenarios. Using semantic knowledge has
resulted in more links between intervention and drugs
as compared to conditions and diseases since drugs
tend to have more semantic equivalences (brand
names and generic names). The use of semantic
matching however is useful particularly for matching
disease names that are abbreviated. The overall
numbers of the links show that although there is some
overlap between the links discovered using semantic
matching and string matching, there is a significant
number of links that are not found using a single
matching technique.
Overall
Linked
Diff
Entity #
%
+53.9% 11,527 +36.5%
Diff
%
1,913 +120.6%
also means that the integration only has to happen one
time.
A distinctive feature of LinkedCT data space is the
large number of links to other existing data sources.
We demonstrated how state-of-the-art approximate
string matching and ontology-based semantic
matching can be used for discovery of such semantic
links between LinkedCT and several other data
sources. The ratio of external links per entities in
LinkedCT is more than two orders of magnitude
larger than the average ratio in the existing linked
data sources (that are part of the Linking Open Data
community project at W3C). This shows the
effectiveness of our proposed techniques in semantic
interlinking of trials data and web data sources in
general. Due to the declarative nature of our
implementation methods, our approach is also useful
for matching electronic medical records (EMRs) with
trials data. We therefore plan to develop a flexible
declarative framework for semantic link discovery
over medical databases.
References
1.
Conclusion
In this paper, we briefly discussed the challenges
involved in creating the LinkedCT data space, the
first open RDF data source of clinical trials. The data
is published using the principles of linked data. Our
approach can be used for the publication of other
medical datasets as well. Currently, when clinical
researchers and pharmaceutical companies bring
copies of data within their organizations for
integration, they each need to have experts who
understand the connectivity across data sets.
However, with the Linked Data approach, this
responsibility is shifted to the data providers. This is
a much more efficient approach, as the data providers
are the individuals who understand their data best. It
1,334
Link
#
2.
Barrett J, What's behind clinical trial data?(Global
News). Journal of Applied Clinical Trials 18.1 (Jan
2009): 22(1).
Knowledge@Wharton, Eli Lilly's Sidney Taurel:
'Tailored Therapeutics' -- the Pharmaceutical Industry's
Next
Blockbuster?
March
13,
2008
http://knowledge.wharton.upenn.edu/article.cfm?articleid=1915
3.
4.
5.
6.
7.
Berners-Lee T, Linked Data - Design Issues, 07-2006,
http://www.w3.org/DesignIssues/LinkedData.html
Elmagarmid AK, Ipeirotis PG, Verykios VS , Duplicate
record detection: A survey. IEEE Transactions on
Knowledge and Data Engineering 2007;19(1), 1–16.
Mi, M, Clinical trials database: linking patients to
medical research <http//clinicaltrials.gov>. Journal of
Consumer Health on the Internet, 2005; 9(3): 59-67.
Bizer C, Seaborne A, D2RQ - Treating non-RDF
databases as virtual RDF graphs, Proceedings of the 3rd
International Semantic Web Conference (posters), 2004.
Hassanzadeh O, Benchmarking declarative approximate
selection predicates. Master’s thesis, University of
Toronto, Feb. 2007.
| 5cs.CE
|
Vector Quantization using the Improved
Differential Evolution Algorithm for Image
Compression
1
Sayan Nag*
1
Department of Electrical Engineering,
Jadavpur University
Kolkata
India
*
Corresponding Author
Abstract- Vector Quantization (VQ) is a popular image compression technique with a simple decoding
architecture and high compression ratio. Codebook designing is the most essential part in Vector Quantization.
Linde–Buzo–Gray (LBG) is a traditional method of generation of VQ Codebook which results in lower PSNR
value. A Codebook affects the quality of image compression, so the choice of an appropriate codebook is a must.
Several optimization techniques have been proposed for global codebook generation to enhance the quality of
image compression. In this paper, a novel algorithm called IDE-LBG is proposed which uses Improved
Differential Evolution Algorithm coupled with LBG for generating optimum VQ Codebooks. The proposed
IDE works better than the traditional DE with modifications in the scaling factor and the boundary control
mechanism. The IDE generates better solutions by efficient exploration and exploitation of the search space.
Then the best optimal solution obtained by the IDE is provided as the initial Codebook for the LBG. This
approach produces an efficient Codebook with less computational time and the consequences include excellent
PSNR values and superior quality reconstructed images. It is observed that the proposed IDE-LBG find better
VQ Codebooks as compared to IPSO-LBG, BA-LBG and FA-LBG.
Keywords – Image Compression, Vector Quantization, Codebook, Improved Differential Evolution (IDE)
Algorithm, Linde-Buzo-Gray (LBG) Algorithm, Improved Particle Swarm Optimization (IPSO) Algorithm, Bat
Algorithm (BA), Firefly Algorithm (FA).
______________________________________
*
Corresponding Author, Email: [email protected]
1. Introduction
The major role of Image compression lies in medical sciences, internet browsing and navigation applications,
TV broadcasting and so on. With the advancement of science and technology, the storage space requirement and the
need for reduction in transmission time for digital images are becoming essential concerns. The transmission
bandwidth being limited creates a problem. An efficient and proper image compression technique is of prime
requirement to tackle the problem of limited bandwidth. This area of study attracted many researchers over the past
decades who have suggested several techniques for image compression. Vector Quantization (VQ) technique
outperforms other techniques such as pulse code modulation (PCM), differential PCM (DPCM), Adaptive DPCM
which belongs to the class of scalar quantization methods. Vector quantization (VQ) [1, 2], one of the most popular
lossy image compression methods is primarily a c-means clustering approach widely used for image compression as
well as pattern recognition [3, 4], speech recognition [5], face detection [6] speech and image coding because of its
advantages which include its simple decoding architecture and the high compression rate it provides with low
distortion- hence its popularity. Vector Quantization involves the following three steps: (i) Firstly, the image in
consideration is divided into non-overlapping blocks commonly known as input vectors. (ii) Next, a set of
representative image blocks from the input vectors is selected referred to as a codebook and each representative image
vector is referred to as a code-word. (iii) Finally, each of the input vectors is approximately converted to a code-word
in the codebook, the corresponding index of which is then transmitted.
Codebook training is considered sacrosanct in the process of Vector Quantization, because a codebook
largely affects the quality of image compression. Such significance of Codebook training gave new impetus to many
researchers leading to the proliferation of researches to design codebook using several projected approaches. Vector
Quantization methods are categorized into two classes: crisp and fuzzy [7]. Sensitive to codebook initialization, Crisp
VQ follows a hard decision making process. Of this type, C-Means algorithm is the most representative of all. Linde
et al. propounded the famous Linde-Buzo-Gray (LBG) algorithm in this respect. Starting with the smallest possible
codebook size the gradual parlay in the size occurs by using a splitting procedure [8]. By implanting specific functions
(utility measures) in the learning process the performance of the LBG algorithm is significantly enhanced. Fuzzy VQ
incorporates fuzzy C-Means algorithm. The approach is similar to that of fuzzy cluster analysis. Each training vector
is posited to be a property of multiple clusters with membership values which makes the learning a soft decision
making process [9]. The drawback of LBG Algorithm is that it gets stuck in a local optimum. In order to overcome
such problem, a clustering algorithm known as enhanced LBG is proposed [10]. By changing the lower utility codewords to the one with higher utility the problem of local optimal with LBG is overcome.
Recently researches involve the association of the evolutionary optimization algorithms with the LGB
algorithm. This conglomeration for design the codebook has a significant impact in improving the results of the LGB
algorithm. Rajpoot, et al. formulated a codebook using an Ant Colony Optimization (ACO) algorithm [11] yielding
fascinating results thereby clearly depicting the improvement in the results obtained by the hybridization of ACO with
LBG. Codebook is optimized abreast appointing the vector coefficients in a bidirectional graph; next, a appropriate
implementation of placing pheromones on the edges of the graph is expounded in the ACO-LBG Algorithm. The
speed of convergence of the ACO_LBG algorithm is proliferated by Tsaia et al. by discarding the redundant
calculations [12]. Particle Swarm Optimization (PSO) is also an adaptive swarm optimization approach based on
updating the global best (gbest) and local best (lbest) solutions. The ease of improvement and fast convergence to an
expected solution attracted many researchers to apply PSO for solving optimization problems. Particle swarm
optimization along with vector quantization [13] overcomes the drawbacks of the LBG algorithm. Evolutionary fuzzy
particle swarm optimization algorithm [14] is an efficient and robust algorithm in terms of performances compared to
that of the LBG learning algorithms. Quantum particle swarm algorithm (QPSO) is yet another PSO put forward by
Wang et al. for solving the 0–1 knapsack problem [15]. Applied together with LBG, the QPSO-LBG algorithm
outperforms LBG algorithm as well.
Yang et al. devised a new method for image compression which involves an Evolutionary clustering based vector
quantization primarily focused on One-Step Gradient Descent Genetic Algorithm (OSGD-GA). The particular
algorithm is formulated for optimizing the codebooks of the low-frequency wavelet coefficient by expounding the
consequence of every coefficient involved abreast employing fuzzy membership values for automatic clustering [16].
The Firefly Algorithm (FA) is an efficient Swarm Intelligence tool which is largely applied to many engineering
design problems nowadays. Firefly Algorithm is inspired by the social activities of fireflies. In presence of a firefly
with brighter intensity, the one with lower intensity value move towards the former and in the absence of the brighter
firefly then one with lower intensity moves in a random fashion. Horng successfully applied a newly developed Firefly
Algorithm (FA) for designing the codebook for vector quantization [17]. An efficient algorithm by the name of Honey
Bee Mating Optimization (HBMO) is applied for codebook generation yielding a good quality reconstructed image
with modicum of distortion in Image Compression Problems [18]. Dynamic programming is used as an efficient
optimization technique for layout optimization of interconnection networks by Tripathy et al. [19]. Tsolakis et al. [20]
presented a fast fuzzy vector quantization technique for compression of gray scale images. Based on a crisp relation,
an input block is assigned to more than one code-word following a fuzzy vector quantization method [21]. LBG
Algorithm alone cannot guarantee optimum results since the initialization of the codebook has a significantly impact
on the quality of results. Thus, an attempt has been made in this respect where an Improved PSO-LBG (IPSO-LBG)
applied to design codebook for Vector Quantization showed fascinating results when compared to other algorithms
like QPSO-LBG and PSO-LBG [22]. Contextual region is encoded giving high priority with high resolution
Contextual vector quantization (CVQ) algorithm [23]. Huanga et al. devised a dynamic learning vector– scalar
quantization for compressing ECG images [24]. Bat Algorithm (BA) is yet another advanced nature inspired algorithm
which is based on the echolocation behavior of bats with variation in the pulse rates of loudness and emission. BA in
association with LBG is applied as BA-LBG for fast vector quantization in image compression [25]. Results obtained
from BA-LBG when compared with QPSO-LBG, HBMO-LBG, PSO-LBG and FA-LBG showed that BA-LBG
outperforms the other mentioned algorithms.
Optimization problems are applied in various domains of Engineering Design, Structural Optimization,
Economics and Scheduling Assignments. These problems have some mathematical models and Objective Functions.
Two varieties of such problems exist: Unconstrained (without constraints) and Constrained (with Constraints)
involving both continuous as well as discrete variables. They are usually non-linear in nature. The task of finding the
optimal solutions is difficult with several restraints being active at the global optima. Traditional methods for solving
these problems are Gradient Descent, Dynamic Programming and Newton Methods. But they are computationally
inefficient. This mark the advent of the meta-heuristic algorithms which provide feasible solutions in a reasonable
amount of time. There list of meta-heuristics include Genetic Algorithm (GA) [26], Particle Swarm Optimization
(PSO) [27], Gravitational Search Algorithm (GSA) [28], Ant Colony Optimization (ACO) [29, 30], Stimulated
Annealing (SA) [31, 32], Plant Propagation Algorithm (PPA) [33, 34] and so on [35, 36].
Differential Evolution (DE) is one such metaheuristic algorithm used for fast convergence and less
computation time. It optimizes a problem by trying to improve a candidate solution through generations based on a
fitness function and using a crossover strategy which is different from that of the Genetic Algorithm. DE is modified
to Improved DE (IDE) and used to generate high quality codebook with LBG for vector quantization for image
compression. A comparison with other algorithms like IPSO, BA and FA clearly reveals that IDE outperforms all of
the other mentioned algorithm in terms of efficiency as judged from the PSNR values of the images.
The paper is organized as follows: In Section 2, codebook design using generalized LBG Vector Quantization
Algorithm is discussed. In Section 3, the proposed method of IDE-LBG Algorithm is presented. Section 4 includes
the experimental results, discussions and comparisons. Eventually, Section 5 contains the concluding part.
2. Codebook Design for Vector Quantization
The Vector Quantization (VQ) is a block (vector) coding technique which is to be optimized for image
compression. Designing a proper codebook lessens the distortion between reconstructed image and original image
with a modicum of computational time. Let us consider an Image of size N x N which is to be vector quantized. It is
𝑁
𝑁
then subdivided into Nb blocks with size n × n pixels, where, 𝑁𝑏 = ( ∗ ). These subdivided image blocks are known
𝑛
𝑛
as training vectors each of which are of size n × n pixels are represented with Xi (where i = 1, 2, 3,. . ., Nb). In the
process of Codebook design, a Codebook has a set of code words, where the ith Code-word is represented as Ci (where
i = 1, 2, ..., Nc) where Nc is the total number of Code-words in Codebook. Every subdivided image vector is
approximately represented by the index of a Code-word using a closest match technique. The approach is based on
the minimum Euclidean distance between the vector and corresponding code-words. In the decoding phase, the same
codebook is used to translate the index back to its corresponding Code-word for image reconstruction. Then the
distortion between the actual and the reconstructed image is calculated. The distortion D between training vectors and
the codebook is given as:
𝐷=
1
𝑁𝑐
𝑐
∑𝑁
∑𝑁 𝑏
𝑗=1 𝑖=1 𝑢𝑖𝑗 . ‖𝑋𝑖 − 𝐶𝑗 ‖
2
(1)
Subject to the following restraints:
𝑁
𝑐
𝐷 = ∑𝑗=1
𝑢𝑖𝑗 = 1, ∀ 𝑖 ∈ {1, 2, … , 𝑁𝑏 }
1, iff 𝑋𝑖 ∈ 𝑗𝑡ℎ cluster
0, otherwise
𝑢𝑖𝑗 = {
(2)
(3)
The other two necessary conditions which are to be followed for an optimal vector quantizer are:
The partition 𝑅𝑗 ∀ 𝑗 = 1,2,3, … , 𝑁𝑐 must meet the following criterion:
The Code-word Cj should be defined as the centroid of Rj such that,
𝑅𝑗 ͻ {𝑥 ∈ 𝑋: 𝑑(𝑐, 𝐶𝑗 ) < 𝑑(𝑥, 𝐶𝑘 ), ∀ 𝑘 ≠ 𝑗}
𝐶𝑗 =
1
𝑁𝑗
𝑁
𝑗
∑𝑖=1
𝑥𝑖 , ∀ 𝑥𝑖 ∈ 𝑅𝑗
(4)
(5)
where Nj is the net number of vectors which belong to Rj. The procedure for the most commonly used method for
codebook design in Vector Quantization, the Lloyd-Buzo-Gray (LBG) Algorithm is provided as follows:
Steps:
1.
2.
3.
4.
5.
Initialization: Iteration counter is set as m = 1 and the initial distortion D1 = 0. Initial codebook C1 is of size
N.
One block (vector) from the set of training vectors is selected and compared with all code-words from
codebook, Cm-1 = {Yi} to find the closest code-word using nearest neighbor condition and then the code-word
set is added.
Once all the blocks are grouped into sets of code-words using the above mentioned step, the centroids of all
the code-word sets are evaluated to give an improved codebook Cm.
Next, the total distortion of all code-word sets from the respective centroid points are calculated. The average
distortion is given as Dm.
If |(Dm-1 - Dm)| ≤ ε, the algorithm terminates and the current codebook is recorded as the final answer.
Otherwise, m = m + 1, the current codebook is used as the initial codebook for step 1 and the entire process
is then repeated.
LBG Algorithm ensures the reduction in distortion with increase in iteration number. But there is no assurance
regarding the resulting codebook, whether it will become an optimum one or not. It is seen that the initial condition
of the codebook has a significant effect on the results. So the initial codebook selection is of prime importance in the
LBG Algorithm. Thus we propose to combine an efficient modified heuristic Improved Differential Evolution (IDE)
with LBG Algorithm, thereby giving a state-of-the-art IDE-LBG Algorithm and the problem initial codebook selection
is tackled with this attempt.
3. Proposed IDE-LBG Vector Quantization A:gorithm
Differential Evolution (DE) is a powerful population-based efficient global optimization technique which was
posited by Storn et al. [37]. It has been successfully applied to diverse fields including pattern recognition,
communication and so on. In our algorithm we have used the DE/current to best/1 scheme. We have modified DE to
IDE with certain modifications in the mutation strategy and in the boundary control strategy. In the mutation strategy
we have considered a different scaling factor. The boundary control strategy is improved such that when the limits are
crossed then based on a probability there can be either of the two possibilities: (i) the solutions are adjusted within the
boundary limits by making the values equal to the values of the limits, or, (ii) by randomly generating a fresh candidate
within the bounds and replacing the one which violated the limits.
The training image vectors or blocks are demarcated into different groups. One block from each group is selected
at random as the initial candidates. A candidate is constructed by Nc × 16 pixels, where Nc is the size of a codebook.
In the process, all the candidates of the entire population are moved nonlinearly to cover all searching space, a process
known as exploration. In each generation, G, IDE uses the mutation and crossover operations to engender a trial vector
Ui,G for each individual vector XiG, known as target vector in the present population. The mutation and crossover
operations along with the fitness function are described below:
a.
b.
Mutation:
There exists an associated mutant vector, Vi,G for each target vector XiG at generation G in the current
population. The DE/current to best/1 scheme dictates the mutation operations as follows:
𝐺
𝑉𝑖𝐺 = 𝑋𝑖𝐺 + 𝐹 . (𝑋𝑏𝑒𝑠𝑡
− 𝑋𝑖𝐺 ) + 𝐹. (𝑋𝑟𝐺1 − 𝑋𝑟𝐺2 )
(6)
where r1 and r2 are random integers which are essentially mutually different and lies in the range [1, NP]. F
is called the Weighting Factor, F = 3.randn, where randn ~ N(0,1), N being the standard normal
𝐺
distribution. 𝑋𝑏𝑒𝑠𝑡
is by far the fittest individual in the population at generation G.
Crossover Operation:
Post-mutation period, a binomial crossover strategy is adopted to generate a trial vector Ui,G = {u1iG, u2iG,
….., udiG,}, where d is the dimension. CR is the crossover operator, a value between 0 and 1. We have
considered CR = 0.9.
𝑈𝑗𝑖𝐺 = {
c.
𝑣𝑗𝑖𝐺 , 𝑖𝑓 (𝑟𝑎𝑛𝑑𝑗 (0, 1) ≤ 𝐶𝑅, 𝑜𝑟, 𝑗 = 𝑗𝑟𝑎𝑛𝑑
(7)
𝑥𝑗𝑖𝐺 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Fitness function:
Size of the testing images is N x N, 𝐼 is the recompressed image pixels, and 𝐼 ̅ is the compressed image pixels.
The fitness function is given in the following set of equations (8-10). First Mean Square Error (MSE) is
calculated, then Peak Signal to Noise Ratio (PSNR) is evaluated which is nothing but the fitness function of
the problem.
𝑀𝑆𝐸 =
𝑁
̅ 2
∑𝑁
𝑖=1 ∑𝑗=1(𝐼𝑖,𝑗 − 𝐼𝑖,𝑗 )
(8)
𝑁×𝑁
𝑃𝑆𝑁𝑅 = 10 × log(
2552
)
(9)
𝑓𝑓𝑖𝑡𝑛𝑒𝑠𝑠 = max(𝑃𝑆𝑁𝑅)
(10)
𝑀𝑆𝐸
The detailed steps for IDE-LBG Algorithm for finding optimal Codebook are described as follows:
1.
2.
3.
Parameters and Functions: Fitness function = ffitness, iteration parameter = iter = 1, number of generations =
NGen, codebook size = Nc, and population size = NP.
Testing images divided into Nb non-overlapping image blocks or input vectors (bk ; k = 1, 2, …, Nb). They
are put into a training set tSet, which of each block is n × n pixels, the total pixels of each image block is
evaluated and the blocks are sorted in the increasing order of pixel number. Here we have considered n = 4.
𝑁
Nb blocks are separated into Nc groups according to the index number of blocks, each group having 𝑏 number
𝑁𝑐
of blocks. Blocks from groups are selected at random to form the Code-words to form the initial population
of size NP.
4.
5.
6.
7.
8.
At current iteration G, the individual candidates of the population are updated using the mutation and
crossover strategies using equations (6)-(7). Boundary control is applied so that the solutions remain within
the limits of the search space.
𝐺
𝐺+1
The fitness of the off-springs are calculated and compared. If any individual is better than 𝑋𝑏𝑒𝑠𝑡
, 𝑋𝑏𝑒𝑠𝑡
is
𝐺+1
𝐺
updated with value equal to that of the fittest individual, otherwise 𝑋𝑏𝑒𝑠𝑡 = 𝑋𝑏𝑒𝑠𝑡 .
𝐺𝑒𝑛
If current value of iteration parameter iter reaches NGen, then 𝑋𝑏𝑒𝑠𝑡
is considered as the optimal solution
obtained to be used as an initial codebook for the LBG algorithm, otherwise iter = iter + 1 and the process is
repeated again from step 4.
𝐺𝑒𝑛
Global best candidate solution 𝑋𝑏𝑒𝑠𝑡
obtained from IDE is used as an initial codebook for LBG Algorithm
where each Code-word essentially represents the centroid point of current Code-word set.
The LBG Algorithm is then applied whose steps are provided in detail in the previous section (Section 2).
4. Experimental Results
In this section, we performed computational experiments and compared the results to show the efficacy of our
newly propounded IDE-LBG approach for VQ for Image Compression. All the simulations are carried out in
MATLAB 2013a in a workstation with Intel Core i3 2.9 GHz processor. Five tested 512 × 512 grayscale images,
namely, ‘‘LENA’’, ‘‘PEPPER’’, ‘‘BABOON’’, ‘‘GOLDHILL’’ and ‘‘LAKE’’, were selected to be used as training
images to train different size of Codebooks and as the basis of comparison of proffered IDE-LBG with other
algorithms. The selected images are compressed with IDE-LBG, IPSO-LBG, BA-LBG and FA-LBG and the results
are compared. The testing image which is to be compressed is subdivided into non-overlapping image blocks of size
4 × 4 pixels. Each subdivided image block is a training vector of (4 x 4) 16 dimensions. The training vectors to be
512
encoded are (
2
×
512
2
) 16384 in number. The subdivided image blocks are sorted by the total pixels of each image
block and are grouped by the codebook size. Six different codebook sizes are used for comparison with values 8, 16,
32, 64, 128, 256. The values considered are: ε = 0.001, NP = 20, are NGen = 10, for the training process. The members
of the initial population of IDE is chosen at random from each group. Eventually, the best solution (Codebook)
obtained by the IDE is used as the initial Codebook for the LBG algorithm-the problem of LBG with random
initialization as mentioned before is overcome. We have considered bitrate per pixel (bpp) to evaluate the data size of
the compressed image for different codebook sizes and the Peak Signal to Noise Ratio (PSNR) values are calculated
for individual codebook sizes. Then the plots of PSNR values versus corresponding bpp values give a clear picture of
the quality of reconstructed images with different algorithms considered, thus giving an idea about the effectiveness
of the proposed IDE-LBG Algorithm.
𝑏𝑝𝑝 =
log2 𝑁𝑐
𝑘
(11)
where Nc is the size of the Codebook and k is the size of the non-overlapping image block.
The total number of runs considered is 10. Table 1 shows the average PSNR values in ten runs (averaged)
obtained by the IDE-LBG, IPSO-LBG, BA-LBG, and FA-LBG algorithms for different codebook sizes (Nc = 8, 16,
32, 64, 128 and 256). Horng et al. [18] simulated five different algorithms in C++6.0, with population size = 100 and
number of iterations = 100. PSNR values obtained by different algorithms are also reported in the same paper, but
since the platform and the value of parameters, namely, the iteration number and the population size is different, there
will be some dissimilarity in the final results obtained for PSNR values of FA-LBG in this paper with that obtained in
[17]. The results from the Table 1 (plots are in Fig. 3) confirmed that the fitness (PSNR) of the five test images using
the IDE-LBG algorithm is higher than the IPSO-LBG, BA-LBG, and FA-LBG algorithm, yet the difference in PSNR
values of IDE-LBG and IPSO-LBG is not very distinct for most of Codebook sizes for various images. Five
reconstructed images of Codebook size 128 and block size 16 are also displayed in Fig. 2. It is clearly visible that the
reconstructed image quality obtained using the IDE-LBG algorithm has the edge over the IPSO-LBG, BA-LBG and
FA-LBG algorithms.
Table 1. PSNR values of image compression for 5 different 512 x 512 images with 6 different Codebook sizes.
Nc
8
16
32
64
128
256
Method
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
(a)LENA
Lena
25.82
25.75
24.20
24.01
27.19
27.03
25.67
25.42
28.50
28.38
25.89
25.73
29.39
29.32
26.47
26.22
30.45
30.45
28.23
27.75
31.47
31.46
28.94
28.49
(b)BABOON
Baboon
20.89
20.33
19.19
19.17
21.11
21.02
20.17
20.14
21.28
21.23
20.24
20.26
22.84
22.72
21.86
21.74
24.09
24.02
22.43
22.21
24.75
24.66
22.92
22.58
(c)GOLDHILL
Goldhill
26.97
26.12
25.01
24.38
27.32
27.11
26.23
25.89
28.83
28.14
26.39
26.36
28.62
28.44
27.26
27.24
29.92
29.80
28.30
28.18
30.63
30.62
29.19
28.93
Pepper
25.79
25.78
24.72
24.64
26.51
26.40
25.95
25.31
27.61
27.23
26.58
26.53
29.25
28.95
28.69
28.21
30.84
30.61
29.90
29.71
31.66
31.62
30.54
30.44
(d)PEPPERS
Barbara
25.32
25.27
24.35
23.58
26.03
25.96
24.88
24.49
27.58
27.55
26.38
26.36
28.18
28.07
27.98
27.85
29.94
29.92
29.01
28.98
30.48
30.41
29.34
29.23
(e)BARBARA
Fig1. The five testing Images of size 512 x 512.
(a)LENA reconstructed
(b)BABOON reconstructed
(c)GOLDHILL reconstructed
(d)PEPPERS reconstructed
(e)BARBARA reconstructed
Fig2. The five reconstructed Images of Codebook size 128 and Block size 16.
IDE-LBG
IPSO-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
BA-LBG
FA-LBG
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
30.38
24.17
31
29.38
30
PSNR in db
PSNR in db
28
27
PSNR in db
23.17
29
22.17
21.17
28.38
27.38
26.38
26
20.17
25.38
25
24
19.17
0.1875 0.25 0.3125 0.375 0.4375 0.5
Bitrate
(a)PSNR of LENA
IDE-LBG
IPSO-LBG
24.38
0.1875 0.25 0.3125 0.375 0.4375 0.5
Bitrate
0.1875 0.25 0.3125 0.375 0.4375 0.5
Bitrate
(b)PSNR of BABOON
BA-LBG
FA-LBG
(c)PSNR of GOLDHILL
IDE-LBG
IPSO-LBG
BA-LBG
FA-LBG
31.6
29.57
30.6
28.57
PSNR in dB
PSNR in db
29.6
28.6
27.6
26.6
27.57
26.57
25.57
24.57
25.6
24.6
23.57
0.1875
0.25
0.3125
0.375
Bitrate
(d)PSNR of PEPPERS
0.4375
0.5
0.1875
0.25
0.3125
0.375
Bitrate
(e)PSNR of BARBARA
Fig 3. Average PSNR of six VQ methods for 5 test images of size 512 x 512
0.4375
0.5
5. Conclusion
In this study, an Improved Differential Evolution (IDE) Algorithm based codebook training has been presented
for Image Compression. The Peak signal to noise ratio (PSNR) of vector quantization is optimized by using IDE-LBG
Algorithm where PSNR is considered as the fitness function for the optimization problem. The algorithm parameters
have been tuned suitably for efficient codebook design. The proposed IDE-LBG Algorithm is tested on five standard
512 x 512 grayscale images. Results reveal the potency of the proposed algorithm when compared to IPSO-LBG, BALBG and FA-LBG Algorithms. It is observed that using the proposed IDE-LBG Algorithm the PSNR values and the
quality of reconstructed image obtained are much better than that obtained from the other algorithms in comparison
for six different Codebook sizes.
References
[1]
R.M. Gray, Vector quantization, IEEE Signal Process. Mag. 1 (2) (1984) 4–29.
[2]
D. Ailing, C. Guo, An adaptive vector quantization approach for image segmentation based on SOM
network, Neurocomputing 149 (2015) 48–58.
[3]
H.B. Kekre, Speaker recognition using vector quantization by MFCC and KMCG clustering algorithm, in:
IEEE International Conferences on Communication, Information & Computing Technology (ICCICT),
IEEE, Mumbai, 2012, pp. 1–5.
[4]
C.W. Tsai, C.Y. Lee, M.C. Chiang, C.S. Yang, A fast VQ codebook generation algorithm via pattern
reduction, Pattern Recognit. Lett. 30 (2009) 653–660.
[5]
S.K. Frank, R.E. Aaron, J. Hwang, R.R. Lawrence, Report: a vector quantization approach to speaker
recognition, AT&T Tech. J. 66 (2) (2014) 14–16.
[6]
C.H. Chan, M.A. Tahir, J. Kittler, M. Pietikäinen, Multiscale local phase quantization for robust
component-based face recognition using kernel fusion of multiple descriptors, IEEE Trans. Pattern Anal.
Mach. Intell. 35 (5) (2013) 1164–1177.
[7]
G.E. Tsekouras, D. Darzentas, I. Drakoulaki, A.D. Niros, Fast fuzzy vector quantization, in: IEEE
International Conference on Fuzzy Systems (FUZZ), IEEE, Barcelona, 2010, pp. 1–8.
[8]
Y. Linde, A. Buzo, R.M. Gray, An algorithm for vector quantizer design, IEEE Trans. Commun. 28 (1)
(1980) 702–710.
[9]
G.E. Tsekouras, D.M. Tsolakis, Fuzzy clustering-based vector quantization for image compression, in: A.
Chatterjee, P. Siarry (Eds.), Computational Intelligence in Image Processing, Springer Berlin, Heidelberg,
2012, pp. 93–105.
[10]
G. Patane, M. Russo, The enhanced LBG algorithm, Neural Netw. 14 (2002) 1219–1237.
[11]
A. Rajpoot, A. Hussain, K. Saleem, Q. Qureshi, A novel image coding algorithm using ant colony system
vector quantization, in: International Workshop on Systems, Signals and Image Processing, Poznan,
Poland, 2004, pp. 13–15.
[12]
C.W. Tsaia, S.P. Tsengb, C.S. Yangc, M.C. Chiangb, PREACO: a fast ant colony optimization for
codebook generation, Appl. Soft Comput. 13 (2013) 3008–3020.
[13]
M. Kumar, R. Kapoor, T. Goel, Vector quantization based on self-adaptive particle swarm optimization,
Int. J. Nonlinear Sci. 9 (3) (2010) 311–319.
[14]
H.M. Feng, C.Y. Chen, F. Ye, Evolutionary fuzzy particle swarm optimization vector quantization learning
scheme in image compression, Expert Syst. Appl. 32 (2007) 213–222.
[15]
Y. Wang, X.Y. Feng, X.Y. Huang, D.B. Pu, W.G. Zhou, Y.C. Liang, et al., A novel quantum swarm
evolutionary algorithm and its applications, Neurocomputing 70 (2007) 633–640.
[16]
S. Yang, R. Wu, M. Wang, L. Jiao, Evolutionary clustering based vector quantization and SPIHT coding
for image compression, Pattern Recognit. Lett. 31 (2010) 1773–1780.
[17]
M.H. Horng, Vector quantization using the firefly algorithm for image compression, Expert Syst. Appl. 39
(1) (2012) 1078–1091.
[18]
M.H. Horng, T.W. Jiang, Image vector quantization algorithm via honey bee mating optimization, Expert
Syst. Appl. 38 (3) (2011) 1382–1392.
[19]
P.K. Tripathy, R.K. Dash, C.R. Tripathy, A dynamic programming approach for layout optimization of
interconnection networks, Eng. Sci. Technol. 18 (2015) 374–384.
[20]
D. Tsolakis, G.E. Tsekouras, A.D. Niros, A. Rigos, On the systematic development of fast fuzzy vector
quantization for gray scale image compression, Neural Netw. 36 (2012) 83–96.
[21]
G.E. Tsekouras, A fuzzy vector quantization approach to image compression, Appl. Math. Comput. 167 (1)
(2005) 539–5605.
[22]
Ping-Yi C, Tsai JT, Chou JH, Ho WH, Shi HY, Chen SH. Improved PSO-LBG to design VQ codebook.
InSICE Annual Conference (SICE), 2013 Proceedings of 2013 Sep 14 (pp. 876-879). IEEE.
[23]
S.M. Hosseini, A. Naghsh-Nilchi, Medical ultrasound image compression using contextual vector
quantization, Comput. Biol. Med. 42 (2012) 743–750.
[24]
B. Huanga, Y. Wanga, J. Chen, ECG compression using the context modeling arithmetic coding with
dynamic learning vector–scalar quantization, Biomed. Signal Process. Control 8 (2013) 59–65.
[25]
Chiranjeevi Karri, Umaranjan Jena, Fast vector quantization using a Bat algorithm for image compression,
In Engineering Science and Technology, an International Journal, Volume 19, Issue 2, 2016, Pages 769781, ISSN 2215-0986, https://doi.org/10.1016/j.jestch.2015.11.003.
[26]
S.-J. Wu and P.-T. Chow, Genetic algorithms for nonlinear mixed discrete-integer optimization problems
via meta-genetic parameter optimization, Engineering Optimization, vol. 24, no. 2, pp. 137–159, 1995.
[27]
R. Eberhart and J. Kennedy, A new optimizer using particle swarm theory, in Proceedings of the 6th
International Symposium on Micro Machine and Human Science (MHS ’95), pp. 39– 43, IEEE, Nagoya,
Japan, October 1995.
[28]
Rashedi, E., Nezamabadi-pour, H., Saryazdi, S., GSA: A Gravitational Search Algorithm 179(13), 2232–
2248 (2009).
[29]
M. Dorigo and G. D. Caro, Ant algorithms for discrete optimization, Artificial Life, vol. 5, no. 3, (1999),
pp. 137-172.
[30]
M. Dorigo and L. M. Gambardella, Ant colony system: a cooperative learning approach to the traveling
salesman problem, IEEE Transactions on Evolutionary Computation, vol. 1, no. 1, (1997), pp. 53-66.
[31]
C. Zhang and H.-P. Wang, Mixed-discrete nonlinear optimization with simulated annealing, Engineering
Optimization, vol. 21, no. 4, pp. 277–291, 1993.
[32]
E. H. L. Aarts, J. H. M. Korst, and P. J. M. van Laarhoven, Simulated annealing, in Local Search in
Combinatorial Optimization, pp. 91–120, 1997.
[33]
Nag S. Adaptive Plant Propagation Algorithm for Solving Economic Load Dispatch Problem. arXiv
preprint arXiv:1708.07040. 2017 Aug 4.
[34]
Nag S. A Type II Fuzzy Entropy Based Multi-Level Image Thresholding Using Adaptive Plant Propagation
Algorithm. arXiv preprint arXiv:1708.09461. 2017 Aug 23.
[35]
JC Bansal, H Sharma and SS Jadon, Artificial bee colony algorithm: a survey. Int. J. of Advanced
Intelligence Paradigms 5.1 (2013): 123-159.
[36]
X.S. Yang, Firefly algorithms for multimodal optimization, LNCS, vol. 5792, pp. 169–178. Springer,
Heidelberg (2009).
[37]
Storn, R., Price, K., Differential evolution-a simple and efficient heuristic for global optimization over
continuous spaces. Journal of Global Optimization 11, 341–359 (1997).
| 9cs.NE
|
A large neighbourhood based heuristic for two-echelon routing
problems
arXiv:1505.08003v2 [cs.DS] 4 Oct 2016
Ulrich Breunig
Department of Business Administration
University of Vienna, Austria
[email protected]
Verena Schmid
Christian Doppler Laboratory for Effcient Intermodal Transport Operations
University of Vienna, Austria
[email protected]
Richard F. Hartl
Department of Business Administration
University of Vienna, Austria
[email protected]
Thibaut Vidal
Departemento de Informtica
Pontifcia Universidade Catlica do Rio de Janeiro, Brazil
[email protected]
October 5, 2016
Abstract
In this paper, we address two optimisation problems arising in the context of city logistics and two-level
transportation systems. The two-echelon vehicle routing problem and the two-echelon location routing
problem seek to produce vehicle itineraries to deliver goods to customers, with transits through intermediate
facilities. To efficiently solve these problems, we propose a hybrid metaheuristic which combines enumerative
local searches with destroy-and-repair principles, as well as some tailored operators to optimise the selections
of intermediate facilities. We conduct extensive computational experiments to investigate the contribution
of these operators to the search performance, and measure the performance of the method on both problem
classes. The proposed algorithm finds the current best known solutions, or better ones, for 95% of the twoechelon vehicle routing problem benchmark instances. Overall, for both problems, it achieves high-quality
solutions within short computing times. Finally, for future reference, we resolve inconsistencies between
different versions of benchmark instances, document their differences, and provide them all online in a
unified format.
1
Introduction
The traffic of vehicles is a major nuisance in densely populated areas. Trucks disturb peoples’ well-being
by emitting noise and air pollution. As the amount of goods in transit increases, a proper planning of road
networks and facility locations becomes critical to mitigate congestion. To face these challenges, algorithmic
tools have been developed to optimise city logistics at several levels: considering traffic regulation, itineraries
and network design choices. Boosting the efficiency of goods transportation from suppliers to customers
1
presents important challenges for different planning horizons. On the operational level, efficient itineraries must
be found for the available vehicles from day to day, e.g., reducing the travelled distance. On a tactical level,
the overall fleet size, vehicle dimensions, capacities and characteristics are of interest. Larger trucks are more
efficient in terms of cost per shipped quantity, whereas smaller vehicles are more desirable in city centres: they
emit less noise, and only need smaller parking spots. Finally, the clever selection of locations for production
sites, warehouses, and freight terminals is a typical strategic decision.
In this article, we consider the problem of jointly determining good routes to deliver goods to customers,
and at which intermediate facilities a switch from larger trucks to smaller city freighters should happen. This
problem is challenging, due to the combination of these two families of combinatorial decisions.
To address this problem, we propose a simple metaheuristic, which combines local and large neighbourhood
search with the ruin and recreate principle. The method is conceptually simple and fast, exploiting a limited
subset of neighbourhoods in combination with a simple strategy for closing and opening intermediate facilities.
We conduct extensive computational experiments on two-echelon vehicle routing problem (2EVRP) and twoechelon location routing problem with single depot (2ELRPSD) instances to investigate the contribution of
these operators, and measure the performance of the method on both problem classes. For the 2EVRP, this
algorithm is able to reach or outperform 95% of the best known solutions. In general, for both problems,
high-quality solutions are attained in short computing times. As such, this algorithm will serve as a good basis
for future developments on more complex and realistic two-tiered delivery problems.
The paper is organised as follows. The problems are described in Section 2 and an overview of the related
literature is given in Section 3. Mathematical formulations are presented in Section 4. Section 5 describes the
proposed algorithm. Section 6 reviews current available benchmark instances and examines the performance of
the proposed method. Section 7 concludes.
2
Problem description
Vehicle routing problems (VRPs) are a class of combinatorial optimisation problems, which aim to find good
itineraries to service a number of customers with a fleet of vehicles. The 2EVRP is a variant of a VRP, which
exploits the different advantages of small and large vehicles in an integrated delivery system. The goal is
to design an efficient distribution chain, organised in two levels: trucks operate on the first level between a
central depot and several selected intermediate distribution facilities, called satellites. The second level also
includes the satellites – because both levels are interconnected there – as well as the end-customers. Small city
freighters are operated between satellites and customers. The depot supplies sufficient quantities to satisfy all
customer demands. The products are directly transferred from trucks to city freighters at satellite locations.
These city freighters will perform the deliveries to the final customers. Any shipment or part of shipment has
to transit through exactly one satellite, and the final delivery to the customer is done in one block. As such,
split deliveries are not allowed for city freighters. The quantity (“demand”) of goods shipped to a satellite is
not explicitly given, but evaluated as the sum of all customer demands served with city freighters originating
from this satellite. Depending on the second level itineraries, split deliveries can occur on the first level since
the total quantity needed at one satellite can exceed the capacity of one truck.
Finding good combined decisions for routing and intermediate facility openings is significantly more difficult
than in well-studied settings such as the capacitated vehicle routing problem (CVRP). The special case of a
2EVRP with only one satellite can be seen as a VRP (Cuda et al., 2015; Perboli et al., 2011). The first level of
the 2EVRP reduces to a CVRP with split deliveries. The structure of the second level is a multi-depot vehicle
routing problem (MDVRP), where the depots correspond to the satellite locations (Jepsen et al., 2012). Those
two levels have to be synchronised with each other. The 2EVRP is a generalisation of the classical VRP and is
thus NP-hard.
Figure 1 shows different set-ups for goods distribution. The depot is represented by a triangle, satellites
by squares, and customers by circles. Figures 1a and 1b show graphical representations of a split delivery
vehicle routing problem (SDVRP) and a MDVRP respectively. Figures 1c–1e represent feasible solutions for
the 2EVRP: with split deliveries occurring at one of the satellites, without split deliveries, and in Figure 1e a
solution where only a subset of satellites is used.
The proposed algorithm has primarily been designed for the 2EVRP, and then tested on the 2ELRPSD,
which includes additional tactical decisions. The basic structure of the 2ELRPSD is very similar to the 2EVRP.
2
(a) SDVRP / first level of
2EVRP
(c) 2EVRP I
(b) MDVRP / second level of
2EVRP
(d) 2EVRP II
(e) 2EVRP III
Figure 1: Subproblems related to the 2EVRP and different solutions depending on which intermediate facilities
are used
The main difference is that it corresponds to a more tactical planning since only potential locations for depots
or satellites are known and the use of any location results in opening costs. In contrast with the 2EVRP, the
fleet size is unbounded, but fixed costs are counted for the use of each vehicle. The classical benchmark sets
from the literature include different costs per mile for large first level trucks and small city freighters, unlike in
2EVRP benchmark instances, where mileage costs are identical for all vehicles. Finally, split deliveries are
not allowed at both levels. We thus applied our algorithm to 2ELRPSD instances, where location decisions
have to be taken at the secondary facilities. Following the notations of Boccia et al. (2011) we focus on 3/T /T
problems.
3
Literature Review
Jacobsen and Madsen (1980) were amongst the first to introduce a two-echelon distribution optimisation
problem. They proposed a three stage heuristic to solve the daily distribution of newspapers in Denmark, but
no mathematical model was designed. Several possible transfer points were considered to transfer newspapers
from one vehicle to another. The solution to this problem consists of three layers of decisions: the number
and location of transfer points, the tours from the printing office and the tours from the transfer points to the
retailers. An improved solution algorithm for the same problem can be found in Madsen (1983). Following the
nomenclature of the authors and the classification in the recent survey on two-echelon routing problems by
Cuda et al. (2015), the problem includes location decisions. Nevertheless, from today’s point of view, it cannot
be categorised as a two-echelon location routing problem (2ELRP) as there are no opening costs associated
with the use of intermediate facilities and retailers can also be served directly from the printing office without
using intermediate nodes. In addition, the problem cannot be categorised as a 2EVRP since first-level split
deliveries are not allowed.
Two-echelon vehicle routing problem Crainic et al. (2004) used data from Rome to study an integrated
urban freight management system. As large trucks cannot pass through the narrow streets in the city centre,
3
they used intermediate facilities to redistribute loads from large trucks to smaller vehicles. The city was
divided into several commercial and external zones, and a mathematical location-allocation formulation was
proposed and solved using a commercial solver. A comparison between solutions for delivering goods lead to
the conclusion that intermediate facilities reduce the use of large trucks significantly, and more work is done by
smaller city freighters.
Crainic et al. (2009) formulated a time dependent version of the problem, including time windows at the
customers. To our knowledge, there are no test instances or solution approaches for this variant so far. Crainic
et al. (2010) studied the impact of different two-tiered transportation set-ups on total cost. According to their
results, the 2EVRP can yield better solutions than the VRP if the depot is not located within the customer
area but externally. Perboli et al. (2011) introduced a flow-based mathematical formulation and generated
three sets of instances for the 2EVRP with a maximum of 50 customers and four satellites, based on VRP
instances. Their branch-and-cut approach is able to solve instances with up to 21 customers to optimality.
Perboli and Tadei (2010) solved additional instances and reduced the optimality gap on others by means of
new cutting rules.
Crainic et al. (2011) solved the 2EVRP with a multi-start heuristic. The method first assigns customers to
satellites heuristically, and then solves the remaining VRPs with an exact method. In a perturbation step, the
assignment of customers to satellites is changed, then the problem is solved again, until a number of iterations
is reached.
Jepsen et al. (2012) presented a branch-and-cut method, solving 47 out of 93 test instances to optimality,
34 of them for the first time. The authors have been the first to consider a constraint on the number of vehicles
per satellite, although it was already specified before in the existing data set. This additional constraint had
not been taken into account by previous publications.
Hemmelmayr et al. (2012) developed a metaheuristic based on adaptive large neighbourhood search (ALNS)
with a variety of twelve destroy and repair operators. This approach tends to privilege accuracy (high quality
solutions) over simplicity and flexibility (Cordeau et al., 2002). The authors also introduced new larger test
instances with up to 200 customers and five to ten satellites. Note that the results of ALNS on the problem
instances with 50 customers cannot be compared with the proven optimal solutions by Jepsen et al. (2012),
since the algorithm does not consider a limit on the number of vehicles per satellite, but rather a constraint
on the total number of vehicles. For most problem instances, this algorithm found the current best known
solution or improved it.
Santos et al. (2014) implemented a branch-and-cut-and-price algorithm, which relies on a reformulation
of the problem to overcome symmetry issues. They also introduced several classes of valid inequalities. The
algorithm performs well in comparison to other exact methods, and they reported solutions for instances with
up to 50 customers.
Baldacci et al. (2013) presented a promising exact method to solve the 2EVRP. They decomposed the
problem into a limited set of MDVRPs with side constraints. Detailed results and comparisons with previous
publications were provided, as they considered both variants on the instances with 50 customers: with and
without the constraint on vehicles per satellite. They also introduced a new set of instances with up to 100
customers.
Recently Zeng et al. (2014) published a greedy randomized adaptive search procedure, combined with
a route-first cluster-second splitting algorithm and a variable neighbourhood descent. They also provide a
mathematical formulation. They provide quite good results, although unfortunately their algorithm was tested
only with instances comprising up to 50 customers.
Two-echelon location routing problem The capacitated 2ELRP is by far the most studied version of
the 2ELRP. Many papers consider location decisions only at the second stage, either because the use of depots
is an outcome of the first level routing optimisation, or they consider problems with only one single depot
location which is known a priori (2ELRPSD).
Laporte (1988) presented a general analysis of location routing problems and multi-layered problem variants.
They compared several mathematical formulations and their computational performance. In a slightly different
context, Laporte and Nobert (1988) formulated a vehicle flow model for the 2ELRP. The locations of the
depots are assumed to be fixed and unchangeable, such that the location decisions only occur for the satellites.
Following the notation by Boccia et al. (2011) they analysed 3/R/R, 3/R/T , 3/T /R, and 3/T /T problem
settings.
4
Boccia et al. (2011) provided three mathematical formulations for the 2ELRP using one-, two-, and three
indexed variables inspired from VRP and MDVRP formulations. A commercial solver was used to solve some
instances generated by the authors, comparing two of the formulations in terms of speed and quality.
Nguyen et al. (2012a) introduced two new sets of instances for the 2ELRPSD. They implemented a GRASP
with path relinking and a learning process and provided detailed results. In Nguyen et al. (2012b) the authors
improved their findings on the same instances by using a multi-start iterated local search.
Contardo et al. (2012) proposed a branch-and-cut algorithm, which is based on a two-indexed vehicle
flow formulation, as well as an ALNS heuristic. Both solution approaches were applied to one set of 2ELRP
instances, and two sets of 2ELRPSD instances, outperforming previous heuristics.
Schwengerer et al. (2012) extended a variable neighbourhood search (VNS) solution approach for the
location routing problem from Pirkwieser and Raidl (2010) and applied it to several instance sets, including
the two aforementioned ones with a single depot.
For further details on both problem classes, we refer to the recent survey by Cuda et al. (2015). The previous
literature review shows that, on the side of exact methods, the best approaches still cannot consistently solve
(in practicable time) 2EVRP instances with more than 50 customers to optimality. On the side of heuristics,
only few methods have been designed and tested to deal with instances with more than 50 delivery locations.
To the best of our knowledge, only one heuristic has reported, to this date, computational results on larger
2EVRP instances (Hemmelmayr et al., 2012). Hence, there is a need for a more fine-grained study of solution
methods for larger problems, as well as for simpler approaches able to efficiently deal with the two families
of decisions related to routing and intermediate facilities selection. The proposed method has been designed
to cope with these challenges. We developed a technique which performs very well on the classic benchmark
instances and, in the meantime, uses fewer and simpler neighbourhood structures than previously published
algorithms. During our research, we finally found inconsistencies regarding different benchmark instances used
in previous papers. Some slightly different instances have also been referenced with the same name. Thus we
collected the different versions and make them available online with unique names in a uniform file format, as
described in Section 6.1.
4
Mathematical model
Different mathematical formulations have been proposed for the 2EVRP (Perboli et al., 2011; Jepsen et al.,
2012; Baldacci et al., 2013; Santos et al., 2013, 2014) and for the 2ELRP (Boccia et al., 2011; Contardo et al.,
2012). In this section, we display compact formulations based on the model of Cuda et al. (2015).
The 2EVRP can be defined on a weighted undirected graph G = (N, E), where the set of vertices N
consists of the depot {0}, the set of possible satellite locations S = {1, . . . , |S|} and the set of customers
C = {|S| + 1, . . . , |S| + |C|}. The set of edges E is divided into two subsets, representing the first and second
echelon respectively. Set E 1 = {(i, j) : i < j, i, j ∈ {0} ∪ S} represents the edges which can be traversed by
first-level vehicles: those connecting the depot to the satellites, and those interconnecting satellites with each
other. The set of edges E 2 = {(i, j) : i < j, i, j ∈ S ∪ C, (i, j) ∈
/ S × S} is used for the second level, and
corresponds to possible trips between satellite and customers or pairs of customers.
A fleet of v 1 homogeneous trucks with capacity Q1 is located at the depot. A total of v 2 homogeneous city
freighters are available, each with a given capacity of Q2 . They can be located at any satellite s ∈ S. Still, the
number of city freighters at one satellite is limited to vs2 .
The set R1 contains all possible routes starting from the depot and delivering a given sequence of customers,
then returning to the depot again. Similarly each route r in the set of secondary routes R2 starts at a satellite
s ∈ S, visits one or several customers in C, and returns again to satellite s. Each customer c ∈ C has a demand
of dc units. Each unit of freight shipped through a satellite induces a handling cost hs .
Given a secondary route r ∈ R2 and a customer c ∈ C, the parameter
βrc ∈ {0, 1} is equal to 1 if and
P
only if customer c is visited in route r, and 0 otherwise. Let dr = c∈C:c∈r dc ≤ Q2 denote the total demand
of customers visited in route r, and pr represents the cost of each route r ∈ R1 ∪ R2 . The binary variables
xr ∈ {0, 1} with r ∈ R1 ∪ R2 take the value 1 if and only if route r is in the solution. Finally, each decision
variable qrs ≥ 0 with r ∈ R1 , s ∈ S ∩ r, gives the load on the truck on route r that has to be delivered to
satellite s.
5
X
min
p r xr +
r∈R1 ∪R2
X
s∈S
hs
X
qrs
(1)
r∈R1
s.t.
X
xr ≤ v 1
(2)
xr ≤ v 2
(3)
r∈R1
X
r∈R2
X
xr ≤ vs2
s∈S
(4)
r ∈ R1
(5)
s∈S
(6)
c∈C
(7)
r ∈ R1 ∪ R2
(8)
r∈R2 :s∈r
X
qrs ≤ Q1 xr
s∈S∩r
X
dr xr
r∈R2 :s∈r
r∈R1
X
X
qrs =
βrc xr = 1
r∈R2
xr ∈ {0, 1}
1
qrs ≥ 0
r ∈ R ,s ∈ S ∩ r
(9)
The objective function (1) sums up routing costs for all routes on both levels and handling costs per
unit moved through each satellite. Constraints (2) and (3) set the number of available vehicles for trucks
and city freighters, respectively. The number of city freighters per satellite is constrained by Constraint (4).
Constraints (5) ensure that the maximum capacity of the trucks is not exceeded. Constraints (6) link the
quantities of goods between the first and the second level. They guarantee that the incoming goods equal the
outgoing goods at the satellites. As there are no split deliveries allowed on the second level, Constraints (7)
ensure that each customer is visited exactly once. The domains of the decision variables are defined by
Constraints (8) and (9).
The mathematical model for the 2ELRPSD is similar to the previous model, but needs some adjustments.
Each satellite s ∈ S has a given opening cost fs and a capacity of ks units of freight. There is an unbounded
number of vehicles available at both levels. Therefore, Constraints (2) to (4) are not needed. The routing of
a vehicle incurs fixed costs of f 1 for each truck, and f 2 for each used city freighter. Each binary parameter
αrs is equal to 1 if and only if satellite s is visited on route r, and 0 otherwise. If satellite s is opened in the
solution, binary variable ys takes value 1, and 0 otherwise.
P
1
P Using2 ConstraintsP(5) to (9), the objective function needs to be changed to min r∈R1 (f + pr )xr +
r∈R2 (f + pr )xr +
s∈S fs ys to consider fixed and mileage-based vehicle costs for both
P levels separately, as
well as opening costs for satellites. The capacity limit at the satellites is imposed by r∈R1 qrs ≤ ks ys for all
s ∈ S. If a satellite
s has been selected to be open, then the delivery by exactly one truck is guaranteed by
P
Constraints r∈R1 αrs xr = ys for all s ∈ S.
5
Solution method
The proposed metaheuristic follows the basic structure of a large neighbourhood search (LNS), which was
first introduced by Shaw (1998). An initial feasible solution is iteratively destroyed and repaired in order to
gradually improve the solution. Such a ruin and recreate approach (Schrimpf et al., 2000) has been successfully
applied to multiple variants of vehicle routing problems in the past (see, e.g., Pisinger and Ropke 2010). The
destruction of parts of a previous solution (ruin) gives freedom to create a new and better solution (recreate).
Algorithm 1 shows the basic structure of the proposed method.
At each iteration of the proposed method, 1) a partial solution destruction is performed on the routes of
the second level; 2) then the second level is repaired and improved by means of local search, and finally 3) the
6
Algorithm 1: LNS-2E
1
2
3
4
5
6
7
S best ← S ← localSearch(repair(instance))
g←0
repeat
for i ← 0 to imax do
S temp ← localSearch(repair(destroy(S, g)))
if Satellite was opened/closed during previous destroy phase then
g←0
10
11
g ←g+1
9
12
13
14
15
16
17
/* reset grace period */
temp
if cost(S
) < cost(S) then
S ← S temp
i←0
8
/* initial solution */
/* accept better solution */
/* reset re-start period */
if cost(S) < cost(S best ) then
S best ← S
/* store best solution */
else
S ← localSearch(repair(instance))
/* re-start:
new solution */
until time > timemax
return S best
first level is reconstructed with a simple heuristic. As such, the first level is constructed from scratch in every
iteration, but since the number of nodes in the first-level sub-problem is relatively small, this simple heuristic
already finds an optimal or near-optimal solution.
Each of the destroy phases performs all the destroy operators sequentially as they are described in
Section 5.1. One single repair mechanism is used for solution reconstruction and also to obtain an initial
solution (Line 1 and 5). This procedure is described in Section 5.2. Afterwards, as needed quantities at the
satellites are known, the first level is reconstructed as described in Section 5.3.
The choices of intermediate facilities may change as a consequence of the repair operator, or through
dedicated destroy operators which temporarily close or re-open some possible locations for intermediate facilities.
If a change in open or closed satellites has recently taken place, the status of another satellite will not be
changed (Line 5) for a number of iterations that we call grace period (g is reset to 0 in Line 7).
We then put emphasis on a strong local search phase, exploiting well-known procedures like 2-opt (Croes,
1958), 2-opt*, or simple relocate and swap moves. Relocate shifts a node before or after one of the closest
neighbours, if costs are improved. Swap explores the exchange of one node with one of the neighbour nodes,
as well as exchanging one node with two successive neighbour nodes. This local search phase is applied after
the destroy and repair operators. In order to reduce the complexity, moves are only attempted between close
customers, as done in the granular search by Toth and Vigo (2003). Further information on these moves can
be found in the survey by Vidal et al. (2013).
If a better solution is obtained, it is accepted as the new incumbent solution (Line 9). If no improvement
can be found for a large number of iterations, then the algorithm will restart from a new initial solution, even
if the objective value is worse (Line 15).
In general, our algorithm requires less local and large neighbourhood operators than the ALNS proposed
by Hemmelmayr et al. (2012). The destruction operator parameters are also selected randomly, since the
method performed equally well, during our computational experiments, without need for a more advanced
adaptive scoring system. In the following, we describe the sets of destroy and repair operators, as well as the
management of the decisions related to the first level.
7
5.1
Destroy operators
Our algorithm relies on different destroy operators which are all invoked at each iteration in sequential order.
They are applied only to the second level. All of them, except the open all satellites neighbourhood, select
nodes which are removed from the current solution.
The first four destroy operators are used at each iteration. The last two ones, which change the status
of a satellite to closed or open again, are only invoked if g in Line 5 of Algorithm 1 has exceeded the grace
period gmax (i.e. no change in open/closed satellites has taken place recently). The destroy operators are now
described, in their order of use. When applicable, all random samples are uniformly distributed within their
given interval.
Related node removal A seed customer is randomly chosen. A random number of its Euclidean closest
customers as well as the seed customer are removed from the current solution and added to the list of nodes to
re-insert. This operator receives a parameter p1 , which denotes the maximum percentage of nodes to remove.
At most dp1 · |C|e nodes are removed, with |C| being the overall number of customers.
Biased node removal First, the removal cost of each customer is computed: the savings associated to a
removal of node j, located between i and k, is given by δj = cik − cij − cjk , where cij denotes the travel cost
from node i to node j. The probability of selection of a node for removal is then linearly correlated with the
delta evaluation value. The higher the gain after removal, the more likely it will be selected and removed. In
every destroy phase, a random percentage of customers from the interval [0, p2 ] is removed.
Random route removal
Randomly selects routes and removes all containing customers, adding them to the
X dc
list of nodes to re-insert. This operator randomly selects a number of routes in the interval [0, dp3 ·
e].
Q2
c∈C
Remove single node routes This operator removes all routes which contain only one single customer. In
the case of the 2EVRP there is a limited number of overall vehicles available, and thus removing the short
routes allows to use a vehicle originating from a different satellite in the next repair phase. This operator is
used with a probability of p̂4 .
The last two destroy operators can be used at most at each gmax iterations. During this grace period after
satellite selection has been actively altered, none of these two operators will be executed.
Close satellite Chooses a random satellite. If the satellite can be closed and the remaining open ones
still can provide sufficient capacity for a feasible solution, the chosen satellite is closed temporarily. All the
customers, which are assigned to it, are removed and added to the list of nodes to re-insert. The satellite stays
closed until it is opened again in a later phase. This operation is chosen with a probability of p̂5 , given variable
g has already exceeded the grace period. If this operator has been executed, g is reset to 0.
Open all satellites This neighbourhood makes all previously closed satellites available again. It comes into
p̂5
effect with a probability |S|
, and thus it depends on the same parameter as the close satellite operator and
the number of satellites. This operator can only be executed if g > gmax , i.e. outside the grace period. Its
execution resets g to 0.
5.2
Repair operator, randomisation and initial solution
At each repair phase, the insertion of the nodes is done in random order. This repair mechanism can sometimes
fail, if one customer remains with a higher demand than the largest free capacity available on any vehicle. In
this exceptional case, the repair process is restarted, and the nodes are inserted by decreasing demands to
preserve feasibility.
Repair is achieved with a simplified cheapest insertion heuristic. All nodes are sequentially inserted at their
cheapest possible position in the solution. The main difference with the classic cheapest insertion heuristic is
8
that the method does not aim to insert the node with the lowest increase in total costs, but just takes the next
candidate from the list and inserts it, in order to reduce complexity and enhance solution diversity. It is a
simple and greedy heuristic.
Both the initial solution and every partial solution are always repaired by the same operator. The initial
solution can be seen as a “completely destroyed” solution.
After the second level has been repaired, the local search procedure is performed: 2-opt on each of the
routes, 2-opt* on all routes originating at the same satellite. The algorithm then tries to relocate single nodes,
swap one node with another and to swap two nodes with one other, within a limited neighbourhood of the τ
closest nodes, again accepting only improvements. This procedure stops when no improving move exists in the
entire neighbourhood. After this procedure, the delivery quantity of each of the satellites is known, and the first
level can be constructed using the same insertion heuristic as for the second level and performing local search.
5.3
Reconstruction of the first level
For the 2EVRP, it is essential to allow satellites to be delivered by several trucks. In particular, if the demanded
quantity at a satellite is larger than a full truckload and no other satellite is available, then there would be
no feasible solution. To reconstruct a first level solution, we propose a very simple preprocessing step. Any
satellite with a demand larger than a full truckload is virtually duplicated into nodes with demands equal to a
truckload, until the remaining demand is smaller than Q1 . The same insertion procedure as the repair operator
is used to generate a first level solution. This creates back-and-forth trips to the virtual nodes with demands
equal to a full truckload, and completes the solution analogously for the remaining nodes.
Usually there are few nodes associated with the SDVRP on the first level. The largest benchmark instances
from literature so far contain only ten satellites at most. This very simple policy enabled to find nearly-optimal
first level solutions for most considered instances with limited computational effort. Finally, note that in the
considered 2ELRPSD instances, the capacity of a satellite is never larger than the trucks’ capacity. Therefore,
split deliveries are not generated during reconstruction.
6
Computational Experiments
This section describes the currently available sets of instances for the 2EVRP (in Section 6.1) and the used
instances for the 2ELRPSD (Section 6.2) and attempts to resolve some inconsistencies. The calibration of the
method is described in Section 6.3. The computational results and the comparisons with other state-of-the-art
algorithms are discussed in Sections 6.4 and 6.5. Finally, Section 6.6 analyses the sensitivity of the method
with respect to several key parameters and design choices.
6.1
Benchmark Instances for the two-echelon vehicle routing problem
When looking at the literature, it may appear that there are six unique sets of benchmark instances. However,
due to inconsistencies with respect to constraints, nomenclature or locations, we identified in fact several
different subsets. In what follows, we explain the differences and provide high quality solutions for them. We
consider five different sets of benchmark instances from literature. Sets 2 and 3 were proposed by Perboli et al.
(2011) and have been generated based on the instances for the CVRP by Christofides and Eilon. Different
customers were chosen and converted into satellites. They also proposed the small Set 1 instances, with just
twelve customers and two satellites, which we did not consider. Set 4 was proposed by Crainic et al. (2010); all
of them were downloaded from OR-Library (Beasley, 2014).
The instance Sets 2 to 5 as used in Hemmelmayr et al. (2012) were also communicated to us by email
(Hemmelmayr, 2013). We noticed a few key differences with the ones available from Beasley (2014).
Set 6 instances were provided from the authors (Baldacci, 2013).
All distances are Euclidean, and computed with double precision. Note that handling costs are set to 0 for
all sets except 6b. We will now explain the characteristics of these instance sets in detail, and propose unique
names for the sets to overcome existing inconsistencies:
9
Set 2 There are two different versions in circulation: Please note that the instances with 50 customers in the
OR-Library contain a mistake1 . This can be resolved by exchanging Q1 and Q2 capacity values, which is also
the way we treated them, like previous authors did.
The names of the instances downloaded from Beasley (2014) and used by Hemmelmayr et al. (2012) were
the same, but the instances with 50 customers included different locations for the satellites. For future reference
we provide both versions, and we rename the instances with less than 50 customers to Set 2a, the Hemmelmayr
(2013) version of 50 customer Set 2 instance files to Set 2b, and the OR-Library version will be called Set 2c.
Table 3 shows the characteristics of all Set 2 instances.
Instance names used by Baldacci et al. (2013) have the satellite numbers incremented by one. Apart from
that they are identical with what we received from Hemmelmayr (2013). For example Set 2a instance named
E-n51-k5-s2-17 (Satellites 2 and 17) corresponds to E-n51-k5-s3-18 in the result tables of Baldacci et al. (2013).
We provide both versions (OR-Library with corrected capacities as well as the ones received by Hemmelmayr)
with distinguishable names at https://www.univie.ac.at/prolog/research/TwoEVRP.
Set 3 There are also two different versions of the Set 3 instances in circulation. We collected and solved
all of them and distinguished between different versions and identified inconsistencies. Again, the sources
Beasley (2014) and Hemmelmayr (2013) were identical for instances with 21 and 32 customers, but different for
instances with 50 customers. In the case of Set 3, the filenames for the different instances were also different,
so there is no need to introduce new distinguishable names.
The only difference between Set 3 instances with 50 customers from the two sources is the location of
the depot. The locations of satellites and customers, as well as the vehicles and demands are identical. All
Set 3 instances from Hemmelmayr (2013) place the depot at coordinates (0,0), whereas the files of Beasley
(2014) have the depot located at (30,40). Table 13 shows which instances correspond to each other, apart from
satellite location.
Please also note that like in Set 2, the Set 3 instances with 50 customers from the OR Library also have the
capacities of the two vehicle types interchanged (see Footnote 1). This has been corrected in the files which we
provide online.
For easier referencing, we also divide the instances in three parts. Set 3a includes all instances with less
than 50 customers, Set 3b the larger instances which have been used by Hemmelmayr et al. (2012), and Set 3c
the larger instances as they are available at the OR-Library, and have been used by Baldacci et al. (2013),
among others.
Set 4 These instances have been treated differently in literature, either with a limit on the number of second
level vehicles allowed per satellite or only considering a total number of vehicles, with no limitations on the
distribution amongst satellites. As proposed in Baldacci et al. (2013), we solved both versions and follow their
nomenclature: Set 4a with the limit per satellite, and Set 4b when the constraint of vehicles per satellite is
relaxed.
Set 5 This set of instances has been proposed by Hemmelmayr et al. (2012). To the best of our knowledge
they were the only ones to report solutions on all instances of that set. Baldacci et al. (2013) were able to find
solutions on the small instances with only five satellites.
Set 6 To the best of our knowledge, solutions on these instances have only been reported in Baldacci et al.
(2013). Set 6 includes two subsets: Set 6a, with hs = 0, and Set 6b, which considers different handling costs
per freight unit at each of the satellites.
Table 1 displays an overview of the characteristics of the individual sets. It lists the number of instances in
the according set and subset with number of customers (C), satellites (S), trucks (T), city freighters (CF) and
available city freighters per satellite vs2 . Column hC shows if the handling costs are non-zero. The source of
the instance sets is also provided (Hemmelmayr, 2013; Beasley, 2014; Baldacci, 2013).
1 First level vehicles have a capacity of Q1 = 160 units, and second level vehicles, which by design are supposed to be smaller
than level 1 trucks, have a capacity of Q2 = 400 units. For instances with twoP
satellites for example, there are 3 trucks available.
They can ship a maximum of 3 ∗ 160 = 480 units. Overall customers’ demand c∈C dc is larger than 480 units, so there is per se
no feasible solution for those instances.
10
Table 1: Characteristics and Sources of Instance Sets
Set Subset Inst.
2
a
b
c
3
a
b
c
4
a
b
5
6
a
b
6.2
C
S
T
hC vs2 HCC OR-Library Baldacci
CF
6
6
6
3
6
3
21
32
50
50
50
50
2
2
2
4
2
4
3
3
3
4
3
4
4
4
5
5
5
5
-
•
•
•
•
6
6
6
6
21
32
50
50
2
2
2
2
3
3
3
3
4
4
5
5
-
•
•
•
18
18
18
18
18
18
50
50
50
50
50
50
2
3
5
2
3
5
3
3
3
3
3
3
6
6
6
6
6
6
4
3
2
-
•
•
•
6 100
6 200
6 200
5
10
10
5 [15,32]
5 [17,35]
5 [30,63]
-
•
•
•
2
3
4
2
3
4
-
9 50 [4,6]
9 75 [4,6]
9 100 [4,6]
9 50 [4,6]
9 75 [4,6]
9 100 [4,6]
50
75
100
50
75
100
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Benchmark Instances for the two-echelon location routing problem
The proposed algorithm was originally designed for the 2EVRP, nevertheless we also tested it on benchmark
instances for the 2ELRPSD. Two sets, called “Nguyen” and “Prodhon” are available at http://prodhonc.
free.fr/Instances/instances0_us.htm. They present some small errors or unclear descriptions, which are
documented in Appendix 8.3.
6.3
Parameters
The parameters of the proposed method have been calibrated using meta-calibration: the problem of finding
good parameters is assimilated to a black-box optimisation problem, in which the method parameters are
the decision variables, and the objective function is simulated by running the method on a set of training
instances, containing five randomly selected instances, for each set. To perform a fast optimisation we rely on
the covariance matrix adaptation evolution strategy (CMA-ES) by Hansen (2006). The source code (in Java)
is available at https://www.lri.fr/~hansen/cmaes_inmatlab.html.
The performance of our algorithm is rather insensitive to changes in parameters for the small instances, but
the rules for closing and opening satellites have to be adjusted to the number of overall available satellites.
Our calibration experiments have been conduced for each instance set, independently, and then we searched for
one compromise setting for the parameters that yields satisfying results for all different benchmark instances.
The calibration results are displayed in Table 2, as well as the average value, standard deviation, and the
compromise value which was used for the runs reported in Section 6.4.
The size of the limited neighbourhood for the local search relocate and swap moves was also determined by
CMA-ES. This parameter always converged to τ = 25 already in early stages of the tuning process, and thus
relocate and swap moves are attempted only for nodes within the radius including the 25 Euclidean closest
nodes.
11
Table 2: Parameter values obtained by meta-calibration
2EVRP
2ELRPSD
Set 2 Set 3 Set 4a Set 4b Set 5 Set 6a Set 6b Nguyen Prodhon Mean Std. Dev. compromise
p1
p2
p3
p̂4
p̂5
6.4
0.35
0.18
0.09
0.33
0.06
0.39
0.20
0.07
0.73
0.01
0.32
0.12
0.14
0.19
0.29
0.31
0.07
0.16
0.09
0.24
0.29
0.80
0.14
0.32
0.14
0.33
0.52
0.21
0.21
0.28
0.26
0.78
0.19
0.41
0.26
0.14
0.50
0.28
0.37
0.21
0.34
0.19
0.17
0.57
0.20
0.30
0.37
0.16
0.36
0.19
0.07
0.27
0.06
0.18
0.09
0.20
0.35
0.25
0.50
0.20
Computational Results
As done in previous literature, we performed five independent runs on each of the 2EVRP benchmark instances
and 20 runs on the 2ELRPSD instances. The code is written in Java with JDK 1.7.0 51 and tested on a Intel
E5-2670v2 CPU at 2.5 GHz with 3 GB RAM. The code was executed single threaded on one core. We compare
the performance of our method on the 2EVRP instances with the hybrid GRASP +VND by Zeng et al. (2014)
and the ALNS by Hemmelmayr et al. (2012), when applicable; as well as the currently best known solutions
for each instance from the literature. We also show the results of the algorithm on the 2ELRP with single
depot and compare with the VNS of Schwengerer et al. (2012). We describe the data of the following tables in
general and discuss results in detail on each of the instance sets separately.
Tables 3 to 11 show the characteristics and detailed results for each instance. The columns C, S, T and
CF display the main characteristics of the instance, where C is the number of customers, S is the number of
satellites, T and CF the number of available trucks and city freighters, respectively. The last two columns are
not applicable for Tables 10 to 11, as they correspond to 2ELRPSD instances with unbounded fleet size.
The next columns display the results of the proposed method (LNS-2E), and methods by Hemmelmayr
et al. (2012) (HCC), Zeng et al. (2014) (ZXXS) for the 2EVRP when applicable, and Schwengerer et al. (2012)
(SPR) for the 2ELRPSD. The average objective value of five runs is given in column Avg. 5. Column Best 5
shows the best solution found within these five runs, and Best gives the best objective value found during all
experiments, including parameter calibration. Following the work of Schwengerer et al. (2012), we also used
average and best of 20 for the 2ELRPSD for better comparison.
Column t reports the average overall runtime of the algorithms in seconds, and t* the average time when the
best solution was found. For easier comparison we chose a simple time limit for termination of our algorithm: 60
seconds for instances with up to 50 customers, and 900 seconds for larger ones. Our time measure corresponds
to the wall-clock time of the whole execution of the program, including input and output, computation of
the distance matrix, and other pre-processing tasks. Hemmelmayr et al. (2012) and Schwengerer et al. (2012)
report CPU times (which may be slightly smaller than wall clock times). Zeng et al. (2014) only report the
time when the best solution was found, but no overall runtime of the algorithm.
BKS refers to the best known solution of that instance. Best known solutions are highlighted in boldface
when found by the algorithm, and new BKS are also underlined. We highlight an instance with an asterisk
after BKS if the best known solution of the instance is known to be optimal from previous literature.
Tables 3 and 4 provide detailed results on the instances of Set 2 and Set 3. HCC, ZXXS and our algorithm
find the best known solutions at every run. The solutions have been proven to be optimal for all the instances
except Set 2c and Set 3b. To the best of our knowledge, we are the first ones to report solutions on the 2c
instances obtained from Beasley (2014). For Set 3c, the optimal objective values are derived from Baldacci
et al. (2013) and Jepsen et al. (2012), but no results from HCC or ZXXS are available. Summarising Tables 3
and 4, we can conclude that Sets 2 and 3 are easy in the sense that all runs of all algorithms always found the
optimal or best known solution.
The instances of Set 4 have been addressed in various ways in the literature. Jepsen et al. (2012) considered
a limit on the number of city freighters available at each satellite, HCC and ZXXS did not impose this limit,
and instead considered the limit on the total number of city freighters only. Baldacci et al. (2013) addressed
both variants of the instances to compare their results to both previous results, introducing a new nomenclature:
Set 4a for the instances including the limit of vehicles per satellite, and Set 4b when this limit is relaxed.
Tables 5 and 6 display the results on Set 4a and 4b instances. 102 out of the 108 instances have been solved
to optimality by Baldacci et al. (2013). Nevertheless we observed small differences of objective values with
12
our solutions (up to a 0.006% difference). This could be explained by a different rounding convention (we
use double precision), or by the small optimality gap of Cplex. As a consequence, bold fonts were used for
BKS within 0.006% precision. In all these cases the underlying solution is identical, just the objective value is
marginally different. We can see from Tables 5 and 6 that in all cases our best solution corresponds to the
optimal or best known solution. Only in 3 and 2 instances of Set 4a and 4b, respectively, some of the runs gave
slightly worse solutions. On average, in instance Set 4b our results are slightly better than those of the other
heuristics.
To the best of our knowledge, Hemmelmayr et al. (2012) were the only authors who published results on
the large Set 5 instances with 10 satellites to this date. Baldacci et al. (2013) report solutions on the small
Set 5 instances (100 customers/5 satellites), improving three out of six instances to optimality. The algorithm
of Hemmelmayr et al. (2012) was evaluated with a limit of 500 iterations. We compare our results in Table 7
and were able to improve the best known solutions on 9 of the 18 instances, depicted with an underlined BKS
value. Known optimal solutions are retrieved at least once within the five performed test runs.
Tables 8 and 9 report the results for the instances of Set 6a and b. From the 54 instances, all except 13
solutions have been proven to be optimal, and on nine of those remaining LNS-2E was able to find better
solutions. Best solutions were found typically after less than three minutes.
Table 3: Results for Set 2 Instances
HCC
Instance
ZXXS
LNS-2E
C
S T CF Avg. 5 t(s) t*(s) Avg. 5 t*(s) Avg. 5
Best 5
Best
Set 2a1,2
E-n22-k4-s6-17
E-n22-k4-s8-14
E-n22-k4-s9-19
E-n22-k4-s10-14
E-n22-k4-s11-12
E-n22-k4-s12-16
E-n33-k4-s14-22
E-n33-k4-s1-9
E-n33-k4-s2-13
E-n33-k4-s3-17
E-n33-k4-s4-5
E-n33-k4-s7-25
Avg.
21
21
21
21
21
21
32
32
32
32
32
32
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
4
4
417.07
384.96
470.60
371.50
427.22
392.78
779.05
730.16
714.63
707.48
778.74
756.85
577.59
37
34
35
37
31
36
85
74
64
58
77
53
51
0
0
0
0
0
0
0
0
0
0
3
0
0
417.07
384.96
470.60
371.50
427.22
392.78
730.16
714.63
707.48
778.74
756.85
779.05
577.59
0
0
0
0
0
0
0
0
0
1
0
0
0
417.07
384.96
470.60
371.50
427.22
392.78
779.05
730.16
714.63
707.48
778.74
756.85
577.59
417.07
384.96
470.60
371.50
427.22
392.78
779.05
730.16
714.63
707.48
778.74
756.85
577.59
417.07
384.96
470.60
371.50
427.22
392.78
779.05
730.16
714.63
707.48
778.74
756.85
577.59
60
60
60
60
60
60
60
60
60
60
60
60
60
1
1
1
2
2
1
1
1
1
1
1
1
1
417.07*
384.96*
470.60*
371.50*
427.22*
392.78*
779.05*
730.16*
714.63*
707.48*
778.74*
756.85*
577.59
Set 2b1
E-n51-k5-s11-19
E-n51-k5-s11-19-27-47
E-n51-k5-s2-17
E-n51-k5-s2-4-17-46
E-n51-k5-s27-47
E-n51-k5-s32-37
E-n51-k5-s4-46
E-n51-k5-s6-12
E-n51-k5-s6-12-32-37
Avg.
50
50
50
50
50
50
50
50
50
2
4
2
4
2
2
2
2
4
3
4
3
4
3
3
3
3
4
5
5
5
5
5
5
5
5
5
581.64
527.63
597.49
530.76
538.22
552.28
530.76
554.81
531.92
549.50
182
147
100
154
136
141
173
149
150
148
6
1
7
1
1
1
0
2
0
2
597.49
530.76
554.81
581.64
538.22
552.28
530.76
531.92
527.63
549.50
1
0
1
4
1
1
1
1
1
1
581.64
527.63
597.49
530.76
538.22
552.28
530.76
554.81
531.92
549.50
581.64
527.63
597.49
530.76
538.22
552.28
530.76
554.81
531.92
549.50
581.64
527.63
597.49
530.76
538.22
552.28
530.76
554.81
531.92
549.50
60
60
60
60
60
60
60
60
60
60
1
4
3
3
1
2
3
4
2
2
581.64*
527.63*
597.49*
530.76*
538.22*
552.28*
530.76*
554.81*
531.92*
549.50
Set 2c2
E-n51-k5-s11-19
E-n51-k5-s11-19-27-47
E-n51-k5-s2-17
E-n51-k5-s2-4-17-46
E-n51-k5-s27-47
E-n51-k5-s32-37
E-n51-k5-s4-46
E-n51-k5-s6-12
E-n51-k5-s6-12-32-37
Avg.
50
50
50
50
50
50
50
50
50
2
4
2
4
2
2
2
2
4
3
4
3
4
3
3
3
3
4
5
5
5
5
5
5
5
5
5
617.42
530.76
601.39
601.39
530.76
752.59
702.33
567.42
567.42
607.94
617.42
530.76
601.39
601.39
530.76
752.59
702.33
567.42
567.42
607.94
617.42
530.76
601.39
601.39
530.76
752.59
702.33
567.42
567.42
607.94
60
60
60
60
60
60
60
60
60
60
3
1
3
4
3
5
4
5
6
4
617.42
530.76
601.39
601.39
530.76
752.59
702.33
567.42
567.42
607.94
1 included in Hemmelmayr (2013)
2 included in Beasley (2014)
13
t(s) t*(s)
BKS
Table 4: Results for Set 3 Instances
HCC
ZXXS
LNS-2E
Instance
C
S T CF Avg. 5 t(s) t*(s) Avg. 5 t*(s) Avg. 5
Best 5
Best
Set 3a1,2
E-n22-k4-s13-14
E-n22-k4-s13-16
E-n22-k4-s13-17
E-n22-k4-s14-19
E-n22-k4-s17-19
E-n22-k4-s19-21
E-n33-k4-s16-22
E-n33-k4-s16-24
E-n33-k4-s19-26
E-n33-k4-s22-26
E-n33-k4-s24-28
E-n33-k4-s25-28
Avg.
21
21
21
21
21
21
32
32
32
32
32
32
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
4
4
526.15
521.09
496.38
498.80
512.81
520.42
672.17
666.02
680.36
680.37
670.43
650.58
591.30
43
44
49
43
26
34
76
77
84
77
88
63
59
0
0
0
0
0
0
3
0
0
0
0
0
0
526.15
521.09
496.38
498.80
512.81
520.42
672.17
666.02
680.36
680.37
670.43
650.58
591.30
0
0
0
0
0
0
0
0
0
0
0
0
0
526.15
521.09
496.38
498.80
512.80
520.42
672.17
666.02
680.36
680.36
670.43
650.58
591.30
526.15
521.09
496.38
498.80
512.80
520.42
672.17
666.02
680.36
680.36
670.43
650.58
591.30
526.15
521.09
496.38
498.80
512.80
520.42
672.17
666.02
680.36
680.36
670.43
650.58
591.30
60
60
60
60
60
60
60
60
60
60
60
60
60
2
2
1
1
4
2
4
1
1
1
2
1
2
526.15*
521.09*
496.38*
498.80*
512.80*
520.42*
672.17*
666.02*
680.36*
680.36*
670.43*
650.58*
591.30
Set 3b1
E-n51-k5-s12-18
E-n51-k5-s12-41
E-n51-k5-s12-43
E-n51-k5-s39-41
E-n51-k5-s40-41
E-n51-k5-s40-43
Avg.
50
50
50
50
50
50
2
2
2
2
2
2
3
3
3
3
3
3
5
5
5
5
5
5
690.59
683.05
710.41
728.54
723.75
752.15
714.75
147
133
217
155
154
158
161
4
38
1
18
17
15
16
690.59
683.05
710.41
728.54
723.75
752.15
714.75
1
1
1
4
3
9
3
690.59
683.05
710.41
728.54
723.75
752.15
714.75
690.59
683.05
710.41
728.54
723.75
752.15
714.75
690.59
683.05
710.41
728.54
723.75
752.15
714.75
60
60
60
60
60
60
60
8
11
5
7
5
12
8
690.59
683.05
710.41
728.54
723.75
752.15
714.75
Set 3c2
E-n51-k5-s13-19
E-n51-k5-s13-42
E-n51-k5-s13-44
E-n51-k5-s40-42
E-n51-k5-s41-42
E-n51-k5-s41-44
Avg.
50
50
50
50
50
50
2
2
2
2
2
2
3
3
3
3
3
3
5
5
5
5
5
5
560.73
564.45
564.45
746.31
771.56
802.91
668.40
560.73
564.45
564.45
746.31
771.56
802.91
668.40
560.73
564.45
564.45
746.31
771.56
802.91
668.40
60
60
60
60
60
60
60
10
3
2
5
15
16
9
560.73*
564.45*
564.45*
746.31*
771.56*
802.91*
668.40
1 included in Hemmelmayr (2013)
2 included in Beasley (2014)
14
t(s) t*(s)
BKS
Table 5: Results for Set 4a Instances (with constraint on the
number of city freighters per satellite)
LNS-2E
Inst.
C
Set 4a
1
50
2
50
3
50
4
50
5
50
6
50
7
50
8
50
9
50
10
50
11
50
12
50
13
50
14
50
15
50
16
50
17
50
18
50
19
50
20
50
21
50
22
50
23
50
24
50
25
50
26
50
27
50
28
50
29
50
30
50
31
50
32
50
33
50
34
50
35
50
36
50
37
50
38
50
39
50
40
50
41
50
42
50
43
50
44
50
45
50
46
50
47
50
48
50
49
50
50
50
51
50
52
50
53
50
54
50
Avg.
S T CF
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
Avg. 5
Best 5
Best
1569.42
1438.32
1570.43
1424.04
2193.52
1279.89
1458.60
1363.76
1450.25
1407.65
2052.21
1209.46
1481.80
1393.64
1489.92
1389.20
2088.48
1227.68
1564.66
1272.98
1577.82
1281.83
1807.35
1282.69
1522.40
1167.47
1481.56
1210.46
1722.06
1211.63
1490.32
1199.05
1508.32
1233.96
1718.42
1228.95
1528.73
1169.20
1520.92
1199.42
1667.96
1194.54
1439.67
1045.14
1451.48
1088.79
1587.29
1082.21
1434.88
1083.16
1398.03
1125.69
1567.79
1127.66
1420.05
1569.42
1438.32
1570.43
1424.04
2193.52
1279.89
1458.60
1363.76
1450.25
1407.65
2047.43
1209.46
1481.80
1393.64
1489.92
1389.20
2088.48
1227.68
1564.66
1272.98
1577.82
1281.83
1807.35
1282.69
1522.40
1167.47
1481.56
1210.46
1722.00
1211.63
1490.32
1199.05
1508.32
1233.96
1718.42
1228.95
1528.73
1169.20
1520.92
1199.42
1667.96
1194.54
1439.67
1045.14
1450.95
1088.79
1587.29
1082.21
1434.88
1083.16
1398.03
1125.69
1567.79
1127.66
1419.95
1569.42
1438.32
1570.43
1424.04
2193.52
1279.89
1458.60
1363.76
1450.25
1407.65
2047.43
1209.46
1481.80
1393.64
1489.92
1389.20
2088.48
1227.68
1564.66
1272.98
1577.82
1281.83
1807.35
1282.69
1522.40
1167.47
1481.56
1210.46
1722.00
1211.63
1490.32
1199.05
1508.32
1233.96
1718.42
1228.95
1528.73
1169.20
1520.92
1199.42
1667.96
1194.54
1439.67
1045.14
1450.95
1088.79
1587.29
1082.21
1434.88
1083.16
1398.03
1125.69
1567.79
1127.66
1419.95
15
t(s) t*(s)
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
4
16
8
7
10
0
2
29
5
6
3
8
7
1
16
2
15
1
3
25
2
3
8
0
4
1
42
3
31
12
7
24
14
7
41
0
25
15
17
2
9
19
14
24
3
2
15
23
14
23
21
6
21
15
12
BKS
1569.42*
1438.33*
1570.43*
1424.04*
2193.52*
1279.87*
1458.63*
1363.74*
1450.27*
1407.64*
2047.46*
1209.42*
1481.83*
1393.61*
1489.94*
1389.17*
2088.49*
1227.61*
1564.66*
1272.97*
1577.82*
1281.83*
1807.35*
1282.68*
1522.42*
1167.46*
1481.57*
1210.44*
1722.04
1211.59*
1490.34
1199.00*
1508.30
1233.92*
1718.41
1228.89*
1528.73*
1169.20*
1520.92*
1199.42*
1667.96*
1194.54*
1439.67*
1045.13*
1450.96*
1088.77*
1587.29*
1082.20*
1434.88*
1083.12*
1398.05*
1125.67*
1567.77*
1127.61*
1419.94
Table 6: Results for Set 4b Instances (vs2 = v 2 )
HCC
Inst. C S T CF
Set 4b
1
50
2
50
3
50
4
50
5
50
6
50
7
50
8
50
9
50
10
50
11
50
12
50
13
50
14
50
15
50
16
50
17
50
18
50
19
50
20
50
21
50
22
50
23
50
24
50
25
50
26
50
27
50
28
50
29
50
30
50
31
50
32
50
33
50
34
50
35
50
36
50
37
50
38
50
39
50
40
50
41
50
42
50
43
50
44
50
45
50
46
50
47
50
48
50
49
50
50
50
51
50
52
50
53
50
54
50
Avg.
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
Avg. 5
Best
1569.42
1441.02
1570.43
1424.04
2194.11
1279.87
1458.63
1360.32
1450.27
1360.56
2059.88
1209.42
1481.83
1393.61
1489.94
1387.83
2088.49
1227.61
1546.28
1272.97
1577.82
1281.83
1652.98
1282.68
1440.84
1167.46
1447.79
1210.44
1561.81
1211.59
1440.86
1199.00
1478.86
1233.92
1570.80
1228.89
1528.81
1163.07
1520.92
1165.24
1652.98
1190.17
1408.95
1035.32
1406.43
1058.97
1564.41
1074.50
1435.28
1065.25
1387.72
1103.76
1545.73
1113.62
1401.39
1569.42
1438.33
1570.43
1424.04
2194.11
1279.87
1458.63
1360.32
1450.27
1360.56
2059.88
1209.42
1481.83
1393.61
1489.94
1387.83
2088.49
1227.61
1546.28
1272.97
1577.82
1281.83
1652.98
1282.68
1440.68
1167.46
1444.50
1210.44
1559.82
1211.59
1440.86
1199.00
1478.86
1233.92
1570.72
1228.89
1528.73
1163.07
1520.92
1163.04
1652.98
1190.17
1406.11
1035.03
1403.10
1058.11
1559.82
1074.50
1434.88
1065.25
1387.51
1103.42
1545.73
1113.62
1400.96
ZXXS
t(s) t*(s)
235
155
183
130
614
99
169
205
204
174
648
205
220
189
173
147
625
94
171
99
155
127
175
110
154
96
163
143
178
132
144
102
159
93
182
123
143
88
158
84
150
95
151
109
144
74
185
83
140
92
138
102
148
90
169
6
43
3
11
63
2
6
5
46
1
101
44
25
6
9
6
165
3
25
12
61
2
5
2
53
0
12
7
102
5
37
11
16
4
116
6
55
15
33
20
12
31
60
30
104
17
103
2
81
16
46
47
37
2
32
Avg. 5
Best
1569.42
1438.33
1570.43
1429.04
2193.52
1279.87
1408.57
1360.32
1403.53
1360.56
2059.41
1209.42
1450.93
1393.61
1466.83
1387.83
2088.49
1227.61
1546.28
1272.97
1577.82
1281.83
1652.98
1282.68
1408.57
1167.46
1454.63
1210.44
1555.56
1211.59
1441.07
1199.00
1478.86
1233.92
1570.72
1228.89
1528.98
1163.07
1520.92
1163.04
1652.98
1190.17
1406.11
1035.03
1402.03
1058.11
1557.04
1074.50
1434.88
1065.25
1387.51
1103.42
1545.73
1113.62
1397.69
1569.42
1438.33
1570.43
1424.04
2193.52
1279.87
1408.57
1360.32
1403.53
1360.56
2059.41
1209.42
1450.93
1393.61
1466.83
1387.83
2088.49
1227.61
1546.28
1272.97
1577.82
1281.83
1652.98
1282.68
1408.57
1167.46
1444.51
1210.44
1552.66
1211.59
1440.86
1199.00
1478.86
1233.92
1570.72
1228.89
1528.73
1163.07
1520.92
1163.04
1652.98
1190.17
1406.11
1035.03
1402.03
1058.11
1552.66
1074.50
1434.88
1065.25
1387.51
1103.42
1545.73
1113.62
1397.27
16
LNS-2E
t*(s)
1
40
1
73
34
1
16
4
4
1
4
6
2
1
1
6
27
1
25
57
19
28
3
1
17
7
39
2
30
1
29
10
14
11
4
2
43
1
16
7
9
71
24
7
27
7
25
1
38
2
3
22
4
12
16
Avg. 5
Best 5
Best
1569.42
1438.32
1570.43
1424.04
2193.52
1279.89
1408.58
1360.32
1403.53
1360.54
2054.60
1209.46
1450.95
1393.64
1466.84
1387.85
2088.48
1227.68
1546.28
1272.98
1577.82
1281.83
1652.98
1282.69
1408.58
1167.47
1444.49
1210.46
1552.66
1211.63
1440.85
1199.05
1478.87
1233.96
1570.73
1228.95
1528.73
1163.07
1520.92
1163.04
1652.98
1190.17
1407.09
1035.05
1401.87
1058.10
1552.66
1074.51
1434.88
1065.30
1387.51
1103.47
1545.76
1113.66
1397.21
1569.42
1438.32
1570.43
1424.04
2193.52
1279.89
1408.58
1360.32
1403.53
1360.54
2047.43
1209.46
1450.94
1393.64
1466.84
1387.85
2088.48
1227.68
1546.28
1272.98
1577.82
1281.83
1652.98
1282.69
1408.58
1167.47
1444.49
1210.46
1552.66
1211.63
1440.85
1199.05
1478.87
1233.96
1570.73
1228.95
1528.73
1163.07
1520.92
1163.04
1652.98
1190.17
1406.10
1035.05
1401.87
1058.10
1552.66
1074.51
1434.88
1065.30
1387.51
1103.47
1545.76
1113.66
1397.06
1569.42
1438.32
1570.43
1424.04
2193.52
1279.89
1408.58
1360.32
1403.53
1360.54
2047.43
1209.46
1450.94
1393.64
1466.84
1387.85
2088.48
1227.68
1546.28
1272.98
1577.82
1281.83
1652.98
1282.69
1408.58
1167.47
1444.49
1210.46
1552.66
1211.63
1440.85
1199.05
1478.87
1233.96
1570.73
1228.95
1528.73
1163.07
1520.92
1163.04
1652.98
1190.17
1406.10
1035.05
1401.87
1058.10
1552.66
1074.51
1434.88
1065.30
1387.51
1103.47
1545.76
1113.66
1397.06
t(s) t*(s)
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
15
6
7
7
18
0
17
8
11
1
2
20
10
3
2
12
13
5
34
11
16
4
4
1
20
16
20
6
39
5
15
21
15
11
10
6
27
17
25
5
14
9
13
42
21
16
24
4
9
33
19
17
21
5
14
BKS
1569.42*
1438.33*
1570.43*
1424.04*
2193.52*
1279.87*
1408.57*
1360.32*
1403.53*
1360.56*
2047.46*
1209.42*
1450.93*
1393.61*
1466.83*
1387.83*
2088.49*
1227.61*
1546.28*
1272.97*
1577.82*
1281.83*
1652.98*
1282.68*
1408.57*
1167.46*
1444.50*
1210.44*
1552.66*
1211.59*
1440.86*
1199.00*
1478.86*
1233.92*
1570.72*
1228.89*
1528.73
1163.07*
1520.92*
1163.04*
1652.98*
1190.17*
1406.11*
1035.03*
1401.87*
1058.11*
1552.66*
1074.50*
1434.88
1065.25*
1387.51*
1103.42*
1545.73*
1113.62*
1397.04
Table 7: Results for Set 5 Instances
HCC
Instance
C
Set 5
100-5-1
100-5-1b
100-5-2
100-5-2b
100-5-3
100-5-3b
100-10-1
100-10-1b
100-10-2
100-10-2b
100-10-3
100-10-3b
200-10-1
200-10-1b
200-10-2
200-10-2b
200-10-3
200-10-3b
Avg.
100
100
100
100
100
100
100
100
100
100
100
100
200
200
200
200
200
200
S T CF Avg. 5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
32
15
32
15
30
16
35
18
33
18
32
17
62
30
63
30
63
30
1588.73
1126.93
1022.29
789.05
1046.67
828.99
1137.00
928.01
1009.49
773.58
1055.28
861.88
1626.83
1239.79
1416.87
1018.57
1808.24
1208.38
1138.14
Best
LNS-2E
t(s) t*(s)
1565.45 353
1111.34 397
1016.32 406
782.25 340
1045.29 352
828.99 391
1130.23 429
916.48 476
990.58 356
768.61 432
1043.25 415
850.92 418
1574.12 888
1201.75 692
1374.74 1072
1003.75 1058
1787.73 916
1200.74 1217
1121.81 589
Avg. 5
116 1566.87
45 1111.93
117 1017.94
170
783.07
80 1045.29
127 828.54
136 1132.11
262
922.85
232 1014.61
157
786.64
209 1053.55
29
858.72
207 1598.46
374 1217.23
496 1406.16
221 1016.05
305 1809.44
478 1206.85
209 1132.02
Best 5
Best
t(s) t*(s)
1564.46
1108.62
1016.32
782.25
1045.29
828.54
1125.53
916.25
1012.14
781.27
1049.77
854.90
1580.34
1191.59
1366.36
1008.46
1797.80
1202.21
1124.01
1564.46
1108.62
1016.32
782.25
1045.29
828.54
1124.93
916.25
1002.15
774.11
1048.53
854.90
1556.79
1187.62
1365.74
1002.85
1793.99
1197.90
1120.62
900
900
900
900
900
900
900
900
900
900
900
900
900
900
900
900
900
900
900
Table 8: Results for Set 6a Instances (no handling costs)
LNS-2E
Instance
C
Set 6a
A-n51-4
A-n51-5
A-n51-6
A-n76-4
A-n76-5
A-n76-6
A-n101-4
A-n101-5
A-n101-6
B-n51-4
B-n51-5
B-n51-6
B-n76-4
B-n76-5
B-n76-6
B-n101-4
B-n101-5
B-n101-6
C-n51-4
C-n51-5
C-n51-6
C-n76-4
C-n76-5
C-n76-6
C-n101-4
C-n101-5
C-n101-6
Avg.
50
50
50
75
75
75
100
100
100
50
50
50
75
75
75
100
100
100
50
50
50
75
75
75
100
100
100
S T CF
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
2
2
2
3
3
3
4
4
4
2
2
2
3
3
3
4
4
4
2
2
2
3
3
3
4
4
4
Avg. 5
Best 5
Best
t(s) t*(s)
50
652.00
652.00
652.00
60
50
663.41
663.41
663.41
60
50
662.51
662.51
662.51
60
75
985.98
985.95
985.95 900
75
981.19
979.15
979.15 900
75
971.65
970.20
970.20 900
100 1194.38 1194.17 1194.17 900
100 1215.89 1211.40 1211.40 900
100 1161.91 1158.97 1155.96 900
50
563.98
563.98
563.98
60
50
549.23
549.23
549.23
60
50
556.32
556.32
556.32
60
75
793.97
792.73
792.73 900
75
784.27
784.19
784.19 900
75
775.75
774.24
774.17 900
100
939.79
939.21
939.21 900
100
971.27
969.13
969.13 900
100
961.91
960.29
960.29 900
50
689.18
689.18
689.18
60
50
723.12
723.12
723.12
60
50
697.00
697.00
697.00
60
75
1055.61 1054.89 1054.89 900
75 1115.32 1115.32 1115.32 900
75
1066.88 1060.52 1060.52 900
100 1305.94 1302.16 1302.16 900
100 1307.24 1306.27 1305.82 900
100 1292.10 1284.48 1284.48 900
912.51
911.11
910.98 620
17
19
37
23
128
286
233
267
414
154
6
55
39
320
190
160
377
161
88
13
19
46
339
113
474
236
141
446
177
BKS
652.00*
663.41*
662.51*
985.95*
979.15*
970.20*
1194.17*
1211.38*
1155.96
563.98*
549.23*
556.32*
792.73*
783.93*
774.17*
939.21*
967.82*
960.29*
689.18*
723.12*
697.00*
1054.89*
1115.32*
1060.52
1302.16
1305.82
1284.48
910.92
201
176
75
152
131
62
567
424
471
416
105
175
730
588
534
721
675
523
374
BKS
1564.46*
1108.62
1016.32*
782.25
1045.29*
828.54
1124.93
916.25
990.58
768.61
1043.25
850.92
1556.79
1187.62
1365.74
1002.85
1787.73
1197.90
1118.81
Table 9: Results for Set 6b Instances (hs 6= 0)
LNS-2E
Instance
C
Set 6b
A-n51-4
A-n51-5
A-n51-6
A-n76-4
A-n76-5
A-n76-6
A-n101-4
A-n101-5
A-n101-6
B-n51-4
B-n51-5
B-n51-6
B-n76-4
B-n76-5
B-n76-6
B-n101-4
B-n101-5
B-n101-6
C-n51-4
C-n51-5
C-n51-6
C-n76-4
C-n76-5
C-n76-6
C-n101-4
C-n101-5
C-n101-6
Avg.
50
50
50
75
75
75
100
100
100
50
50
50
75
75
75
100
100
100
50
50
50
75
75
75
100
100
100
S T CF
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
4
5
6
2
2
2
3
3
3
4
4
4
2
2
2
3
3
3
4
4
4
2
2
2
3
3
3
4
4
4
50
50
50
75
75
75
100
100
100
50
50
50
75
75
75
100
100
100
50
50
50
75
75
75
100
100
100
Avg. 5
Best 5
Best
t(s) t*(s)
744.24
811.51
930.11
1385.51
1519.86
1666.28
1884.48
1723.06
1795.36
653.09
672.10
767.13
1094.52
1218.12
1328.90
1500.80
1398.05
1455.05
866.58
943.12
1050.42
1439.39
1745.49
1759.40
2076.36
1974.39
1867.45
1343.36
744.24
811.51
930.11
1385.51
1519.86
1666.06
1883.79
1714.58
1793.76
653.09
672.10
767.13
1094.52
1218.12
1328.90
1500.55
1398.05
1453.54
866.58
943.12
1050.42
1438.96
1745.39
1756.54
2073.84
1967.51
1861.50
1342.20
744.24
811.51
930.11
1385.51
1519.86
1666.06
1883.79
1711.95
1791.44
653.09
672.10
767.13
1094.52
1218.12
1326.76
1500.55
1395.32
1453.54
866.58
943.12
1050.42
1438.96
1745.39
1756.54
2064.86
1964.63
1861.50
1341.39
60
60
60
900
900
900
900
900
900
60
60
60
900
900
900
900
900
900
60
60
60
900
900
900
900
900
900
620
17
49
31
26
71
533
283
318
100
21
51
58
44
19
329
82
317
460
22
12
18
84
281
358
378
285
401
172
BKS
744.24*
811.52*
930.11*
1385.51*
1519.86*
1666.06*
1881.44*
1709.06
1777.69
653.09*
672.10*
767.13*
1094.52*
1218.13*
1326.76*
1500.55
1395.32
1450.39
866.58*
943.12*
1050.42*
1438.96*
1745.39*
1756.54*
2064.86
1964.63
1861.50
1340.57
To test the robustness of our algorithm we applied it also to the 2ELRPSD instances without any further
changing of operators or tuning. The results can be seen in Tables 10 and 11. BKS are derived from Schwengerer
et al. (2012); Nguyen et al. (2012a,b); Contardo et al. (2012). Also on this problem class LNS-2E is competitive,
with solutions being on average within 0.6% of the solutions found by the state-of-the-art VNS by Schwengerer
et al. (2012) (SPR).
Table 10: Results for 2ELRPSD Instances Set Nguyen
SPR
Instance
C
Set Nguyen
25-5N
25
25-5Nb
25
25-5MN
25
25-5MNb
25
50-5N
50
50-5Nb
50
50-5MN
50
50-5MNb
50
50-10N
50
50-10Nb
50
50-10MN
50
50-10MNb
50
100-5N
100
100-5Nb
100
100-5MN
100
100-5MNb 100
100-10N
100
100-10Nb
100
100-10MN 100
100-10MNb 100
200-10N
200
200-10Nb
200
200-10MN 200
200-10MNb 200
Avg.
S
Avg. 20
5
5
5
5
5
5
5
5
10
10
10
10
5
5
5
5
10
10
10
10
10
10
10
10
80370.00
64562.00
78947.00
64438.00
137815.00
110204.40
123484.00
105687.00
115725.00
87345.00
135519.00
110613.00
200685.05
164508.10
206567.40
166357.35
214585.60
155790.60
203798.05
170791.25
349584.15
264228.90
332207.50
292036.65
163993.75
LNS-2E
Best 20 t(s) t*(s)
80370
64562
78947
64438
137815
110094
123484
105401
115725
87315
135519
110613
193228
158927
204682
165744
209952
155489
201275
170625
345267
256171
323801
287076
161938
76
91
61
89
116
132
125
202
143
176
144
218
168
258
184
315
223
352
229
347
641
907
453
944
275
Avg. 20
Best 20
Best
t(s) t*(s)
3
80370.00
80370
80370
60
0
64562.00
64562
64562
60
1
78947.00
78947
78947
60
0
64438.00
64438
64438
60
35 137815.00 137815 137815
60
51
110981.85 110094 110094
60
41 123484.00 123484 123484
60
48
105783.45 105401 105401
60
24
117325.55 115725 115725
60
66
88212.00
87520
87315
60
10
138241.35 135519 135519
60
13
111520.80 110613 110613
60
101
193806.85 193229 193229 900
144
159064.10 158927 158927 900
159
204876.10 204682 204682 900
247
165795.35 165744 165744 900
167
216265.50 210799 209952 900
251
161273.30 155489 155489 900
163
204396.15 201275 201275 900
283
172202.45 170625 170625 900
525
359948.65 350680 350680 900
791
260698.20 257191 257191 900
441
329486.45 324279 324279 900
843
297857.50 293339 290702 900
184
164472.98 162531 162377 480
18
5
16
6
4
25
22
4
16
20
23
14
22
239
315
105
252
344
442
324
268
758
748
777
778
230
BKS
80370
64562
78947
64438
137815
110094
123484
105401
115725
87315
135519
110613
193228
158927
204682
165744
209952
155489
201275
170625
345267
256171
323801
287076
161938
Table 11: Results for 2ELRPSD Instances Set Prodhon
SPR
Instance
C
Set Prodhon
20-5-1
20
20-5-1b
20
20-5-2
20
20-5-2b
20
50-5-1
50
50-5-1b
50
50-5-2
50
50-5-2b
50
50-5-2BIS
50
50-5-2bBIS 50
50-5-3
50
50-5-3b
50
100-5-1
100
100-5-1b
100
100-5-2
100
100-5-2b
100
100-5-3
100
100-5-3b
100
100-10-1
100
100-10-1b
100
100-10-2
100
100-10-2b
100
100-10-3
100
100-10-3b
100
200-10-1
200
200-10-1b
200
200-10-2
200
200-10-2b
200
200-10-3
200
200-10-3b
200
Avg.
S
Avg. 20
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
89075.00
61863.00
84489.50
61033.80
130859.30
101548.00
131825.00
110332.00
122599.00
105935.50
128436.00
104006.00
318667.00
257436.35
231340.00
194812.70
245334.90
195586.20
357381.40
300239.15
304931.20
264592.00
312701.25
261577.90
552488.90
448095.45
513673.40
432687.00
529578.00
404431.25
245251.87
Best 20
LNS-2E
t(s)
89075
63
61863
83
84478
62
60838 125
130843
80
101530 128
131825
97
110332 198
122599 112
105696 198
128379
80
104006 131
318225 226
256991 301
231305 204
194763 240
244470 174
195381 180
352694 233
298186 299
304507 248
264092 307
311447 227
260516 303
548730 1009
445791 635
497451 1158
422668 730
527162 970
402117 592
243599 313
t*(s)
Avg. 20
Best 20
Best
t(s) t*(s)
2
89075.00
89075
89075
60
0
61863.00
61863
61863
60
11
84478.00
84478
84478
60
0
60838.00
60838
60838
60
16
131454.00 131085 130843
60
35
101669.20 101530 101530
60
11
131827.00 131825 131825
60
12 110332.00 110332 110332
60
91 122599.00 122599 122599
60
155
105707.85 105696 105696
60
9
128614.50 128379 128379
60
6 104006.00 104006 104006
60
153
319268.60 318399 318399 900
220
257686.40 256991 256888 900
131
231488.85 231305 231305 900
202
194800.35 194763 194729 900
124
245178.75 244071 244071 900
111
195123.20 194110 194110 900
167
362648.70 354525 352122 900
194
312451.60 299758 298298 900
194
307937.60 304909 304438 900
208
265814.85 264173 263876 900
141
318952.10 311699 310930 900
218
265442.40 262932 261566 900
748
564159.80 550672 550672 900
576
456952.40 448188 447113 900
832
499499.45 498486 498397 900
696
428912.35 422967 422877 900
903
568539.15 534271 533174 900
557
425078.20 417686 417429 900
224
248413.28 244720 244395 564
4
2
19
5
24
14
16
14
3
18
10
7
303
268
229
176
377
327
401
540
520
310
512
437
725
692
656
601
668
700
286
BKS
89075
61863
84478
60838
130843
101530
131825
110332
122599
105696
128379
104006
318134
256878
231305
194728
244071
194110
351243
297167
304438
263873
310200
260328
548703
445301
497451
422668
527162
401672
243363
Finding high-quality solutions is increasingly difficult as the problem size grows, and different instance
characteristics influence solution quality. Figure 2 displays boxplots of 2EVRP instances grouped together by
number of customers (2a), or number of satellites (2b), similar customer distribution (2c) and similar satellite
distribution (2d) using gaps of the average value of five runs to BKS.
Problem difficulty quickly grows with the number of satellites, as well as with the number of customers, as
plotted in the upper part of Figure 2. The number of samples for the different classes varies a lot: There are
only 18 instances with 75 customers, but 165 instances with 50 customers, which explains the large number of
outliers for those instances. Crainic et al. (2010) provide a detailed overview on the distribution of customers
and satellites in instances of Set 4. There are three distribution patterns for customers: random, with equally
distributed nodes; centroids, where more customers are located in six centroids in a central zone, and some
customers closer together in four outer areas, representing suburbs. In the quadrant pattern customers are
arranged in conglomerations of higher density in each of the four quadrants. The three patterns for satellites
are: random, where satellites are randomly placed on a ring around the customers; sliced, with the satellites
distributed more evenly on the ring around customers; and forbidden zone: a random angle on the ring
was chosen where no satellites could be used, to simulate various conditions like cities located near lakes or
mountains. Figure 2c shows that instances with customers located in centroids are harder to solve. On the
other hand, the distribution pattern of the satellites doesn’t have a large impact on solution quality.
6.5
Graphical Example
Several structurally different 2EVRP solutions can have similar objective values. We discuss and visualise
this with the help of a demonstrative graphical example. Also the solution differences for an instance with or
without constraining the number of city freighters per satellite are pointed out.
The selection of the correct subset of satellites to use is crucial. A graphical representation of different
solutions of the Set 4 instance 38 is provided in Figure 3. The locations of the depot (square), satellites
(triangles) and customers (circles) is the same for each solution. Nevertheless, the obtained vehicle routes are
substantially different given different subsets of open satellites.
19
Figure 3a represents the optimal solution to the Set 4a instance 38. A maximum of vs2 = 2 city freighters
per satellite are available. The total demand of all customers sums up to 20206 units of freight. The capacity
Q2 of a city freighter is 5000 units. Any feasible solution needs at least d 20206
5000 e = 5 city freighters, and thus at
least d5/2e = 3 satellites have to be used. Two city freighter routes are located on the left of the figure, and
three city freighter routes on the right hand side, where two city freighters leave from the same satellite, and a
third one from a close by satellite.
Considering the instance as in Set 4b, with a global number of city freighters available but no constraints
on the distribution amongst satellites (vs2 = v 2 ), the optimal solution is displayed in Figure 3b. Three city
freighters can leave the same satellite, as is the case on the left hand side. Only two of the five satellites have
to be used.
LNS-2E starts with the construction of routes at the second level. In early stages of the optimisation
process, all customers are likely assigned to their closest satellites. Without neighbourhoods that impact the
selection of satellites, the algorithm would likely be trapped in a local optimum such as the one of Figure 3c. If
partial routes originating at the bottom right satellite exist, customers may be sequentially inserted in those
routes and the bottom left satellite may not be opened. The solution displayed in Figure 3d is often obtained.
It has the best cost considering only the second level, but long truck routes on the first level set this gain off,
leading to a worse solution overall.
Figure 3e shows the best found solution if only the top right satellite is open. The cost differences from one
solution to the next one are small, although the solution itself is fundamentally different. If the two satellites
at the bottom are both selected for closure at the same time, then the algorithm finds the optimal solution
within seconds.
We tried different strategies to evaluate the chances of a satellite to be included in the best solution:
taking into account a delta evaluation on the truck route, or combining this value with the total units shipped
through this satellite; or the absolute distance from the depot. We observed that closing satellites randomly
is a straightforward and very simple approach, which performs quite well on average over all the different
benchmark instances, whereas other techniques present advantages and disadvantages in several special cases.
For further research, we suggest to shift the cost structure towards more realistic scenarios. In the classic
2EVRP as we considered it, the cost of large trucks and small city freighters is the same. For instance Sets 2
to 4, the capacity of a truck is 2.5 times higher than of a city freighter. For Set 5 instances this ratio gets up to
more than 14, so one can safely assume that the operating cost of a truck will be higher than of a city freighter
in more realistic set-ups. This has not yet been taken into account in previous publications on the 2EVRP,
and would lead to large differences between the solution costs of Figure 3.
6.6
Sensitivity Analysis
Sensitivity analyses were conducted to evaluate the impact of major parameters and components of the method.
In particular, we evaluate the impact of disabling single destroy operators or local search procedures at a time
and provide the average objective value over five runs of all benchmark instances. Table 12 shows the average
objective values for instances of each class and its average deviation (Gap (%)) when disabling elements of the
algorithm.
The elimination of the open all satellites (no open) neighbourhood has a strong impact on solution quality.
As closed satellites are only opened again in a re-start phase, the algorithm is likely to be trapped in a local
optimum.
If no satellites are forced to be closed, there is no need to open up any satellite again. In this case, solution
quality deteriorates by 1.15% on average (no close). For some instances, satellites located very far away from
the depot will not be used in the best solution, but on the second level it seems to be beneficial to use them for
customers close by. The impact on the 2ELRPSD, especially the instance Set Prodhon is quite strong, as the
fixed costs of a satellite are not taken into account on the second level. If a satellite with high opening costs is
located close to many customers, it will very likely be used, although it would pay off to use a cheaper one
further away. A similar behaviour was discussed in Section 6.5 and Figure 3.
Some techniques work better on smaller instances, while others perform better on the larger instances. For
some cases we even observed small improvements when an operator was not used. Not using biased node
1
hinges depict approximately the first and third quartile of the solutions. The whiskers extend to ±1.58 Q3√−Q
. (The R
n
Foundation for Statistical Computing, 2014). See the Appendix 8.2 for a detailed definition.
2 The
20
3.5
3.5
●
●
●
●
3.0
2.5
2.5
3.0
●
●
1.0
●
2.0
Gap (%)
●
●
1.5
2.0
1.5
●
●
1.0
Gap (%)
●
●
●
●
●
●
●
21
32
50
0.0
●
●
●
●
●
●
●
●
●
0.5
●
●
0.0
0.5
●
●
75
100
200
●
●
●
●
●
●
●
●
●
2
3
Customers
●
●
4
5
6
10
Satellites
(a) number of Customers
(b) number of Satellites
●
●
3.5
3.5
●
●
●
●
2.5
2.5
3.0
●
3.0
●
●
●
2.0
●
●
●
●
●
●
●
0.5
●
Centroids
Quadrants
●
0.0
●
●
0.0
0.5
●
1.0
1.5
●
1.5
Gap (%)
2.0
●
1.0
Gap (%)
●
●
●
Random
Forbidden
Customer Characteristics
●
●
●
●
●
●
●
Random
Sliced
●
●
Satellite Characteristics
(c) Set 4: Customer distribution
(d) Set 4: Satellite distribution
Figure 2: Boxplots2 of solution quality for instances grouped by number of customers/satellites and distribution
characteristics
21
(a) max. 2 CF/Sat, 1169.20*, optimal for
Set 4a
(b) 2 Satellites used, 1163.07*, optimal
for Set 4b
(d) 4 Satellites used, 1226.43
(c) 3 Satellites used, 1210.81
(e) 1 Satellite used, 1255.98
Figure 3: Different Solutions of Instance 38 from Set 4, depending on satellite openings and vs2
22
removal for example is needed to robustly find optimal solutions on the smaller instance sets (2-4) but yields
improvements on the larger sets.
For local search techniques, we also observe differences on methods for small or large instance sets respectively.
Removing classic 2-opt has a small impact. The repair mechanism finds already high quality single routes in
terms of 2-optimality. Eliminating the inter-tour operator 2-opt* has a stronger effect on larger Set 5 and Set 6
instances, which contain more routes than the smaller instances. The relocate neighbourhood is essential for
instances with more than 50 customers, whereas exchanging two nodes against one can have a negative effect
on those larger instances. We still decided to keep it in the design of the algorithm, as this local search is
needed to find optimal solutions for smaller instances. Of course the algorithm could be fine-tuned for specific
applications or instance sizes.
Table 12: Sensitivity Analysis and contribution of individual components
Set
Sensitivity Analysis
base no related no biased no route
no single no close
2
578.27
3
641.44
4a
1420.05
4b
1397.21
5
1132.02
6a
912.51
6b
1343.36
Prodhon 248413.28
Nguyen 164492.82
Avg.
46703.44
Set
base no restart
2
578.27
3
641.44
4a
1420.05
4b
1397.21
5
1132.02
6a
912.51
6b
1343.36
Prodhon 248413.28
Nguyen 164492.82
Avg.
46703.44
7
0.21%
0.31%
0.25%
0.16%
0.69%
0.25%
0.15%
0.95%
0.61%
0.40%
0.08%
0.13%
0.11%
0.10%
0.06%
0.11%
0.07%
0.43%
-0.08%
0.11%
0.02%
0.06%
0.02%
-0.01%
-0.07%
-0.03%
0.04%
-0.31%
0.05%
-0.03%
0.17%
0.27%
0.10%
0.18%
0.31%
0.07%
0.06%
-0.55%
0.28%
0.10%
0.02%
0.06%
0.04%
0.05%
0.35%
-0.01%
0.01%
-0.36%
0.19%
0.04%
0.07%
0.07%
0.53%
1.14%
0.39%
1.20%
1.15%
4.65%
1.13%
1.15%
no open
2.07%
1.66%
1.37%
1.39%
2.35%
1.36%
0.97%
3.58%
1.38%
1.79%
no 2opt no 2opt* no relocate no swap no swap 2
0.04%
0.05%
0.05%
0.03%
0.31%
0.02%
0.05%
-0.29%
-0.20%
0.01%
0.10%
0.09%
0.03%
0.27%
0.53%
0.15%
0.17%
0.30%
0.21%
0.21%
0.00%
0.06%
-0.01%
0.04%
0.47%
0.24%
0.12%
0.56%
0.40%
0.21%
0.00%
0.08%
0.01%
0.02%
0.29%
-0.03%
0.05%
-0.51%
0.22%
0.01%
0.00%
0.18%
-0.01%
0.03%
0.09%
0.02%
-0.01%
0.00%
0.03%
0.04%
Conclusion
We presented a very simple and fast LNS heuristic for the 2EVRP and 2ELRPSD. LNS-2E makes use of one
repair operator and only a few destroy operators. The proposed method finds solutions of higher quality than
existing algorithms, while being fast and conceptually simpler. The impact of various parameters and design
choices was highlighted. Meta-calibration techniques were used to set the parameters to good values, which
were subsequently verified during sensitivity analyses. Different techniques were attempted to open or close
satellites, and thus explore various combinations of design choices. At the end, a simple randomised approach
for fixing satellites, assorted with a minimum number of iterations without change of this decision led to good
and robust results on a wide range of benchmark instances. LNS-2E was able to improve 18 best known
solutions on the 49 2EVRP instances for which no proven optimal solution exists so far. Having resolved the
inconsistencies on the different sets of benchmark instances used in literature paves the way for future research
on this topic, which will focus on solving rich city logistics problems, and shifting the cost structure to a more
realistic scenario. It will be interesting to examine the implications of using higher operating costs for larger
trucks than for smaller city freighters, approaching more realistic and real-life transportation problems.
Online Resources
All necessary data can be found in the online section at https://www.univie.ac.at/prolog/research/
TwoEVRP. All instances have been transformed into a uniform format and can be downloaded. Set 4 instances
used to have negative and real x/y coordinates (with a maximum of two positions after decimal space). We
23
added a fixed factor to shift them only to be positive and multiplied them by 100 to be able to use positive
integers. This does not change the solution, but note that the objective should be adjusted by a factor 100.
We also provide detailed results on the new found best known solutions, both in human readable text and a
graphical representation.
Acknowledgement
This work is partially funded by the Austrian Climate and Energy Fund within the “Electric Mobility Flagship
projects” program under grant 834868 (project VECEPT). The financial support by the Austrian Federal
Ministry of Science, Research and Economy and the National Foundation for Research, Technology and
Development is gratefully acknowledged.
We would like to thank Roberto Baldacci, Vera Hemmelmayr, Sandro Pirkwieser, Martin Schwengerer and
Gianfranco Guastaroba for providing us detailed information, and/or solutions on the instances solved by their
methods.
We would also like to thank the two anonymous reviewers for their helpful and constructive comments
which helped us to improve the manuscript.
8
8.1
Appendix
Nomenclature of Set 3 instances
Table 13: Set 3 instances with 50 customers: identical except for depot coordinates
Set 3b1
depot at (0,0)
E-n51-k5-
Set 3c2
depot at (30,40)
12-18
12-41
12-43
39-41
40-41
40-43
E-n51-k5-
13-19
13-42
13-44
40-42
41-42
41-44
1 included in Hemmelmayr (2013)
2 included in Beasley (2014)
Hemmelmayr et al. (2012) report a best solution with total costs of 690.59 for instance E-n51-k5-s12-18,
which corresponds exactly to the solution found by our algorithm. On the other hand Jepsen et al. (2012)
report an objective value of 560.73 for an instance of that name. This corresponds exactly to the objective
value found by our algorithm for the instance E-n51-k5-s13-19 from Beasley (2014). So we assume Jepsen et al.
(2012) used the same instances as are provided from Beasley (2014), but the IDs of satellites in the names have
to be increased by one. Both versions exist in literature, as we described in Section 6.1. The only difference
between Sets 3b and 3c is the location of the depot. Table 13 shows, which instances correspond to each other,
apart from depot coordinates.
8.2
Definition of boxplots in Figure 2
The two hinges are versions of the first (Q1 ) and third quartile (Q3 ), i.e., close to quantile (x, c(1,3)/4). The
hinges equal the quartiles for odd n (where n ← length(x)) and differ for even n. Whereas the quartiles only
equal observations for n = (1 mod 4), the hinges do so additionally for n = (2 mod 4), and are in the middle
of two observations otherwise. They are based on asymptotic normality of the median and roughly equal sample
sizes for the two medians being compared, and are said to be rather insensitive to the underlying distributions of
√ , where Interquartile Range IQR = Q3 − Q1 . This gives roughly
the samples. The notches extend to ± 1.58 IQR
n
24
a 95% confidence interval for the difference in two medians. (The R Foundation for Statistical Computing,
2014)
8.3
2ELRPSD instances
The instances for the 2ELRPSD were downloaded at the homepage3 of Caroline Prodhon. If the description
file is renamed to “.doc” it can be opened in human readable format by Word. We found some small typos in
the description: At the end of the Prodhon files first the fixed cost of a second level vehicle (F2) is given, then
the fixed cost of a truck on first level (F1).
The description of the cost structure also seems to contain an error: According to the description, second
level vehicles operate at higher cost per distance than first level trucks, which should be the other way round,
obviously. This is also the way we treated the instances, and we believe previous authors did, too. The cost
linking point A to point B was calculated as shown in Table 14. Please note that first level vehicles do not
operate at exactly twice the Euclidean distance, due to the use of the ceil function after multiplication by
factor 2.
Table 14: Distance Matrix Calculation for the 2ELRPSD
Instance set
first level d
p
(xA
second level d
p
(xA
Prodhon
Nguyen
p
− xB )2 + (yA − yB )2 ∗ 100 ∗ 2e d (xA − xB )2 + (yA − yB )2 ∗ 10 ∗ 2e
p
− xB )2 + (yA − yB )2 ∗ 100e
d (xA − xB )2 + (yA − yB )2 ∗ 10e
In the instance file 200-10-3b, the capacity of first level vehicles is missing, and thus we used 5000 (as this
is the value used for all the other Prodhon instance files). Some customers in the 2ELRPSD instances have a
demand of 0. This case was not explicitly dealt with in the papers: our algorithm still plans an itinerary for a
city freighter which will visit the customer, but does not deliver any quantity of goods.
Acronyms
2EVRP two-echelon vehicle routing problem
2ELRP two-echelon location routing problem
2ELRPSD two-echelon location routing problem with single depot
ALNS adaptive large neighbourhood search
CMA-ES covariance matrix adaptation evolution strategy
CVRP capacitated vehicle routing problem
LNS large neighbourhood search
MDVRP multi-depot vehicle routing problem
SDVRP split delivery vehicle routing problem
VNS variable neighbourhood search
VRP vehicle routing problem
Bibliography
References
R. Baldacci. personal communication. email, May 20th 2013.
R. Baldacci, A. Mingozzi, R. Roberti, and R. Wolfler Calvo. An Exact Algorithm for the Two-Echelon
Capacitated Vehicle Routing Problem. Operations Research, 61(2):298–314, 2013. doi: 10.1287/opre.1120.
1153.
3 http://prodhonc.free.fr/Instances/instances0_us.htm
25
J. E. Beasley. OR-Library, accessed Oct. 6th 2014. URL http://people.brunel.ac.uk/~mastjjb/jeb/
orlib/files/vrp2e.rar.
M. Boccia, T. G. Crainic, A. Sforza, and C. Sterle. Location-Routing Models for Designing a Two-Echelon
Freight Distribution System. Technical report, Cirrelt, 2011.
C. Contardo, V. C. Hemmelmayr, and T. G. Crainic. Lower and upper bounds for the two-echelon capacitated
location-routing problem. Computers & Operations Research, 39:3185–3199, 2012.
J.-F. Cordeau, M. Gendreau, G. Laporte, J.-Y. Potvin, and F. Semet. A guide to vehicle routing heuristics.
Journal of the Operational Research Society, 53(5):512–522, 2002. doi: 10.1057/palgrave/jors/2601319.
T. G. Crainic, N. Ricciardi, and G. Storchi. Advanced Freight Transportation Systems for Congested Urban
Areas. Transportation Research Part C: Emerging Technologies, 12(2):119–137, 2004. doi: 10.1016/j.trc.2004.
07.002.
T. G. Crainic, N. Ricciardi, and G. Storchi. Models for Evaluating and Planning City Logistics Systems.
Transportation Science, 43(4):432–454, 2009. doi: 10.1287/trsc.1090.0279.
T. G. Crainic, G. Perboli, S. Mancini, and R. Tadei. Two-Echelon Vehicle Routing Problem: a Satellite Location
Analysis. Procedia - Social and Behavioral Sciences, 2(3):5944–5955, 2010. doi: 10.1016/j.sbspro.2010.04.009.
T. G. Crainic, S. Mancini, G. Perboli, and R. Tadei. Multi-Start Heuristics for the Two-Echelon Vehicle Routing
Problem. In P. Merz and J.-K. Hao, editors, Evolutionary Computation in Combinatorial Optimization,
volume 6622, pages 179–190. Springer Berlin Heidelberg, Berlin Heidelberg, 2011.
G. A. Croes. A method for solving traveling-salesman problems. Operations Research, 6(6):791–812, 1958. doi:
10.1287/opre.6.6.791.
R. Cuda, G. Guastaroba, and M. G. Speranza. A Survey on Two-Echelon Routing Problems. Computers &
Operations Research, 55:185–199, 2015. doi: 10.1016/j.cor.2014.06.008.
N. Hansen. The CMA evolution strategy: a comparing review. In J. Lozano, P. Larranaga, I. Inza, and
E. Bengoetxea, editors, Towards a new evolutionary computation. Advances on estimation of distribution
algorithms, pages 75–102. Springer Berlin Heidelberg, 2006.
V. C. Hemmelmayr. personal communication. email, March 21st 2013.
V. C. Hemmelmayr, J.-F. Cordeau, and T. G. Crainic. An Adaptive Large Neighborhood Search Heuristic for
Two-Echelon Vehicle Routing Problems Arising in City Logistics. Computers & Operations Research, 39(12):
3215–3228, 2012. doi: 10.1016/j.cor.2012.04.007.
S. Jacobsen and O. Madsen. A Comparative Study of Heuristics for a Two-Level Routing-Location Problem.
European Journal of Operational Research, 5(6):378–387, 1980. doi: 10.1016/0377-2217(80)90124-1.
M. Jepsen, S. Spoorendonk, and S. Ropke. A Branch-and-Cut Algorithm for the Symmetric Two-Echelon
Capacitated Vehicle Routing Problem. Transportation Science, 47(1):23–37, 2012. doi: 10.1287/trsc.1110.0399.
G. Laporte. Location routing problems. In B. L. Golden and A. A. Assad, editors, Vehicle Routing: Methods
and Studies, pages 163–197. Elsevier Science Publishers B.V., North-Holland, 1988.
G. Laporte and Y. Nobert. A vehicle flow model for the optimal design of a two-echelon distribution system. In
H. A. Eiselt and G. Pederzoli, editors, Advances in Optimization and Control: Proceedings of the Conference
“Optimization Days 86” Held at Montreal, Canada, April 30 – May 2, 1986, pages 158–173, Berlin, Heidelberg,
1988. Springer Berlin Heidelberg. doi: 10.1007/978-3-642-46629-8 11.
O. B. Madsen. Methods for solving combined two level location-routing problems of realistic dimensions.
European Journal of Operational Research, 12(3):295–301, mar 1983. doi: 10.1016/0377-2217(83)90199-6.
26
V. P. Nguyen, C. Prins, and C. Prodhon. Solving the two-echelon location routing problem by a GRASP
reinforced by a learning process and path relinking. European Journal of Operational Research, 216(1):
113–126, 2012a. doi: 10.1016/j.ejor.2011.07.030.
V. P. Nguyen, C. Prins, and C. Prodhon. A multi-start iterated local search with tabu list and path relinking
for the two-echelon location-routing problem. Engineering Applications of Artificial Intelligence, 25(1):56–71,
2012b. doi: 10.1016/j.engappai.2011.09.012.
G. Perboli and R. Tadei. New Families of Valid Inequalities for the Two-Echelon Vehicle Routing Problem.
Electronic Notes in Discrete Mathematics, 36:639–646, 2010.
G. Perboli, R. Tadei, and D. Vigo. The Two-Echelon Capacitated Vehicle Routing Problem: Models and
Math-Based Heuristics. Transportation Science, 45(3):364–380, 2011. doi: 10.1287/trsc.1110.0368.
S. Pirkwieser and G. R. Raidl. Variable Neighborhood Search coupled with ILP-based Very Large Neighborhood
Searches for the (Periodic) Location-Routing Problem. In Lecture Notes in Computer Science (including
subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume 6373 LNCS,
pages 174–189, Berlin, Heidelberg, 2010. Springer Berlin Heidelberg. doi: 10.1007/978-3-642-16054-7 13.
D. Pisinger and S. Ropke. Large Neighborhood Search. In M. Gendreau and J.-Y. Potvin, editors, Handbook of
Metaheuristics, volume 146, pages 399–419. Springer, 2010.
F. A. Santos, A. Salles da Cunha, and G. R. Mateus. Branch-and-price algorithms for the Two-Echelon Capacitated Vehicle Routing Problem. Optimization Letters, 7:1537–1547, 2013. doi: 10.1007/s11590-012-0568-3.
F. A. Santos, G. R. Mateus, and A. Salles da Cunha. A Branch-and-Cut-and-Price Algorithm for the TwoEchelon Capacitated Vehicle Routing Problem. Transportation Science, 2014. doi: 10.1287/trsc.2013.0500.
G. Schrimpf, J. Schneider, H. Stamm-Wilbrandt, and G. Dueck. Record Breaking Optimization Results
Using the Ruin and Recreate Principle. Journal of Computational Physics, 159:139–171, 2000. doi:
10.1006/jcph.1999.6413.
M. Schwengerer, S. Pirkwieser, and G. R. Raidl. A variable neighborhood search approach for the twoechelon location-routing problem. In J.-K. Hao and M. Middendorf, editors, Evolutionary Computation
in Combinatorial Optimization, volume 7245 of Lecture Notes in Computer Science, pages 13–24. Springer
Berlin Heidelberg, 2012. doi: 10.1007/978-3-642-29124-1 2.
P. Shaw. Using Constraint Programming and Local Search Methods to Solve Vehicle Routing Problems. In
M. Maher and J.-F. Puget, editors, Principles and Practice of Constraint Programming, volume 1520 of Lecture
Notes in Computer Science, pages 417–431. Springer Berlin Heidelberg, 1998. doi: 10.1007/3-540-49481-2\ 30.
The R Foundation for Statistical Computing. R help documentation on boxplot.stats, version 3.1.2, 2014. URL
http://web.mit.edu/r/r_v3.0.1/lib/R/library/grDevices/html/boxplot.stats.html. retrieved Feb.
20th 2015.
P. Toth and D. Vigo. The Granular Tabu Search and its Application to the Vehicle-Routing Problem. INFORMS
Journal on Computing, 15(4):333–346, 2003.
T. Vidal, T. G. Crainic, M. Gendreau, and C. Prins. Heuristics for Multi-Attribute Vehicle Routing Problems:
A Survey and Synthesis. European Journal of Operational Research, 231(1):1–21, 2013.
Z.-y. Zeng, W.-s. Xu, Z.-y. Xu, and W.-h. Shao. A Hybrid GRASP+VND Heuristic for the Two-Echelon
Vehicle Routing Problem Arising in City Logistics. Mathematical Problems in Engineering, 2014:1–11, 2014.
doi: 10.1155/2014/517467.
27
| 8cs.DS
|
A
An Abstract Interpretation-based Model of Tracing Just-In-Time
Compilation
arXiv:1411.7839v2 [cs.PL] 28 Oct 2015
STEFANO DISSEGNA, University of Padova
FRANCESCO LOGOZZO, Facebook Inc.
FRANCESCO RANZATO, University of Padova
Tracing just-in-time compilation is a popular compilation technique for the efficient implementation of dynamic languages, which is commonly used for JavaScript, Python and PHP. It relies on two key ideas.
First, it monitors program execution in order to detect so-called hot paths, i.e., the most frequently executed
program paths. Then, hot paths are optimized by exploiting some information on program stores which
is available and therefore gathered at runtime. The result is a residual program where the optimized hot
paths are guarded by sufficient conditions ensuring some form of equivalence with the original program.
The residual program is persistently mutated during its execution, e.g., to add new optimized hot paths or
to merge existing paths. Tracing compilation is thus fundamentally different from traditional static compilation. Nevertheless, despite the practical success of tracing compilation, very little is known about its
theoretical foundations. We provide a formal model of tracing compilation of programs using abstract interpretation. The monitoring phase (viz., hot path detection) corresponds to an abstraction of the trace semantics of the program that captures the most frequent occurrences of sequences of program points together with
an abstraction of their corresponding stores, e.g., a type environment. The optimization phase (viz., residual
program generation) corresponds to a transform of the original program that preserves its trace semantics
up to a given observation as modeled by some abstraction. We provide a generic framework to express dynamic optimizations along hot paths and to prove them correct. We instantiate it to prove the correctness
of dynamic type specialization and constant variable folding. We show that our framework is more general
than the model of tracing compilation introduced by Guo and Palsberg [2011] which is based on operational
bisimulations. In our model we can naturally express hot path reentrance and common optimizations like
dead-store elimination, which are either excluded or unsound in Guo and Palsberg’s framework.
Categories and Subject Descriptors: D.2.4 [Software Engineering]: Software/Program Verification – correctness proofs, formal methods; D.3.4 [Programming Languages]: Processors – compilers, optimization;
F.3.2 [Logics and Meanings of Programs]: Semantics of Programming Languages – program analysis
Additional Key Words and Phrases: Tracing JIT compilation, abstract interpretation, trace semantics
1. INTRODUCTION
Efficient traditional static compilation of popular dynamic languages like JavaScript,
Python and PHP is very hard if not impossible. In particular, these languages present so many dynamic features which make all traditional static
analyses used for program optimization very imprecise. Therefore, practical implementations of dynamic languages should rely on dynamic information in
The work of Francesco Logozzo was carried out while being affiliated with Microsoft Research, Redmond,
WA, USA. The work of Francesco Ranzato was partially supported by Microsoft Research Software Engineering Innovation Foundation 2013 Award (SEIF 2013) and by the University of Padova under the PRAT
projects BECOM and ANCORE.
Author’s addresses: S. Dissegna and F. Ranzato, Dipartimento di Matematica, University of Padova, Padova,
Italy; F. Logozzo, Facebook Inc., Seattle, WA, USA.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights
for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component
of this work in other works requires prior specific permission and/or a fee. Permissions may be requested
from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or [email protected].
c YYYY ACM 0164-0925/YYYY/01-ARTA $15.00
DOI:http://dx.doi.org/10.1145/0000000.0000000
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:2
S. Dissegna et al.
order to produce an optimized version of the program. Tracing just-in-time
(JIT) compilation (TJITC) [Bala et al. 2000; Bauman et al. 2015; Bebenita et al. 2010;
Böhm et al. 2011; Bolz et al. 2009; Bolz et al. 2011; Gal et al. 2006; Gal et al. 2009;
Pall 2005; Schilling 2013] has emerged as a valuable implementation and optimization
technique for dynamic languages (and not only, e.g. Java [Häubl and Mössenböck 2011;
Häubl et al. 2014; Inoue et al. 2011]). For instance, the Facebook HipHop virtual machine for PHP and the V8 JavaScript engine of Google Chrome use some form
of tracing compilation [Adams et al. 2014; Facebook Inc. 2013; Google Inc. 2010]. The
Mozilla Firefox JavaScript engine used to have a tracing engine, called TraceMonkey, which has been later substituted by whole-method just-in-time compilation engines (initially JägerMonkey and then IonMonkey) [Mozilla Foundation 2010;
Mozilla Foundation 2013].
The Problem. Tracing JIT compilers leverage runtime profiling of programs to detect
and record often executed paths, called hot paths, and then they optimize and compile
only these paths at runtime. A path is a linear sequence (i.e., no loops or join points are
allowed) of instructions through the program. Profiling may also collect information
about the values that the program variables may assume during the execution of that
path, which is then used to specialize/optimize the code of the hot path. Of course, this
information is not guaranteed to hold for all the subsequent executions of the hot path.
Since optimizations rely on that information, the hot path is augmented with guards
that check the profiled conditions, such as, for example, variable types and constant
variables. When a guard fails, execution jumps back to the old, non-optimized code.
The main hypotheses of tracing compilers, confirmed by the practice, are: (i) loop bodies
are the most interesting code to optimize, so they only consider paths inside program
loops; and (ii) optimizing straight-line code is easier than a whole-method analysis
(involving loops, goto, etc.).
Hence, tracing JIT compilers look quite different than traditional compilers. These
differences raise some natural questions on trace compilation: (i) what is a viable formal model, which is generic yet realistic enough to capture the behavior of real optimizers? (ii) which optimizations are sound? (iii) how can one prove their soundness? In
this paper we answer these questions.
Our formal model is based on program trace semantics [Cousot 2002] and abstract
interpretation [Cousot and Cousot 1977; Cousot and Cousot 2002]. Hot path detection
is modeled just as an abstraction of the trace semantics of the program, which only retains: (i) the sequences of program points which are repeated more than some threshold; (ii) an abstraction of the possible program stores, e.g., the type of the variables instead of their concrete values. As a consequence, a hot path does not contain loops nor
join points. Furthermore, in the hot path, all the correctness conditions (i.e., guards)
are explicit, for instance before performing integer addition, we should check that the
operands are integers. If the guard condition is not satisfied then the execution leaves
the hot path, reverting to the non-optimized code. Guards are essentially elements of
some abstract domain, which is then left as a parameter in our model. The hot path is
then optimized using standard compilation techniques—we only require the optimization to be sound.
We define the correctness of the residual (or extracted) program in terms of an abstraction of its trace semantics: the residual program is correct if it is indistinguishable, up to some abstraction of its trace semantics, from the original program. Examples of abstractions are the program store at the exit of a method, or the stores at loop
entry and loop exit points.
Main Contributions. This paper puts forward a formal model of TJITC whose key fea-
tures are as follows:
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:3
– We provide the first model of tracing compilation based on abstract interpretation
of trace semantics of programs.
– We provide a more general and realistic framework than the model of TJITC by
Guo and Palsberg [2011] based on program bisimulations: we employ a less restrictive correctness criterion that enables the correctness proof of practically implemented optimizations; hot paths can be annotated with runtime information on
the stores, notably type information; optimized hot loops can be re-entered.
– We formalize and prove the correctness of type specialization of hot paths.
Our model focusses on source-to-source program transformations and optimizations
of a low level imperative language with untyped global variables, which may play the
role of intermediate language of some virtual machine. Our starting point is that program optimizations can be seen as transformations that lose some information on the
original program, so that optimizations can be viewed as approximations and in turn
can be formalized as abstract interpretations. More precisely, we rely on the insight
by Cousot and Cousot [2002] that a program source can be seen as an abstraction of
its trace semantics, i.e. the set of all possible execution sequences, so that a sourceto-source optimization can be viewed as an abstraction of a transform of the program
trace semantics. In our model, soundness of program optimizations is defined as program equivalence w.r.t. an observational abstract interpretation of the program trace
semantics. Here, an observational abstraction induces a correctness criterion by describing what is observable about program executions, so that program equivalence
means that two programs are indistinguishable by looking only at their observable
behaviors.
A crucial part of tracing compilation is the selection of the hot path(s) to optimize.
This choice is made at runtime based on program executions, so it can be seen once
again as an abstraction of trace semantics. Here, a simple trace abstraction selects
cyclic instruction sequences, i.e. loop paths, that appear at least N times within a
single execution trace. These instruction sequences are recorded together with some
property of the values assumed by program variables at that point, which is represented as an abstract store belonging to a suitable store abstraction, which in general
depends on the successive optimizations to perform.
A program optimization can be seen as an abstraction of a semantic transformation
of program execution traces, as described by Cousot and Cousot [2002]. The advantage
of this approach is that optimization properties, such as their correctness, are easier to
prove at a semantic level. The optimization itself can be defined on the whole program
or, as in the case of real tracing JIT compilers, can be restricted to the hot path. This
latter restriction is achieved by transforming the original program so that the hot path
is extracted, i.e. made explicit: the hot path is added to the program as a path with no
join points that jumps back to the original code when execution leaves it. A guard is
placed before each command in this hot path that checks if the necessary conditions,
as selected by the store abstraction, are satisfied. A program optimization can then be
confined to the hot path only, making it linear, by ignoring the parts of the program
outside it. The guards added to the hot path allows us to retain precision.
We apply our TJITC model to type specialization. Type specialization is definitely
the key optimization for dynamic languages such as Javascript [Gal et al. 2009], as
they make available generic operations whose execution depends on the type of runtime values of their operands. Moreover, as a further application of our model, we
consider the constant variable folding optimization along hot paths, which relies on
the standard constant propagation abstract domain [Wegman and Zadeck 1991].
Related Work. A formal model for tracing JIT compilation has been put forward by
Guo and Palsberg [2011] at POPL symposium. It is based on operational bisimulation
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:4
S. Dissegna et al.
[Milner 1995] to describe the equivalence between source and optimized programs. We
show how this model can be expressed within our framework through the following
steps: Guo and Palsberg’s language is compiled into ours; we then exhibit an observational abstraction which is equivalent to Guo and Palsberg’s correctness criterion;
finally, after some minor changes that address a few differences in path selection, the
transformations performed on the source program turn out to be the same. Our framework overcomes some significant limitations in Guo and Palsberg’s model. The bisimulation equivalence model used in [Guo and Palsberg 2011] implies that the optimized
program has to match every change to the store made by the original program, whereas
in practice we only need this match to hold in certain program points and for some
variables, such as in output instructions. This limits the number of real optimizations
that can be modeled in this framework. For instance, dead store elimination is proven
unsound in [Guo and Palsberg 2011], while it is implemented in actual tracing compilers [Gal et al. 2009, Section 5.1]. Furthermore, their formalization fails to model some
important features of actual TJITC implementation: (i) traces are mere linear paths
of instructions, i.e., they cannot be annotated with store properties; (ii) hot path selection is completely non-deterministic, since they do not model a selection criterion;
and, (iii) once execution leaves an optimized hot path the program will not be able to
re-enter it.
It is also worth citing that abstract interpretation of program trace semantics roots at the foundational work by Cousot [1997; 2002] and has
been widely used as a technique for defining a range of static program
analyses [Barbuti et al. 1999; Colby and Lee 1996; Handjieva and Tzolovski 1998;
Logozzo 2009; Rival and Mauborgne 2007; Schmidt 1998; Spoto and Jensen 2003].
Also, Rival [2004] describes various program optimizations as the trace abstractions
they preserve. In the Cousot and Cousot terminology [Cousot and Cousot 2002], Rival’s approach corresponds to offline transformations whereas tracing compilation is
an online transformation.
Structure. The rest of the paper is organized as follows. Sections 2 and 3 contain
some necessary background: the language considered in the paper and its operational
trace semantics are defined in Section 2, while Section 3 recalls some basic notions of
abstract interpretation, in particular for defining abstract domains of program stores.
Hot paths are formally defined in Section 4 as a suitable abstract interpretation of
program traces, while Section 5 defines the program transform for extracting a given
hot path. The correctness of the hot path extraction transform is defined and proved
correct in Section 6, which also introduces in Subsection 6.2 program optimizations
along hot paths together with a methodology for proving their correctness. Section 7
applies our model of hot path optimization to type specialization of untyped program
commands, while Section 8 describes an application to constant variable folding along
hot paths. Nested hot paths and the corresponding program transform for their extraction are the subject of Section 9. Section 10 provides a thorough formal comparison of
our model with Guo and Palsberg [2011]’s framework for tracing compilation. Finally,
Section 11 concludes, also discussing some directions for future work.
This is an expanded and revised version of the POPL symposium article
[Dissegna et al. 2014] including all the proofs.
2. LANGUAGE AND CONCRETE SEMANTICS
2.1. Notation
Given a finite set X of objects, we will use the following notation concerning sequences:
ǫ is the empty sequence; X + is the set of nonempty finite sequences of objects of X;
X ∗ , X + ∪ {ǫ}; if σ ∈ X ∗ then |σ| denotes the length of σ; indices of objects in a
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:5
nonempty sequence σ ∈ X + start from 0 and thus range in the interval [0, |σ|) ,
[0, |σ| − 1]; if σ ∈ X + and i ∈ [0, |σ|) then σi ∈ X (or σ(i)) denotes the i-th object in σ; if
σ ∈ X ∗ and i, j ∈ [0, |σ|) then σ[i,j] ∈ X ∗ denotes the subsequence σi σi+1 . . . σj , which is
therefore the empty sequence if j < i, while if k ∈ N then σk ∈ X ∗ denotes the suffix
σk σk+1 . . . σ|σ|−1 , which is the empty sequence when k ≥ |σ|.
If f : X → Y is any function then its collecting version f c : ℘(X) → ℘(Y ) is defined
pointwise by f c (S) , {f (x) ∈ Y | x ∈ S}, and when clear from the context, by a slight
abuse of notation, it is sometimes denoted by f itself.
2.2. Syntax
We consider a basic low level language with untyped global variables, a kind
of elementary dynamic language, which is defined through the notation used in
[Cousot and Cousot 2002]. Program commands range in C and consist of a labeled action which specifies a next label (Ł is the undefined label, where the execution becomes
stuck: it is used for defining final commands).
Labels:
Values:
Variables:
Expressions:
Boolean Expressions:
Actions:
Commands:
L∈L
Ł 6∈ L
v ∈ Value
x ∈ Var
Exp ∋ E ::= v | x | E1 + E2
BExp ∋ B ::= tt | ff | E1 ≤ E2 | ¬B | B1 ∧ B2
A ∋ A ::= x := E | B | skip
C ∋ C ::= L : A → L′ (with L′ ∈ L ∪{Ł})
For any command L : A → L′ , we use the following notation:
lbl(L : A → L′ ) , L,
act(L : A → L′ ) , A,
suc(L : A → L′ ) , L′ .
Commands L : B → L′ whose action is a Boolean expression are called conditionals.
A program P ∈ ℘(C) is a (possibly infinite, at least in theory) set of commands. In
order to be well-formed, if a program P includes a conditional C ≡ L : B → L′ then P
must also include a unique complement conditional L : ¬B → L′′ , which is denoted by
cmpl(C) or C c , where ¬¬B is taken to be equal to B, so that cmpl(cmpl(C)) = C. The set
of well-formed programs is denoted by Program. In our examples, programs P will be
deterministic, i.e., for any C1 , C2 ∈ P such that lbl(C1 ) = lbl(C2 ): (1) if act(C1 ) 6= act(C2 )
then C1 = cmpl(C2 ); (2) if act(C1 ) = act(C2 ) then C1 = C2 . We say that two programs
P1 and P2 are equal up to label renaming, denoted by P1 ∼
= P2 , when there exists a
suitable renaming for the labels of P1 that makes P1 equal to P2 .
2.3. Transition Semantics
The language semantics relies on values ranging in Value, possibly undefined values
ranging in Valueu , truth values in Bool, possibly undefined truth values ranging in
Boolu and type names ranging in Types, which are defined as follows:
Value , Z ∪ Char∗
Bool , {true, false}
Valueu , Z ∪ Char∗ ∪{undef}
Boolu , {true, false, undef}
Types , {Int, String, Undef, ⊤T , ⊥T }
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:6
S. Dissegna et al.
E : Exp → Store → Valueu
EJvKρ , v
EJxKρ , ρ(x)
EJE1 Kρ +Z EJE2 Kρ
EJE1 + E2 Kρ , EJE1 Kρ · EJE2 Kρ
undef
if type(EJEi Kρ) = Int
if type(EJEi Kρ) = String
otherwise
B : BExp → Store → Boolu
BJttKρ , true
BJ ff Kρ , false
EJE1 Kρ ≤Z EJE2 Kρ
BJE1 ≤ E2 Kρ , ∃σ ∈ String . EJE2 Kρ = (EJE1 Kρ)·σ
undef
BJ¬BKρ , ¬BJBKρ
if type(EJEi Kρ) = Int
if type(EJEi Kρ) = String
otherwise
BJB1 ∧ B2 Kρ , BJB1 Kρ ∧ BJB2 Kρ
A : A → Store → Store ∪ {⊥}
AJskipKρ , ρ
ρ[x/EJEKρ] if EJEKρ 6= undef
AJx := EKρ ,
⊥
if EJEKρ = undef
ρ if BJBKρ = true
AJBKρ ,
⊥ if BJBKρ ∈ {false, undef}
Fig. 1. Semantics of program expressions and actions.
where Char is a nonempty finite set of characters and undef is a distinct symbol. The
mapping type : Valueu → Types provides the type of any possibly undefined value:
if v ∈ Z
Int
type(v) , String if v ∈ Char∗
Undef if v = undef
The type names ⊥T and ⊤T will be used in Section 7 as, respectively, top and bottom
type, that is, subtype and supertype of all types.
Let Store , Var → Valueu denote the set of possible stores on variables in Var, where
ρ(x) = undef means that the store ρ is not defined on a program variable x ∈ Var.
Hence, let us point out that the symbol undef will be used to represent both store
undefinedness and a generic error when evaluating an expression (e.g., additions and
comparisons between integers and strings), two situations which are not distinguished
in our semantics. A store ρ ∈ Store will be denoted by [x/ρ(x)]ρ(x)6=undef , thus omitting
undefined variables, while [ ] will denote the totally undefined store. If P ∈ Program
then vars(P ) denotes the set of variables in Var that occur in P , so that StoreP ,
vars(P ) → Valueu is the set of possible stores for P .
The semantics of expressions E, Boolean expressions B and program actions A is
standard and goes as defined in Fig. 1. Let us remark that:
(i) the binary function +Z denotes integer addition, ≤Z denotes integer comparison,
while · is string concatenation;
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:7
(ii) logical negation and conjunction in Boolu are extended in order to handle undef as
follows: ¬undef = undef and undef ∧ b = undef = b ∧ undef;
(iii) ρ[x/v] denotes a store update for the variable x with v ∈ Value;
(iv) the distinct symbol ⊥ 6∈ Valueu is used to denote the result of: AJx := EKρ when the
evaluation of the expression E for ρ generates an error; AJBKρ when the evaluation
of the Boolean expression B for ρ is either false or generates an error.
With a slight abuse of notation we also consider the collecting versions of the semantic
functions in Fig. 1, which are defined as follows:
E : Exp → ℘(Store) → ℘(Valueu )
EJEKS , {EJEKρ ∈ Valueu | ρ ∈ S}
B : BExp → ℘(Store) → ℘(Store)
BJBKS , {ρ ∈ S | BJBKρ = true}
A : A → ℘(Store) → ℘(Store)
AJAKS , {AJAKρ | ρ ∈ S, AJAKρ ∈ Store}
Let us point out that, in the above collecting versions, if EJEKρ = undef then EJEK{ρ} =
{undef} and AJx := EK{ρ} = ∅, while if BJBKρ ∈ {false, undef} then BJBK{ρ} = ∅ and
AJBK{ρ} = ∅.
Generic program states are pairs of stores and commands: State , Store × C. We
extend the previous functions lbl, act and suc to be defined on states, meaning that
they are defined on the command component of a state. Also, store(s) and cmd (s) return, respectively, the store and the command of a state s. The transition semantics
S : State → ℘(State) is a relation between generic states defined as follows:
Shρ, Ci , {hρ′ , C ′ i ∈ State | ρ′ ∈ AJact(C)K{ρ}, suc(C) = lbl(C ′ )}.
If P is a program then StateP , StoreP ×P is the set of possible states of P . Given
P ∈ Program, the program transition relation SJP K : StateP → ℘(StateP ) between
states of P is defined as:
SJP Khρ, Ci , {hρ′ , C ′ i ∈ StateP | ρ′ ∈ AJact(C)K{ρ}, C ′ ∈ P, suc(C) = lbl(C ′ )}.
Let us remark that, according to the above definition, if C ≡ L : A → L′ , C1 ≡ L′ : B →
L′′ and C1c ≡ L′ : ¬B → L′′′ are all commands in P and ρ′ ∈ AJAKρ then we have that
SJP Khρ, Ci = {hρ′ , C1 i, hρ′ , C1c i}.
A state s ∈ StateP is stuck for P when SJP Ks = ∅. Let us point that:
(i) If the conditional command of a state s = hρ, L : B → L′ i ∈ StateP is such that
BJBKρ = false then s is stuck for P because there exists no store ρ′ ∈ AJBK{ρ} = ∅.
(ii) If the command of a state s = hρ, L : A → Łi ∈ StateP has the undefined label Ł as
next label then s is stuck for P .
(iii) We have a stuck state s when an error happens. E.g., this is the case for an undefined
evaluation of an addition as in s = h[y/3, z/foo], L : x := y + z → L′ i and for an
undefined evaluation of a Boolean expression as in s = h[y/3, z/foo], L : y ≤ x → L′ i.
Programs typically have an entry point, and this is modeled through a distinct initial
label Lι ∈ L from which execution starts. StateιP , {hρ, Ci | lbl(C) = Lι } denotes the
set of possible initial states for P .
2.3.1. Trace Semantics. A partial trace is any nonempty finite sequence of generic program states which are related by the transition relation S. Hence, the set Trace of
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:8
S. Dissegna et al.
partial traces is defined as follows:
Trace , {σ ∈ State+ | ∀i ∈ [1, |σ|). σi ∈ Sσi−1 }.
The partial trace semantics of P ∈ Program is in turn defined as follows:
TJP K = TraceP , {σ ∈ (StateP )+ | ∀i ∈ [1, |σ|). σi ∈ SJP Kσi−1 }.
A trace σ ∈ TraceP is complete if for any state s ∈ StateP , σs 6∈ TraceP and sσ 6∈
TraceP . Observe that TraceP contains all the possible partial traces of P , complete
traces included. Let us remark that a trace σ ∈ TraceP does not necessarily begin with
an initial state, namely it may happen that σ0 6∈ StateιP . Traces of P starting from
initial states are denoted by
Tι JP K = TraceιP , {σ ∈ TraceP | σ0 ∈ StateιP }.
Also, a complete trace σ ∈ TraceιP such that suc(σ|σ|−1 ) = Ł corresponds to a terminating run of the program P .
Example 2.1. Let us consider the program Q below written in some while-language:
x := 0;
while (x ≤ 20) do
x := x + 1;
if (x%3 = 0) then x := x + 3;
Its translation as a program P in our language is given below, with Lι = L0 , where,
with a little abuse, we assume an extended syntax that allows expressions like
x%3 = 0.
P = C0 ≡ L0 : x := 0 → L1 ,
C1 ≡ L1 : x ≤ 20 → L2 , C1c ≡ L1 : ¬(x ≤ 20) → L5 ,
C2 ≡ L2 : x := x + 1 → L3 ,
C3 ≡ L3 : (x%3 = 0) → L4 , C3c ≡ L3 : ¬(x%3 = 0) → L1
C4 ≡ L4 : x := x + 3 → L1 , C5 ≡ L5 : skip → Ł
Its trace semantics from initial states TraceιP includes the following complete traces,
where [ ] is the initial totally undefined store.
h[ ], C0 ih[x/0], C1c i
h[ ], C0 ih[x/0], C1 ih[x/0], C2 ih[x/1], C3 i
h[ ], C0 ih[x/0], C1 ih[x/0], C2 ih[x/1], C3c ih[x/1], C1c i
···
···
h[ ], C0 ih[x/0], C1 ih[x/0], C2 ih[x/1], C3c ih[x/1], C1 i · · · h[x/21], C4 ih[x/24], C1 i
h[ ], C0 ih[x/0], C1 ih[x/0], C2 ih[x/1], C3c ih[x/1], C1 i · · · h[x/21], C4 ih[x/24], C1c ih[x/24], C5 i
Observe that the last trace corresponds to a terminating run of P .
3. ABSTRACTIONS
3.1. Abstract Interpretation Background
In
standard
abstract
interpretation
[Cousot and Cousot 1977;
Cousot and Cousot 1979], abstract domains, also called abstractions, are specified by Galois connections/insertions (GCs/GIs for short) or, equivalently, adjunctions.
Concrete and abstract domains, hC, ≤C i and hA, ≤A i, are assumed to be complete
lattices which are related by abstraction and concretization maps α : C → A and
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:9
γ : A → C such that, for all a and c, α(c) ≤A a ⇔ c ≤C γ(a). A GC is a GI when
α ◦ γ = λx.x. It is well known that a join-preserving α uniquely determines, by
adjunction, γ as follows: γ(a) = ∨{c ∈ C | α(c) ≤A a}; conversely, a meet-preserving γ
uniquely determines, by adjunction, α as follows: α(c) = ∧{a ∈ A | c ≤C γ(a)}.
Let f : C → C be some concrete monotone function—for simplicity, we consider 1-ary
functions—and let f ♯ : A → A be a corresponding monotone abstract function defined
on some abstraction A related to C by a GC. Then, f ♯ is a correct abstract interpretation of f on A when α ◦ f ⊑ f ♯ ◦ α holds, where ⊑ denotes the pointwise ordering
between functions. Moreover, the abstract function f A , α ◦ f ◦ γ : A → A is called the
best correct approximation of f on A because any abstract function f ♯ is correct iff
f A ⊑ f ♯ . Hence, for any A, f A plays the role of the best possible approximation of f on
the abstraction A.
3.2. Store Abstractions
As usual in abstract interpretation [Cousot and Cousot 1977], a store property is modeled by some abstraction Store♯ of ℘(Store) which is formalized through a Galois connection:
(αstore , h℘(Store), ⊆i, hStore♯ , ≤i, γstore ).
Given a program P , when Store♯ is viewed as an abstraction of h℘(StoreP ), ⊆i we
emphasize it by adopting the notation Store♯P . A store abstraction Store♯P also induces
a state abstraction State♯P , Store♯P ×P and, in turn, a trace abstraction defined by
Trace♯P , (State♯P )∗ .
3.2.1. Nonrelational Abstractions. Nonrelational store abstractions (i.e., relationships between program variables are not taken into account) can be easily designed by a standard pointwise lifting of some value abstraction. Let Value♯ be an abstraction of sets of
possibly undefined values in ℘(Valueu ) as formalized by a Galois connection
(αvalue , h℘(Valueu ), ⊆i, hValue♯ , ≤Value♯ i, γvalue ).
The abstract domain Value♯ induces a nonrelational store abstraction
ρ♯ ∈ Store♯value , hVar → Value♯ , ⊑i
where ⊑ is the pointwise ordering induced by ≤Value♯ : ρ♯1 ⊑ ρ♯2 iff for all x ∈ Var,
ρ♯1 (x) ≤Value♯ ρ♯2 (x). Hence, the bottom and top abstract stores are, respectively,
♯
λx.⊥Value♯ and λx.⊤Value♯ . The abstraction map α⊑
value : ℘(Store) → Storevalue is defined
as follows:
α⊑
value (S) , λx.αvalue ({ρ(x) ∈ Valueu | ρ ∈ S})
⊑
The corresponding concretization map γvalue
: Store♯value → ℘(Store) is defined, as recalled in Section 3.1, by adjunction from the abstraction map α⊑
value and it is easy to
check that it can be given as follows:
⊑
γvalue
(ρ♯ ) = {ρ ∈ Store | ∀x ∈ Var . ρ(x) ∈ γvalue (ρ♯ (x))}.
Let us observe that:
(i) α⊑
value (∅) = λx.αvalue (∅) = λx.⊥Value♯ because αvalue (∅) = ⊥Value♯ always holds in a
GC;
(ii) α⊑
value ({[ ]}) = λx.αvalue ({undef});
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:10
S. Dissegna et al.
⊑
(iii) if γvalue (⊥Value♯ ) = ∅, ρ♯ ∈ Store♯value and ρ♯ (x) = ⊥Value♯ then γvalue
(ρ♯ ) = ∅;
⊑
(iv) if γvalue (⊥Value♯ ) = {undef} then γvalue
(λx.⊥Value♯ ) = {[ ]}.
Example 3.1 (The constant propagation abstraction). The constant propagation (see [Wegman and Zadeck 1991]) lattice hCP, i is depicted below.
⊤
· · · v−2
v−1
v0
v1
v2 · · ·
⊥
where {vi }i∈Z is any enumeration of Valueu , thus undef is included. Abstraction
αcp : ℘(Valueu ) → CP and concretization γcp : CP → ℘(Valueu ) functions are defined
as follows:
if a = ⊥
⊥ if S = ∅
∅
αcp (S) , vi if S = {vi }
γcp (a) , {vi }
if a = vi
⊤ otherwise
Valueu if a = ⊤
and give rise to a GI (αcp , h℘(Valueu ), ⊆i, hCP, i, γcp ). The corresponding nonrelational
˙ where αCP : ℘(Store) → CPst and
store abstraction is denoted by CPst , hVar → CP, i,
γCP : CPst → ℘(Store) denote the abstraction and concretization maps. For example,
for Var = {x, y, z, w} and omitting the bindings v/undef also in abstract stores, we have
that:
αCP ({[x/2, y/foo, z/1], [x/2, y/bar]}) = [x/2, y/⊤, z/⊤],
γCP ([x/2, y/⊤, w/foo]) = {ρ ∈ Store | ρ(x) = 2, ρ(y) ∈ Valueu , ρ(z) = undef, ρ(w) = foo},
γCP ([x/2, y/⊤, w/⊥]) = ∅.
4. HOT PATH SELECTION
A loop path is a sequence of program commands which is repeated in some execution of
a program loop, together with a store property which is valid at the entry of each command in the path. A loop path becomes hot when, during the execution, it is repeated
at least a fixed number N of times. In a TJITC, hot path selection is performed by a
loop path monitor that also records store properties (see, e.g., [Gal et al. 2009]). Here,
hot path selection is not operationally defined, it is instead semantically modeled as
an abstraction map over program traces, i.e., program executions.
Given a program P and therefore its trace semantics TraceP , we first define a mapping loop : TraceP → ℘(TraceP ) that returns all the loop paths in some execution trace
of P . More precisely, a loop path is a proper substring (i.e., a segment) τ of a program
trace σ such that:
(1) the successor command in σ of the last state in τ exists and coincides with the
command – or its complement, when this is the last loop iteration – of the first
state in τ ;
(2) there is no other such command within τ (otherwise the sequence τ would contain
multiple iterations);
(3) the last state of τ performs a backward jump in the program P .
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:11
To recognize backward jumps, we consider a topological order on the control flow graph
of commands in P , denoted by ⋖. This leads to the following formal definition:
loop(hρ0 , C0 i · · · hρn , Cn i) , hρi , Ci ihρi+1 , Ci+1 i · · · hρj , Cj i | 0 ≤ i ≤ j < n, Ci ⋖ Cj ,
suc(Cj ) = lbl(Ci ), ∀k ∈ (i, j]. Ck 6∈ {Ci , cmpl(Ci )} .
Let us remark that a loop path
hρi , Ci i · · · hρj , Cj i ∈ loop(hρ0 , C0 i · · · hρn , Cn i)
may contain some sub-loop path, namely it may happen that loop(hρi , Ci i · · · hρj , Cj i) 6=
∅ so that some commands Ck , with k ∈ [i, j], may occur more than once in
hρi , Ci i · · · hρj , Cj i; for example, this could be the case of a while loop whose body includes a nested while loop.
We abuse notation by using αstore to denote a map αstore : TraceP → Trace♯P which
“abstracts” a program trace τ into Trace♯P by abstracting the sequence of stores occurring in τ :
αstore (hρ0 , C0 i · · · hρn , Cn i) , hαstore ({ρ0 }), C0 i · · · hαstore ({ρn }), Cn i.
Given a static integer parameter N > 0, we define a function
hot N : TraceP → ℘(Trace♯P )
which returns the set of Store♯ -abstracted loop paths appearing at least N times in
some program trace. In order to count the number of times a loop path appears within
a trace we need an auxiliary function count : Trace♯P × Trace♯P → N such that count(σ, τ )
yields the number of times an abstract path τ occurs in an abstract trace σ:
′
count(ha0 , C0 i · · · han , Cn i, hb0 , C0′ i · · · hbm , Cm
i) ,
n−m
X 1 if hai , Ci i · · · hai+m , Ci+m i = hb0 , C ′ i · · · hbm , C ′ i
0
m
0 otherwise
i=0
Hence, hot
N
can be defined as follows:
hot N (σ ≡ hρ0 , C0 i · · · hρn , Cn i) , hai , Ci i · · · haj , Cj i | ∃hρi , Ci i · · · hρj , Cj i ∈ loop(σ) s.t.
αstore (hρi , Ci i · · · hρj , Cj i) = hai , Ci i · · · haj , Cj i,
count(αstore (σ), hai , Ci i · · · haj , Cj i) ≥ N .
♯
Finally, an abstraction map αN
hot : ℘(TraceP ) → ℘(TraceP ) collects the results of applying hot N to a set of traces:
[
αN
hot N (σ).
hot (T ) ,
σ∈T
A N -hot path hp in a program P is therefore any hp ∈ αN
hot (TraceP ) and is compactly
denoted as hp = ha0 , C0 , ..., an , Cn i. Let us observe that if the hot path corresponds to
the body of some while loop then its first command C0 is a conditional, namely C0 is
the Boolean guard of the while loop. We define the successor function next for indices
in a hot path ha0 , C0 , ..., an , Cn i as follows: next , λi. i = n ? 0 : i + 1. For a N -hot
path ha0 , C0 , ..., an , Cn i ∈ αN
hot (TraceP ), for any i ∈ [0, n], if Ci is a conditional command
Li : Bi → Lnext(i) then throughout the paper its complement Cic = cmpl(Ci ) will be also
denoted by Li : ¬Bi → Lcnext(i) .
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:12
S. Dissegna et al.
Example 4.1. Let us consider the program P in Example 2.1 and a trivial one-point
store abstraction Store♯ = {⊤}, where all the stores are abstracted to the same abstract
store ⊤, i.e., αstore = λS.⊤. Here, we have two 2-hot paths in P , that is, it turns out
that α2hot (TraceP ) = {hp1 , hp2 } where:
hp1 = h⊤, C1 ≡ L1 : x ≤ 20 → L2 , ⊤, C2 ≡ L2 : x := x + 1 → L3 ,
⊤, C3c ≡ L3 : ¬(x%3 = 0) → L1 i;
hp2 = h⊤, C1 ≡ L1 : x ≤ 20 → L2 , ⊤, C2 ≡ L2 : x := x + 1 → L3 ,
⊤, C3 ≡ L3 : (x%3 = 0) → L4 , ⊤, C4 ≡ x := x + 3 → L1 i.
Therefore, the hot paths hp1 and hp2 correspond, respectively, to the cases where the
Boolean test (x%3 = 0) fails and succeeds. Observe that the maximal sequence of
different values assumed by the program variable x is as follows:
? 7→ 0 7→ 1 7→ 2 7→ 3 7→ 6 7→ 7 7→ 8 7→ 9 7→ 12 7→ 13 7→ 14 7→ 15 7→ 18 7→ 19 7→ 20 7→ 21 7→ 24
Hence, if σ is the complete terminating trace of P in Example 2.1 then it turns out that
count(αstore (σ), hp1 ) = 8 and count(αstore (σ), hp2 ) = 4.
5. TRACE EXTRACTION
For any abstract store a ∈ Store♯ , a corresponding Boolean expression denoted by
guard Ea ∈ BExp is defined (where the notation Ea should hint at an expression which
is induced by the abstract store a), whose semantics is as follows: for any ρ ∈ Store,
true if ρ ∈ γstore (a)
BJguard Ea Kρ ,
false if ρ 6∈ γstore (a)
In turn, we also have program actions guard Ea ∈ A such that:
ρ if ρ ∈ γstore (a)
AJguard Ea Kρ ,
⊥ if ρ 6∈ γstore (a)
Let P be a program and hp = ha0 , C0 , ..., an , Cn i ∈ αN
hot (TraceP ) be a hot path on some
♯
store abstraction Store . We define a syntactic transform of P where the hot path hp is
explicitly extracted from P . This is achieved by a suitable relabeling of each command
Ci in hp which is in turn preceded by the conditional guard Eai induced by the corresponding store property ai . To this aim, we consider three injective relabeling functions
ℓ : [0, n] → L1
l : [1, n] → L2
(·) : L → L
(∗)
where L1 , L2 and L are pairwise disjoint sets of fresh labels, so that labels(P )∩(L1 ∪L2 ∪
L) = ∅. The transformed program extr hp (P ) for the hot path hp is defined as follows
and a graphical example of this transform is depicted in Fig. 2.
Definition 5.1 (Trace extraction transform). The trace extraction transform of P
for the hot path hp = ha0 , C0 , ..., an , Cn i is:
extr hp (P ) , P r {C0 } ∪ {cmpl(C0 ) | cmpl(C0 ) ∈ P }
∪ {L0 : act(C0 ) → L1 } ∪ {L0 : ¬act(C0 ) → Lc1 | cmpl(C0 ) ∈ P } ∪ stitch P (hp)
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:13
.
.
.
L0
guard Ea0
.
.
.
L0
B0
ℓ0
B0
Lc1
B0
L0
Lc1
Lc1
l1
guard Ea1
L1
A1
L1
A1
ℓ1
L2
A1
L2
.
.
.
.
.
.
l2
.
.
.
Ln
An
ln
guard Ean
Ln
An
ℓn
An
Fig. 2. An example of trace extraction transform: on the left, a hot path hp with commands in pink (in
black/white: loosely dotted) shapes; on the right, the corresponding trace transform extr hp (P ) with new
commands in blue (in black/white: densely dotted) shapes.
where the stitch of hp into P is defined as follows:
stitch P (hp) , {L0 : guard Ea0 → ℓ0 , L0 : ¬guard Ea0 → L0 }
∪ {ℓi : act(Ci ) → li+1 | i ∈ [0, n − 1]}∪{ℓn : act(Cn ) → L0 }
∪ {ℓi : ¬act(Ci ) → Lcnext(i) | i ∈ [0, n], cmpl(Ci ) ∈ P }
∪ {li : guard Eai → ℓi ,
li : ¬guard Ea
i
→ Li | i ∈ [1, n]}.
The new command L0 : guard Ea0 → ℓ0 is therefore the entry conditional of the
stitched hot path stitch P (hp), while any command C ∈ stitch P (hp) such that suc(C) ∈
labels(P ) ∪ L is a potential exit (or bail out) command of stitch P (hp).
L EMMA 5.2. If P is well-formed then, for any hot path hp, extr hp (P ) is well-formed.
P ROOF. Recall that a program is well-formed when for any its conditional command
it also includes a unique complement conditional. It turns out that extr hp (P ) is wellformed because: (1) P is well-formed; (2) for each conditional in Pnew = extr hp (P ) r P =
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:14
S. Dissegna et al.
stitch P (hp) ∪ {L0 : act(C0 ) → L1 } ∪ {L0 : ¬act(C0 ) → Lc1 | cmpl(C0 ) ∈ P } we also have
a unique complement conditional in Pnew . Moreover, observe that if P is deterministic
then extr hp (P ) still is deterministic.
Let us remark that the stitch of the hot path hp into P is always a linear sequence
of different commands, namely, stitch P (hp) does not contain loops nor join points. Furthermore, this happens even if the hot path hp does contain some inner sub-loop. Technically, this is achieved as a consequence of the fact that the above relabeling functions
ℓ and l are required to be injective. Hence, even if some command C occurs more than
once inside hp, e.g., Ci = C = Cj for some i, j ∈ [0, n − 1] with i 6= j, then these
multiple occurrences of C in hp are transformed into differently labeled commands in
stitch P (hp), e.g., because ℓi 6= ℓj and li+1 6= lj+1 .
Let us now illustrate the trace extraction transform on a first simple example.
Example 5.3. Let us consider the program P in Example 2.1 and the hot path hp =
h⊤, C1 , ⊤, C2 , ⊤, C3c i in Example 4.1 (denoted there by hp1 ), where stores are abstracted
to the trivial one-point abstraction Store♯ = {⊤}. Here, for any ρ ∈ Store, we have
that BJguard E⊤ Kρ = true. The trace extraction transform of P w.r.t. hp is therefore as
follows:
extr hp (P ) = P r {C1 , C1c } ∪ {L1 : x ≤ 20 → L2 , L1 : ¬(x ≤ 20) → L5 } ∪ stitch P (hp)
where
stitch P (hp) = {H0 ≡ L1 : guard E⊤ → ℓ0 , H0c ≡ L1 : ¬guard E⊤ → L1 }
∪ {H1 ≡ ℓ0 : x ≤ 20 → l1 , H1c ≡ ℓ0 : ¬(x ≤ 20) → L5 }
∪ {H2 ≡ l1 : guard E⊤ → ℓ1 , H2c ≡ l1 : ¬guard E⊤ → L2 }
∪ {H3 ≡ ℓ1 : x := x + 1 → l2 }
∪ {H4 ≡ l2 : guard E⊤ → ℓ2 , H4c ≡ l2 : ¬guard E⊤ → L3 }
∪ {H5 ≡ ℓ2 : ¬(x%3 = 0) → L1 , H5c ≡ ℓ2 : (x%3 = 0) → L4 }.
The flow graph of extr hp (P ) is depicted in Figure 3, while a higher level representation
using while-loops and gotos is as follows:
x := 0;
L1 : while guard E⊤ do
if ¬(x ≤ 20) then goto L5
if ¬guard E⊤ then goto L2
x := x + 1;
if ¬guard E⊤ then goto L3
if (x%3 = 0) then goto L4
if ¬(x ≤ 20) then goto L5
L2 : x := x + 1;
L3 : if ¬(x%3 = 0) then goto L1
L4 : x := x + 3;
goto L1 ;
L5 : skip;
6. CORRECTNESS
As advocated by Cousot and Cousot [2002, par. 3.8], correctness of dynamic program
transformations and optimizations should be defined with respect to some observational abstraction of program trace semantics: a dynamic program transform is correct
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:15
L0
C0
x := 0
L1
H0
H0c
guard E⊤
L1
C1
ℓ0
H1
H1c
x ≤ 20
C1c
x ≤ 20
L5
L2
l1
H2
H2c
C2
x := x + 1
C5
skip
Ł
guard E⊤
L3
ℓ1
H3
x := x + 1
C3
C3c
l2
H4
H4c
guard E⊤
x%3 = 0
L4
C4 x := x + 3
ℓ2
H5
H5c
¬(x%3 = 0)
Fig. 3. The flow graph of the trace extraction transform extr hp (P ) in Example 5.3, where commands of
stitch P (hp) are in blue (in black/white: densely dotted) shapes, while commands of the source program P
are in pink (in black/white: loosely dotted) shapes.
when, at some level of abstraction, the observation of the execution of the subject program is equivalent to the observation of the execution of the transformed/optimized
program.
Store Changes Abstraction. The approach by Guo and Palsberg [2011] to tracing compilation basically relies on a notion of correctness that requires the same store changes
to happen in both the transformed/optimized program and the original program. This
can be easily encoded by an observational abstraction αsc : ℘(Trace) → ℘(Store∗ ) of
partial traces that observes store changes in execution traces:
sc : Trace → Store∗
ε
ρ
sc(σ) ,
sc(hρ,
C1 iσ ′ )
ρ0 sc(hρ1 , C1 iσ ′ )
if σ
if σ
if σ
if σ
=ε
= hρ, Ci
= hρ, C0 ihρ, C1 iσ ′
= hρ0 , C0 ihρ1 , C1 iσ ′ , ρ0 6= ρ1
αsc (T ) , {sc(σ) | σ ∈ T }
Since the function αsc obviously preserves arbitrary set unions, as recalled in Section 3.1, it admits a right adjoint γsc : ℘(Store∗ ) → ℘(Trace) defined as γsc (S) , ∪{T ∈
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:16
S. Dissegna et al.
℘(Trace) | αsc (T ) ⊆ S}, that gives rise to a GC (αsc , h℘(Trace), ⊆i, h℘(Store∗ ), ⊆i, γsc ).
By a slight abuse of notation, αsc is also used as an abstraction of the partial trace
semantics of a given program P , that is, αsc : ℘(TraceP ) → ℘(Store∗P ), which, clearly,
gives rise to a corresponding GC (αsc , h℘(TraceP ), ⊆i, h℘(Store∗P ), ⊆i, γsc ).
Output Abstraction. The store changes abstraction αsc may be too strong in practice. This can be generalized to any observational abstraction of execution traces
αo : h℘(Trace), ⊆i → hA, ≤A i (which gives rise to a GC). As a significant example,
one may consider an output abstraction that demands to have the same stores (possibly restricted to some subset of program variables) only at some specific output points.
For example, in a language with no explicit output primitives, as that considered by
Guo and Palsberg [2011], one could be interested just in the final store of the program
(when it terminates), or in the entry and exit stores of any loop containing an extracted hot path. If we consider a language including a distinct primitive command
“put X ” that “outputs” the value of program variables ranging in some set X then we
may want to have the same stores for variables in X at each output point put X . In
this case, optimizations should preserve the same sequence of outputs, i.e. optimizations should not modify the order of output commands. More formally, this can be
achieved by adding a further sort of actions: put X ∈ A, where X ⊆ Var is a set of
program variables. The semantics of put X obviously does not affect program stores,
i.e., AJput X Kρ , ρ. Correspondingly, if StoreX denotes stores on variables ranging in
X then the following output abstraction αout : ℘(Trace) → ℘(Store∗X ) of partial traces
observes program stores at output program points only:
out : Trace → Store∗X
ε
out (σ) , out (σ ′ )
ρ|X out(σ ′ )
if σ = ε
if σ = sσ ′ ∧ act(s) 6= put X
if σ = hρ, L : put X → L′ iσ ′
αout (T ) , {out(σ) | σ ∈ T }
where ρ|X denotes the restriction of the store ρ to variables in X . Similarly to αsc , here
again we have a GC (αo , h℘(Trace), ⊆i, h℘(Store∗X ), ⊆i, γo ).
Example 6.1 (Dead store elimination). This approach based on a generic observational abstraction enables to prove the correctness of program optimizations that
are unsound in Guo and Palsberg [2011]’s framework based on the store changes abstraction, such as dead store elimination. For example, in a program fragment such
as
while (x ≤ 0) do
z := 0;
x := x + 1;
z := 1;
one can extract the hot path hp = hx ≤ 0, z := 0, x := x + 1, z := 1i (here we ignore
store abstractions) and perform dead store elimination of the command z := 0 by optimizing hp to hp′ = hx ≤ 0, x := x + 1, z := 1i. As observed by Guo and Palsberg [2011,
Section 4.3], this is clearly unsound in bisimulation-based correctness because this hot
path optimization does not output bisimilar code. By contrast, this optimization can be
made sound by choosing and then formalizing an observational abstraction of program
traces which requires to have the same stores at the beginning and at the exit of loops
containing an extracted hot path, while outside of hot paths one could still consider
the store changes abstraction.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:17
Observational Abstraction. One can generalize the store changes abstraction αsc by
considering any observational abstraction αo : h℘(Trace), ⊆i → hA, ≤A i which is less
precise (i.e., more approximate) than αsc : this means that for any T1 , T2 ∈ ℘(Trace),
if αsc (T1 ) = αsc (T2 ) then αo (T1 ) = αo (T2 ), or, equivalently, for any T ∈ ℘(Trace),
γsc (αsc (T )) ⊆ γo (αo (T )). Informally, this means that αo abstracts more information
than αsc . As an example, when considering programs with output actions, the following abstraction αosc : ℘(Trace) → ℘(Store∗X ) observes store changes at output program
points only:
osc : Trace → Store∗X
ε
ρ|X
osc(hρ, L1 : put X → L′1 iσ ′ )
osc(σ) ,
osc(hρ, L1 : A1 → L′1 iσ ′ )
′
1 iσ )
ρ0|X osc(hρ1 , C
osc(hρ1 , C1 iσ ′ )
if σ
if σ
if σ
if σ
if σ
if σ
= ε or σ = hρ, Ci, act(C) 6= put X
= hρ, Ci, act(C) = put X ,
= hρ, C0 ihρ, L1 : A1 → L′1 iσ ′ , act(C0 ) = put X
= hρ, C0 ihρ, L1 : A1 → L′1 iσ ′ , act(C0 ) 6= put X
= hρ0 , C0 ihρ1 , C1 iσ ′ , ρ0 6= ρ1 , act(C0 ) = put X
= hρ0 , C0 ihρ1 , C1 iσ ′ , ρ0 6= ρ1 , act(C0 ) 6= put X
αosc (T ) , {osc(σ) | σ ∈ T }
Clearly, it turns out that αosc is more approximate than αsc since osc(σ) records a store
change ρ0 ρ1 only when the two contiguous subsequences of commands whose common
stores are ρ0 and ρ1 contain among them at least a put command.
6.1. Correctness of Trace Extraction
It turns out that the observational correctness of the hot path extraction transform in
Definition 5.1 can be proved w.r.t. the observational abstraction αsc of store changes.
T HEOREM 6.2 (C ORRECTNESS OF TRACE EXTRACTION ). For any P ∈ Program
and hp ∈ αN
hot (TraceP ), we have that αsc (TJextr hp (P )K) = αsc (TJP K).
This is the crucial result concerning the correctness of our hot path extraction transform. We will show in Section 10.5 (see Theorem 10.12) that the correctness of the hot
path extraction strategy defined in [Guo and Palsberg 2011] can be proved by a simple
adaptation of the proof technique that we will use here.
In order to prove Theorem 6.2, we need to define some suitable “dynamic” transformations of execution traces. Let us fix a hot path hp = ha0 , C0 , ..., an , Cn i ∈ αN
hot (TraceP )
(w.r.t. some store abstraction) and let Php , extr hp (P ) denote the corresponding transform of P given by Definition 5.1. We first define a mapping trout
hp of the execution traces
of the program P into execution traces of the transformed program Php that unfolds
the hot path hp (or any prefix of it) according to the hot path extraction strategy given
by Definition 5.1: a function application trout
hp (τ ) should replace any occurrence of the
hot path hp in the execution trace τ ∈ TraceP with its corresponding guarded and suitably relabeled path obtained through Definition 5.1. More precisely, Fig. 4 provides the
definitions for the following two functions:
trout
hp : TraceP → TracePhp
∗
trin
hp : TraceP → (StateP ∪ StatePhp )
Let us first describe how the trace transform trout
hp works. A function application
trout
(sσ)
on
a
trace
sσ
of
P
—the
superscript
out
hints that the first state s of the
hp
trace sσ is still outside of the hot path hp so that trout
hp (sσ) could either enter into the
transform of hp or remain outside of hp—triggers the unfolding of the hot path hp in
Php when the first state s is such that:
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:18
S. Dissegna et al.
(i) s = hρ, C0 i, where C0 is the first command of hp;
(ii) the entry conditional guard Ea0 of stitch P (hp) is satisfied in the store ρ of the state
s = hρ, C0 i, that is, αstore ({ρ}) ≤ a0 .
If the unfolding for the trace hρ, C0 iσ is actually started by applying trout
hp (hρ, C0 iσ)
then:
(iii) the first state hρ, C0 i is unfolded into the following sequence of two states of Php :
hρ, L0 : guard Ea0 → ℓ0 ihρ, ℓ0 : act(C0 ) → l1 i;
(iv) in turn, the unfolding of the residual trace σ is carried on by applying trin
hp (σ).
in
Let us now focus on the function trin
hp . A function application trhp (sσ)—here the superscript in suggests that we are currently inside the hot path hp so that trin
hp (sσ) could
either exit from the unfolding of hp or advance with the unfolding of hp—carries on the
unfolding of hp as a trace in Php when the current state s is such that:
(i) s = hρ, Ci i, where i ∈ [1, n − 1], meaning that the command Ci is strictly inside hp,
i.e., Ci is different from the first command C0 and the last command Cn of hp;
(ii) the guarded conditional guard Eai is satisfied in the store ρ of the state s = hρ, Ci i,
that is, αstore ({ρ}) ≤ ai .
If one of these two conditions does not hold then the trace transformation trin
hp (hρ, Ci iσ),
after a suitable unfolding step for hρ, Ci i, jumps back to the “outside of hp” modality by
progressing with trout
hp (σ).
Example 6.3. Consider the transform Php of Example 5.3 for the program P in
Example 2.1 w.r.t. the hot path hp = h⊤, C1 , ⊤, C2 , ⊤, C3c i. In particular, we refer to the
notation Hi , Hic used to denote the commands in the stitch of hp into P . Consider the
following trace fragment τ ∈ TraceP :
τ = h[x/3], C0 ih[x/0], C1 ih[x/0], C2 ih[x/1], C3c ih[x/1], C1 ih[x/1], C2 ih[x/2], C3c i
h[x/2], C1 ih[x/2], C2 ih[x/3], C3 ih[x/3], C4 i
Then, we have that the dynamic transformation trout
hp (τ ) acts as follows:
in
out
trout
hp (τ ) = h[x/3], C0 itrhp (τ1 ) = h[x/3], C0 ih[x/0], H0 ih[x/0], H1 itrhp (τ2 )
in
trin
hp (τ2 ) = h[x/0], H2 ih[x/0], H3 itrhp (τ3 )
in
trin
hp (τ3 ) = h[x/1], H4 ih[x/1], H5 itrhp (τ4 )
···
trin
hp (τ9 )
out
c
= trin
hp (h[x/3], C3 ih[x/3], C4 i) = h[x/3], H4 ih[x/3], H5 itrhp (h[x/3], C4 i)
= h[x/3], H4 ih[x/3], H5c ih[x/3], C4 itrout
hp (ǫ)
= h[x/3], H4 ih[x/3], H5c ih[x/3], C4 i
Summing up, using the colors in the flow graph of Php in Fig. 3 and representing traces
as sequences of commands only, we have that:
τ ≡ C0 → C1 → C2 → C3c → C1 → C2 → C3c → C1 → C2 → C3 → C4
trout
hp (τ ) ≡ C0 → H0 → H1 → H2 → H3 → H4 → H5 → H0 → H1 → H2 →
→ H3 → H4 → H5 → H0 → H1 → H2 → H3 → H4 → H5c → C4
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:19
hp = ha0 , C0 , ..., an , Cn i is a given hot path
trout
hp (ǫ) , ǫ
hρ, L0 : guard Ea0 → ℓ0 ihρ, ℓ0 : act(C0 ) → l1 i trin
hp (σ)
if s = hρ, C0 i, αstore ({ρ}) ≤ a0
hρ, L0 : ¬guard Ea0 → L0 ihρ, L0 : act(C0 ) → L1 i trout
hp (σ)
if
s
=
hρ,
C
i,
α
0
store ({ρ}) 6≤ a0
trout
hρ, L0 : guard Ea0 → ℓ0 ihρ, ℓ0 : ¬act(C0 ) → Lc1 i trout
hp (sσ) ,
hp (σ)
if
s
=
hρ,
cmpl(C
0 )i, αstore ({ρ}) ≤ a0
hρ, L : ¬guard E → L ihρ, L : ¬act(C ) → Lc i trout (σ)
0
a0
0
0
0
1
hp
if
s
=
hρ,
cmpl(C
)i, αstore ({ρ}) 6≤ a0
0
s · trout
otherwise
hp (σ)
trin
hp (ǫ) , ǫ
hρ, li : guard Eai → ℓi ihρ, ℓi : act(Ci ) → li+1 i trin
hp (σ)
if s = hρ, Ci i, i ∈ [1, n − 1], αstore ({ρ}) ≤ ai
hρ, ln : guard Ean → ℓn ihρ, ℓn : act(Cn ) → L0 i trout
hp (σ)
if
s
=
hρ,
C
i,
α
({ρ})
≤ an
n
store
hρ, li : ¬guard Eai → Li ihρ, Ci i trout
hp (σ)
in
if
s
=
hρ,
Ci i, i ∈ [1, n], αstore ({ρ}) 6≤ ai
trhp (sσ) ,
hρ, li : guard Eai → ℓi ihρ, ℓi : ¬act(Ci ) → Lcnext(i) i trout
hp (σ)
if
s
=
hρ,
cmpl(C
)i,
i
∈
[1,
n], αstore ({ρ}) ≤ ai
i
hρ, li : ¬guard Eai → Li ihρ, cmpl(Ci )i trout
hp (σ)
if
s
=
hρ,
cmpl(C
i )i, i ∈ [1, n], αstore ({ρ}) 6≤ ai
s · trout
otherwise
hp (σ)
in
Fig. 4. Definitions of trout
hp and trhp .
where red boxes denote commands of τ and trout
hp (τ ) outside of the hot path hp, black
boxes with red commands denote commands of τ inside hp, while black boxes with blue
out
commands denote commands of trout
hp (τ ) in stitch P (hp). Hence, trhp (τ ) carries out the
unfolding of the hot path hp for the execution trace τ of P , and therefore provides an
execution trace of the transformed program Php .
It turns out that trout
hp maps traces of P into traces of Php and does not alter store
change sequences.
out
L EMMA 6.4. trout
hp is well-defined and for any σ ∈ TraceP , sc(trhp (σ)) = sc(σ).
P ROOF. We first show that: (1) trout
hp is well-defined, i.e., for any σ ∈ TraceP ,
out
trhp (σ) ∈ TracePhp , and (2) for any σ ∈ TraceP , if cmd (σ0 ) 6∈ {C0 , cmpl(C0 )} then
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:20
S. Dissegna et al.
hp = ha0 , C0 , ..., an , Cn i is a given hot path
rtrhp (ǫ) , ǫ
rtrhp (sσ) ,
hstore(s), Ci i
rtrhp (σ)
hρ, C0 irtrhp (σ)
hρ, C c irtrhp (σ)
0
hρ, Ci irtrhp (σ)
hρ, Cic irtrhp (σ)
hρ, Cn irtrhp (σ)
s · rtrhp (σ)
if σ = ǫ, act(s) ∈ {guard Eai , ¬guard Eai }, i ∈ [1, n]
if σ 6= ǫ, act(s) ∈ {guard Eai , ¬guard Eai }, i ∈ [1, n]
if s = hρ, L0 : act(C0 ) → L1 i
if s = hρ, L0 : ¬act(C0 ) → Lc1 i
if s = hρ, ℓi : act(Ci ) → li+1 i, i ∈ [1, n − 1]
if s = hρ, ℓi : ¬act(Ci ) → Lcnext(i) i, i ∈ [1, n]
if s = hρ, ℓn : act(Cn ) → L0 i
otherwise
Fig. 5. Definition of rtrhp .
trin
hp (σ) ∈ TracePhp . In order to prove these two points, it is enough an easy induction
on the length of the execution trace σ and to observe that:
out
(i) for the first four clauses that define trout
hp (sσ) in Fig. 4 we have that trhp (sσ) =
out
′
′ ′′
′ ′′ in
s′ s′′ trout
hp (σ) or trhp (sσ) = s s trhp (σ), where s is a guard command of Php and s s
is in turn a legal sub-execution trace of Php ;
(ii) for the last clause that defines trout
hp (sσ) in Fig. 4 we have that cmd (s) 6∈
out
{C0 , cmpl(C0 )}, hence s is a legal state in Php and, in turn, trout
hp (sσ) = s · trhp (σ) is
a trace of Php ;
in
(iii) for the clauses 1, 2 and 4 that define trin
hp (sσ) in Fig. 4 we have that trhp (sσ) =
in
′ ′′ out
′
′′
s′ s′′ trin
hp (σ) or trhp (sσ) = s s trhp (σ), where s is a guard command and s is an
′ ′′
action command such that s s is a legal sub-execution trace of Php ;
in
(iv) for the clauses 3 and 5 that define trin
hp (sσ) in Fig. 4 we have that trhp (sσ) =
in
′
′
′
s strhp (σ) where s is a guard command and s s turns out to be a legal sub-execution
trace of Php ;
(v) for the last clause that defines trin
hp (sσ) in Fig. 4 we have that cmd (s) 6∈
{Ci , cmpl(Ci ) | i ∈ [1, n]}; by hypothesis, cmd (s) 6∈ {C0 , cmpl(C0 )}, so that cmd (s) 6∈
{Ci , cmpl(Ci ) | i ∈ [0, n]}, hence s is a legal state in Php and in turn trin
hp (sσ) =
s · trout
(σ)
is
a
trace
of
P
;
hp
hp
out
(vi) trin
hp (sσ) is never recursively called by a function application trhp (s0 sσ) when
cmd (s) ∈ {C0 , cmpl(C0 )}.
′ ′′
Then, it is immediate to check from the definitions in Fig. 4 that if trout
hp (sσ) = s s τ
′
′′
then store(s) = store(s ) = store(s ). Therefore, for any σ ∈ TraceP , we obtain that
sc(trout
hp (σ)) = sc(σ).
Vice versa, it is a simpler task to define a reverse transformation function rtrhp
that “decompiles” an execution trace σ of Php into an execution trace of P by removing
guarded commands in σ, as generated by the hot path hp, and by mapping the relabeled
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:21
commands of hp in σ back to their corresponding source commands of hp. This function
rtrhp : TracePhp → TraceP is correctly defined by the clauses in Fig. 5 and it preserves
store change sequences.
L EMMA 6.5. rtrhp is well-defined and for any σ ∈ TracePhp , sc(rtrhp (σ)) = rtrhp (σ).
P ROOF. We show that rtrhp is well-defined, i.e., for any σ ∈ TracePhp , rtrhp (σ) ∈
TraceP . This follows by an easy induction on the length of the execution trace σ by
observing that:
(i) the first clause that defines rtrhp (sσ) in Fig. 5 is an extremal base case where sσ = s
and the command action of s is a guard command guard Eai (or its complement);
in this case, we simply retain the store of s and pick the command Ci of P .
(ii) the clause 2 of rtrhp (sσ) in Fig. 5 simply removes the states whose commands are
some guard Eai ; since guard Eai does not alter stores, this removal preserves the
sequence of store changes.
(iii) the clauses 3-7 of rtrhp (sσ) in Fig. 5 map a state s of Php whose command Hi is
a relabeled action act(Ci ) or ¬act(Ci ) of the hot path hp to a corresponding state
of P that has the same store(s) and whose command is: Ci for act(Ci ) and Cic for
¬act(Ci ); here, we observe that since guards in σ are removed, by induction, these
definitions allow us to obtain that sσ is mapped to a legal trace of P that does not
alter the sequence of store changes.
(iv) the clause 8 of rtrhp (sσ) in Fig. 5 states that if s is already a state of P then it is
left unchanged.
Hence, the above points also show that the sequence of store changes is not affected by
rtrhp , i.e., for any σ ∈ TracePhp , sc(rtrhp (σ)) = sc(σ).
Example 6.6. We carry on Example 6.3 by considering the following trace fragment
σ ∈ TracePhp , where the transformed program Php is in Example 5.3:
σ = h[x/2], H4 ih[x/2], H5 ih[x/2], H0 ih[x/2], H1 ih[x/2], H2 ih[x/2], H3 ih[x/3], H4 i
h[x/3], H5c ih[x/3], C4 ih[x/6], C1 i
Here, the decompilation of σ back into an execution trace of P through rtrhp yields:
rtrhp (σ) = rtrhp (σ1 ) = h[x/2], C3c irtrhp (σ2 ) = h[x/2], C3c irtrhp (σ3 )
= h[x/2], C3c ih[x/2], C1 irtrhp (σ4 ) = h[x/2], C3c ih[x/2], C1 irtrhp (σ5 )
= h[x/2], C3c ih[x/2], C1 ih[x/2], C2 irtrhp (σ6 )
= h[x/2], C3c ih[x/2], C1 ih[x/2], C2 irtrhp (σ7 )
= h[x/2], C3c ih[x/2], C1 ih[x/2], C2 ih[x/3], C3 irtrhp (σ8 )
= h[x/2], C3c ih[x/2], C1 ih[x/2], C2 ih[x/3], C3 ih[x/3], C4 irtrhp (σ9 )
= h[x/2], C3c ih[x/2], C1 ih[x/2], C2 ih[x/3], C3 irtrhp (σ8 )
= h[x/2], C3c ih[x/2], C1 ih[x/2], C2 ih[x/3], C3 ih[x/3], C4 ih[x/6], C1 i
Indeed, h[x/2], C3c ih[x/2], C1 ih[x/2], C2 ih[x/3], C3 ih[x/3], C4 ih[x/6], C1 i is a well-defined
execution trace of P .
We are now in the position to prove Theorem 6.2.
P ROOF OF T HEOREM 6.2. With an abuse of notation for rtrhp , let us define two
functions trhp : ℘(TraceP ) → ℘(TracePhp ) and rtrhp : ℘(TracePhp ) → ℘(TraceP ) which
out
are the collecting versions of trout
hp and rtrhp , that is, trhp (T ) , {trhp (σ) | σ ∈ T } and
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:22
S. Dissegna et al.
rtrhp (T ) , {rtrhp (σ) | σ ∈ T }. As consequences of the above lemmata, we have the
following properties.
(A)
(B)
(C)
(D)
αsc ◦ trhp = αsc : by Lemma 6.4.
trhp (TJP K) ⊆ TJPhp K: because, by Lemma 6.4, trout
hp is well-defined.
αsc ◦ rtrhp = αsc : by Lemma 6.5.
rtrhp (TJPhp K) ⊆ TJP K: because, by Lemma 6.5, rtrhp is well-defined.
We therefore obtain:
αsc (TJP K) =
αsc (trhp (TJP K)) ⊆
αsc (TJPhp K) =
αsc (rtrhp (TJPhp K)) ⊆
αsc (TJP K)
[By point (A)]
[By point (B)]
[By point (C)]
[By point (D)]
and this closes the proof.
6.2. Correctness of Hot Path Optimizations
Guarded hot paths are a key feature of our tracing compilation model and are meant
to be dynamically recorded by a hot path monitor. An abstract guard for a command
C of some stitched hot path stitch P (hp) encodes a property of program stores which is
represented as an element of an abstract domain Store♯ and is guaranteed to hold at
the entry of C. This information on program stores, as encapsulated by the abstract
guards in stitch P (hp), can then be used in hot path optimizations, namely, to optimize
the commands in hp.
We follow a modular approach for proving the correctness of hot path optimizations.
A hot path optimization O should optimize P along some hot path hp of P , by relying on
the abstract store information recorded in hp, while leaving unchanged the commands
outside of hp. Hence, in our framework, fixed P ∈ Program, an optimization O is defined
to be a program transform of the commands in stitch P (hp), that is,
O : {stitch P (hp) | hp ∈ αN
hot (TraceP )} → Program
where Program may allow new optimized expressions and/or actions introduced by O,
as it will be the case of type-specific additions +Type in the type specialization optimization described in Section 7. Let P¬hp , extr hp (P ) r stitch P (hp) denote the commands
outside of the stitched hot path. Then, the corresponding full optimization Ofull of the
whole program P w.r.t. the hot path hp should extract and simultaneously optimize hp,
namely, this is defined by
Ofull (P, hp) , P¬hp ∪ O(stitch P (hp))
where Ofull (P, hp) is required to be a well-formed program, i.e., Ofull (P, hp) ∈ Program.
This full optimization Ofull (P, hp) has to be proved correct w.r.t. some observational abstraction αo : ℘(TraceP ) → A of program traces, which is assumed to be more abstract
than the store changes abstraction αsc (cf. Section 6). Then, this full optimization is
correct for αo when:
αo (TJOfull (P, hp)K) = αo (TJP K).
Since Theorem 6.2 ensures that the unoptimized trace extraction transform is already
correct for the store changes abstraction αsc , which is more precise than αo , the intuition is that in order to prove the correctness of Ofull w.r.t. αo , it is enough to focus
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:23
on the correctness of the optimization O along the stitched hot path stitch P (hp). This
therefore leads to the following definition of correctness for a hot path optimization.
Definition 6.7 (Correctness of hot path optimization). O is correct for the observational abstraction αo if for any P ∈ Program and for any hp ∈ αN
hot (TraceP ),
αo (TJO(stitch P (hp))K) = αo (TJstitch P (hp)K).
In order to prove that this correctness of a hot path optimization implies the correctness of the corresponding full optimization, we define two functions
to : Tracestitch P (hp) → TraceO(stitch P (hp))
tdo : TraceO(stitch P (hp)) → Tracestitch P (hp)
which must be well-defined, i.e. they have to map well-formed traces into well-formed
traces, and, intuitively, encode the effect of optimizing (function to) and de-optimizing
(function tdo) execution traces along a stitched hot path. Since Tracestitch P (hp) ⊆
Traceextr hp (P ) and TraceO(stitch P (hp)) ⊆ TraceOfull (P,hp) , we then extend to and tdo to two
functions
to full : Traceextr hp (P ) → TraceOfull (P,hp)
tdo full : TraceOfull (P,hp) → Traceextr hp (P )
which simply apply to and tdo to maximal subtraces, respectively, in Tracestitch P (hp)
and TraceO(stitch P (hp)) , while leaving unchanged the remaining states. Let us formalize this idea. If σ ∈ Traceextr hp (P ) is nonempty and, for some k ∈ [0, |σ|), cmd (σk ) ∈
stitch P (hp) then σ[k,nst ] denotes the maximal subtrace of σ beginning at σk which
belongs to Tracestitch P (hp) , that is, the index nst ≥ k is such that: (1) cmd (σnst ) ∈
stitch P (hp), (2) if nst < |σ| − 1 then cmd (σnst +1 ) 6∈ stitch P (hp), (3) for any j ∈ [k, nst ],
cmd (σj ) ∈ stitch P (hp). Analogously, if τ ∈ TraceOfull (P,hp) is nonempty and cmd (τk ) ∈
O(stitch P (hp)) then τ[k,nst ] denotes the maximal subtrace of τ beginning at τk which
belongs to TraceO(stitch P (hp)) . Then, the formal definition of to full goes as follows:
if σ = ǫ
ǫ
if σ 6= ǫ, cmd (σ0 ) 6∈ stitch P (hp)
to full (σ) , σ0 to full (σ1 )
to(σ
[0,nst ] ) to full (σ(nst +1) ) if σ 6= ǫ, cmd (σ0 ) ∈ stitch P (hp)
and analogously for tdo full . Since to and tdo are supposed to be well-defined, it turns
out that to full and tdo full are well-defined once we make the weak and reasonable assumption that to and tdo do not modify the entry (which is always L0 ) and exit labels
of the stitched hot path. This assumption, e.g., for to can be formalized as follows:
if σ ∈ Tracestitch P (hp) and to(σ) = τ then (i) if lbl(σ0 ) = L0 then lbl(τ0 ) = L0 ; (ii) if
suc(σ|σ|−1 ) = L′ 6∈ labels(P ) then suc(τ|τ |−1) = L′ . In the following, to full and tdo full are
also used to denote their corresponding collecting functions defined on sets of traces.
L EMMA 6.8. Assume that αo ◦ to full = αo = αo ◦ tdo full . If O is correct for αo then
Ofull is correct for αo .
P ROOF. We have that:
αo (TJOfull (P, hp)K) =
αo (tdo full (TJOfull (P, hp)K)) ⊆
αo (TJextr hp (P )K) =
αo (to full (TJextr hp (P )K)) ⊆
αo (TJOfull (P, hp)K)
[By αo ◦ tdo full = αo ]
[Since tdo full is well-defined]
[By αo ◦ to full = αo ]
[Since to full is well-defined]
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:24
S. Dissegna et al.
Thus, αo (TJOfull (P, hp)K) = αo (TJextr hp (P )K). By Theorem 6.2, αsc (TJextr hp (P )K) =
αsc (TJP K), so that, since αsc is more precise than αo , αo (TJextr hp (P )K) = αo (TJP K),
and, in turn, αo (TJOfull (P, hp)K) = αo (TJP K).
We will see in Sections 7 and 8 two significant examples of hot path optimizations,
namely, type specialization and constant folding.
7. TYPE SPECIALIZATION
One key optimization for dynamic languages like JavaScript and PHP is type specialization, that is, the use of type-specific primitives in place of generic untyped operations whose runtime execution can be costly. As a paradigmatic example, a generic
addition operation could be defined on more than one type, so that the runtime environment must check the type of its operands and execute a different operation depending
on these types: this is the case of the addition operation in JavaScript (see its runtime
semantics in the ECMA-262 standard [Ecma International 2015, Section 12.7.3.1])
and of the semantics of + in our language as given in Section 2.3. Of course, type
specialization avoids the overhead of dynamic type checking and dispatch of generic
untyped operations. When a type is associated to each variable before the execution
of a command in some hot path, this type environment can be used to replace generic
operations with type-specific primitives. In this section, we show that type specialization can be viewed as a particular hot path optimization which can be proved correct
according to our definition in Section 6.2.
7.1. Type Abstraction
Let us recall that the set of type names is Types = {⊤T , Int, String, Undef, ⊥T }, which
can be viewed as the following finite lattice hTypes, ≤t i:
⊤T
Int
String
Undef
⊥T
The abstraction αtype : ℘(Valueu ) → Types and concretization γtype
℘(Valueu ) functions are defined as follows:
⊥T
if S = ∅
∅
if T
if ∅ 6= S ⊆ Z
if T
Int
Z
αtype (S) , String if ∅ 6= S ⊆ Char∗
γtype (T ) , Char∗
if T
Undef
if
∅
=
6
S
=
{undef}
{undef}
if
T
⊤T
otherwise
Valueu
if T
: Types →
= ⊥T
= Int
= String
= Undef
= ⊤T
Thus, αtype (S) provides the smallest type in hTypes, ≤t i for a set S of values. In particular, given v ∈ Valueu , αtype ({v}) coincides with type(v). Following the approach
described in Section 3.2.1, we then consider a simple nonrelational store abstraction
for types
Storet , hVar → Types, ≤˙ t i
where ≤˙ t is the standard pointwise lifting of ≤t , so that λx.⊥T and λx.⊤T are, respectively, the bottom and top abstract stores in Storet . The abstraction and concretization
maps αstore : ℘(Store) → Storet and γstore : Storet → ℘(Store) are defined as a straight
instantiation of the definitions in Section 3.2.1.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:25
The abstract type semantics Et : Exp → Storet → Types of expressions is defined
as the best correct approximation of the concrete collecting semantics E : Exp →
℘(Store) → ℘(Value) on the type abstractions Storet and Types, i.e.,
Et JEKρt , αtype (EJEKγstore (ρt )).
Hence, this definition leads to the following equalities:
Et JvKρt = type(v)
Et JxKρt = ρt (x)
⊥T
t
E
JE1 Kρt
Et JE1 + E2 Kρt =
Undef
⊤T
if ∃i. Et JEi Kρt = ⊥T
else if Et JE1 Kρt = Et JE2 Kρt ∈ {Int, String}
else if ∀i. Et JEi Kρt < ⊤T
otherwise
For instance, we have that:
Et Jx + yK[x/ String, y/⊥T ] = αtype (EJx + yK∅) = αtype (∅) = ⊥T
Et Jx + yK[x/ String, y/ String] = αtype (EJx + yK{ρ | ρ(x), ρ(y) ∈ Char∗ }) =
= αtype (Char∗ ) = String,
Et Jx + yK[x/ Int, y/ String] = αtype (EJx + yK{ρ | ρ(x) ∈ Z, ρ(y) ∈ Char∗ }) =
αtype ({undef}) = Undef,
Et Jx + yK[x/ Int, y/⊤T ] = αtype (EJx + yK{ρ | ρ(x) ∈ Z, ρ(y) ∈ Valueu }) =
αtype (Z ∪ {undef}) = ⊤T
Being defined as best correct approximation, it turns out that the abstract type semantics Et of expressions is correct by definition.
C OROLLARY 7.1. If ρ ∈ γstore (ρt ) then EJEKρ ∈ Et JEKρt .
According to Section 5, for any abstract type store (that we also call type environment) [xi /Ti | xi ∈ Var] ∈ Storet we consider a corresponding Boolean action guard
denoted by
guard x0 : T0 , . . . , xn : Tn ∈ BExp
whose corresponding action semantics is automatically induced, as defined in Section 5, by the Galois connection (αstore , ℘(Store), Storet , γstore ): for any ρ ∈ Store,
ρ if ρ ∈ γstore ([xi /Ti | xi ∈ Var])
AJguard x0 : T0 , ..., xn : Tn Kρ ,
⊥ otherwise
ρ if ∀i. ρ(xi ) ∈ γtype (Ti )
=
⊥ ∃i. ρ(xi ) 6∈ γtype (Ti )
For example, we have that:
AJguard x : String, y
AJguard x : String, y
AJguard x : String, y
AJguard x : String, y
: StringK[x/foo, y/bar] = [x/foo, y/bar],
: ⊤T K[x/foo, y/3] = [x/foo, y/3],
: ⊤T K[x/1, y/3] = ⊥,
: UndefK[x/foo] = [x/foo].
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:26
S. Dissegna et al.
7.2. Type Specialization of Hot Paths
Let us consider some hot path hp = hρt0 , C0 , . . . , ρtn , Cn i ∈ αN
hot (TraceP ) on the type abstraction hStoretP , ≤˙ t i, where each ρti is therefore a type environment for P . Thus, in
the transformed program extr hp (P ), the stitched hot path stitch P (hp) contains n + 1
typed guards, that, for any i ∈ [0, n], we simply denote as guard ρti . Typed guards allow
us to perform type specialization of commands in the stitched hot path. In order to
keep the notation simple, we only focus on type specialization of addition operations
occurring in assignments, while one could also consider an analogous type specialization of Boolean comparisons in conditional commands. This is defined as a program
transform that instantiates most type-specific addition operations in place of generic
untyped additions by exploiting the type information dynamically recorded by typed
guards in stitch P (hp). Note that if C ∈ stitch P (hp) and act(C) ≡ x := E1 + E2 then
C ≡ ℓi : x := E1 + E2 → L′ , for some i ∈ [0, n], where L′ ∈ {li+1 , L0 }. Let Ct denote the
extended set of commands which includes type specific additions +Int and +String and,
in turn, let Programt denote the possibly type-specialized programs with commands
ranging in Ct . The semantic function E for expressions is then updated to type specific
additions as follows:
EJE1 Kρ +Z EJE2 Kρ if type(EJEi Kρ) = Int
EJE1 +Int E2 Kρ ,
undef
otherwise
EJE1 Kρ · EJE2 Kρ if type(EJEi Kρ) = String
EJE1 +String E2 Kρ ,
undef
otherwise
Given a hot path hp = hρt0 , C0 , . . . , ρtn , Cn i, the type specialization function
tshp : stitch P (hp) → Ct is defined as follows:
′
if Et JE1 + E2 Kρti = Int
ℓi : x := E1 +Int E2 → L
′
′
tshp (ℓi : x := E1 + E2 → L ) , ℓi : x := E1 +String E2 → L if Et JE1 + E2 Kρti = String
ℓi : x := E1 + E2 → L′
otherwise
tshp (C) , C
if C 6≡ ℓi : x := E1 + E2 → L′
Hence, if a typed guard guard ρti preceding a command ℓi : x := E1 + E2 → L′ allows
us to derive abstractly on Storet that E1 and E2 have the same type (Int or String) then
the addition E1 + E2 is accordingly type specialized. This function allows us to define
the hot path type specialization optimization
t
Ots : {stitch P (hp) | hp ∈ αN
hot (TraceP )} → Program
simply by
Ots (stitch P (hp)) , {tshp (C) | C ∈ stitch P (hp)}.
In turn, as described in Section 6.2, this induces the full type specialization optimization
ts
Ofull
(P, hp) , extr hp (P ) r stitch P (hp) ∪ Ots (stitch P (hp)).
ts
(P, hp) is also called typed trace extraction since it extracts and simultaneously
Ofull
type specializes a typed hot path hp in a program P . The correctness of this program
optimization can be proved for the store changes observational abstraction by relying
on Lemma 6.8.
T HEOREM 7.2 (C ORRECTNESS OF TYPED TRACE EXTRACTION ). For any typed
ts
hot path hp ∈ αN
hot (TraceP ), we have that αsc (TJOfull (P, hp)K) = αsc (TJP K).
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:27
P ROOF. Let td : TraceOts (stitch P (hp)) → Tracestitch P (hp) be the following type despecialization function, where Type is either Int or String:
td(ǫ) , ǫ
′
hρ, ℓi : x := E1 + E2 → L i
td(sσ) , hρ, ℓi : x := E1 + E2 → L′ i · td(σ)
s · td(σ)
if s = hρ, ℓi : x := E1 +Type E2 → L′ i,
type(EJE1 + E2 Kρ) 6= Type
if s = hρ, ℓi : x := E1 +Type E2 → L′ i,
type(EJE1 + E2 Kρ) = Type
otherwise
Let us explain the first defining clause of td(sσ), i.e., s = hρ, ℓi : x := E1 +Type E2 → L′ i
and type(EJE1 + E2 Kρ) 6= Type. These conditions can never hold in an inductive call
of the function td: in fact, when td(sσ) is recursively called by td(s′ sσ), we necessarily
have that s′ = hρ, li : guard ρti → ℓi i, so that ρ ∈ γstore (ρti ), and, in turn, by Corollary 7.1,
EJE1 + E2 Kρ ∈ Et JE1 + E2 Kρt , which implies type(EJE1 + E2 Kρ) = Type, which is a
contradiction. Thus, the first defining clause of td(sσ) only applies to type specialized
traces in TraceOts (stitch P (hp)) whose first state is s = hρ, ℓi : x := E1 +Type E2 → L′ i: in
this case, we necessarily have that σ = ǫ, because AJE1 +Type E2 Kρ = undef so that
Ss = ∅. This clarifies the definition of td in this particular case. Also, observe that
in this case, sc(td(s)) = sc(s) trivially holds. In all the remaining cases, it is clear
that td maps type specialized traces into legal unspecialized traces of stitch P (hp) since
labels are left unchanged. Moreover, sc ◦ td = sc holds, in particular because in the
second defining clause of td(sσ), the condition type(EJE1 + E2 Kρ) = Type guarantees
that EJE1 + E2 Kρ = EJE1 +Type E2 Kρ.
On the other hand, we define a trace specialization function sp : Tracestitch P (hp) →
TraceOts (stitch P (hp)) as follows:
sp(ǫ) , ǫ
sp(hµ0 , H0 i· · ·hµk , Hk i) ,
hµ0 , tshp (H0 )i
if tshp (H0 ) ≡ ℓi : x := E1 +Type E2 → L′ ,
µ0 6∈ γstore (ρti )
hµ0 , tshp (H0 )i· · ·hµk , tshp (Hk )i
otherwise
Let us comment on this definition. If σ ∈ Tracestitch P (hp) and σ 6= ǫ then it may
happen that the first state hµ0 , H0 i of σ is such that the command H0 is ℓi : x :=
E1 + E2 → L′ and, since Et JE1 + E2 Kρti = Type (Int or String), H0 is type specialized to tshp (H0 ) ≡ ℓi : x := E1 +Type E2 → L′ , while the store µ0 is not approximated by the abstract store ρti , i.e., µ0 6∈ γstore (ρti ). Thus, in this case, the trace in
Ots (stitch P (hp)) beginning at hµ0 , tshp (H0 )i is stuck, because the concrete semantics
of addition is EJE1 +Type E2 Kµ0 = undef, and in turn AJx := E1 +Type E2 Kµ0 = ⊥,
so that we necessarily have to define sp(σ) = hµ0 , tshp (H0 )i. Otherwise, sp(σ) simply type specializes through tshp all the commands (actually, addition expressions)
occurring in σ. Here, it turns out that sp is well-defined, i.e. sp(σ) is a legal trace of
Ots (stitch P (hp)), because any state hρ, ℓi : x := E1 + E2 → L′ i of σ is always preceded by the state hρ, li : guard ρti → ℓi i and ρ ∈ γstore (ρti ) must hold. Thus, by
Corollary 7.1, EJE1 + E2 Kρ ∈ Et JE1 + E2 Kρt = Type, so that AJx := E1 +Type E2 Kρ =
AJx := E1 + E2 Kρ holds. Consequently, the trace fragment
sp(hρ, li : guard ρti → ℓi ihρ, ℓi : x := E1 + E2 → L′ i) =
hρ, li : guard ρti → ℓi ihρ, ℓi : x := E1 +Type E2 → L′ i
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:28
S. Dissegna et al.
is legal in Ots (stitch P (hp)). Furthermore, let us also observe that sc ◦ ts = sc trivially
holds.
Thus, following the scheme in Section 6.2, these two functions td and ts allow us to
ts (P,hp) → Traceextr
ts (P,hp)
define tdfull : TraceOfull
and tsfull : Traceextr hp (P ) → TraceOfull
hp (P )
such that αsc ◦ tdfull = αsc = αsc ◦ tsfull , so that the thesis follows by Lemma 6.8.
Example 7.3. Let us consider the following sieve of Eratosthenes in a Javascriptlike language—this is taken from the running example in [Gal et al. 2009]—where
primes is an array initialized with 100 true values:
for (var i = 2; i < 100; i = i + 1) do
if (!primes[i]) then continue
for (var k = i + i; k < 100; k = k + i) do primes[k] = false
With a slight abuse, we assume that our language is extended with arrays and Boolean
values ranging in the type Bool. The semantics of read and store for arrays is standard:
first, the index expression is checked to be in bounds, then the value is read or stored
into the array. If the index is out of bounds then the corresponding action command
gives ⊥, that is, we assume that the program generates an error (e.g., it is aborted).
The above program is encoded in our language as follows:
P = C0 ≡ L0 : i := 2 → L1 , C1 ≡ L1 : i < 100 → L2 , C1c ≡ L1 : ¬(i < 100) → L8 ,
C2 ≡ L2 : primes[i] = tt → L3 , C2c ≡ L2 : ¬(primes[i] = ff ) → L7 ,
C3 ≡ L3 : k := i + i → L4 , C4 ≡ L4 : k < 100 → L5 , C4c ≡ L4 : ¬(k < 100) → L7 ,
C5 ≡ L5 : primes[k] := ff → L6 , C6 ≡ L6 : k := k + i → L4 ,
C7 ≡ L7 : i := i + 1 → L1 , C8 ≡ L8 : skip → Ł .
Let us consider the following type environment
ρt , {primes[n]/ Bool, i/ Int, k/ Int} ∈ Storet
where primes[n]/ Bool is a shorthand for primes[0]/ Bool, . . . , primes[99]/ Bool. Then the
first traced 2-hot path on the type abstraction Storet is hp1 , hρt , C4 , ρt , C5 , ρt , C6 i. As
a consequence, the typed trace extraction of hp1 yields:
ts
P1 , Ofull
(P, hp1 )
= P r {C4 , C4c } ∪ {L4 : k < 100 → L5 , L4 : ¬(k < 100) → L7 } ∪ Ots (stitch P (hp1 ))
where:
Ots (stitch P (hp1 )) = H0 ≡ L4 : guard (primes[n] : Bool, i : Int, k : Int) → ℓ0 ,
H0c ≡ L4 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L4 ,
H1 ≡ ℓ0 : k < 100 → l1 , H1c ≡ ℓ0 : ¬(k < 100) → L7 ,
H2 ≡ l1 : guard (primes[n] : Bool, i : Int, k : Int) → ℓ1 ,
H2c ≡ l1 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L5 ,
H3 ≡ ℓ1 : primes[k] := ff → l2 ,
H4 ≡ l2 : guard (primes[n] : Bool, i : Int, k : Int) → ℓ2 ,
H4c ≡ l2 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L6 ,
H5 ≡ ℓ2 : k := k +Int i → L4 .
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:29
8. CONSTANT VARIABLE FOLDING
Constant variable folding, a.k.a. constant propagation [Wegman and Zadeck 1991], is
a standard and well-known program optimization, whose goal is to detect which program variables at some program point are constant on all possible executions and
then to propagate these constant values as far forward through the program as possible. Guo and Palsberg [2011] show how to define this optimization along hot paths and
then prove its correctness. As a significant example, we show here how to specify and
prove the correctness w.r.t. the store changes abstraction αsc of this simple hot path
optimization according to the approach defined in Section 6.2.
The constant propagation store abstraction CPst and its corresponding GI
(αCP , ℘(Store), CPst , γCP ) have been defined in Example 3.1. Following Section 5, any
abstract store [xi /ai | xi ∈ Var] ∈ CPst , where, as usual, the bindings xi /undef are
omitted, defines a corresponding guard x0 : a0 , . . . , xn : an ∈ BExp whose semantics is
induced by the GI (αCP , ℘(Store), CPst , γCP ), as defined in Section 5: for any ρ ∈ Store,
ρ if ρ ∈ γCP ([xi /ai | xi ∈ Var])
AJguard x0 : a0 , ..., xn : an Kρ ,
⊥ otherwise
ρ if ∀i. ρ(xi ) ∈ γcp (ai )
=
⊥ ∃i. ρ(xi ) 6∈ γcp (ai )
Therefore, we have that:
AJguard x : 2, y
AJguard x : 2, y
AJguard x : 2, y
AJguard x : 2, y
AJguard x : 2, y
AJguard x : 2, y
: fooK[x/2, y/3] = ⊥,
: fooK[x/2, y/foo, z/4] = ⊥,
: fooK[x/2] = ⊥,
: fooK[x/2, y/foo] = [x/2, y/foo],
: ⊤K[x/2, y/foo] = [x/2, y/foo],
: ⊤K[x/2] = [x/2].
Let us consider some hot path hp = hρc0 , C0 , . . . , ρcn , Cn i ∈ αN
hot (TraceP ) on the constant propagation abstraction CPst , where each ρci is therefore an abstract store in
CPst , whose corresponding guard in stitch P (hp) will be denoted by guard ρci . The
constant value information encoded in these guards is used to define the variable
folding in the stitched hot path. Following Guo and Palsberg [2011, Section 2.4], let
FV : ℘(C) → ℘(Var) denote the function that returns the “free” variables occurring
in some set of commands (in particular, a well-defined program), i.e., FV(P ) is the set
of variables occurring in P which are never-assigned-to in some command of P . As
in Guo and Palsberg [2011], constant variable folding is restricted to expressions E of
some assignment x := E and is defined as a program transform which exploits the
constant information recorded by abstract guards in stitch P (hp). The constant folding
function cfhp : stitch P (hp) → C is defined as follows:
cfhp (ℓi : x := E → L′ ) ,
′
ℓi : x := E[y1 /vy1 , ..., yk /vyk ] → L if {y1 , ..., yk } = {y ∈ vars(E) ∩ FV(stitch P (hp)) |
6 ∅
ρci (y) = vy ∈ Value} =
′
ℓi : x := E → L
otherwise
cfhp (C) , C
if C 6≡ ℓi : x := E → L′
where E[y1 /vy1 , ..., yk /vyk ] denotes the standard synctatic substitution of variables
yj ∈ vars(E) with constant values ρci (yj ) = vyj ∈ Value. Hence, when the abstract
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:30
S. Dissegna et al.
guard guard ρci which precedes an assignment ℓi : x := E → L′ tells us that a free variable y occuring in the expression E is definitely a constant value vy ∈ Value then cfhp
performs the corresponding variable folding in E. Thus, the hot path constant folding
optimization is defined by
Ocf (stitch P (hp)) , {cfhp (C) | C ∈ stitch P (hp)}
cf
and, in turn, this induces the full constant folding optimization Ofull
(P, hp). The correctness of this constant folding optimization can be proved for the store changes observational abstraction
T HEOREM 8.1 (C ORRECTNESS OF CONSTANT FOLDING OPTIMIZATION ).
For
any hot path hp ∈ αN
hot (TraceP ) w.r.t. the constant propagation store abstraction CPst ,
cf
αsc (TJOfull
(P, hp)K) = αsc (TJP K).
This proof is omitted, since it follows the same pattern of Theorem 7.2 for the correctness of typed trace extraction, in particular it relies on Lemma 6.8.
Example 8.2. Let us consider the following program written in a while-language:
x := 0; a := 2;
while (x ≤ 15) do
if (x ≤ 5) then x := x + a
else {a := a + 1; x := x + a; }
whose translation as P ∈ Program goes as follows:
P = C0 ≡ L0 : x := 0 → L1 , C1 ≡ L1 : a := 2 → L2
C2 ≡ L2 : x ≤ 15 → L3 , C2c ≡ L2 : ¬(x ≤ 15) → L7 ,
C3 ≡ L3 : x ≤ 5 → L4 , C3c ≡ L3 : ¬(x ≤ 5) → L5 ,
C4 ≡ L4 : x := x + a → L2 , C5 ≡ L5 : a := a + 1 → L6
C6 ≡ L6 : x := x + a → L2 , C7 ≡ L7 : skip → Ł
The first traced 2-hot path for the abstraction CPst is:
hp = h[x/⊤, a/2], C2 , [x/⊤, a/2], C3 , [x/⊤, a/2], C4 i.
In fact, the initial prefix of the complete trace of P which corresponds to the terminating run of P is as follows:
h[ ], C0 ih[x/0], C1 ih[x/0, a/2], C2 ih[x/0, a/2], C3 ih[x/0, a/2], C4 ih[x/2, a/2], C2 i
h[x/2, a/2], C3 ih[x/2, a/2], C4 ih[x/4, a/2], C2 ih[x/4, a/2], C3 ih[x/4, a/2], C4 i
so that hp ∈ α2hot (TraceP ). Hence, the constant folding optimization Ocf along hp provides:
cf
Ofull
(P, hp) = P r{C2 , C2c } ∪ {L2 : x ≤ 15 → L3 , L2 : ¬(x ≤ 15) → L7 } ∪ Ocf (stitch P (hp))
where:
Ocf (stitch P (hp)) = H0 ≡ L2 : guard [x : ⊤, a : 2] → ℓ0 , H0c ≡ L2 : ¬guard [x : ⊤, a : 2] → L2 ,
H1
H2
H3
H4
≡ ℓ0 : x ≤ 15 → l1 , H1c ≡ ℓ0 : ¬(x ≤ 15) → L7 ,
≡ l1 : guard [x : ⊤, a : 2] → ℓ1 , H2c ≡ l1 : ¬guard [x : ⊤, a : 2] → L3 ,
≡ ℓ1 : x ≤ 5 → l2 , H3c ≡ ℓ1 : ¬(x ≤ 5) → L5 ,
≡ l2 : guard [x : ⊤, a : 2] → ℓ2 , H4c ≡ l2 : ¬guard [x : ⊤, a : 2] → L4 ,
H5 ≡ ℓ2 : x := x + 2 → L2 .
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:31
Therefore, this hot path optimization allows us to fold the constant value 2 for the
variable a, in the hot path command H5 ≡ ℓ2 : x := x + 2 → L2 .
9. NESTED HOT PATHS
Once a first hot path hp1 has been extracted by transforming P to P1 , extr hp1 (P ),
it may well happen that a new hot path hp2 in P1 contains hp1 as a nested sub-path.
Following TraceMonkey’s trace recording strategy [Gal et al. 2009], we attempt to nest
an inner hot path inside the current trace: during trace recording, an inner hot path
is called as a kind of “subroutine”, this executes a loop to a successful completion and
then returns to the trace recorder that may therefore register the inner hot path as
part of a new hot path.
In order to handle nested hot paths, we need a more general definition of hot path
which takes into account previously extracted hot paths and a corresponding program
transform for extracting nested hot paths. Let P be the original program and let P ′ be
a hot path transform of P so that P ′ r P contains all the commands (guards included)
in the hot path. We define a function hotcut : TraceP ′ → (StateP ′ )∗ that cuts from an
execution trace σ of P ′ all the states whose commands appear in some previous hot
path hp except for the entry and exit states of hp:
if σ = ǫ
ǫ
hotcut(σ) , hotcut (hρ1 , C1 ihρ3 , C3 iσ ′ ) if σ = hρ1 , C1 ihρ2 , C2 ihρ3 , C3 iσ ′ & C1 , C2 , C3 6∈ P
otherwise
σ0 hotcut (σ1 )
In turn, we define outerhot N : TraceP ′ → ℘((State♯P ′ )∗ ) as follows:
outerhot N (σ) , {hai , Ci i · · · haj , Cj i ∈ (State♯P ′ )∗ | ∃hρi , Ci i · · · hρj , Cj i ∈ loop(hotcut(σ))
such that i ≤ j, αstore (hρi , Ci i · · · hρj , Cj i) = hai , Ci i · · · haj , Cj i,
count(αstore (hotcut (σ)), hai , Ci i · · · haj , Cj i) ≥ N }.
Clearly, when P ′ = P it turns out that hotcut = λσ.σ so that outerhot N = hot N . We
define the usual collecting version of outerhot N on ℘(TraceP ′ ) as the abstraction map
N
N
′
αN
outerhot , λT.∪σ∈T outerhot (σ). Then, αouterhot (TJP K) provides the set of N -hot paths
′
in P .
Example 9.1. Let us consider again Example 5.3, where Store♯ is the trivial one-point store abstraction {⊤}. In Example 5.3, we first extracted hp1 =
h⊤, C1 , ⊤, C2 , ⊤, C3c i by transforming P to P1 , extr hp (P ). We then consider the following trace in TJP1 K:
σ =h[ ], C0 ih[x/0], H0 ih[x/0], H1 ih[x/0], H2 ih[x/0], H3 ih[x/1], H4 ih[x/1], H5 i · · · h[x/2], H3 i
h[x/3], H4 ih[x/3], H5c ih[x/3], C4 ih[x/6], H0 i · · · h[x/9], H5c ih[x/9], C4 ih[x/12], H0 i · · ·
Thus, here we have that
hotcut (σ) = h[ ], C0 ih[x/0], H0 ih[x/3], H5c ih[x/3], C4 ih[x/6], H0 ih[x/9], H5c ih[x/9], C4 i · · ·
so that hp2 = h⊤, H0 , ⊤, H5c , ⊤, C4 i ∈ α2outerhot (TJP1 K). Hence, hp2 contains a nested hot
path, which is called at the beginning of hp2 and whose entry and exit commands are,
respectively, H0 and H5c .
′
′
Let hp = ha0 , C0 , . . . , an , Cn i ∈ αN
outerhot (TJP K) be a N -hot path in P , where, for all
i ∈ [0, n], we assume that Ci ≡ Li : Ai → Lnext(i) . Let us note that:
– If for all i ∈ [0, n], Ci ∈ P then hp actually is a hot path in P , i.e., hp ∈ αN
hot (TJP K).
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:32
S. Dissegna et al.
– Otherwise, there exists some Ck 6∈ P . If Ci ∈ P and Ci+1 6∈ P then Ci+1 is the entry
command of some inner hot path; on the other hand, if Ci 6∈ P and Ci+1 ∈ P then Ci
is the exit command of some inner hot path.
The transform of P ′ for extracting hp is then given as the following generalization of
Definition 5.1.
Definition 9.2 (Nested trace extraction transform). The nested trace extraction
transform of P ′ for the hot path hp = ha0 , C0 , . . . , an , Cn i is:
extr hp (P ′ ) , P
(1)
r ({C0 | C0 ∈ P } ∪ {cmpl(C0 ) | cmpl(C0 ) ∈ P })
(2)
∪ {H0 : act(C0 ) → L1 | C0 ∈ P } ∪ {H0 : ¬act(C0 ) → Lc1 | cmpl(C0 ) ∈ P }
(3)
(4)
(5)
∪ {L0 : guard Ea0 → ~0 , L0 : ¬guard Ea → H0 | C0 ∈ P }
∪ {~i : act(Ci ) → hi+1 | i ∈ [0, n − 1], Ci , Ci+1 ∈ P } ∪ {~n : act(Cn ) → L0 | Cn ∈ P }
∪ {~i : ¬act(Ci ) → Lcnext(i) | i ∈ [0, n], Ci , cmpl(Ci ) ∈ P }
(6)
(7)
(8)
(9)
∪ {hi : guard Eai → ~i , hi : ¬guard Eai → Li | i ∈ [1, n], Ci ∈ P }
∪ {~i : act(Ci ) → Li+1 | i ∈ [0, n − 1], Ci ∈ P, Ci+1 6∈ P }
r {Ci | i ∈ [0, n − 1], Ci 6∈ P, Ci+1 ∈ P }
∪ {Li : act(Ci ) → hi+1 | i ∈ [0, n − 1], Ci 6∈ P, Ci+1 ∈ P }
where we define stitch P ′ (hp) , (3) ∪ (4) ∪ (5) ∪ (6) ∪ (7) ∪ (9).
Let us observe that:
– Clauses (1)–(6) are the same clauses of the trace extraction transform of Definition 5.1, with the additional constraint that all the commands Ci of hp are required
to belong to the original program P . This is equivalent to ask that any Ci is not the
entry or exit command of a nested hot path inside hp, i.e., Ci 6∈ P ′ r P . In Definition 5.1, where no previous hot path extraction is assumed, any command Ci of hp
belongs to P , so that this constraint is trivially satisfied.
– Clause (7) where Ci ∈ P and Ci+1 6∈ P , namely next(Ci ) is the call program point of
a nested hot path nhp and Ci+1 is the entry command of nhp, performs a relabeling
that allows to neatly nest nhp in hp.
– Clauses (8)–(9) where Ci 6∈ P and Ci+1 ∈ P , i.e., Ci is the exit command of a nested
hot path nhp that returns to the program point lbl(Ci+1 ), performs the relabeling of
suc(Ci ) in Ci in order to return from nhp to hp;
– H0 , ~i and hi are meant to be fresh labels, i.e., they have not been already used in
P ′.
Example 9.3. Let us go on with Example 9.1. The second traced hot path in
α2outerhot (TJP1 K) is:
hp2 = h⊤, H0 ≡ L1 : guard E⊤ → ℓ0 ,
⊤, H5c ≡ ℓ2 : (x%3 = 0) → L4 , ⊤, C4 ≡ L4 : x := x + 3 → L1 i.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:33
According to Definition 9.2, trace extraction of hp2 in P1 yields the following transform:
extr hp2 (P1 ) ,
[by clause (8)]
P1 r {H5c }
[by clause (9)]
∪ {ℓ2 : (x%3 = 0) → h2 }
[by clause (6)]
∪ {h2 : guard E⊤ → ~2 , h2 : ¬guard E⊤ → L4 }
[by clause (4)]
∪ {~2 : x := x + 3 → L1 }
where we used the additional fresh labels h2 and ~2 .
Example 9.4. Let us consider again Example 7.3. After the trace extraction of hp1
that transforms P to P1 , a second traced 2-hot path is the following:
hp2 , hρt , C1 , ρt , C2 , ρt , C3 , ρt , H0 , ρt , H1c , ρt , C7 i
where ρt = {primes[n]/ Bool, i/ Int, k/ Int} ∈ Storet . Thus, hp2 contains a nested hot
path which is called at suc(C3 ) = L4 and whose entry and exit commands are, respectively, H0 and H1c . Here, typed trace extraction according to Definition 9.2 provides the
following transform of P1 :
ts
P2 , Ofull
(P1 , hp2 ) = P1 r {C1 , C1c } ∪
H0 : i < 100 → L2 , H0 : ¬(i < 100) → L8 ,
H6 ≡ L1 : guard (primes[n] : Bool, i : Int, k : Int) → ~0 ,
H6c ≡ L1 : ¬guard (primes[n] : Bool, i : Int, k : Int) → H0 ,
H7 ≡ ~0 : i < 100 → h1 , H7c ≡ ~0 : ¬(i < 100) → L8 ,
H8 ≡ h1 : guard (primes[n] : Bool, i : Int, k : Int) → ~1 ,
H8c ≡ h1 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L2 ,
H9 ≡ ~1 : primes[i] = t t → h1 , H9c ≡ ~1 : ¬(primes[i] = t t) → L7 ,
H10 ≡ h2 : guard (primes[n] : Bool, i : Int, k : Int) → ~2 ,
c
H10
≡ h2 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L3 ,
H11 ≡ ~2 : k := i +Int i → L4
r {H1c } ∪ (H1c )′ ≡ ℓ0 : ¬(k < 100) → h3 ,
H12 ≡ h3 : guard (primes[n] : Bool, i : Int, k : Int) → ~3 ,
c
H12
≡ h3 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L7 ,
H13 ≡ ~3 : i := i +Int 1 → L1 .
Finally, a third traced 2-hot path in P2 is hp3 , hρt , H6 , ρt , H9c , ρt , C7 i which contains
a nested hot path which is called at the beginning of hp3 and whose entry and exit
commands are, respectively, H6 and H9c . Here, typed trace extraction of hp3 yields:
ts
P3 , Ofull
(P2 , hp3 ) = P2 r {H9c } ∪ (H9c )′ ≡ ~1 : ¬(primes [i] = t t) → j2 ,
j2 : guard (primes[n] : Bool, i : Int, k : Int) → 2 ,
j2 : ¬guard (primes[n] : Bool, i : Int, k : Int) → L7 ,
2 : i := i +Int 1 → L1 .
We have thus obtained the same three trace extraction steps described by
Gal et al. [2009, Section 2]. In particular, in P1 we specialized the typed addition operation k +Int i, in P2 we specialized i +Int i and i +Int 1, while in P3 we specialized once
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:34
S. Dissegna et al.
again i +Int 1 in a different hot path. Thus, in P3 all the addition operations occurring
in assignments have been type specialized.
10. COMPARISON WITH GUO AND PALSBERG’S FRAMEWORK
A formal model for tracing JIT compilation has been put forward at POPL 2011 symposium by Guo and Palsberg [2011]. Its main distinctive feature is the use of a bisimulation relation [Milner 1995] to model the operational equivalence between source and
optimized programs. In this section, we show how this model can be expressed within
our framework.
10.1. Language and Semantics
Guo and Palsberg [2011] rely on a simple imperative language (without jumps and)
with while loops and a so-called bail construct. Its syntax is as follows:
E ::= v | x | E1 + E2
B ::= tt | ff | E1 ≤ E2 | ¬B | B1 ∧ B2
Cmd ∋ c ::= skip; | x := E; | if B then S | while B do S | bail B to S
Stm ∋ S ::= ǫ | cS
where ǫ stands for the empty string. Thus, any statement S ∈ Stm is a (possibly empty)
sequence of commands cn , with n ≥ 0. We follow Guo and Palsberg [2011] in making an
abuse in program syntax by assuming that if S1 , S2 ∈ Stm then S1 S2 ∈ Stm, where S1 S2
denotes a simple string concatenation of S1 and S2 . We denote by StateGP , Store × Stm
the set of states for this language. The baseline small-step operational semantics →B
⊆ StateGP × StateGP is standard and is given in continuation-style (where K ∈ Stm):
hρ, ǫi 6→B
hρ, skip; Ki →B hρ, Ki
hρ, x := E; Ki →B hρ[x/EJEKρ], Ki
hρ, Ki
if BJBKρ = false
hρ, (if B then S)Ki →B
hρ, SKi if BJBKρ = true
hρ, (while B do S)Ki →B hρ, (if B then (S while B do S)) Ki
hρ, Ki if BJBKρ = false
hρ, (bail B to S)Ki →B
hρ, Si if BJBKρ = true
The relation →B is clearly deterministic and we denote by
TraceGP , {σ ∈ State+
GP | ∀i ∈ [0, |σ| − 1). σi →B σi+1 }
the set of generic program traces for Guo and Palsberg’s language. Then, given a program S ∈ Stm, so that StoreS , vars(S) → Valueu denotes the set of stores for S, its
partial trace semantics is
TGP JSK = TraceGP
, {σ ∈ TraceGP | σ0 = hρ, Si, ρ ∈ StoreS }.
S
Notice that, differently from our trace semantics, a partial trace of the program S
always starts from an initial state, i.e., hρ, Si.
10.2. Language Compilation
Programs in Stm can be compiled into Program by resorting to an injective labeling
function l : Stm → L that assigns different labels to different statements.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:35
Definition 10.1 (Language compilation). The “first command” compilation function C : Stm → ℘(C) is defined as follows:
C(ǫ) , {l(ǫ) : skip → Ł}
C S ≡ (skip; K) , {l(S ′ ) : skip → l(K)}
C S ′ ≡ (x := E; K) , {l(S ′ ) : x := E → l(K)}
C S ′ ≡ ((if B then S)K) , {l(S ′ ) : B → l(SK), l(S ′ ) : ¬B → l(K)}
C S ′ ≡ ((while B do S)K) , {l(S ′ ) : skip → l((if B then (S while B do S))K)}
C S ′ ≡ ((bail B to S)K) , {l(S ′ ) : B → l(S), l(S ′ ) : ¬B → l(K)}
′
Then, the full compilation function C : Stm → ℘(C) is recursively defined by the following clauses:
C(ǫ) , C(ǫ)
C(skip; K) , C(skip; K) ∪ C(K)
C(x := E; K) , C(x := E; K) ∪ C(K)
C((if B then S)K) , C((if B then S)K) ∪ C(SK) ∪ C(K)
C((while B do S)K) , C((while B do S)K) ∪ C((if B then (S while B do S))K)
C((bail B to S)K) , C((bail B to S)K) ∪ C(S) ∪ C(K)
Given S ∈ Stm, l(S) is the initial label of C(S), while Ł is, as usual, the undefined label
where the execution becomes stuck.
It turns out that the recursive function C is well-defined—the easy proof is standard
and is omitted, let us just observe that C((while B do S)K) is a base case—so that, for
any S ∈ Stm, C(S) is a finite set of commands. Let us observe that, by Definition 10.1,
if hρ, Si →B hρ′ , S ′ i then C(S ′ ) ⊆ C(S) (this can be proved through an easy structural
induction on S). Consequently, if hρ, Si →∗B hρ′ , S ′ i then C(S ′ ) ⊆ C(S).
Example 10.2. Consider the following program S ∈ Stm in Guo and Palsberg’s
syntax:
x := 0;
while B1 do x := 1;
x := 2;
bail B2 to x := 3;
x := 4;
S is then compiled in our language by C in Definition 10.1 as follows:
C(S) = l(S) : x := 0 → lwhile , lwhile : skip → lifwhile ,
lifwhile : B1 → l1 , lifwhile : ¬B1 → l2 , l1 : x := 1 → lwhile ,
l2 : x := 2 → lbail , lbail : B2 → l3 , lbail : ¬B2 → l4 ,
l3 : x := 3 → lǫ , l4 : x := 4 → lǫ , lǫ : skip → Ł .
Notice that in the command lbail : B2 → l3 , the label l3 stands for l(x := 3; ) so that
C(x := 3; ) ≡ l3 : x := 3 → lǫ , i.e., after the execution of x := 3 the program terminates.
Correctness for the above compilation function C means that for any S ∈ Stm:
(i) C(S) ∈ Program and (ii) program traces of S and C(S) have the same store sequences.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:36
S. Dissegna et al.
C s (hρ, ǫi) , hρ, l(ǫ) : skip → Łi
C s (hρ, S ≡ (skip; K)i) , hρ, l(S) : skip → l(K)i
C s (hρ, S ≡ (x := E; K)i) , hρ, l(S) : x := E → l(K)i
hρ, l(S) : B → l(S ′ K)i if BJBKρ = true
s
′
C (hρ, S ≡ ((if B then S )K)i) ,
hρ, l(S) : ¬B → l(K)i if BJBKρ = false
C s (hρ, S ≡ ((while B do S ′ )K)i) , hρ, l(S) : skip → l((if B then (S ′ while B do S ′ ))K)i
hρ, l(S) : B → l(S ′ )i
if BJBKρ = true
s
′
C (hρ, S ≡ ((bail B to S )K)i) ,
hρ, l(S) : ¬B → l(K)i if BJBKρ = false
Fig. 6. Definition of the state compile function C s : StateGP → State.
In the proof we will make use of a “state compile” function C s : StateGP → State
as defined in Figure 6. In turn, C s allows us to define a “trace compile” function
C t : TGP JSK → Tι JC(S)K which applies state-by-state the function C s to traces as follows:
C t (ǫ) , ǫ;
C t (sτ ) , C s (s)C t (τ ).
L EMMA 10.3.
(1) hρ, Si →B hρ′ , S ′ i ⇔ C s (hρ′ , S ′ i) ∈ S(C s (hρ, Si))
(2) C t is well-defined.
P ROOF. We show the equivalence (1) by structural induction on S ∈ Stm.
[S ≡ ǫ]: Trivially true, since hρ, Si 6→B and Shρ, l(ǫ) : skip → Łi = ∅.
[S ≡ x := E; K] (⇒): If hρ, x := E; Ki →B hρ[x/EJEKρ], Ki, C s (hρ, x := E; Ki) =
hρ, l(S) : x := E → l(K)i and C s (hρ[x/EJEKρ], Ki) = hρ[x/EJEKρ], l(K) : A → l(S ′ )i
for some action A and statement S ′ , then, by definition of the transition semantics S,
hρ[x/EJEKρ], l(K) : A → l(S ′ )i ∈ Shρ, l(S) : x := E → l(K)i.
(⇐): If hρ′′ , Ci = C s (hρ′ , S ′ i) ∈ Shρ, l(S) : x := E → l(K)i then: (1) EJEKρ 6= undef ,
(2) ρ′′ = ρ[x/EJEKρ], and therefore ρ′ = ρ[x/EJEKρ]; (3) lbl(C) = l(K), and therefore
S ′ = K. Hence, hρ, x := E; Ki →B hρ[x/EJEKρ], Ki = hρ′ , S ′ i.
[S ≡ skip; K] Analogous to S ≡ x := E; K.
[S ≡ (if B then T )K] (⇒): Assume that BJBKρ = false, so that
hρ, (if B then T )Ki →B hρ, Ki, C s (hρ, (if B then T )Ki) = hρ, l(S) : ¬B → l(K)i and
C s (hρ, Ki) = hρ, l(K) : A → l(T ′ )i for some A and T ′ ∈ Stm. Hence, by definition of S,
hρ, l(K) : A → l(T ′ )i ∈ Shρ, l(S) : ¬B → l(K)i. On the other hand, if BJBKρ = true then
hρ, (if B then T )Ki →B hρ, T Ki, C s (hρ, (if B then T )Ki) = hρ, l(S) : B → l(T K)i and
C s (hρ, T Ki) = hρ, l(T K) : A → l(T ′ )i for some A and T ′ . Hence, hρ, l(T K) : A → l(T ′ )i ∈
Shρ, l(S) : B → l(T K)i.
(⇐): Assume that BJBKρ = false, so that C s (hρ, (if B then T )Ki) = hρ, l(S) : ¬B →
l(K)i, and hρ′′ , Ci = C s (hρ′ , S ′ i) ∈ Shρ, l(S) : ¬B → l(K)i. Hence: (1) ρ′′ = ρ and therefore ρ′ = ρ; (2) lbl(C) = l(K), and therefore S ′ = K. Hence, hρ, (if B then T )Ki →B
hρ, Ki = hρ′ , S ′ i. On the other hand, if BJBKρ = true then C s (hρ, (if B then T )Ki) =
hρ, l(S) : B → l(T K)i and hρ′′ , Ci = C s (hρ′ , S ′ i) ∈ Shρ, l(S) : B → l(T K)i. We thus have
that: (1) ρ′′ = ρ and therefore ρ′ = ρ; (2) lbl(C) = l(T K), and therefore S ′ = T K. Hence,
hρ, (if B then T )Ki →B hρ, T Ki = hρ′ , S ′ i.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:37
[S ≡ (while B do T )K] (⇒): We have that hρ, (while B do T )Ki →B
hρ, (if B then (T while B do T )) Ki and C s (hρ, (while B do T )Ki) = hρ, l(S) :
skip → l((if B then (T while B do T )) K)i. If BJBKρ = true then
C s (hρ, (if B then (T while B do T )) Ki) = hρ, l((if B then (T while B do T )) K) :
B → l(T (while B do T )K)i; on the other hand, if BJBKρ = false then
C s (hρ, (if B then (T while B do T )) Ki) = hρ, l((if B then (T while B do T )) K) :
¬B → l(K)i. In both cases, we have that:
hρ, l((if B then (T while B do T )) K) : B → l(T (while B do T )K)i,
hρ, l((if B then (T while B do T )) K) : B → l(K)i
∈ Shρ, l(S) : skip → l((if B then (T while B do T )) K)i.
(⇐): If hρ′′ , Ci = C s (hρ′ , S ′ i) ∈ Shρ, l(S) : skip → l((if B then (T while B do T )) K)i
then: (1) ρ′′ = ρ, and therefore ρ′ = ρ; (2) lbl(C) = l((if B then (T while B do T )) K),
and therefore S ′ = (if B then (T while B do T )) K. Hence, hρ, (while B do T )Ki →B
hρ, (if B then (T while B do T )) Ki = hρ′ , S ′ i.
[S ≡ (bail B to T )K] Analogous to S ≡ (if B then T )K.
Let us now turn to point (2). By the ⇒ implication of the equivalence (1), we have
that if τ ∈ TGP JSK then C t (τ ) ∈ TJC(S)K: this can be shown by an easy induction on
the length of τ and by using the fact that if C t (τ ) = hρ0 , C0 ihρ1 , C1 i · · · hρn , Cn i then, for
any i, Ci ∈ C(S). Moreover, since l(S) is the initial label of the compiled program C(S)
and lbl(C0 ) = l(S), we also notice that C t (τ ) ∈ Tι JC(S)K. Therefore, C t is a well-defined
function.
Let st : TraceGP ∪ Trace → Store∗ be the function that returns the store sequence of
any trace, that is:
st (ǫ) , ǫ and
st(hρ, Siσ) , ρ · st(σ).
Also, given a set X of traces, let αst (X) , {st(σ) | σ ∈ X}. Then, correctness of the
compilation function C goes as follows:
T HEOREM 10.4 (C ORRECTNESS OF LANGUAGE
C(S) ∈ Program and αst (TGP JSK) = αst (Tι JC(S)K).
COMPILATION ).
If S ∈ Stm then
P ROOF. We define a “trace de-compile” function Dt : Tι JC(S)K → TGP JSK as follows.
Consider a trace σ = hρ0 , C0 i · · · hρn , Cn i ∈ Tι JC(S)K, so that lbl(C0 ) = l(S), for any
i ∈ [0, n], Ci ∈ C(S) and for any i ∈ [0, n), hρi+1 , Ci+1 i ∈ SJC(S)Khρi , Ci i. Since lbl(C0 ) =
l(S), by definition of C s , we have that hρ0 , C0 i = C s (hρ0 , Si). Then, since hρ1 , C1 i ∈
SJC(S)K(C s (hρ0 , Si)), there exists S1 ∈ Stm such that lbl(C1 ) = l(S1 ), so that, hρ1 , C1 i =
C s (hρ1 , S1 i). Hence, from C s (hρ1 , S1 i) ∈ SJC(S)K(C s (hρ0 , Si)), by the implication ⇐ of
Lemma 10.3 (1), we obtain that hρ0 , Si →B hρ1 , S1 i. Thus, an easy induction allows us
to show that for any i ∈ [1, n] there exists Si ∈ Stm such that
hρ0 , Si →B hρ1 , S1 i →B · · · →B hρn , Sn i
and C s (hρi , Si i) = hρi , Ci i. We therefore define Dt (σ) , hρ0 , Sihρ1 , S1 i · · · hρn , Sn i ∈
TGP JSK. Moreover, we notice that st(Dt (σ)) = st(σ). Let us also observe that st◦C t = st,
since C t does not affect stores.
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:38
S. Dissegna et al.
Summing up, we obtain:
αst (TGP JSK) =
αst (C t (TGP JSK)) ⊆
αst (Tι JC(S)K) =
αst (Dt (Tι JC(S)K)) ⊆
αst (TGP JSK)
[since st ◦ C t = st]
[by Lemma 10.3 (2), C t is well-defined]
[since st ◦ Dt = st]
[since Dt is well-defined]
and this closes the proof.
10.3. Bisimulation
Correctness of trace extraction in [Guo and Palsberg 2011] relies on a notion of bisimulation relation, parameterized by program stores. Let us recall this definition. If
hρ, Si →B hρ, S ′ i then this “silent” transition that does not change the store is also deτ
noted by hρ, Si →B hρ, S ′ i. Moreover, for the assignment transition hρ, x := E; Ki →B
hρ[x/EJEKρ], Ki, if δ = [x/EJEKρ] denotes the corresponding store update of ρ then this
δ
transition is also denoted by hρ, x := E; Ki →B hρ[x/EJEKρ], Ki. Let Act , {δ | δ is a
store update} ∪ {τ }. Then, for a nonempty sequence of actions s = a1 · · · an ∈ Act + , we
define:
s
hρ, Si ⇒B hρ′ , S ′ i iff
τ ∗
a
τ ∗
τ ∗
a
τ ∗
n
′
′
hρ, Si →B ◦ →1 B ◦ →B · · · →B ◦ →
B ◦ →B hρ , S i,
namely, there may be any number of silent transitions either in front of or following
a
any ai -transition →i B . Moreover, if s ∈ Act + is a nonempty sequence of actions then
∗
ŝ ∈ Act denotes the possibly empty sequence of actions where all the occurrences of τ
are removed.
Definition 10.5 ([Guo and Palsberg 2011]). A relation R ⊆ Store × Stm × Stm is a
bisimulation when R(ρ, S1 , S2 ) implies:
a
â
a
â
(1) if hρ, S1 i →B hρ′ , S1′ i then hρ, S2 i ⇒B hρ′ , S2′ i, for some hρ′ , S2′ i such that R(ρ′ , S1′ , S2′ );
(2) if hρ, S2 i →B hρ′ , S2′ i then hρ, S1 i ⇒B hρ′ , S1′ i, for some hρ′ , S1′ i such that R(ρ′ , S1′ , S2′ ).
S1 is bisimilar to S2 for a given ρ ∈ Store, denoted by S1 ≈ρ S2 , if R(ρ, S1 , S2 ) for some
bisimulation R.
τ
τ̂
Let us remark that if hρ, S1 i → hρ′ , S1′ i then τ̂ = ǫ, so that hρ, S2 i ⇒ hρ, S2 i ≡ hρ, S2 i is
allowed to be the matching (empty) transition sequence.
It turns out that bisimilarity can be characterized through an abstraction of traces
that observes store changes. By a negligible abuse of notation, the store changes
function sc : Trace → Store∗ defined in Section 6 is applied to GP traces, so that
sc : Trace ∪ TraceGP → Store∗ . In turn, given ρ ∈ Store, the function αρsc : ℘(TraceGP ) →
℘(Store∗ ) is then defined as follows:
αρsc (X) , {sc(τ ) ∈ Store∗ | τ ∈ X, ∃S, τ ′. τ = hρ, Siτ ′ }.
It is worth remarking that αρsc is a weaker abstraction than αsc defined in Section 6,
that is, for any X, Y ∈ ℘(TraceGP ), αsc (X) = αsc (Y ) ⇒ αρsc (X) = αρsc (Y ) (while the
converse does not hold in general).
T HEOREM 10.6. For any S1 , S2 ∈ Stm, ρ ∈ Store, we have that S1 ≈ρ S2 iff
αρsc (TGP JS1 K) = αρsc (TGP JS2 K).
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:39
P ROOF. (⇒): We prove that if R(ρ, S1 , S2 ) holds for some bisimulation R then
αρsc (TGP JS1 K) ⊆ αρsc (TGP JS2 K) (the reverse containment is symmetric), that is, if
sc(τ ) ∈ Store∗ for some τ ∈ TGP JS1 K such that τ = hρ, S1 iτ ′ then there exists
some ψ ∈ TGP JS2 K such that ψ = hρ, S2 iψ ′ and sc(τ ) = sc(ψ). Let us then consider
τ ∈ TGP JS1 K such that τ = hρ, S1 iτ ′ . If τ ′ = ǫ then we pick hρ, S2 i ∈ TGP JS2 K so that
sc(hρ, S1 i) = ρ = sc(hρ, S2 i). Otherwise, τ = hρ, S1 iτ ′ ∈ TGP JS1 K, with ǫ 6= τ ′ = τ ′′ hµ, Si.
We prove by induction on |τ ′ | ≥ 1 that there exists ψ = hρ, S2 iψ ′′ hµ, T i ∈ TGP JS2 K such
that sc(τ ) = sc(ψ) and R(µ, S, T ).
a
(|τ ′ | = 1): In this case, τ = hρ, S1 ihµ, Si ∈ TGP JS1 K, so that hρ, S1 i →B hµ, Si. Since,
â
by hypothesis, R(ρ, S1 , S2 ) holds, we have that hρ, S2 i ⇒B hµ, T i, for some T , and
R(µ, S, T ). Let ψ ∈ TGP JS2 K be the trace corresponding to the sequence of transitions
â
â
hρ, S2 i ⇒B hµ, T i. Then, by definition of ⇒B , we have that sc(τ ) = sc(ψ), and, by definition of bisimulation, R(µ, S, T ) holds.
(|τ ′ | > 1): Here, τ ′ = τ ′′ hµ, Si and τ = hρ, S1 iτ ′ ∈ TGP JS1 K, with |τ ′′ | = |τ ′ | − 1 ≥ 1.
Hence, τ ′′ = τ ′′′ hη, U i. By inductive hypothesis, there exists ψ = hρ, S2 iψ ′′ hη, V i ∈
TGP JS2 K such that sc(hρ, S1 iτ ′′′ hη, U i) = sc(hρ, S2 iψ ′′ hη, V i) and R(η, U, V ). Since
a
â
hη, U i →B hµ, Si and R(η, U, V ) holds, we obtain that hη, V i ⇒B hµ, T i, for some T ,
and R(µ, S, T ) holds. Let hη, V i · · · hµ, T i be the sequence of states corresponding to
â
the sequence of transitions hη, V i ⇒B hµ, T i so that we pick hρ, S2 iψ ′′ hη, V i · · · hµ, T i ∈
â
TGP JS2 K. The condition R(µ, S, T ) already holds. Moreover, by definition of ⇒B , we
have that sc(hη, U ihµ, Si) = sc(hη, V i · · · hµ, T i), and therefore we obtain sc(τ ) =
sc(hρ, S1 iτ ′′′ hη, U ihµ, Si) = sc(hµ, S2 iψ ′′ hη, V i · · · hµ, T i).
(⇐): We first observe the following property (∗), which is a straight consequence of
the fact that →B is a deterministic relation: If S ∈ Stm and σ, τ ∈ TJSK are such that
σ0 = hµ, Si = τ0 and |τ | ≤ |σ| then there exists some ψ such that σ = τ ψ.
Given ρ ∈ Store, we assume that αρsc (TGP JS1 K) = αρsc (TGP JS2 K) and we then define the
following relation R:
R , {(ρ, S1 , S2 )} ∪ {(µ, T1 , T2 ) | hρ, S1 i · · · hµ, T1 i ∈ TJS1 K, hρ, S2 i · · · hµ, T1 i ∈ TJS2 K,
sc(hρ, S1 i · · · hµ, T1 i) = sc(hρ, S2 i · · · hµ, T1 i)}.
We show that R is a bisimulation, so that R(ρ, S1 , S2 ) follows.
a
(case A) Assume that hρ, S1 i →B hρ′ , S1′ i. Then, since hρ, S1 ihρ′ , S1′ i ∈ TJS1 K and
ρ
ρ
αsc (TGP JS1 K) = αsc (TGP JS2 K), we have that there exists τ = hρ, S2 i · · · ∈ TJS2 K such
that sc(hρ, S1 ihρ′ , S1′ i) = sc(τ ). Hence, τ necessarily has the following shape:
τ = hρ, S2 ihρ, U1 i · · · hρ, Un ihρ′ , V1 i · · · hρ′ , Vm i
where n ≥ 0 (n = 0 means that hρ, U1 i · · · hρ, Un i is indeed the empty sequence) and m ≥
â
1. This therefore means that hρ, S2 i ⇒B hρ′ , Vm i, so that, by definition of R, R(ρ′ , S1′ , Vm )
holds.
(case B) Assume now that R(µ, T1 , T2 ) holds because δ = hρ, S1 i · · · hµ, T1 i ∈ TJS1 K,
a
σ = hρ, S2 i · · · hµ, T2 i ∈ TJS2 K and sc(δ) = sc(σ). Hence, let us suppose that hµ, T1 i →B
′
′
′
′
ρ
ρ
hµ , T1 i. Then, since δhµ , T1 i ∈ TJS1 K and αsc (TGP JS1 K) = αsc (TGP JS2 K), we have that
there exists τ = hρ, S2 i · · · ∈ TJS2 K such that sc(δhµ′ , T1′ i) = sc(τ ).
(case B1). If |τ | ≤ |σ| then, by the property (∗) above, σ = τ ψ, for some ψ. Hence,
sc(τ ) = sc(δhµ′ , T1′ i) is a prefix of sc(σ) = sc(δ). Consequently, sc(δhµ′ , T1′ i) can
be a prefix of sc(δ) only if sc(δhµ′ , T1′ i) = sc(δ), so that the action a is τ and
τ
µ′ = µ, that is, hµ, T1 i →B hµ, T1′ i. We thus consider the empty transition sequence
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:40
S. Dissegna et al.
(T1 ) hρ, (if B then (S while B do S)) Ki →T hρ, (while B do S)K, ǫ, S(while B do S)Ki
if BJBKρ = true
(T2 ) hρ, Kw , t, skip; Ki →T hρ, Kw , t(skip;), Ki
(T3 ) hρ, Kw , t, x := E; Ki →T hρ[x/EJEKρ], Kw , t(x := E;), Ki
(
hρ, Kw , t(bail B to (SK)), Ki if BJBKρ = false
(T4 ) hρ, Kw , t, (if B then S)Ki →T hρ, K , t(bail ¬B to K), SKi if BJBKρ = true
w
hρ, Kw , t(skip; ), (if B then (S while B do S)) Ki
if Kw 6≡ (while B do S)K
(T5 ) hρ, Kw , t, (while B do S)Ki →T
hρ, O(while B do t, ρ)Ki
if Kw ≡ (while B do S)K
(T6 ) hρ, Kw , t, Si →T hρ′ , S ′ i if Kw 6≡ S and hρ, Si →B hρ′ , S ′ i
Fig. 7. Definition of the tracing relation →T .
τ̂
hµ, T2 i ⇒ hµ, T2 i, so that from sc(δhµ, T1′ i) = sc(σ), by definition of R we obtain that
R(µ, T1′ , T2 ) holds.
(case B2). If |τ | > |σ| then, by (∗) above, τ = σψ, for some ψ, i.e., τ =
σ · · · hµ′′ , T2′ i, for some µ′′ and T2′ . Since sc(hρ, S2 i · · · hµ, T2 i) = sc(hρ, S1 i · · · hµ, T1 i)
and sc(hρ, S2 i · · · hµ, T2 i · · · hµ′′ , T2′ i) = sc(hρ, S1 i · · · hµ, T1 ihµ′ , T1′ i), we derive that
a
µ′′ = µ′ and hµ, T2 i ⇒ hµ′′ = µ′ , T2′ i. By definition of R, R(µ′ , T1′ , T2′ ) holds.
This closes the proof.
10.4. Hot Paths
Let us recall the set of rules that define the tracing transitions in Guo and Palsberg [2011] model. Let tStateGP , Store × Stm × Stm × Stm denote the set of states
in trace recording mode, whose components are, respectively, the current store, the entry point of the recorded trace (this is always a while statement), the current trace (i.e.,
a sequence of commands) and the current program to be evaluated. In turn, StateeGP ,
StateGP ∪ tStateGP denotes the corresponding extended notion of state, which encompasses the trace recording mode. Then, the relation →T ⊆ StateeGP × StateeGP is defined by the clauses in Figure 7, where O : Stm × Store → Stm is a “sound” optimization function that depends on a given store. Correspondingly, the trace semantics TGP JSK ⊆ (StateeGP )+ of a program S ∈ Stm is naturally extended to the relation
→B,T , →B ∪ →T ⊆ StateeGP × StateeGP .
Let us notice that in Guo and Palsberg’s model of hot paths:
(i) By clause (T1 ), trace recording is always triggered by an unfolded while loop, and
the loop itself is not included in the hot path.
(ii) By clause (T4 ), when we bail out of a hot path t through a bail command, we cannot
anymore re-enter into t.
(iii) By clause (T5 )—the second condition of this clause is called stitch rule in
[Guo and Palsberg 2011]—the store used to optimize a hot path t is recorded at
the end of the first loop iteration. This is a concrete store which is used by O to
optimize the stitched hot path while B do t.
(iv) Hot paths actually are 1-hot paths according to our definition, since, by clause (T1 ),
once the first iteration of the traced while loop is terminated, trace recording necessarily discontinues.
(v) There are no clauses for trace recording bail commands. Hence, when trying to
trace a loop that already contains a nested hot path, by clause (T6 ), trace recording
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:41
is aborted when a bail command is encountered. In other terms, in contrast to our
approach described in Section 9, nested hot paths are not allowed.
(vi) Observe that when tracing a loop while B do S whose body S does not contain
branching commands, i.e. if or while statements, it turns out that the hot path t
coincides with the body S, so that while B do t ≡ while B do S, namely, in this
case the hot path transform does not change the subject while loop.
In the following, we show how this hot path extraction model can be formalized
within our trace-based approach. To this aim, we do not consider optimizations of hot
paths, which is an orthogonal issue here, so that we assume that O performs no optimization, that is, O(while B do t, ρ) = while B do t.
A sequence of commands t ∈ Stm is defined to be a GP hot path for a program
Q ∈ Stm when we have the following transition sequence:
hρ, Qi →∗B,T hρ′ , (while B do S)Ki →∗B,T hρ′′ , (while B do S)K, t, (while B do S)Ki.
Since the operational semantics →B,T is given in continuation-style, without loss of
generality, we assume that the program Q begins with a while statement, that is Q ≡
(while B do S)K. Guo and Palsberg’s hot loops can be modeled in our framework by
exploiting a revised loop selection map loop GP : Trace → ℘(C+ ) defined as follows:
loop GP (hρ0 , C0 i · · · hρn , Cn i) , Ci Ci+1 · · · Cj | 0 ≤ i ≤ j < n, Ci ⋖ Cj ,
suc(Cj ) = lbl(Ci ), ∀k ∈ (i, j]. Ck 6∈ {Ci , cmpl(Ci )} .
Thus, loop GP (τ ) contains sequences of commands without store. The map αGP
hot :
+
GP
℘(Trace) → ℘(C ) then lifts loop GP to sets of traces as usual: αhot (T ) , ∪τ ∈T loop GP (τ ).
Then, let us consider a GP hot path t as recorded by a transition sequence τ :
τ , hρ, S0 ≡ (while B do S)Ki →B
hρ, S1 ≡ (if B then (S while B do S)) Ki →T
hρ, (while B do S)K, ǫ, S2 ≡ S(while B do S)Ki →T
· · · →T
hρ′ , (while B do S)K, t′ , Sn i →T
hρ′′ , (while B do S)K, t, Sn+1 ≡ (while B do S)Ki
(‡)
where BJBKρ = true. Hence, the Si ’s occurring in τ are the current statements to be evaluated. With a negligible abuse of notation, we assume that τ ∈
TGP J(while B do S)KK, that is, the arrow symbols →B and →T are taken out of the sequence τ . By Lemma 10.3 (2), we therefore consider the corresponding execution trace
C t (τ ) of the compiled program C((while B do S)K), where the state compile function
C s in Figure 6, when applied to states in trace recording mode, is assumed to act on the
current store and the program to be evaluated, that is, C s (hρ, Kw , t, Si) = C s (hρ, Si). We
thus obtain:
C t (τ ) , hρ, C0 ≡ l((while B do S)K) : skip → l((if B then (S while B do S)) K)i
hρ, C1 ≡ l((if B then (S while B do S)) K) : B → l(S (while B do S)K)i
hρ, C2 ≡ l(S(while B do S)K) : A2 → l(T )i
···
hρ′ , Cn ≡ l(Sn ) : An → l((while B do S) K)i
hρ′′ , Cn+1 ≡ l((while B do S)K) : skip → l((if B then (S while B do S)) K)i.
We therefore obtain a hot path hpt = C0 C1 · · · Cn ∈ loop GP (C t (τ )), i.e. hpt ∈
ι
αGP
hot (T JC((while B do S)K)K), where lbl(C0 ) = l((while B do S)K) = suc(Cn ). This
is a consequence of the fact that for all k ∈ (0, n], Ck cannot be the entry command C0
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:42
S. Dissegna et al.
or its complement command, because, by the stitch rule of clause (T5 ), Sn+1 is necessarily the first occurrence of (while B do S)) K as current program to be evaluated
in the trace τ , so that, for any k ∈ (0, n], lbl(Ck ) 6= l((while B do S)K). We have thus
shown that any GP hot path arising from a trace τ generates a corresponding hot path
extracted by our selection map loop GP on the compiled trace C t (τ ):
L EMMA 10.7. Let Qw ≡ (while B do S)K. If t is a GP hot path for Qw where
τ ≡ hρ, Qw i →∗B,T hρ′ , Qw , t, Qw i is the transition sequence (‡) that records t, then there
ι
exists a hot path hpt = C0 C1 · · · Cn ∈ αGP
hot (T JC(Qw )K) such that, for any i ∈ [0, n],
lbl(Ci ) = l(Si ), and, in particular, lbl(C0 ) = l(Qw ) = suc(Cn ).
Example 10.8. Let us consider the while statement Qw of the program in Example 2.1:
Qw ≡ while (x ≤ 20) do (x := x + 1; (if (x%3 = 0) then x := x + 3; ))
This program is already written in Guo and Palsberg language, so that Qw is a well
formed statement in Stm. The tracing rules in Figure 7 yield the following trace t for
Qw :
t ≡ x := x + 1; bail (x%3 = 0) to (x := x + 3; Qw ).
On the other hand, the compiled program C(Qw ) ∈ ℘(C) is as follows:
C(Qw ) = D0 ≡ lwhile : skip → lifwhile ,
D1 ≡ lifwhile : (x ≤ 20) → l1 , D1c ≡ lifwhile : ¬(x ≤ 20) → lǫ ,
D2 ≡ l1 : x := x + 1 → lif ,
D3 ≡ lif : (x%3 = 0) → l2 , D3c ≡ lif : ¬(x%3 = 0) → lwhile ,
D4 ≡ l2 : x := x + 3 → lwhile , D5 ≡ lǫ : skip → Ł ,
where labels have the following meaning:
lwhile , l(Qw )
lifwhile , l(if (x ≤ 20) then (x := x + 1; (if (x%3 = 0) then x := x + 3; ) Qw ))
l1 , l(x := x + 1; (if (x%3 = 0) then x := x + 3; ) Qw ))
lif , l((if (x%3 = 0) then x := x + 3; ) Qw ))
l2 , l(x := x + 3; Qw ).
Hence, in correspondence with the trace t, we obtain the hot path hpt = D0 D1 D2 D3c ∈
ι
αGP
hot (T JC(Qw )K). In turn, this hot path hpt corresponds to the 2-hot path hp1 consisting
of the analogous sequence of commands, which has been selected in Example 4.1.
10.5. GP Trace Extraction
In the following, we conform to the notation used in Section 5 for our trace extraction
transform. Let us consider a while program Qw ≡ (while B do S)K ∈ Stm and its
compilation Pw , C(Qw ) ∈ ℘(C). Observe that, by Definition 10.1 of compilation C, a
hot path C0 · · · Cn ∈ αGP
hot (TJPw K) for the compiled program Pw always arises in correspondence with some while loop while B ′ do S ′ occurring in Qw and therefore has
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:43
necessarily the following shape:
C0 ≡ l((while B ′ do S ′ )J) : skip → l((if B ′ then (S ′ while B ′ do S ′ )) J)
C1 ≡ l((if B ′ then (S ′ (while B ′ do S ′ ))) J) : B ′ → l(S ′ (while B ′ do S ′ ) J)
C2 ≡ l(S ′ (while B ′ do S ′ ) J) : A2 → l(T3 )
···
Cn ≡ l(Tn ) : An → l((while B ′ do S ′ )J)
The GP hot path extraction scheme for Qw described by the rules in Figure 7 can be
defined in our language by the following simple transform of Pw .
Definition 10.9 (GP trace extraction transform). The GP trace extraction transGP
form extr GP
hp (Pw ) of Pw for the hot path hp = C0 C1 · · · Cn ∈ αhot (TJPw K) is defined as
follows:
(1) If for any i ∈ [2, n], cmpl(Ci ) 6∈ Pw then extr GP
hp (Pw ) , Pw ;
(2) Otherwise:
extr GP
hp (Pw ) , Pw ∪ {ℓi : act(Ci ) → ℓnext(i) | i ∈ [0, n]}
∪ {ℓi : ¬act(Ci ) → Lcnext(i) | i ∈ [0, n], cmpl(Ci ) ∈ Pw }.
Clearly, extr GP
hp (P ) remains a well-formed program. Also observe that the case (1) of
Definition 10.9 means that the traced hot path hp does not contain conditional commands (except from the entry conditional C1 ) and therefore corresponds to point (vi) in
Section 10.4.
Example 10.10. Let us consider the programs Qw and C(Qw ) of Example 10.8 and
the hot path hpt = D0 D1 D2 D3c ∈ αGP
hot (TJC(Qw )K) which corresponds to the trace t ≡
x := x + 1; bail (x%3 = 0) to (x := x + 3; Qw ) of Qw . Here, the GP trace extraction of
hpt , according to Definition 10.9, provides the following program transform:
extr GP
hpt (C(Qw )) , D0 ≡ lwhile : skip → lifwhile , D1 ≡ lifwhile : (x ≤ 20) → l1 ,
D1c ≡ lifwhile : ¬(x ≤ 20) → lǫ , D2 ≡ l1 : x := x + 1 → lif ,
D3 ≡ lif : (x%3 = 0) → l2 , D3c ≡ lif : ¬(x%3 = 0) → lwhile ,
D4 ≡ l2 : x := x + 3 → lwhile , D5 ≡ lǫ : skip → Ł ∪
ℓ0 : skip → ℓ1 , ℓ1 : x ≤ 20 → ℓ2 , ℓ1 : ¬(x ≤ 20) → lǫ ,
ℓ2 : x := x + 1 → ℓ3 , ℓ3 : ¬(x%3 = 0) → ℓ0 , ℓ3 : (x%3 = 0) → l2 .
On the other hand, the stitch rule (T5 ) transforms Qw into the following program Qt :
while (x ≤ 20) do
x := x + 1;
bail (x%3 = 0) to (x := x + 3; Qw )
whose compilation yields the following program:
C(Qt ) = lwhilet : skip → lifwhilet , lifwhilet : (x ≤ 20) → l1t , lifwhilet : ¬(x ≤ 20) → lǫ ,
l1t : x := x + 1 → lbail , lbail : (x%3 = 0) → lbailtrue , lbail : ¬(x%3 = 0) → lwhilet ,
lwhile : skip → lifwhile , lifwhile : (x ≤ 20) → l1 , lifwhile : ¬(x ≤ 20) → lǫ ,
l1 : x := x + 1 → lif , lif : (x%3 = 0) → l2 , lif : ¬(x%3 = 0) → lwhile ,
lbailtrue : x := x + 3 → lwhile , lǫ : skip → Ł
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:44
S. Dissegna et al.
with the following new labels:
lwhilet , l(while (x ≤ 20) t)
lifwhilet , l(if (x ≤ 20) then (t (while (x ≤ 20) t)))
l1t , l(t (while (x ≤ 20) t))
lbail , l((bail (x%3 = 0) to (x := x + 3; Qw ))(while (x ≤ 20) t))
while observe that lbailtrue , l(x := x + 3; Qw ) = l2 . It is then immediate to check that
the programs C(Qt ) and extr GP
hpt (C(Qw )) are equal up to the following label renaming of
extr GP
(C(Q
)):
w
hpt
{ℓ0 7→ lwhilet , ℓ1 7→ lifwhilet , ℓ2 7→ l1t , ℓ3 7→ lbail }.
The equivalence of this GP trace extraction with the stitch of hot paths by
Guo and Palsberg [2011] goes as follows.
T HEOREM 10.11 (E QUIVALENCE WITH GP TRACE EXTRACTION ). Let t be a GP
trace such that hρ, (while B do S)Ki →∗B,T hρ′ , (while B do S)K, t, (while B do S)Ki
ι
and let hpt ∈ αGP
hot (T JC((while B do S)K)K) be the corresponding GP hot path as determined by Lemma 10.7. Then, C((while B do t)K) ∼
= extr GP
hpt (C((while B do S)K)).
P ROOF. Let the GP hot path t be recorded by the following transition sequence for
hρ, (while B do S)Ki:
hρ, S−2 ≡ (while B do S)Ki →B
hρ, S−1 ≡ (if B then (S while B do S)) Ki →T
[with BJBKρ = true]
hρ0 , ρ, (while B do S)K, t0 ≡ ǫ, S0 ≡ S(while B do S)Ki →T
hρ1 , (while B do S)K, t1 ≡ c1 , S1 i →T
· · · →T
hρn , (while B do S)K, tn ≡ tn−1 cn , Sn i →T
hρn+1 , ρ′ , (while B do S)K, t ≡ tn cn+1 , Sn+1 ≡ (while B do S)Ki
where n ≥ 0, so that the body S is assumed to be nonempty, i.e., S 6= ǫ (there is no
loss of generality since for S = ǫ the result trivially holds). Hence, t = c1 ...cn cn+1 , for
some commands ci ∈ Cmd, and the corresponding hot path hpt ≡ H−2 H−1 H0 ...Hn as
determined by Lemma 10.7 is as follows:
H−2 , l(S−2 ) : skip → l(S−1 )
H−1 , l(S−1 ) : B → l(S0 )
[because BJBKρ = true]
H0 , l(S0 ) : A0 → l(S1 )
H1 , l(S1 ) : A1 → l(S2 )
···
Hn , l(Sn ) : An → l(S−2 )
where the action Ai , with i ∈ [0, n], and the command ci+1 depend on the first command
of the statement Si as follows (this range of cases will be later referred to as (∗)):
(1) Si ≡ skip; J ⇒ Ai ≡ skip & ci+1 ≡ skip; & Si+1 ≡ J
(2) Si ≡ x := E; J ⇒ Ai ≡ x := E & ci+1 ≡ x := E; & Si+1 ≡ J
(3) Si ≡ (if B ′ then S ′ )J & BJB ′ Kρi = true ⇒
Ai ≡ B ′ & ci+1 ≡ bail ¬B ′ to J & Si+1 ≡ S ′ J
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:45
(4) Si ≡ (if B ′ then S ′ )J & BJB ′ Kρi = false ⇒
Ai ≡ ¬B ′ & ci+1 ≡ bail B ′ to (S ′ J) & Si+1 ≡ J
(5) Si ≡ (while B ′ do S ′ )J & (while B ′ do S ′ )J 6= (while B do S)K ⇒
Ai ≡ skip & ci+1 ≡ skip; & Si+1 ≡ (if B ′ then (S ′ (while B ′ do S ′ )))J
If, for any i ∈ [0, n], Hi is not a conditional command then, by case (1) of Definition 10.9,
we have that extr GP
hpt (C((while B do S)K)) = C((while B do S)K). Also, for any i ∈
[0, n], Ai is either a skip or an assignment, so that ci+1 = Ai , and, in turn, t = S. Hence,
(while B do t)K ≡ (while B do S)K, so that the thesis follows trivially.
Thus, we assume that Hk , with k ∈ [0, n], is the first conditional command occuring in
the sequence H0 ...Hn . Case (2) of Definition 10.9 applies, so that:
extr GP
hpt (C((while B do S)K)) = C((while B do S)K) ∪
{ℓ−2 : skip → ℓ−1 , ℓ−1 : B → ℓ0 , ℓ−1 : ¬B → l(K),
ℓ0 : A0 → ℓ1 , ..., ℓn : An → ℓ−2 } ∪
{ℓi : ¬Ai → l(Snext(i) )c | i ∈ [0, n], Ai ∈ BExp}.
Moreover, we have that:
C((while B do S)K) =
l((while B do S)K) : skip → l((if B then (S (while B do S))) K),
l((if B then (S (while B do S))) K) : B → l(S(while B do S)K),
l((if B then (S (while B do S))) K) : ¬B → l(K)
∪ C(S(while B do S)K) ∪ C(K)
C((while B do t)K) =
l((while B do t)K) : skip → l((if B then (t (while B do t))) K),
l((if B then (t (while B do t))) K) : B → l(t(while B do t)K),
l((if B then (t (while B do t))) K) : ¬B → l(K)
∪ C(t(while B do t)K) ∪ C(K)
extr GP
We first show that C((while B do t)K) ⊆/∼
hpt (C((while B do S)K)). We consider
=
the following label renaming:
l((while B do t)K) 7→ ℓ−2
l((if B then (t (while B do t))) K) 7→ ℓ−1
l(t(while B do t)K) 7→ ℓ0
extr GP
so that it remains to show that C(t(while B do t)K) ⊆/∼
hpt (C((while B do S)K)).
=
Since t = c1 t′ , with t′ = c2 ...cn+1 , let us analyze the five different cases for the first
command c1 of t.
(i) c1 ≡ x := E;. Thus, S0 ≡ x := E; T (while B do S)K, S1 ≡ T (while B do S)K,
A0 ≡ x := E. In this case,
C(t(while B do t)K) = {l(t(while B do t)K) : x := E → l(t′ (while B do t)K)}
∪ C(t′ (while B do t)K).
Hence, it is enough to consider the relabeling l(t′ (while B do t)K) 7→ ℓ1 and to show
extr GP
that C(t′ (while B do t)K) ⊆/∼
hpt (C((while B do S)K)).
=
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:46
S. Dissegna et al.
(ii) c1 ≡ skip; and S0 ≡ skip; T (while B do S)K. Thus, S1 ≡ T (while B do S)K, so
that A0 ≡ skip. This case is analogous to the previous case (i).
(iii) c1 ≡ skip; and S0 ≡ (while B ′ do S ′ )T (while B do S)K. Thus, S1 ≡
(if B ′ then (S ′ (while B ′ do S ′ )))T (while B do S)K and A0 ≡ skip. Here, we have
that
C(t(while B do t)K) = {l(t(while B do t)K) : skip → l(t′ (while B do t)K)}
∪ C(t′ (while B do t)K).
Again, it is enough to consider the relabeling l(t′ (while B do t)K) 7→ ℓ1 and to show
extr GP
that C(t′ (while B do t)K) ⊆/∼
hpt (C((while B do S)K)).
=
′
(iv) c1 ≡ bail ¬B to (T (while B do S)K), with S0 ≡ (if B ′ then S ′ )T (while B do S)K
and BJB ′ Kρ0 = true, so that S1 ≡ S ′ T (while B do S)K and A0 ≡ B ′ . In this case:
C(t(while B do t)K) = {l(t(while B do t)K) : ¬B ′ → l(T (while B do S)K),
l(t(while B do t)K) : B ′ → l(t′ (while B do t)K)}
∪ C(T (while B do S)K) ∪ C(t′ (while B do t)K),
C(S(while B do S)K) = {l(S(while B do S)K) : B ′ → l(S ′ T (while B do S)K),
l(S(while B do S)K) : ¬B ′ → l(T (while B do S)K)}
∪ C(S ′ T (while B do S)K) ∪ C(T (while B do S)K).
Hence, since l(t(while B do t)K) 7→ ℓ0 and A0 ≡ B ′ , it is enough to consider the
relabeling l(t′ (while B do t)K) 7→ ℓ1 and to show that C(t′ (while B do t)K) ⊆/∼
=
extr GP
(C((while
B
do
S)K)).
hpt
(v) c1 ≡ bail B ′ to (S ′ T (while B do S)K), with S0 ≡ (if B ′ then S ′ )T (while B do S)K
and BJB ′ Kρ0 = false, so that S1 ≡ T (while B do S)K and A0 ≡ ¬B ′ . In this case:
C(t(while B do t)K) = {l(t(while B do t)K) : B ′ → l(S ′ T (while B do S)K),
l(t(while B do t)K) : ¬B ′ → l(t′ (while B do t)K)}
∪ C(S ′ T (while B do S)K) ∪ C(t′ (while B do t)K),
while C(S(while B do S)K) is the same as in the previous point (iv). Hence,
since l(t(while B do t)K) 7→ ℓ0 and A0 ≡ ¬B ′ , it is enough to consider the relabeling l(t′ (while B do t)K) 7→ ℓ1 and to show that C(t′ (while B do t)K) ⊆/∼
=
extr GP
hpt (C((while B do S)K)).
Thus, in order to prove this containment, it remains to show that
′
extr GP
C(t′ (while B do t)K) ⊆/∼
hpt (C((while B do S)K)). If t = ǫ then the con=
GP
tainment boils down to C((while B do t)K) ⊆/∼
extr hpt (C((while B do S)K)) which
=
is therefore proved. Otherwise, t′ = c2 t′′ , so that the containment can be inductively
proved by using the same five cases (i)-(v) above.
Let us now show the reverse containment, that is, extr GP
hpt (C((while B do S)K)) ⊆/∼
=
C((while B do t)K). For the trace t = c1 c2 ...cn+1 , we know by (∗) that each command ci
either is in {skip; , x := E; } or is one of the two following bail commands (cf. cases (3)
and (4) in (∗)):
bail ¬B ′ to (T (while B do S)K),
bail B ′ to (S ′ T (while B do S)K).
Furthermore, at least a bail command occurs in t because there exists at least a conditional command Hk in hpt . Let ck , with k ∈ [1, n + 1], be the first bail command occurring in t. Thus, since the sequence c1 ...ck−1 consists of skip and assignment commands
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:47
only, we have that C(t(while B do t)K) ⊇ C(ck ...cn+1 (while B do t)K). Hence, either
C(ck ...cn+1 (while B do t)K) ⊇ C(T (while B do S)K) or C(ck ...cn+1 (while B do t)K) ⊇
C(T (while B do S)K). In both cases, we obtain that C(ck ...cn+1 (while B do t)K) ⊇
C((while B do S)K), so that C((while B do S)K) ⊆ C(t(while B do t)K) ⊆
C((while B do t)K). Thus, it remains to show that
{ℓ−2 : skip → ℓ−1 , ℓ−1 : B → ℓ0 , ℓ−1 : ¬B → l(K)} ∪
{ℓi : Ai → ℓnext(i) | i ∈ [0, n]} ∪ {ℓi : ¬Ai → l(Snext(i) )c | i ∈ [0, n], Ai ∈ BExp}
is contained in C(t(while B do t)K). We consider the following label renaming:
ℓ−2 →
7 l((while B do t)K)
ℓ−1 →
7 l((if B then (t (while B do t))) K)
ℓ0 →
7 l(t(while B do t)K)
so that it remains to check that for any i ∈ [0, n], the commands ℓi : Ai → ℓnext(i) and
ℓi : ¬Ai → l(Snext(i) )c , when Ai ∈ BExp, are in C(t(while B do t)K). We analyze the
possible five cases listed in (∗) for the action A0 :
(i) A0 ≡ skip because S0 ≡ skip; T (while B do S)K. Here, t = skip; t′ . Hence,
l(t(while B do t)K) : skip → l(t′ (while B do t)K) ∈ C(t(while B do t)K) and
it is enough to use the relabeling ℓ1 7→ l(t′ (while B do t)K).
(ii) A0 ≡ x := e because S0 ≡ x := E; T (while B do S)K, so that t ≡ x := E; t′ .
Analogous to case (i).
(iii) A0 ≡ skip because S0 ≡ (while B ′ do S ′ )T (while B do S)K. Here, t =
skip; t′ . Here, again, l(t(while B do t)K) : skip → l(t′ (while B do t)K) ∈
C(t(while B do t)K), so that it is enough to use the relabeling ℓ1 7→
l(t′ (while B do t)K).
(iv) A0 ≡ B ′ because S0 ≡ (if B ′ then S ′ )T (while B do S)K and BJB ′ Kρ0 = true. Thus,
t = (bail ¬B ′ to (T (while B do S)K))t′ and S1 ≡ T (while B do S)K. Note that
l(S1 )c = l(T (while B do S)K). Hence,
l(t(while B do t)K) : ¬B ′ → l(T (while B do S)K) ∈ C(t(while B do t)K),
l(t(while B do t)K) : B ′ → l(t′ (while B do t)K) ∈ C(t(while B do t)K).
Once again, the relabeling ℓ1 7→ l(t′ (while B do t)K) allows us to obtain that ℓ0 :
B ′ → ℓ1 and ℓ0 : ¬B ′ → l(T (while B do S)K) are in C(t(while B do t)K).
(v) A0 ≡ ¬B ′ because S0 ≡ (if B ′ then S ′ )T (while B do S)K and BJB ′ Kρ0 = false.
Here, t = (bail B ′ to (S ′ T (while B do S)K))t′ and S1 ≡ T (while B do S)K. Note
that l(S1 )c = l(S ′ T (while B do S)K). Hence,
l(t(while B do t)K) : B ′ → l(S ′ T (while B do S)K) ∈ C(t(while B do t)K),
l(t(while B do t)K) : ¬B ′ → l(t′ (while B do t)K) ∈ C(t(while B do t)K).
Thus, through the relabeling ℓ1 7→ l(t′ (while B do t)K) we obtain that ℓ0 : ¬B ′ → ℓ1
and ℓ0 : B ′ → l(S ′ T (while B do S)K) are in C(t(while B do t)K).
This case analysis (i)-(v) for the action A0 can be iterated for all the other actions Ai ,
with i ∈ [1, n], and this allows us to close the proof.
Finally, we can also state the correctness of the GP trace extraction transform for
the store changes abstraction as follows.
T HEOREM 10.12 (C ORRECTNESS OF GP TRACE EXTRACTION ). For any P ∈
GP
Program, hp = C0 · · · Cn ∈ αGP
hot (TJP K), we have that αsc (TJextr hp (P )K) = αsc (TJP K).
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:48
S. Dissegna et al.
The proof of Theorem 10.12 is omitted, since it is a conceptually straightforward
adaptation of the proof technique for the analogous Theorem 6.2 on the correctness
of trace extraction. Let us observe that since αsc is a stronger abstraction than αρsc
and, by Theorem 10.6, we know that αρsc characterizes bisimilarity, we obtain the socalled Stitch lemma in [Guo and Palsberg 2011, Lemma 3.6] as a straight consequence
ρ
of Theorem 10.12: αρsc (TJextr GP
hp (P )K) = αsc (TJP K).
11. CONCLUSION AND FURTHER WORK
This article put forward a formal model of tracing JIT compilation which allows:
(1) an easy definition of program hot paths—that is, most frequently executed program traces; (2) to prove the correctness of a hot path extraction transform of programs; (3) to prove the correctness of dynamic optimizations confined to hot paths,
such as dynamic type specialization along a hot path. Our approach is based on two
main ideas: the use of a standard trace semantics for modeling the behavior of programs and the use of abstract interpretation for defining the notion of hot path as
an abstraction of the trace semantics and for proving the correctness of hot path extraction and optimization. We have shown that this framework is more flexible than
Guo and Palsberg [2011] model of tracing JIT compilation, which relies on a notion of
correctness based on operational program bisimulations, and allows to overcome some
limitations of [Guo and Palsberg 2011] on selection and annotation of hot paths and on
the correctness of optimizations such as dead store elimination. We expect that most
optimizations employed by tracing JIT compilers can be formalized and proved correct
using the proof methodology of our framework.
We see a number of interesting avenues for further work on this topic. As a significant example of optimization implemented by a practical tracing compiler, it would be
worth to cast in our model the allocation removal optimization for Python described
by Bolz et al. [2011] in order to formally prove its correctness. Then, we think that our
framework could be adapted in order to provide a model of whole-method just-in-time
compilation, as used, e.g., by IonMonkey [Mozilla Foundation 2013], the current JIT
compilation scheme in the Firefox JavaScript engine. Finally, the main ideas of our
model could be useful to study and relate the foundational differences between traditional static vs dynamic tracing compilation.
ACKNOWLEDGMENTS
We are grateful to the anonymous referees for their helpful comments.
REFERENCES
K. Adams, J. Evans, B. Maher, G. Ottoni, A. Paroski, B. Simmers, E. Smith, and O. Yamauchi. 2014.
The Hiphop virtual machine. In Proceedings of the 2014 ACM International Conference on Object
Oriented Programming Systems Languages (OOPSLA 2014). ACM, New York, NY, USA, 777–790.
DOI:http://dx.doi.org/10.1145/2660193.2660199
V. Bala, E. Duesterwald, and S. Banerjia. 2000. Dynamo: a transparent dynamic optimization system. In
Proceedings of the ACM SIGPLAN Conference on Programming Language Design and Implementation
(PLDI 2000). ACM, New York, NY, USA, 1–12. DOI:http://dx.doi.org/10.1145/349299.349303
R. Barbuti, N. De Francesco, A. Santone, and G. Vaglini. 1999. Abstract interpretation
of trace semantics for concurrent calculi. Inform. Process. Lett. 70, 2 (1999), 69–78.
DOI:http://dx.doi.org/10.1016/S0020-0190(99)00042-3
S. Bauman, R. Bolz, C.F. Hirschfeld, V. Krilichev, T. Pape, J.G. Siek, and S. Tobin-Hochstadt. 2015. Pycket: A tracing JIT for a functional language. In Proceedings of the 20th ACM SIGPLAN International Conference on Functional Programming (ICFP 2015). ACM, New York, NY, USA, 22–34.
DOI:http://dx.doi.org/10.1145/2784731.2784740
M. Bebenita, F. Brandner, M. Fahndrich, F. Logozzo, W. Schulte, N. Tillmann, and H. Venter. 2010. SPUR: a
trace-based JIT compiler for CIL. In Proceedings of the ACM International Conference on Object Oriented
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
An Abstract Interpretation-based Model of Tracing Just-In-Time Compilation
A:49
Programming Systems Languages and Applications (OOPSLA 2010). ACM, New York, NY, USA, 708–
725. DOI:http://dx.doi.org/10.1145/1869459.1869517
I. Böhm, T.J.K. Edler von Koch, S.C. Kyle, B. Franke, and N. Topham. 2011. Generalized just-in-time trace
compilation using a parallel task farm in a dynamic binary translator. In Proceedings of the 32nd ACM
SIGPLAN Conference on Programming Language Design and Implementation (PLDI 2011). ACM, New
York, NY, USA, 74–85. DOI:http://dx.doi.org/10.1145/1993498.1993508
C.F. Bolz, A. Cuni, M. Fijałkowski, M. Leuschel, S. Pedroni, and A. Rigo. 2011. Allocation removal by
partial evaluation in a tracing JIT. In Proceedings of the 20th ACM SIGPLAN Workshop on Partial Evaluation and Program Manipulation (PEPM 2011). ACM, ACM, New York, NY, USA, 43–52.
DOI:http://dx.doi.org/10.1145/1929501.1929508
C.F. Bolz, A. Cuni, M. Fijalkowski, and A. Rigo. 2009. Tracing the meta-level: PyPy’s tracing JIT compiler. In Proceedings of the 4th Workshop on the Implementation, Compilation, Optimization of ObjectOriented Languages and Programming Systems (ICOOOLPS 2009). ACM, New York, NY, USA, 18–25.
DOI:http://dx.doi.org/10.1145/1565824.1565827
C. Colby and P. Lee. 1996. Trace-based program analysis. In Proceedings of the 23rd ACM SIGACTSIGPLAN Symposium on Principles of Programming Languages (POPL 1996). ACM, New York, NY,
USA, 195–207. DOI:http://dx.doi.org/10.1145/237721.237776
P. Cousot. 1997. Constructive design of a hierarchy of semantics of a transition system by abstract interpretation (Extended Abstract). Electronic Notes in Theoretical Computer Science 6, 0 (1997), 77–102.
DOI:http://dx.doi.org/10.1016/S1571-0661(05)80168-9 Proceedings of the 13th Annual Conference on
Mathematical Foundations of Progamming Semantics (MFPS XIII).
P. Cousot. 2002. Constructive design of a hierarchy of semantics of a transition system by abstract interpretation. Theoretical Computer Science 277, 1-2 (2002), 47–103.
P. Cousot and R. Cousot. 1977. Abstract interpretation: a unified lattice model for static analysis of programs
by construction or approximation of fixpoints. In Proceedings of the 4th ACM SIGACT-SIGPLAN Symposium on Principles of Programming Languages (POPL 1977). ACM, New York, NY, USA, 238–252.
DOI:http://dx.doi.org/10.1145/512950.512973
P. Cousot and R. Cousot. 1979. Systematic design of program analysis frameworks. In Proceedings of the
6th ACM SIGACT-SIGPLAN Symposium on Principles of Programming Languages (POPL 1979). ACM,
New York, NY, USA, 269–282. DOI:http://dx.doi.org/10.1145/567752.567778
P. Cousot and R. Cousot. 2002. Systematic design of program transformation frameworks by
abstract interpretation. In Proceedings of the 29th ACM SIGACT-SIGPLAN Symposium on
Principles of Programming Languages (POPL 2002). ACM, New York, NY, USA, 178–190.
DOI:http://dx.doi.org/10.1145/503272.503290
S. Dissegna, F. Logozzo, and F. Ranzato. 2014. Tracing compilation by abstract interpretation. In Proceedings
of the 41st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL 2014).
ACM, New York, NY, USA, 47–59. DOI:http://dx.doi.org/10.1145/2535838.2535866
Ecma International. 2015. Standard ECMA-262, ECMAScript 2015 Language Specification, 6th Edition.
http://www.ecma-international.org/ecma-262/6.0. (2015).
Facebook Inc. 2013. The HipHop Virtual Machine. (Oct. 2013). https://www.facebook.com/hhvm.
A. Gal, B. Eich, M. Shaver, D. Anderson, D. Mandelin, M.R. Haghighat, B. Kaplan, G. Hoare, B. Zbarsky,
J. Orendorff, J. Ruderman, E.W. Smith, R. Reitmaier, M. Bebenita, M. Chang, and M. Franz. 2009.
Trace-based just-in-time type specialization for dynamic languages. In Proceedings of the 2009 ACM
SIGPLAN Conference on Programming Language Design and Implementation (PLDI 2009). ACM, New
York, NY, USA, 465–478.
A. Gal, C.W. Probst, and M. Franz. 2006. HotPathVM: an effective JIT compiler for resource-constrained
devices. In Proceedings of the 2nd International Conference on Virtual Execution Environments (VEE
2006). ACM, ACM, New York, NY, USA, 144–153. DOI:http://dx.doi.org/10.1145/1542476.1542528
Google Inc. 2010. A new crankshaft for V8. (Dec. 2010). The Chromium Blog.
S. Guo and J. Palsberg. 2011. The essence of compiling with traces. In Proceedings of the 38th ACM SIGACTSIGPLAN Symposium on Principles of Programming Languages (POPL 2011). ACM, New York, NY,
USA, 563–574. DOI:http://dx.doi.org/10.1145/1926385.1926450
M. Handjieva and S. Tzolovski. 1998. Refining static analyses by trace-based partitioning using control
flow. In Proceedings of the 5th International Static Analysis Symposium (SAS 1998) (LNCS), Vol. 1503.
Springer, Berlin, Germany, 200–214. DOI:http://dx.doi.org/10.1007/3-540-49727-7 12
C. Häubl and H. Mössenböck. 2011. Trace-based compilation for the Java HotSpot virtual machine. In Proceedings of the 9th International Conference on Principles and Practice of Programming in Java (PPPJ
2011). ACM, New York, NY, USA, 129–138. DOI:http://dx.doi.org/10.1145/2093157.2093176
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
A:50
S. Dissegna et al.
C. Häubl, C. Wimmer, and H. Mössenböck. 2014. Trace transitioning and exception handling in a tracebased JIT compiler for Java. ACM Trans. Archit. Code Optim. 11, 1, Article 6 (Feb. 2014), 26 pages.
DOI:http://dx.doi.org/10.1145/2579673
H. Inoue, H. Hayashizaki, Peng Wu, and T. Nakatani. 2011. A trace-based Java JIT compiler retrofitted
from a method-based compiler. In 9th Annual IEEE/ACM International Symposium on Code Generation and Optimization (CGO 2011). IEEE Computer Society, Washington, DC, USA, 246–256.
DOI:http://dx.doi.org/10.1109/CGO.2011.5764692
F. Logozzo. 2009. Class invariants as abstract interpretation of trace semantics. Computer Languages, Systems and Structures 35, 2 (2009), 100–142. DOI:http://dx.doi.org/10.1016/j.cl.2005.01.001
R. Milner. 1995. Communication and Concurrency. Prentice Hall, Englewood Cliffs, NJ.
Mozilla Foundation. 2010. TraceMonkey. (Oct. 2010). MozillaWiki.
Mozilla Foundation. 2013. IonMonkey. (May 2013). MozillaWiki.
M. Pall. 2005. The LuaJIT Project. http://luajit.org. (2005).
X. Rival. 2004. Symbolic transfer function-based approaches to certified compilation. In Proceedings of the
31st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL ’04). ACM,
New York, NY, USA, 1–13. DOI:http://dx.doi.org/10.1145/964001.964002
X. Rival and L. Mauborgne. 2007. The trace partitioning abstract domain. ACM Trans. Program. Lang. Syst.
29, 5, Article 26 (2007), 51 pages. DOI:http://dx.doi.org/10.1145/1275497.1275501
T. Schilling. 2013. Trace-based Just-In-Time Compilation for Lazy Functional Programming Languages.
Ph.D. Dissertation. University of Kent, UK.
D.A. Schmidt. 1998. Trace-based abstract interpretation of operational semantics. Lisp Symb. Comput. 10,
3 (1998), 237–271. DOI:http://dx.doi.org/10.1023/A:1007734417713
F. Spoto and T. Jensen. 2003. Class analyses as abstract interpretations of trace semantics. ACM Trans.
Program. Lang. Syst. 25, 5 (2003), 578–630. DOI:http://dx.doi.org/10.1145/937563.937565
M.N. Wegman and F.K. Zadeck. 1991. Constant propagation with conditional branches. ACM Trans. Program. Lang. Syst. 13, 2 (1991), 181–210. DOI:http://dx.doi.org/10.1145/103135.103136
ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Article A, Publication date: January YYYY.
| 6cs.PL
|
Safe Exploration in Continuous Action Spaces
Gal Dalal 1 2 Krishnamurthy Dvijotham 2 Matej Vecerik 2 Todd Hester 2 Cosmin Paduraru 2 Yuval Tassa 2
arXiv:1801.08757v1 [cs.AI] 26 Jan 2018
Abstract
We address the problem of deploying a reinforcement learning (RL) agent on a physical system
such as a datacenter cooling unit or robot, where
critical constraints must never be violated. We
show how to exploit the typically smooth dynamics of these systems and enable RL algorithms
to never violate constraints during learning. Our
technique is to directly add to the policy a safety
layer that analytically solves an action correction formulation per each state. The novelty of
obtaining an elegant closed-form solution is attained due to a linearized model, learned on past
trajectories consisting of arbitrary actions. This
is to mimic the real-world circumstances where
data logs were generated with a behavior policy
that is implausible to describe mathematically;
such cases render the known safety-aware offpolicy methods inapplicable. We demonstrate the
efficacy of our approach on new representative
physics-based environments, and prevail where
reward shaping fails by maintaining zero constraint violations.
1. Introduction
In the past two decades, RL has been mainly explored in
toy environments (Sutton & Barto, 1998) and video games
(Mnih et al., 2015), where real-world applications were
limited to a few typical use-cases such as recommender
systems (Shani et al., 2005). However, RL is recently also
finding its path into industrial applications in the physical world; e.g., datacenter cooling (Evans & Gao, 2016),
robotics (Gu et al., 2016), and autonomous vehicles (Sallab
et al., 2017). In all these use-cases, safety is a crucial concern: unless safe operation is addressed thoroughly and ensured from the first moment of deployment, RL is deemed
incompatible for them.
In real-world applications such as the above, constraints are
an integral part of the problem description, and never vio1
Technion, Israel Institute of Technology 2 Google DeepMind.
Correspondence to: Gal Dalal <[email protected]>.
lating them is often a strict necessity. Therefore, in this
work, we define our goal to be maintaining zero-constraintviolations throughout the whole learning process. Note that
accomplishing this goal for discrete action spaces is more
straightforward than for continuous ones. For instance, one
can pre-train constraint-violation classifiers on offline data
for pruning unsafe actions. However, in our context, this
goal becomes considerably more challenging due to the infinite number of candidate actions. Nevertheless, we indeed manage to accomplish this goal for continuous action
spaces and show to never violate constraints throughout the
whole learning process.
Specifically, we tackle the problem of safe control in physical systems, where certain observable quantities are to be
kept constrained. To illustrate, in the case of datacenter
cooling, temperatures and pressures are to be kept below
respective thresholds at all times; a robot must not exceed
limits on angles and torques; and an autonomous vehicle
must always maintain its distance from obstacles above
some margin. In this work, we denote these quantities as
safety signals. As these are physical quantities, we exploit
their smoothness for avoiding unexpected, unsafe operation. Moreover, we deal with the common situation where
offline logged data are available; it thus can be used to pretrain models for aiding safety from the initial RL deployment moments.
Safe exploration, as depicted above, traditionally requires
access to data generated with some known behavior policy upon which gradual safe updates are performed; see
(Thomas, 2015) for a comprehensive study of such offpolicy methods. Such data are necessary because, unless
assumed otherwise, actions in one state might have catastrophic consequences down the road. Hence, long-term
behavior should be inferred in advance prior to deployment. In contrast, in this work, we eliminate the need
for behavior-policy knowledge as we focus on physical
systems, whose actions have relatively short-term consequences. Obviating behavior-policy knowledge is a key
benefit in our work, as lack of such data is a challenging
yet familiar real-world situation. It is rarely the case that
past trajectories in complex systems were generated using
a consistent behavior policy that can be mathematically described. Such systems are traditionally controlled by humans or sophisticated software whose logic is hard to por-
Safe Exploration in Continuous Action Spaces
tray. Hence, off-policy RL methods are deemed inapplicable in such situations. Contrarily, we show how singlestep transition data can be efficiently exploited for ensuring safety. To demonstrate our method’s independence of
a behavior policy, in our experiments we generate our pretraining data with purely random actions.
Our approach relies on one-time initial pre-training of a
model that predicts the change in the safety signal over a
single time step. This model’s strength stems from its simplicity: it is a first-order approximation with respect to the
action, where its coefficients are the outputs of a state-fed
neural network (NN). We then utilize this model in a safety
layer that is composed directly on top the agent’s policy to
correct the action if needed; i.e., after every policy query,
it solves an optimization problem for finding the minimal
change to the action such that the safety constraints are met.
Thanks to the linearity with respect to actions, the solution
can be derived analytically in closed-form and amounts to
basic arithmetic operations. Thus, our safety layer is both
differentiable and has a trivial three-line software implementation. Note that relating to our safety mechanism as a
‘safety layer’ is purely a semantical choice; it merely is
a simple calculation that is not limited to the nowadays
popular deep policy networks and can be applied to any
continuous-control algorithm (not necessarily RL-based).
2. Related Work
As this work focuses on control problems with continuous
state and action spaces, we limit our comparison to the literature on safe RL in the context of policy optimization that
attempts to maintain safety also during the learning process. Such an example is (Achiam et al., 2017), where
constrained policy optimization was solved with a modified trust-region policy gradient. There, the algorithm’s
update rule projected the policy to a safe feasibility set in
each iteration. Under some policy regularity assumptions,
it was shown to keep the policy within constraints in expectation. As such, it is unsuitable to our use-cases, where
safety must be ensured for all visited states. Another recent
work (Berkenkamp et al., 2017) described control-theoretic
conditions under which safe operation can be guaranteed
for a discretized deterministic control framework. An appropriate Lyapunov function was identified for policy attraction regions if certain Lipschitz continuity conditions
hold. Though under the appropriate conditions safe exploration was guaranteed, knowledge on the specific system
was required. Moreover, a NN may not be Lipschitz continuous with a reasonable coefficient. Lastly, very recent
work (Pham et al., 2017) utilized an in-graph QP solver
first introduced in (Amos & Kolter, 2017). It exhibited an
approach similar in nature to the one we take: solve an optimization problem in the policy-level, for ensuring safety
on a state-wise basis. However, two main differences are
to be noticed. First, the solution there relied on an in-graph
implementation of a full QP solver (Amos & Kolter, 2017),
which runs an iterative interior-point algorithm with each
forward propagation. This is both a challenge to implement
(currently only a pytorch version is available), and computationally expensive. An additional difference from our
work is that in (Pham et al., 2017) expert knowledge was
required to explicitly hand-design the physical constraints
of a robotic arm. Contrarily, in our method no such manual
effort is needed; these dynamics are learned directly from
data, while also being behavior-policy independent.
To summarize, this work is the first, to our knowledge, to
solve the problem of state-wise safety directly at the policy level, while also doing it in a data-driven fashion using arbitrary data logs. Moreover, it can be applied to any
continuous-control algorithm; it is not restricted to a specific RL algorithm or any at all.
3. Definitions
We consider a special case of constrained Markov decision
processes (CMDP) (Altman, 1999), where observed safety
signals should be kept bounded. First, let us denote by [K]
+
the set {1, . . . , K}, and by [x] the operation max{x, 0},
where x ∈ R. A CMDP is a tuple (S, A, P, R, γ, C), where
S is a state space, A is an action space, P : S × A × S →
[0, 1] is a transition kernel, R : S × A → R is a reward
function, γ ∈ (0, 1) is a discount factor, and C = {ci :
S × A → R | i ∈ [K]} is a set of immediate-constraint
functions. Based on that, we also define a set of safety signals C¯ = {c̄i : S → R | i ∈ [K]}. These are per-state observations of the immediate-constraint values, which we introduce for later ease of notation. To illustrate, if c1 (s, a) is
the temperature in a datacenter to be sensed after choosing
a in s, c̄1 (s0 ) is the same temperature sensed in s0 after transitioning to it. In the type of systems tackled in this work,
P is deterministic and determines f s.t. s0 = f (s, a). Thus,
we have c̄i (s0 ) , ci (s, a). For a general non-deterministic
transition kernel, c̄i (s0 ) can be defined as the expectation
over s0 ∼ P (·|s, a). Lastly, let policy µ : S → A be a
stationary mapping from states to actions.
4. State-wise Constrained Policy Optimization
We study safe exploration in the context of policy optimization, where at each state, all safety signals c̄i (·) are upper
bounded by corresponding constants Ci ∈ R:
∞
X
max E[
γ t R(st , µθ (st ))]
θ
s.t.
t=0
c̄i (st ) ≤ Ci ∀i ∈ [K] ,
where µθ is a parametrized policy.
(1)
Safe Exploration in Continuous Action Spaces
We stress that our goal is to ensure state-wise constraints
not only for the solution of (1), but also for its optimization
process. This goal might be intractable in general since for
an arbitrary MDP some actions can have a long-term effect in terms of possible state paths. However, for the types
of physical systems we consider it is indeed plausible that
safety constraints can be ensured by adjusting the action in
a single (or few) time step(s). In the context of our realworld use-cases, cooling system dynamics are governed
by factors such as the first-order heat-transfer differential
equation (Goodwine, 2010), and the second-order Newton
differential equation that governs the water mass transfer.
The latter also governs the movement of a robotic arm or a
vehicle on which one applies forces. In these types of control problems, it is feasible to satisfy state-wise constraints
even in the presence of inertia given reasonable slack in
the choice of Ci . We expand on this further and provide
evidence in Section 7.
5. Linear Safety-Signal Model
Figure 1. Each safety signal ci (s, a) is approximated with a linear model with respect to a, whose coefficients are features of s,
extracted with a NN.
fields, e.g. aircraft design (Liang & Sun, 2013). For a comprehensive study see (Enqvist, 2005).
Given a policy-oblivious set of tuples D = {(sj , aj , s0j )},
we train g(s; wi ) by solving
arg min
wi
Solving (1) is a difficult task, even for the types of systems
listed above. A major contributor to this challenge is the
RL agent’s intrinsic need to explore for finding new and
improved actions. Without prior knowledge on its environment, an RL agent initialized with a random policy cannot ensure per-state constraint satisfaction during the initial
training stages. This statement also holds when the reward
is carefully shaped to penalize undesired states: for an RL
agent to learn to avoid undesired behavior it will have to
violate the constraints enough times for the negative effect
to propagate in our dynamic programming scheme.
In this work, we thus incorporate some basic form of prior
knowledge, based on single-step dynamics. Single-step
transition data in logs is rather common, and, as explained
before, more realistic compared to also knowing behavior policies. We do not attempt to learn the full transition model, but solely the immediate-constraint functions
ci (s, a). While it is attractive to simply approximate them
with NNs that take (s, a) as inputs, we choose a more elegant approach that comes with significant advantages listed
in Subsection 6.1. Namely, we perform the following linearization:
X
c̄i (s0 ) − (c̄i (s) + g(s; wi )> a)
2
, (2)
(s,a,s0 )∈D
where c̄i (s) is assumed to be included in s.
In our experiments, to generate D we merely initialize the
agent in a uniformly random location and let it perform uniformly random actions for multiple episodes. The episodes
terminate when a time limit is reached or upon constraint
violation. The latter corresponds to real-world mechanisms
governing production systems: high-performance, efficient
control is often backed up by an inefficient conservative
policy; when unsafe operation flags are raised, a preemption mechanism is triggered and the conservative policy
kicks in to ensure constraint satisfaction.
Training g(s; wi ) on D is performed once per task as a pretraining phase that precedes the RL training. However, additional continual training of g(s; wi ) during RL training is
also optional. Since in our experiments continual training
showed no benefit compared to solely pre-training, we only
show results of the latter.
6. Safety Layer via Analytical Optimization
where wi are weights of a NN, g(s; wi ), that takes s as
input and outputs a vector of the same dimension as a. This
model is a first-order approximation to ci (s, a) with respect
to a; i.e., an explicit representation of sensitivity of changes
in the safety signal to the action using features of the state.
See Fig. 1 for a visualization.
We now show how to solve problem (1) using the policy
gradient algorithm (Baxter & Bartlett, 2001) via a simple
addition to the policy itself. We experiment with Deep Deterministic Policy Gradient (DDPG) (Lillicrap et al., 2015)
whose policy network directly outputs actions and not their
probabilities. However, our approach is not limited to it
and can be added, as is, to probabilistic policy gradient or
any other continuous-control algorithm.
Remark 1. Linear approximations of non-linear physical systems prove accurate and are well accepted in many
Denote by µθ (s) the deterministic action selected by the
deep policy network. Then, on top of the policy network
c̄i (s0 ) , ci (s, a) ≈ c̄i (s) + g(s; wi )> a,
Safe Exploration in Continuous Action Spaces
Original
policy
Safety layer
Figure 2. A safety layer is composed on top of a deep policy
network, solving an action correction optimization program with
each forward-propagation. Our linearized safety-signal model allows a closed-form solution µ̃θ (s) = arg mina f (s, a, µ(s)) that
reduces to a trivial linear projection.
we compose an additional, last layer, whose role is to solve
arg min 12 ka − µθ (s)k2
(3)
a
s.t. ci (s, a) ≤ Ci ∀i ∈ [K] .
This layer, which we refer to as safety layer, perturbs the
original action as little as possible in the Euclidean norm in
order to satisfy the necessary constraints. Fig. 2 visualizes
its relation to the policy network.
To solve (3) we now substitute our linear model for ci (s, a),
introduced in Section 5, and obtain the quadratic program
a∗ = arg min 21 ka − µθ (s)k2
(4)
a
s.t. c̄i (s) + g(s; wi )> a ≤ Ci ∀i ∈ [K] .
Moreover, for other systems with multiple intersecting constraints, a joint model can be learned. For instance, instead
of treating distances from two walls as two constraints, the
minimum between them can be treated as a single constraint. In rudimentary experiments, this method produced
similar results to not using it, when we grouped two constraints into one, in the first and simplest task in Section 7.
However, jointly modeling the dynamics of more than a
few safety signals with a single g(·; ·) network is a topic requiring careful attention, which we leave for future work.
We now provide the closed-form solution to (4).
Proposition 1. Assume there exists a feasible solution to
∗
(4) denoted by (a∗ , {λ∗i }K
i=1 ), where λi is the optimal Lagrange multiplier associated with the i-th constraint. Also,
assume |{i|λ∗i > 0}| ≤ 1; i.e., at most one constraint is
active. Then
+
g(s; wi )> µθ (s) + c̄i (s) − Ci
∗
(5)
λi =
g(s; wi )> g(s; wi )
and
a∗ = µθ (s) − λ∗i∗ g(s; wi∗ ),
(6)
where i∗ = arg maxi λ∗i .
Proof. As the objective function and constraints in (4) are
convex, a sufficient condition for optimality of a feasible
solution (a∗ , {λ∗i }K
i=1 ) is for it to satisfy the KKT conditions. The Lagrangian of (4) is
L(a,λ) =
Thanks to the positive-definite quadratic objective and linear constraints, we can now find the global solution to this
convex problem. Generally, to solve it one can implement
an in-graph iterative QP-solver such as in (Amos & Kolter,
2017). This would result in a method similar to the one
in (Pham et al., 2017), but with the advantage that all the
physical constraint model is learned directly from data instead of being hand-designed. Alternatively, if the number of active constraints is known to be bounded
by some
m ≤ K, one can exhaustively iterate on all K
m combinations of possibly active constraints and select the optimal
feasible one; this would be reasonable for a small m.
However, in this work, at the expense of one simplifying
assumption, we gain the benefit of obtaining a closed-form
analytical solution to (4) that has a trivial three-lines-ofcode software implementation. The assumption is that no
more than a single constraint is active at a time. As demonstrated in our experiments, this is reasonable to assume
when an agent navigates in a physical domain and avoids
obstacles. As the distance from each obstacle is modeled
as a separate constraint, only a single obstacle is the closest one at a time. Proximity to corners shows to pose no
issues, as can be seen in the plots and videos in Section 7.
1
2 ka
− µθ (s)k2 +
K
X
λi c̄i (s) + g(s; wi )> a − Ci ;
i=1
and hence the KKT conditions at (a∗ , {λ∗i }K
i=1 ) are
∇a L = a∗ − µθ (s) +
K
X
λ∗i g(s; wi ) = 0,
(7)
i=1
λ∗i c̄i (s) + g(s; wi )> a∗ − Ci = 0 ∀i ∈ [K] . (8)
First, consider the case where |{i|λ∗i > 0}| = 1, i.e., λ∗i∗ >
0. We then easily get (6) from (7). Next, from (8) we have
that c̄i∗ (s)+g(s; wi∗ )> a∗ −Ci∗ = 0. Substituting (6) in the
latter gives λ∗i∗ g(s; wi∗ )> g(s; wi∗ ) = g(s; wi∗ )> µθ (s) +
c̄i∗ (s) − Ci∗ . This gives us (5) when i = i∗ . As for i ∈
[K] \ {i∗ }, the corresponding constraints are inactive since
λ∗i = 0. Hence, c̄i (s) + g(s; wi )> a∗ − Ci < 0, making
the fraction in (5) negative, which indeed results in a value
of 0 due to the [·]+ operator and gives us (5) also when
i ∈ [K] \ {i∗ }.
To conclude the proof, consider the second case where
λ∗i = 0 ∀i ∈ [K]. From (7) we have that a∗ = µθ (s).
This gives us (6) since λ∗i∗ = 0. Lastly, (5) holds due to
Safe Exploration in Continuous Action Spaces
the same inactive constraints argument as above, this time
uniformly ∀i ∈ [K].
The solution (6) is essentially a linear projection of the
original action µθ (s) to the “safe” hyperplane with slope
g(s; wi∗ ) and intercept c̄i∗ (s) − Ci∗ . In terms of implementation, it consists of a few primitive arithmetic operations: vector products followed by a ’max’ operation. The
benefits of its simplicity are three-fold: i) it has a trivial,
almost effortless software implementation; ii) its computational cost is negligible; and iii) it is differentiable (almost
everywhere, as is ReLu).
6.1. An Alternative: Additional Loss Term
To stress the prominence of our linear model in solving (3),
we now briefly describe the drawbacks of an alternative
likely choice we initially experimented with: a straightforward (s, a)-fed NN model for ci (s, a). In this case, an
approximate solution to (3) can be obtained by penalizing
the objective for constraint violations and solving the unconstrained surrogate
arg min
a
1
2 ka
2
− µθ (s)k +
K
X
+
λi [ci (s, a) − Ci ]
,
i=1
(9)
where {λi > 0} are now hyper-parameters. Problem (9)
can be solved numerically using gradient descent. However, even though from our experience this approach indeed works (corrects actions to safe ones), it is inferior to
our analytic approach for several reasons:
pre-trained safety layer to the policy network. As mentioned earlier, our RL algorithm of choice for the experiments is DDPG (Lillicrap et al., 2015). In this section we
show that during training, DDPG never violates constraints
and converges faster compared to without our addition.
To mimic the physics-based use-cases described in this
work, where continuous safety signals are observations of
the state ought to be constrained, we set up appropriate simulation domains in Mujoco (Todorov et al., 2012). In these
domains, an object is located in some feasible bounded region. Each of the constraints, therefore, lower bounds the
object’s distance to each of the few boundaries. Though
the lower bound on the distance is zero by the tasks’ definition, in practice we set it to be some small positive value
to allow slack for avoidance actions in the presence of inertia. In all simulations, the episode immediately terminates in the case of a constraint violation. These conditions
comply with the ones in our real-world examples: a datacenter cooling system’s maximal temperature set for the
formulation and algorithm need not be the one originally
defined by the datacetner operator; a lower, more conservative value can be set to allow for some slack. Episode termination corresponds to preemption, followed by a swap to
some backup heuristic each time the conservative temperature cap is reached.
We now introduce our two new Mujoco domains: Ball and
Spaceship, consisting of two tasks each. The dynamics of
Ball and Spaceship are governed by first and second order
differential equations, respectively. As such, they are representative domains to the systems of interest in this paper.
7.1. Ball Domain
1. Running gradient descent with every policy query (i.e.
forward propagation) requires sophisticated in-graph
implementation and is computationally intensive.
2. Since the sensitivity of ci (s, a) varies for different entries in a, different orders of magnitude are observed
in entries of the resulting gradient ∇a ci (s, a). This
causes numerical instabilities and long convergence
times, and requires careful stepsize selection.
3. Solutions to the non-convex (9) are local minima,
which depend on non-reliable convergence of an iterative optimization algorithm, as opposed to the closedform global optimum we obtain for (4).
4. There are K hyper-parameters necessitating tuning.
7. Experiments
Per each task introduced next, we run the initial pretraining phase described in Section 5. We construct D with
1000 random-action episodes per task. We then add our
In the Ball domain, the goal is to bring a ball as close as
possible to a changing target location, by directly setting
the velocity of the ball every 4-th time-step (it is common
for the operational frequency of a typical torque controller
to be less than the environment’s one). Each episode lasts
at most 30 seconds, during which the target is appearing in
a new, uniformly random location every 2 seconds. Let us
(d)
now define the d-dimensional [a, b]-cube B[a,b] = {x|a ≤
xi ≤ b, i = 1, . . . , d}. The feasible region for the ball is
(d)
(d)
B[0,1] , while for the target it is B[0.2,0.8] . If the ball steps
(d)
out of B[0,1] , the episode terminates. To allow some slack
in maneuvering away from the boundaries, we set Ci in (3)
so as to constrain the ball’s feasible region to effectively
(d)
be B[0.1,0.9] . This is for the safety layer to start correcting
the actions once the ball steps out of the latter cube. Let
the ball location, ball velocity and target location respectively be xB , vB , xT ∈ Rd . Then, s = (xB , vB , xT + d ),
where d ∼ N (0, 0.05 · I d×d ); i.e., only a noisy estimate
of the target is observed. The action a = vB ; i.e., it is
Safe Exploration in Continuous Action Spaces
Figure 3. Ball-1D (left) and Ball-3D (right) tasks. The goal is to
keep the green ball as close as possible to the pink target ball by
setting its velocity. The safe region is the [0, 1] interval in Ball-1D
and the [0, 1] cube in Ball-3D; if the green ball steps out of it, the
episode terminates.
to set the velocity of the ball. Actions are taken every 4
time-steps and remain constant in between. The dynamics are governed by Newton’s laws, with a small amount
of damping. The reward has a maximum of 1 when the
ball is exactly at the target and quickly diminishes to 0
+
away from it: R(s, a) = 1 − 10 · kxB − xT k22 . Lastly,
γ = 0.99. Our experiments are conducted on two tasks:
Ball-1D where d = 1, Ball-3D where d = 3. Images of the
two can be found in Fig. 3.
As the ball’s velocity is controlled directly, its dynamics
are governed by a first-order differential equation. This domain thus represents phenomena such as heat transition and
several known control tasks such as maintaining the speed
of a rotating engine, whether it is a classic steam engine or
cruise-control in modern vehicles (Sotomayor et al., 2006).
Figure 4. Spaceship-Corridor (left) and Spaceship-Arena (right)
tasks. The goal is to bring the green spaceship to to the pink
rounded target by controlling its thrust engines. Touching the
walls with the spaceship’s bow terminates the episode.
when reaching the target and 0 elsewhere; and γ = 0.99.
As in the Ball domain, a small gap away from each wall is
incorporated in the choice of Ci in (3). This is for the safety
layer to begin correcting the actions a few time-steps before
the spaceship actually reaches a wall with its bow. For both
tasks, this gap is 0.05, where, for comparison, the distance
between the walls in Corridor task is 1.
Since the control is via forces, the spaceship’s dynamics
are governed by a second-order differential equation. This
domain thus represents situations such as pumping masses
of water for cooling, and actuating objects such as robotic
arms.
7.3. Implementation Details
7.2. Spaceship Domain
In the Spaceship domain, the goal is to bring a spaceship to a fixed target location by controlling its thrust engines. Hence, as opposed to setting velocities in the Ball
domain, here we set the forces. Our first task for this
domain is Spaceship-Corridor, where the safe region is
bounded between two infinite parallel walls. Our second
task, Spaceship-Arena, differs from the first in the shape of
the safe region; it is bounded by four walls in a diamond
form. Images of the two tasks are given in Fig. 4. Episodes
terminate when one of the three events occur: the target is
reached, the spaceship’s bow touches a wall, or the time
limit is reached. Time limits are 15 seconds for Corridor
and 45 seconds for Arena. Relating to the screenshots in
Fig. 4, the spaceship’s initialization location is uniformly
random in the lowest third part of the screen for Corridor,
and the right-most third part of the screen for Arena.
In this domain, the state is the spaceship’s location and velocities; the action a ∈ [−1, 1]2 actuates two thrust engines
in forward/backward and right/left directions; the transitions are governed by the rules of physics where damping
is applied; the reward is sparse: 1000 points are obtained
Per each of the four tasks above, we pre-train the safetysignal model for each of the task’s constraints; i.e., we
solve (2) for each i ∈ [K] using that task’s data-set D.
In a hyper-parameter sweep, we witnessed that the ability
to achieve our end-goal is relatively insensitive to the architecture of g(·; wi ) and a simple one suffices. Namely, for
achieving zero-constraint-violations after plugging g(·; wi )
into the safety layer, a single-layered NN with 10 hidden
neurons was adequate for all tasks. For training we use
Adam optimizer (Kingma & Ba, 2014) with a batch size of
256. We also report that low validation errors for g(·; wi )
don’t necessarily correspond to a better safety layer; it is
the exploratory nature of D that matters. Specifically, we
witnessed the following counter-intuitive behavior: generating D with a policy that fixes a single randomly drawn
action for each episode, as opposed to a random action in
each step, yields better regression accuracy for g(·; wi ) but
lesser safety layer performance.
Based on a hyper-parameter sweep for DDPG, for the actor and critic we use two-hidden-layer NNs of respective
sizes (100, 100) and (500, 500). The rest of the experiment
parameters were taken to be as in (Lillicrap et al., 2015).
Safe Exploration in Continuous Action Spaces
Figure 5. Accumulated constraint violations (lower is better) throughout the training of DDPG+reward shaping, per each task. Plotted
are medians with upper and lower quantiles of 10 seeds. The x-axis corresponds to different choice of M – the margin from limits in
which reward penalty is incurred. The optimal choice of M is colored red.
7.4. Safety Layer versus Reward Shaping
Before exhibiting the performance of our safety layer, we
first relate to a natural alternative approach for ensuring
safety: manipulate the agent to avoid undesired areas by
artificially shaping the reward. This can be done by setting the reward to large negative values in subsets of the
state-space. In our case, those areas are the neighborhood
of the enclosing boundaries. Therefore, for comparison,
we first ran a series of such reward shaping experiments on
DDPG without the safety layer. We set the penalty to be
on the same scale as the original reward: −1 in the Ball domain and −1000 in the spaceship domain. The margin from
boundaries in which the shaping occurs is a parameter M
which we cross-validated per each task. We experimented
with values of M ∈ {0.08, 0.11, 0.14, 0.17, 0.2, 0.23}. Per
each M , we ran DDPG, where each training episode is followed by an evaluation episode for which we count whether
it terminated due to a constraint violation. Fig. 5 gives, per
each M, the median with upper and lower quantiles of accumulated constraint violations of 10 seeds of DDPG runs.
Per each task, we mark the best choice of M in red.
Fig. 5 depicts the drawbacks of reward shaping for ensuring safety. The first fault is the failure to achieve our zeroconstraint-violations goal; all parameter choices resulted in
significant portions of the episodes terminating due to violations. The second drawback is the difficulty of choosing
the correct parameter. As seen in the figure, there is no
clear trend for the dependence on M ; each task has a different ’best value’, and the plots have no structure.
Next, we compare the performance of our safety layer to
that of the best reward shaping choice, and to no reward
shaping at all. Namely, we compare the following alternatives: DDPG, DDPG+reward shaping, and DDPG+safety
layer. For reward shaping we use the same simulation re-
sults of the best M (colored red in Fig. 5). The comparison
outcomes are summarized in Fig. 6. Its top row gives the
sum of discounted rewards per evaluation episode, and the
bottom provides the number of constraint violations, accumulated over all evaluation episodes. The plots show medians along with upper and lower quantiles of 10 seeds.
The most prominent insight from Fig. 6 is that the constraints were never violated with the safety layer. This is
true for all 10 seeds of each of the four tasks. Secondly,
the safety layer dramatically expedited convergence. For
Spaceship, it is, in fact, the only algorithm that enabled
convergence. In contrast, without the safety layer, a significant amount of episodes ended with a constraint violation and convergence was often not attained. This is due
to the nature of our tasks: frequent episode terminations
upon boundary crossing impede the learning process in
our sparse reward environments. However, with the safety
layer, these terminations never occur, allowing the agent to
maneuver as if there were no boundaries. Next, the following domain-specific discussion is in order.
In the Ball domain, since in the 1D task the target is always located in one of two possible directions, its reward
is less sparse compared to 3D. This easy setup allowed
DDPG to converge to a reasonable return value even without reward shaping or the safety layer; in 3D this happened only for the upper quantile of the seeds. In both
tasks, an improvement was obtained with reward shaping.
Nonetheless, DDPG+safety layer obtained the highest discounted return with much faster convergence. As for cumulative constraint violations, as DDPG converged slower
than DDPG+reward shaping, it also stabilized later on a
higher value. All the same, DDPG+safety layer accomplished our safety goal and maintained 0 accumulated constraint violations.
Safe Exploration in Continuous Action Spaces
Figure 6. Convergence and constraint violations of DDPG, DDPG+reward shaping, and DDPG+safety layer, per each task. Plotted are
medians with upper and lower quantiles of 10 seeds. The top row shows the sum of discounted rewards, where the bottom gives the
cumulative number of episodes terminated due to a constraint violation. As shown by the bottom blue curves, with the safety layer,
constraints were never violated.
In the Spaceship domain, the reward is obtained only once
at the target, and the spaceship is re-initialized away from
it with each constraint violation. This setup proves fatal
for DDPG, which was not able to converge to any reasonable policy in both tasks. Surprisingly, reward shaping poses no improvement but rather has an adverse effect: it resulted in highly negative episodic returns. On
the other hand, DDPG+safety layer converged extremely
fast to a high-performing safe policy. This behavior stems
from the closed-region type of tasks; exploring while straying away from the walls allowed the spaceship to quickly
meet the target and then learn how to reach it. With regards to safety, DDPG+safety layer again prevailed where
DDPG and DDPG+reward shaping failed, and maintained
0 accumulated constraint violations.
7.5. Videos
To visualize the safety layer in action, we depict the intensity of its action correction via the magnitude of the dominant Lagrange multiplier λ∗i∗ as calculated in (5). We do
so by coloring the Ball and Spaceship objects in varying
shades of red. We share two videos of the colored agents
showcasing the following.
In Video 1, to track the learning process of an agent with the
safety layer, we recorded several episodes during training
in the Spaceship-Corridor task. These exhibit how the maneuvers produced by the safety layer to dodge the walls are
gradually learned by the agent, as demonstrated by the less
frequent coloring of the spaceship red. Video 1 is found in
https://youtu.be/KgMvxVST-9U. In Video 2, per
each task of the four, we show episodes of i) the first DDPG
iteration (initialized with a random policy) with the safety
layer off; ii) the same random initial policy but with the
safety layer on; and iii) the final iteration of DDPG+safety
layer, to demonstrate the task’s goal. Video 2 is found in
https://youtu.be/yr6y4Mb1ktI.
8. Discussion
In this work, we proposed a state-based action correction mechanism, which accomplishes the goal of zeroconstraint-violations in tasks where the agent is constrained
to a confined region. This is in contrast with the standard
reward shaping alternative, which has failed in achieving
the above goal. The resulting gain is not only in maintaining safety but also in enhanced performance in terms of
reward. This suggests our method promotes more efficient
exploration – it guides the exploratory actions in the direction of feasible policies. Since our solution is stand-alone
and applied directly at the policy level, it is independent of
the RL algorithm used and can be plugged into any other
continuous control algorithm.
Throughout this work, we relate to a preemption mechanism implemented in real-world critical systems, which
halts the RL agent in borderline situations and replaces it
with a safe-operation heuristic. The latter runs until the sys-
Safe Exploration in Continuous Action Spaces
tem is back to being far from operation limits. This heuristic is expected to be conservative and less efficient than the
RL agent; hence, the contribution of our work can also be
interpreted as in reducing operational costs by minimizing
the number of times such takeovers occur.
Lastly, an advantage of our approach over off-policy methods, often considered in industry, is that one does not
need to know the behavior policy used to generate existing data logs. This is thanks to our single-step model,
which we train on trajectories generated with random actions. These trajectories are always within operating limits
(due to episode termination when the limits are crossed)
and are independent of any particular policy whose longterm behavior can be elevated. Nevertheless, they carry
rich information due to their exploratory nature. It is thus
intriguing to study in future work additional types of pretraining data that are typical to specific real-world domains.
References
Achiam, Joshua, Held, David, Tamar, Aviv, and Abbeel,
Pieter. Constrained policy optimization. arXiv preprint
arXiv:1705.10528, 2017.
Altman, Eitan. Constrained Markov decision processes,
volume 7. CRC Press, 1999.
Amos, Brandon and Kolter, J Zico. Optnet: Differentiable optimization as a layer in neural networks. arXiv
preprint arXiv:1703.00443, 2017.
Baxter, Jonathan and Bartlett, Peter L. Infinite-horizon
policy-gradient estimation. Journal of Artificial Intelligence Research, 15:319–350, 2001.
Berkenkamp, Felix, Turchetta, Matteo, Schoellig, Angela P, and Krause, Andreas. Safe model-based reinforcement learning with stability guarantees. arXiv
preprint arXiv:1705.08551, 2017.
Enqvist, Martin. Linear models of nonlinear systems. 2005.
Evans, Richard and Gao, Jim. Deepmind ai reduces google
data centre cooling bill by 40%. DeepMind blog, 20,
2016.
Goodwine, Bill. Engineering differential equations: theory
and applications. Springer Science & Business Media,
2010.
Gu, Shixiang, Holly, Ethan, Lillicrap, Timothy, and Levine,
Sergey. Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. arXiv
preprint arXiv:1610.00633, 2016.
Kingma, Diederik and Ba, Jimmy.
method for stochastic optimization.
arXiv:1412.6980, 2014.
Adam: A
arXiv preprint
Liang, Bin and Sun, Mao. Nonlinear flight dynamics and
stability of hovering model insects. Journal of The Royal
Society Interface, 10(85):20130269, 2013.
Lillicrap, Timothy P, Hunt, Jonathan J, Pritzel, Alexander,
Heess, Nicolas, Erez, Tom, Tassa, Yuval, Silver, David,
and Wierstra, Daan. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971,
2015.
Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David,
Rusu, Andrei A, Veness, Joel, Bellemare, Marc G,
Graves, Alex, Riedmiller, Martin, Fidjeland, Andreas K,
Ostrovski, Georg, et al. Human-level control through
deep reinforcement learning. Nature, 518(7540):529–
533, 2015.
Pham, Tu-Hoa, De Magistris, Giovanni, and Tachibana,
Ryuki. Optlayer-practical constrained optimization for
deep reinforcement learning in the real world. arXiv
preprint arXiv:1709.07643, 2017.
Sallab, Ahmad EL, Abdou, Mohammed, Perot, Etienne,
and Yogamani, Senthil. Deep reinforcement learning
framework for autonomous driving. Electronic Imaging,
2017(19):70–76, 2017.
Shani, Guy, Heckerman, David, and Brafman, Ronen I. An
mdp-based recommender system. Journal of Machine
Learning Research, 6(Sep):1265–1295, 2005.
Sotomayor, Jorge, Mello, Luis Fernando, and Braga, Denis
de Carvalho. Stability and hopf bifurcation in the watt
governor system. arXiv preprint math/0604177, 2006.
Sutton, Richard S and Barto, Andrew G. Reinforcement
learning: An introduction, volume 1. MIT press Cambridge, 1998.
Thomas, Philip S. Safe reinforcement learning. 2015.
Todorov, Emanuel, Erez, Tom, and Tassa, Yuval. Mujoco:
A physics engine for model-based control. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp. 5026–5033. IEEE, 2012.
| 2cs.AI
|
1
Global asymptotic stability
of a PID control system with Coulomb friction
arXiv:1609.09103v1 [cs.SY] 28 Sep 2016
Andrea Bisoffi, Mauro Da Lio, Andrew R. Teel and Luca Zaccarian
Abstract—We propose a model for representing a point mass
subject to Coulomb friction in feedback with a PID controller,
based on a differential inclusion comprising all the possible
magnitudes of static friction during the stick phase. For this
model we study the set of all equilibria and we establish its
global asymptotic stability using a discontinuous Lyapunov-like
function, and a suitable LaSalle’s invariance principle. We finally
use well-posedness of the proposed model to establish useful
robustness results, including an ISS property from a suitable
input in a perturbed context. Simulation results are also given
to illustrate our statements.
I. I NTRODUCTION
F
RICTION in mechanical systems has been investigated
since the times of Leonardo da Vinci, Guillaume Amontons, Charles-Augustin de Coulomb and Arthur-Jules Morin.
Their main findings acknowledge that, for a moving mass, the
friction force is proportional to the normal force through a
kinetic coefficient (Coulomb friction) and presents possibly a
term proportional to the velocity (viscous friction), whereas at
rest the friction force is bounded by the product of the normal
force and a static coefficient, generally greater than the kinetic
coefficient.
Within the control community, the interest in the dynamical
properties of friction had its peak in the 1990’s, and the control
engineering reasons for this interest are lucidly argued in [18,
§1]. These properties have been studied along a modeling
direction, where we mention the Dahl model [10], the LuGre
model [4], [8], the models by Bliman and Sorine [7] and the
Leuven model [25]. The characteristics of all these models are
also detailed in [11]. When a mass moves with steady velocity
and the corresponding friction force is measured, there is a
small interval of velocities near zero where the friction force
decreases before increasing again due to viscous friction and
this behaviour is given the name of Stribeck effect. Another
experimental observation is the distinction between the two
motion regimes of presliding and sliding. During presliding,
the friction force is better described as a function of the
(microsliding) displacement (see [20]), which is intuitively due
Research supported in part by AFOSR grant FA9550-15-1-0155 and NSF
grant ECCS-1508757, by the ANR project LimICoS contract number 12 BS03
005 01, and by grant OptHySYS funded by the University of Trento.
A. Bisoffi and M. Da Lio are with Dipartimento di Ingegneria
Industriale,
University
of
Trento,
Italy
{andrea.bisoffi,
mauro.dalio}@unitn.it
A. R. Teel is with the Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, USA
[email protected]
L. Zaccarian is with CNRS, LAAS, 7 avenue du Colonel Roche, F-31400
Toulouse, France and Université de Toulouse, 7 avenue du Colonel Roche,
31077 Toulouse cedex 4, France, and Dipartimento di Ingegneria Industriale,
University of Trento, Italy [email protected]
to the asperity junctions that build up between the surfaces
and that behave like stiff springs [11]. After a critical value of
displacement (and a break-away force) is reached, the sliding
motion can begin. This property that the friction force is only
position dependent is called rate independence [4], and is to
be found in the Dahl model [18, §4.1], and in the models by
Bliman and Sorine [18, §5.1]. In the latter ones, considering
friction as depending only on the path allows using the theory
of hysteresis operators [16], [27]. On the other hand, rate
dependence holds for the LuGre model [8]. As a final remark,
the LuGre model itself proved to be amenable to theoretical
analysis, as [6] presents necessary and sufficient conditions
for the passivity of its underlying operator from velocity to
friction force.
In this work, we propose to characterize Coulomb friction
in terms of differential inclusions [5], and we apply this
characterization to the case of a point mass under such a
friction force and actuated by a proportional-integral-derivative
(PID) controller. This problem is a classical one in the friction
literature (together with the point mass on a moving belt)
and we will be able to prove the global asymptotic stability
of the attractor having zero velocity, zero position and a
bounded integral error. The use of a set-valued map for
the friction force can be seen as quite natural and is taken
into consideration in [7], [19], [26]: in [26] it is applied
to uncontrolled multi-degree-of-freedom mechanical systems,
in [19] to a PD controlled 1 degree-of-freedom system. The
combination of set-valued friction laws and Lyapunov tools is
also the subject of [17, Chap. 5-6].
To the best of the authors’ knowledge, global asymptotic
stability has not been proved so far. In particular, it was proved
(see [1, Thm. 1] and the related works [2], [3]) that in our
same setting there exists no stick-slip limit cycle (the so-called
hunting phenomenon), which is the detrimental signature of
a stiction greater than the Coulomb friction. As an overall
achievement, Lyapunov tools applied to a differential-inclusion
model enable proving global asymptotic stability of the largest
set of equilibria. Additionally, the established properties and
the regularity of our model imply robustness of asymptotic
stability. This, in turn, allows us to prove an input-to-state stability (ISS) property for the perturbed dynamics, establishing
that more general friction phenomena (including the Stribeck
effect) cause a gradual deterioration of the response, in an ISS
sense. We regard this work as a stepping stone to stiction larger
than Coulomb and to its description through hybrid systems
[13], and to proposing compensation schemes using hybrid
friction laws.
The paper is structured as follows. We present the proposed
model and the main results in Section II. Then Section III
2
contains an illustration by simulation of the established properties. The end of the paper contains the Lyapunov-based proof,
separated into the proof of global attractivity (Section IV) and
of stability (Section V).
Notation. The sign function is defined as: sign(x) := 1 if
x > 0, sign(0) := 0, sign(x) := −1 if x < 0. The saturation
function is defined as: sat(x) := sign(x) if |x| > 1, sat(x) :=
x if |x| ≤ 1. For c 6= 0, the
function x 7→ dzc (x) is defined
as dzc (x) := x − c sat xc . |x| denotes the Euclidean norm
of vector x. h·, ·i defines the scalar product between its two
vector arguments.
velocity v
so = 0
mass m
control action uPID
friction force ff
position s
Fig. 1. Mass under the action of friction and controlled by a PID controller.
II. P ROPOSED MODEL AND MAIN RESULT
A. Derivation of the model
Consider a point mass m described by position s and
velocity v, as in Figure 1. The overall friction force ff acting
on the mass comprises both Coulomb and viscous friction. Its
classical description (see [2, Eq. (3)], or similarly [18, Eq. (5)])
is parametrized by a Coulomb friction constant f¯c > 0 and by
the viscous friction constant αv > 0. The expression of ff
reads
¯
fc sign(v) + αv v, if v 6= 0
ff (fr , v) := fr ,
(1)
if v = 0, |fr | < f¯c
¯
fc sign(fr ),
if v = 0, |fr | ≥ f¯c
where fr is the resultant tangential force. The mass is actuated
by the PID control uPID
Z t
ds(t)
uPID (t) := −k̄p s(t) − k̄i
s(τ )dτ − k̄d
(2)
dt
0
= −k̄p s(t) − k̄i ei (t) − k̄d v(t),
where ei is defined to be the integral of the position error and
is the state of the controller, satisfying ėi = s and ei (0) = 0.
Using Newton’s law, we write the mechanical dynamics ṡ =
v and mv̇ = uPID − ff (uPID , v). The convenient definitions
k̄
f¯c
−αv v
k̄i
v
u := uPID m
, (kp , kv , ki ) := ( mp , k̄d +α
m , m ) and fc := m
yield then
ėi = s
(3a)
ṡ = v
u − fc
v̇ = 0
u + fc
(3b)
if v > 0 or (v = 0, u ≥ fc )
if (v = 0, |u| < fc )
if v < 0 or (v = 0, u ≤ −fc )
u = −kp s − kv v − ki ei ,
where we used that uPID = m u for v = 0.
(3c)
Model (3) arises from a relatively intuitive description of
the mechanical principles behind the model of Figure 1.
Its discontinuous right hand side makes it hard to prove
existence of solutions for any initial conditions, even though
such a property can be shown to hold on a case-by-case
basis. Moreover, it seems to be hard to use dynamics (3) for
establishing some stability properties and certifying that the
position s converges to zero.
In this paper we use the monotone set-valued friction law
[17, Eq. 5.36] for which existence of solutions is structurally
guaranteed. Defining the overall state z := (ei , s, v), this
is equivalent to applying the Filippov [12] or Krasovskii
regularization to the discontinuous dynamics (3) and obtain
s
0
− fc 0 SGN(v)
v
ż ∈
(4a)
−ki ei − kp s − kv v
1
where the function SGN is a set-valued map defined as
(
sign(v), if v 6= 0
SGN(v) :=
(4b)
[−1, 1],
if v = 0.
Note that model (4) recognizes that the Coulomb friction
can be selected as any force in the set [−f¯c , f¯c ] when v
is zero and has magnitude f¯c and direction opposite to v
whenever v 6= 0. One may wonder whether any artificial
solution is introduced by such an enriched description of
the dynamics. (For (3) or (4), we consider a solution to be
any locally absolutely continuous function x that satisfies
respectively ẋ(t) = f (x(t)) or ẋ(t) ∈ F (x(t)) for almost all
t in its domain.) The following result establishes uniqueness
of the solutions to (4), which implies that the unique solution
to (4) must necessarily be the unique solution to (3). Indeed,
dynamics (3) allows for only some selections of v̇ compared
to those allowed by (4), so that any solution to (3) is also a
solution to (4).
Lemma 1: For any initial condition z(0) ∈ R3 , system (4)
has a unique solution defined for all t ≥ 0.
Proof. Existence of solutions follows from [12, §7, Thm. 1]
because the mapping in (4) is outer semicontinuous and locally
bounded with nonempty compact convex values (see also [13,
Prop. 6.10]). Completeness of maximal solutions follows from
local existence and no finite escape times, as (4) can be regarded as a linear system forced by a bounded input. To prove
uniqueness, consider two solutions z1 = (z1,ei , z1,s , z1,v ), z2
both starting at z0 and define δ(t) = (δei (t), δs (t), δv (t)) :=
z1 (t) − z2 (t), for all t ≥ 0. Then, δ(0) = 0 and, for almost
all t ≥ 0,
h0i
δ̇(t) ∈ Aδ δ(t)−fc 0 SGN(z1,v (t))−SGN(z1,v (t)−δv (t)) ,
1
h
0
1
0
0
0
1
−ki −kp −kv
i
with Aδ :=
, whose maximum singular value
is λδ . Therefore we can write for almost all t
2
d |δ(t)|
dt
2
(3d)
M (t) :=
= δ(t)T δ̇(t) ≤ λδ |δ(t)|2 + M (t)
max
δv (t)(f2
f1 ∈fc SGN(z1,v (t))
f2 ∈fc SGN(z1,v (t)−δv (t))
− f1 ).
3
Whether z1,v (t) and z1,v (t) − δv (t) are positive, zero or
negative, by trivial inspection of all the cases it can be shown
that M (t) ≤ 0 for all t ≥ 0. Therefore,
2
d |δ(t)|
dt
2
≤ λδ |δ(t)|2 for almost all t ≥ 0,
and from standard comparison theorems δ(0) = 0 implies
δ(t) = 0 for all t ≥ 0, that is, z1 (t) = z2 (t) for all t ≥ 0.
B. Main result
The advantage in the use of the compact dynamics (4) is
that we may adopt Lyapunov tools to study the asymptotic
stability properties of the rest position under the following
standard assumption (see, e.g., [1]).
Assumption 1: The parameters in (3d) are such that
ki > 0, kp > 0, kv kp > ki .
According to the Routh stability test, Assumption 1 holds
if and only if the origin of the dynamics in (4) with fc = 0 is
globally exponentially stable.
Under Assumption 1, one readily sees that all possible
equilibria of dynamics (4) correspond to (ei , s, v) = (ēi , 0, 0)
with |ēi | ≤ fkci , that is, whenever the mass is at rest at zero
position and the size of the integral error ei is bounded by the
specific threshold fkci . Any of these points is an equilibrium
for (4) because in (4) a value can be selected from fc SGN(0)
such that the (unique) solution maintains ż identically zero.
Note that here we consider the problem of tracking a position
setpoint so = 0, but this is trivially generalized to piecewise
constant setpoints, thanks to the global nature of our results
and a trivial change of coordinates. Denote then the set of
these equilibria as
fc fc
A := (ei , s, v) : s = 0, v = 0, ei ∈ − ,
. (5)
ki ki
Given the attractor A, we prove its global attractivity in
Section IV and its stability in Section V. For the attractivity,
we use a suitable discontinuous Lyapunov-like function and
a nonsmooth version of LaSalle’s invariance principle [15,
§4.2]. These tools are applicable to our scenario because of
the desirable structural properties of the regularization in (4).
Proposition 1: Under Assumption 1, the attractor A in (5) is
1) globally attractive and 2) Lyapunov stable for dynamics (4).
Note that no smaller set could be proven to be globally
attractive because A is a union of equilibria.
Our main result in Theorem 1 establishes relevant robust
stability properties of A for (4) in terms of the behavior of
solutions to the perturbed dynamics1
ż ∈ coF̃ (z + ρ(z)B) + ρ(z)B,
(6)
where ρ : R3 → R≥0 is a suitable non-negative perturbation
function satisfying z ∈
/ A ⇒ ρ(z) > 0. In particular,
by using the equivalences in [13, Chap. 7] (which apply
because system (4) is well-posed from the regularity of F̃
1 F̃ ,
B and co denote respectively the (set-valued) right-hand side of (4a),
the closed unit ball and the closed convex hull of a set.
[13, Thm 6.30] and A is compact), Proposition 1 (see [13,
Thm. 7.21]) implies robust global KL asymptotic stability of
A for (4), namely the existence of β0 ∈ KL such that all
solutions to (6) satisfy |z(t)|A ≤ β0 (|z(0)|A , t) for all t ≥ 0,
in addition to robust global uniform asymptotic stability of A
for (4), namely the property that A is uniformly globally stable
and attractive for (6) (see [13, Def. 3.6]).
Theorem 1: Under Assumption 1, the attractor A in (5) is
robustly globally KL asymptotically stable for dynamics (4)
and robustly uniformly globally asymptotically stable.
An interesting consequence of the robustness result established in Theorem 1 is the semiglobal practical robust
asymptotic stability of attractor A established in [13, Thm 7.12
and Lemma 7.20]. A specific perturbation of interest arises
when selecting a constant scalar ρv ∈ R and perturbing the
friction effect as follows:
i
h0i
h
s
ż ∈ −k e −kv s−k v − fc 0 SGNρv (v)
(7)
1
( i i p v
[sign(v) − |ρv |, sign(v) + |ρv |], if |v| > |ρv |
SGNρv (v) :=
[−1 − |ρv |, 1 + |ρv |],
if |v| ≤ |ρv |.
The general definition of the inflation of a set-valued
mapping for SGN in (4b) is SGN(v + |ρv |B) + |ρv |B. This
inflation coincides with SGNρv in (7), and in the special case
ρv = 0, SGN0 clearly coincides with SGN.
Corollary 1: Under Assumption 1, the attractor A in (5) is
globally input-to-state stable for dynamics (7) from input ρv .
Proof. The solutions
h 0 i to (7) are a subset of the solutions to
ż = Aδ z − fc 0 m, where Aδ was trivially defined in the
1
proof of Lemma 1 and is Hurwitz from Assumption 1, and
m is a locally integrable signal that for the constant scalar
ρv and for all t satisfies m(t) ≤ 1 + |ρv | because, for all t,
SGNρv (v(t)) ≤ 1+|ρv |. From BIBO stability of exponentially
stable linear systems, there exist positive c and λ such that all
solutions satisfy
|z(t)| ≤ ce−λt |z(0)| + c(1 + |ρv |).
(8)
2
From the two distances |z|2A := s2 + v 2 + dzfc /ki (ei ) ,
|z|2 := s2 + v 2 + e2i , we have |z|A ≤ |z| and |z|2 ≤
2
(by splitting into the cases |ei | ≥ fkci and
2|z|2A + 2 fkci
√
|ei | < fkci ), which implies |z| ≤
2 |z|A + fkci . These
relationships between the two distances and (8) imply that
there exist positive constants κ1 , κ2 , κ3 such that all solutions
satisfy
|z(t)|A ≤ |z(t)| ≤ ce−λt |z(0)| + c(1 + |ρv |)
≤ κ1 e−λt |z(0)|A + κ2 + κ3 |ρv |, ∀t ≥ 0.
(9)
Using Theorem 1 and the semiglobal practical robustness of
KL asymptotic stability established in [13, Lemma 7.20], one
can transform the δ- argument of [13, Lemma 7.20] into a
class K function γ` by following similar steps to [15, Lemma
4.5]. Moreover, using a similar approach to [24, Thm. 2]
relating the size of the initial condition and of the input, we
obtain the following:
|z(0)|A ≤
1
δ` , |ρv |
≤ δ` ⇒
|z(t)|A ≤ β` (|z(0)|A , t) + γ` (|ρv |), ∀t ≥ 0, (10)
4
fc SGN(v)
fc SGNρv (v)
Stribeck
III. I LLUSTRATION BY SIMULATION
|ρv |
v
Fig. 2. Stribeck effect is included in the perturbation (7).
TABLE I
PARAMETERS AND EIGENVALUES FOR SIMULATIONS IN S ECTION III.
Case kv
(a)
(b)
kp
ki
Roots
3
4
−6.01, −0.19 ± i0.79
0.66 0.08 −0.8, −0.5, −0.2
6.4
1.5
for some suitable class KL and class K functions β` and
γ` , and for a small enough scalar δ` > 0. Without loss of
generality, consider now using in (10) a small enough δ` such
that (2δ` )−1 ≥ κ2 +κ3 δ` . Introduce function T ? : R≥0 → R≥0
with T ? (s) := max{0, λ−1 log(2δ` κ1 s)}, which satisfies:
κ1 e−λT
?
(s)
s + κ2 + κ3 δ` ≤ δ`−1 ,
∀s ≥ 0.
(11)
Finally, we conclude the proof by establishing the following
(global) ISS bound from ρv :
|z(t)|A ≤ β(|z(0)|A , t) + γ(|ρv |), ∀z(0), ∀ρv , ∀t ≥ 0, (12)
where functions β and γ of class KL and class K, respectively,
are built starting from the following inequalities:
(
κ1 e−λt s + κ2 + κ3 δ` , if s ≥ δ1` , t ≤ T ? (s) (13a)
β(s, t) ≥
b(s, t),
otherwise
(13b)
b(s, t) := max β` (s, max{0, t − T ? (s)}) , κ1 e−λt s (13c)
γ(s) ≥
κ2 + κ3 s, if s ≥ δ`
γ` (s),
if s ≤ δ` .
(13d)
(13e)
The effectiveness of selections (13) for establishing the ISS
bound (12) can be verified case by case.
Case 1 (|ρv | ≥ δ` ): use (9), (13d), and bound κ1 e−λt s in
(13c)-(13b).
Case 2 (|ρv | ≤ δ` and |z(0)|A ≤ δ`−1 ): use (10), (13e), and
bound β` (s, max{0, t − T ? (s)}) in (13c)-(13b).
Case 3 (|ρv | ≤ δ` and |z(0)|A ≥ δ`−1 ): for t ≤ T ? (|z(0)|A )
use (13a) and nonnegativity of γ, whereas for t ≥ T ? (|z(0)|A )
use |z(T ? (|z(0)|A ))|A ≤ δ` −1 (from (9) and (11)) and the
semigroup property of solutions to fall again into case 2
above.
A consequence of Corollary 1 is that the Stribeck effect, which is known to lead to persistent oscillations (the
so-called hunting phenomenon), produces solutions that are
graceful degradations (in the ISS sense) of the asymptotically
stable solutions to the unperturbed dynamics because small
Stribeck deformations lead to graphs included in the graph of
fc SGNρv (v), as shown in Figure 2.
Before we prove our main result, we obtain solutions by
simulation in order to illustrate the typical behaviour of (4)
and the convergence to the attractor. Simulations capture for
each initial condition the unique solution to (4) because of
Lemma 1.
When fc = 0, (4) reduces to a linear system with characteristic polynomial s3 + kv s2 + kp s + ki = 0. From
Assumption 1, its roots have negative real part, but we do not
assume anything about their specific locations in the complex
plane. In [1, Lemmas L1 and L2], each of these possible
locations needed examining as part of a proof relying on the
algebraic expression of solutions. We will not need this in
our subsequent proof, but we base our simulations on two
representative cases, complex conjugate and three distinct real
roots. The corresponding parameters kv , kp , ki and roots are
listed in Table I and the value fc = 1 m/s2 is common to all
simulations.
First, we present the solutions to (4) for different sets of
initial conditions (ēi , s̄, v̄) for the cases (a) and (b) in Table I,
respectively in the left and right top plots of Figure 3. We use
a heavier black line to point out a specific solution and the two
different phases that are visible on it. In particular, there are
time intervals when the mass is in motion (called slip phases
in the friction literature) and others when the mass is at rest
(called stick phases), which can be detected easily from the
velocity being zero on a nonzero time interval. Whenever the
mass is in a slip phase, the PID control acts in the direction
of getting the mass closer to the position setpoint at zero.
During a stick phase starting at ti , the mass is at rest (v =
0) and only the error integral builds up linearly in time as
ei (t) = ei (ti ) + s(ti )(t − ti ) until the control action u is such
to overcome the Coulomb friction, that is, |u| = | − ki ei −
kp s| = fc . So, the closer the mass is to zero position (smaller
s(ti )), the longer it takes the error to build up. This explains
the ramps in the top of Figure 3, and their decreasing slope
and increasing duration. Moreover, we notice that position and
velocity converge to zero, but the error integral does not in
general. For the case of complex conjugate roots (top, left),
the error integral continues to oscillate and these oscillations
enter asymptotically the set [−fc /ki , fc /ki ] as the position
approaches zero. For the case of distinct real roots (top, right),
after a stick phase the position and the velocity converge to
zero only asymptotically, so that v remains always nonzero and
in (4) ei can only approach the equilibria fc /ki or −fc /ki . As
a consequence of the reasoning above (the smaller the position,
the longer it takes to exit a stick phase), solutions converge
asymptotically, but not exponentially.
Second, we present in the left and right center plots of
Figure 3 a phase portrait for the same solutions in the top
plots. In these figures it is evident that solutions converge to
the attractor in (5), as we postulate in item 1) of Proposition 1.
In particular, the two different manners in which solutions
converge to the attractor in the cases of complex conjugate or
distinct real roots are evident here as well. In the left center
plot, we can also appreciate the presence of a strip in the plane
v = 0 expressed by equation −fc ≤ −ki ei − kp s ≤ fc (as in
5
20
1
ei (m s)
ei (m s)
2
0
−1
−2
0
5
10
15
20
25
30
10
0
−10
−20
0
35
1
0
−1
−2
0
20
30
40
50
60
70
80
90
100
5
10
15
20
25
30
10
20
30
40
50
60
70
80
90
100
10
20
30
40
50
t (s)
60
70
80
90
100
0
−1
−2
0
35
2
0.4
1
0.2
v (m/s)
v (m/s)
10
1
s (m)
s (m)
2
0
−1
0
5
10
15
20
25
30
0
−0.2
−0.4
0
35
t (s)
A
0.4
1
0.2
0.5
v (m/s)
v (m/s)
A
1.5
0
−0.5
−1
−1.5
0
−0.2
−0.4
−1.5
−1
−1
−0.5
−0.5
0
0
0.5
1
s (m)
1.5
2
0.5
−2
−1
0
1
s (m)
20
ei (m s)
0.12
25
0.1
20
0.08
V(x(t))
V(x(t))
30
15
0.06
10
0.04
5
0.02
0
0
1
ei (m s)
−20
−10
0
10
5
10
15
20
25
30
35
0
0
10
20
30
40
t
50
t
60
70
80
90
100
Fig. 3. Top: solutions to (4) for different initial conditions. Center: phase portraits for (4) for the same solutions. Bottom: Lyapunov-like function V in (18)
evaluated along the same solutions. All the figures to the left (resp., right) refer to the PID parameters (a) (resp., (b)) in Table I.
[21, Page 7]), that is, the region of the state space where stick
is bound to occur.
Third, we keep the same initial conditions and parameters and we anticipate the evolution along solutions of
the Lyapunov-like function introduced in the next section
(see (18)). In particular, this function is nonincreasing along
solutions, it can be discontinuous (e.g., the left, bottom, violet
curve at t = 0.123 s), and remains constant during stick (as
pointed out by the same heavier black curves).
A. Coordinate change and discontinuous LaSalle function
For the following analysis we adopt a specific change of
coordinates for (4), that is,
φ := −ki ei − kp s,
v := v.
σ̇
−ki v
σ − kp v
ẋ := φ̇ ∈
φ − kv v − fc SGN(v)
v̇
σ
0 0 −ki
0
= 1 0 −kp φ − 0 SGN(v)
v
fc
0 1 −kv
{z
}
| {z }
|
=:A
IV. P ROOF OF P ROP. 1: GLOBAL ATTRACTIVITY
σ := −ki s
The change of coordinates is nonsingular thanks to Assumption 1 (ki , kp strictly positive) and it rewrites (4) as
(14)
(15)
=:b
= Ax − b SGN(v) =: F (x).
In the new coordinates x, the attractor A in (16) can be
expressed as
A = {(σ, φ, v) : |φ| ≤ fc , σ = 0, v = 0}.
(16)
Among other things, the simple expression in (16) allows
6
writing explicitly the distance of a point x from A as
2
|x|2A := inf |x − y| = σ 2 + v 2 + dzfc (φ)2 .
(17)
y∈A
where dzfc (φ) := φ − satfc (φ) is the symmetric scalar
deadzone function returning zero when φ ∈ [−fc , fc ]. Indeed,
(17) follows from separating the cases φ < −fc , |φ| ≤ fc ,
φ > fc and applying the definition given in (17).
Based on the above observation, it is rather intuitive to
introduce the following discontinuous Lyapunov-like function
V (x) :=
=
T kv
σ
ki
v
−1
min
f ∈fc SGN(v)
−1 σ
+ min
|φ − f |2
kp v f ∈fc SGN(v)
(18a)
h σ iT h σ i
φ−f
P φ−f
v
v
where the matrix P is given by
kv
ki
P := 0
−1
0 −1
1 0 .
0 kp
(18b)
Note that for v 6= 0 the minimization in (18a) becomes trivial
because f can take only the value fc sign(v). It is emphasized
that function V is discontinuous. For example, if we evaluate
V along the sequence of points (σi , φi , vi ) = (0, 0, εi ) for
εi ∈ (0, 1) converging to zero, V converges to fc2 , even though
its value at zero is zero. Nevertheless, function V enjoys a
number of useful properties established in the next lemma
whose proof is given in Section IV-C below.
Lemma 2: The Lyapunov-like function in (18) is lower
semicontinuous (lsc) and enjoys the following properties:
1) V (x) = 0 for all x ∈ A and there exists c1 > 0 such
that c1 |x|2A ≤ V (x) for all x ∈ R3 ,
2) there exists c > 0 such that each solution x = (σ, φ, v)
to (15) satisfies for all t2 ≥ t1 ≥ 0
Z t2
V (x(t2 )) − V (x(t1 )) ≤ −c
v(t)2 dt.
(19)
t1
Remark 1: In [1] it is proven that if a solution is in a slip
phase in the nonempty time interval (ti , ti+1 ) (namely, for all
t ∈ (ti , ti+1 ), v(t) 6= 0) and the slip phase is preceded and
followed by a stick phase (namely, there exist δ > 0 such
that, for all t ∈ [ti − δ, ti ] ∪ [ti+1 , ti+1 + δ], v(t) = 0 and
|φ(t)| ≤ fc ), then
|σ(ti+1 )| < |σ(ti )|.
(20)
Instead of using the explicit form of solutions as [1,
Lemma L2] depending on the nature of the eigenvalues of
A, (20) is easily concluded from (19), the definition (18a),
and |φ(ti )| ≤ fc , |φ(ti+1 )| ≤ fc .
y
B. Proof of item 1) of Proposition 1 (global attractivity)
We can now prove the first item of Proposition 1 based on
Lemma 2 and a generalized version for differential inclusions
of the invariance principle [15, §4.2]. The following fact comes
indeed from specializing the result in [23, Thm. 2.10] to our
case, where the differential inclusion (4) has actually unique
solutions defined for all nonnegative times (as established in
Lemma 1). We also select G = R3 , U = R3 in the original
result of [23].
Fact 1: [23] Let ` : R3 → R≥0 be lower semicontinuous
3
and such that `(x) ≥ 0, for all x ∈ R
. If x is a complete
R +∞
and bounded solution to (4) satisfying 0 `(x(t))dt < +∞,
then x converges to the largest forward invariant subset M of
Σ := {x ∈ R3 : `(x) = 0}.
Proof of item 1) of Proposition 1 (global attractivity of A).
The proof exploits Fact 1, where we take `(x) = v 2 .
From Lemma 2, V (x(t)) ≤ V (x(0)) (item 2) and
c1 |x(t)|2A ≤ V (x(t)) (item 1), so that c1 |x(t)|2A ≤ V (x(0))
and consequently all solutions to (15) are bounded (their
completeness is established
in Lemma 1). Apply (19) from 0
Rt
to t, and obtain c 0 v 2 (τ )dτ ≤ V (x(0))−V (x(t)) ≤ V (x(0))
because
≥ 0 from Lemma 2, item 1. Then we have
R t 2 V (x(t))
V (x(0))
, and if t → +∞ we get the required
v
(τ
)dτ
≤
c
0
boundedness of the integral of `(x(·)). Then Fact 1 guarantees
that the solution converges to the largest forward invariant
subset M of Σ = {x : v = 0}. We claim that such a subset
is A. Indeed, M ⊂ Σ implies v = 0 in M. Moreover, σ = 0
in M because each solution starting from v = 0 and σ 6= 0
causes a ramp of φ that eventually reaches |φ| > fc and
drives v away from zero (therefore out of Σ). Finally, in M
we must have |φ| ≤ fc otherwise v would become nonzero
again. Therefore the largest forward invariant set M in Σ is
the attractor A.
C. Proof of Lemma 2
To the end of proving Lemma 2, we note that model (15)
and function (18) suggest that there are three relevant affine
systems and smooth functions associated to the three cases
in (3c) that are worth considering (and will be used in our
proofs). They correspond to
ξ˙ = f1 (ξ) := Aξ − b,
h0 0 0i
ξ˙ = f0 (ξ) := 1 0 0 ξ,
ξ(0) = ξ1 ,
(21a)
ξ(0) = ξ0 ,
(21b)
ξ˙ = f−1 (ξ) := Aξ + b,
ξ(0) = ξ−1 ,
(21c)
000
and, with the definition |ξ|2P := ξ T P ξ,
h σ i2
hσi 2
h σ i2
V1 (ξ) := φ−fc
, V0 (ξ) := 0
, V−1 (ξ) := φ+fc
.
0 P
v
v
P
P
(21d)
Based on the description above, we can state the following
claim relating solutions to (15) and V in (18) to (21). Its proof
mostly relies on straightforward inspection of the various cases
and is given in Appendix A.
Claim 1: There exists c > 0 such that, for each initial
condition (σ̄, φ̄, v̄), one can select k ∈ {−1, 0, 1} and T > 0
satisfying the following:
1) the unique solution ξ = (ξσ , ξφ , ξv ) to the k-th initial
value problem among (21a)-(21c) with initial condition
ξk = (σ̄, φ̄, v̄) coincides in [0, T ] with the unique
solution to (15);
7
2) the solution ξ mentioned above satisfies for all t ∈ [0, T ]
V (ξ(t)) = Vk (ξ(t)),
d
dt Vk (ξ(t))
t1
Building on Claim 1 and Fact 2 we can prove Lemma 2.
Proof of Lemma 2. We show first that V is lower semicontinuous. Define the set-valued mapping
h σ iT h σ i
[
G(x) :=
g(σ, φ, v, f ), g(σ, φ, v, f ) := φ−f P φ−f ,
v
v
f ∈SGN(v)
and consider the additional set-valued mapping (σ, φ, v) ⇒
H(σ, φ, v) := (σ, φ, v, fc SGN(v)). By the very definition of
set-valued mapping, we can write G = g ◦ H (the composition
of g and H), that is, (σ, φ, v) ⇒ g(σ, φ, v, fc SGN(v)) =
G(x). Then, G is outer semicontinuous (osc) by [22, Proposition 5.52, item (b)] because both g and H are osc and
H is locally bounded. Finally, by the definition of distance
d(u, S) between a point u and a closed set S, we can write
V (x) = d(0, G(x)). Then, V is lsc by [22, Proposition 5.11,
item (a)] because G was proven to be osc.
We prove now the properties of V item by hitem. i
kv
T
ki −1
Item 1). There exists g > 0 such that [ σv ]
[ σv ] ≥
−1 kp
g(σ 2 + v 2 ) because the inner matrix is positive definite
2 by
Assumption 1. Moreover, from (18a),
min
φ−f ≥
f ∈fc SGN(v)
2
min
φ−f = dzfc (φ)2 . Therefore, (17) yields V (x) ≥
f ∈[−fc ,fc ]
c1 |x|2A with
fc
≤ −c|ξv (t)|2 . (22)
Additionally, we restate a fact from [14] that is beneficial
to proving Lemma 2. Specifically, we use [14, Theorem 9]
together with the variant in [14, Section 5 (point a.)]. We
also specialize the statement, using the fact that when the
function g is integrable, the standard integral can replace the
upper integral (as noted after [14, Definition 8]). The lower
right Dini derivative D+ h of h is defined as D+ h(t) :=
.
lim inf →0+ h(t+)−h(t)
Fact 2: [14] Given t2 > t1 ≥ 0, suppose that h satisfies
lim inf τ →τ̄ h(τ ) ≥ h(τ̄ ) (i.e., h is lower semicontinuous) and
that l is locally integrable in [t1 , t2 ]. If D+ h(τ ) ≤ l(τ ) for all
τ ∈ [t1 , t2 ], then
Z t2
l(τ )dτ.
h(t2 ) − h(t1 ) ≤
c1 := min{g, 1}.
Item 2). Equation (19) is a mere application of Fact 2 for
h(·) = V (x(·)) and l(·) = −c(v(·))2 where x = (σ, φ, v) is
a solution to (15). So, we need to check that the assumptions
of Fact 2 are verified.
We already established above that V is lsc. Solutions x
to (15) are absolutely continuous functions by definition.
Then, because the composition of a lower semicontinuous
and a continuous function is lower semicontinuous (see [22,
Exercise 1.40]), the Lyapunov-like function (18a) evaluated
along the solutions of (15) is lsc. Since solutions are absolutely
continuous, −cv 2 is locally integrable.
Finally, it was proven in Claim 1, item 1 that for each
initial condition, the unique solution to (15) coincides with the
solution to one of the three affine systems in (21) (numbered
k) on a finite time interval T . Moreover, from Claim 1, item 2
φ
σ
−fc
v
Fig. 4. R is the (closed) blue region in Lemma 3, R̂ is its complement.
V coincides in [0, T ] with the function Vk in (22), which is
differentiable, then V (x(t)) at t = 0 is at least differentiable
from the right and the lower right Dini derivative coincides
with the right derivative. In particular, we established in (22)
that this right derivative is upper bounded by −cv 2 .
V. P ROOF OF P ROP. 1: STABILITY
The Lyapunov-like function introduced in (18) of the previous section is unfortunately not enough to prove stability.
Indeed, its discontinuity on the attractor A prevents us from
obtaining a uniform continuous upper bound depending on
the distance from A. However, a stability bound can be
constructed through an auxiliary function defined as
2
V̂ (x) := 12 k1 σ 2 + 12 k2 dzfc (φ) + k3 |σ||v| + 12 k4 v 2 . (23)
Function V̂ allows establishing bounds in the directions of
discontinuity of V . In particular, we define the two subsets
R := {x : v(φ − sign(v)fc ) ≥ 0}
R̂ := R3 \R
represented in Figure 4. The following holds.
Lemma 3: For suitable positive scalars k1 , . . . , k4 in (23),
there exist positive scalars c1 , c2 , ĉ1 , ĉ2 such that
c1 |x|2A ≤ V (x) ≤ c2 |x|2A ,
ĉ1 |x|2A
◦
≤ V̂ (x) ≤
V̂ (x) :=
ĉ2 |x|2A ,
∀x ∈ R,
hv, fi ≤ 0,
max
v∈∂ V̂ (x),f∈F (x)
∀x ∈ R̂,
∀x ∈ R̂,
(24a)
(24b)
(24c)
where ∂ V̂ (x) denotes the generalized gradient of V̂ at x
(see [9, §1.2]) and F is the set-valued
2 map in (15).
Proof. Note that
min
φ − f = dzfc (φ)2 whenever
f ∈fc SGN(v)
x ∈ R. Since P in (18b) is positive definite and V (x) =
h σ iT h σ i
dzfc (φ)
P dzfc (φ) in R, positive c1 and c2 can be chosen
v
v
to satisfy (24a), using the definition (17). (The lower bound
in (24a) was already established for all x ∈ R3 in Lemma 2,
item 1.) For positive k1 , . . . , k4 and k1 k4 > k32 , the inner
|σ|
|σ| T
k1 0 k3
1 |dz (φ)|
|dzfc (φ)|
0
k
0
matrix in V̂ (x) = 2
is
fc
2
|v|
k3 0 k4
|v|
positive definite and (24b) can be satisfied for the same reason.
To prove (24c), we consider only the set R̂> := R̂∩{x : v >
0} because a parallel reasoning can be followed in R̂∩{x : v <
0}. For x ∈ R̂> , we have v > 0, φ < fc and (15) reduces to
the differential equation
σ̇ = −ki v =: fσ (x)
φ̇ = σ − kp v =: fφ (x)
v̇ = −kv v + φ − fc =: fv (x) ≤ −kv |v| − |dzfc (φ)|.
(25a)
8
Consistently, the max in (24c) is to be checked only for the
singleton f = (fσ (x), fφ (x), fv (x)) which F (x) reduces to for
all x ∈ R̂> . Moreover,
2
d
1
dz
(φ)
= dzfc (φ), ∂ (|σ|) = SGN(σ), (25b)
f
c
dφ 2
VI. C ONCLUSIONS AND FUTURE WORK
In this work we characterized properties of a differential
inclusion model of the feedback interconnection of a sliding
mass with a PID controller under Coulomb friction. We proved
global asymptotic stability of the largest set of closed-loop
equilibria. Due to the regularity properties of the differential
where ∂ (|σ|) denotes the generalized gradient of σ 7→ |σ|. We inclusion model, global asymptotic stability is intrinsically
need then to find suitable positive constants k1 , . . . , k4 satisfy- robust. Additionally, taking as input the size of the inflation
ing k1 k4 > k32 and such that V̂ ◦ (x) is negative semidefinite in of a perturbed model, the dynamics is input-to-state stable,
R̂> . Since in R̂> we have v = |v| and dzfc (φ) = −|dzfc (φ)|, and this perturbation includes the well-known Stribeck effect.
then we get maxζ∈∂|σ| (−ki k3 |v|2 ζ) = ki k3 |v|2 for all x ∈ Future work will address further the case of static friction force
R̂> , which gives in turn:
larger than the Coulomb one and will propose for that setting
compensation schemes relying on the proposed LyapunovV̂ ◦ (x) ≤ [ki k3 |v|2− k4 kv |v|2 ]+[k2 σdzfc (φ)− k3 |σ||dzfc (φ)|] based proof.
+[−k1 ki σv− k3 kv |v||σ|]+[k2 kp |v||dzfc (φ)|−k4 |v||dzfc (φ)|].
R EFERENCES
Since k1 , . . . , k4 are positive by assumption, in
[1] B. Armstrong and B. Amin, “PID control in the presence of static
each pair in brackets the second term is negative
friction: A comparison of algebraic and describing function analysis,”
semidefinite and dominates the first (sign-indefinite
or
Automatica, vol. 32, no. 5, pp. 679–692, 1996.
ki
[2] B. Armstrong-Hélouvry and B. Amin, “PID control in the presence of
nonnegative) term as long as k3 > max kv k1 , k2 and
static friction,” Dept. of Electrical Engineering and Computer Science,
k2
k4 > max kkvi k3 , kp k2 , k31 . With this selection, (24b) and
University of Wisconsin – Milwaukee, Tech. Rep. EE-93–2, 1993.
[3] ——, “PID control in the presence of static friction: exact and describing
(24c) are simultaneously satisfied.
Proof of item 2) of Proposition 1 (stability). Based on the
constants c1 , c2 , ĉ1 , ĉ2 introduced in Lemma 3, the following
stability bound for each solution x to (15)
q
|x(t)|A ≤ cc21 ĉĉ21 |x(0)|A , ∀t ≥ 0
(26a)
is proven by splitting the analysis in two cases.
Case (i): x(t) ∈
/ R, ∀t ≥ 0. Since R ∪ R̂ = R3 , x(t) ∈ R̂ for
all t ≥ 0 and from (24c)
V̂ ◦ (x(t)) ≤ 0, ∀t ≥ 0 ⇒ V̂ (x(t)) ≤ V̂ (x(0)), ∀t ≥ 0.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
Using bound (24b) we obtain
[11]
ĉ1 |x(t)|2A
ĉ2 |x(0)|2A ,
≤ V̂ (x(t)) ≤ V̂ (x(0)) ≤
∀t ≥ 0,
p
which implies (26a) because 1 ≤ c2 /c1 from (24a).
Case (ii): ∃t1 ≥ 0 such that x(t1 ) ∈ R. Consider the
smallest t1 ≥ 0 such that x(t1 ) ∈ R (the existence of such a
smallest time follows from R being closed). Then, following
the analysis of Case (i) for the (possibly empty) time interval
[0, t1 ) and using continuity of solutions, we obtain
ĉ1 |x(t)|2A ≤ ĉ2 |x(0)|2A , ∀t ∈ [0, t1 ].
[12]
[13]
[14]
[15]
[16]
(26b)
[17]
At t1 we apply (24a) (because
x(t1 ) ∈ R) and (26b) to obtain
V (x(t1 )) ≤ c2 ĉĉ21 |x(0)|2A . Finally, by the bounds in items 1
and 2 of Lemma 2,
[18]
c1 |x(t)|2A ≤ V (x(t)) ≤ V (x(t1 )) ≤ c2 ĉĉ21 |x(0)|2A , ∀t ≥ t1 .
(26c)
q
Since cc21 ≥ 1, (26b) implies c1 |x(t)|2A ≤ c2 ĉĉ12 |x(0)|2A , ∀t ∈
[0, t1 ], which proves (26a) when combined with (26c).
[19]
[20]
[21]
function analysis,” in American Control Conference, 1994, vol. 1. IEEE,
1994, pp. 597–601.
K. J. Åström and C. Canudas-de-Wit, “Revisiting the LuGre friction
model,” Control Systems, IEEE, vol. 28, no. 6, pp. 101–114, 2008.
J. P. Aubin and A. Cellina, Differential Inclusions. Set-valued Maps and
Viability Theory. Springer-Verlag, 1984.
N. Barahanov and R. Ortega, “Necessary and sufficient conditions for
passivity of the LuGre friction model,” IEEE Transactions on Automatic
Control, vol. 45, no. 4, pp. 830–832, 2000.
P.-A. Bliman and M. Sorine, “Easy-to-use realistic dry friction
models for automatic control,” in Proceedings of 3rd European
Control Conference, vol. 113, 1995, pp. 267–272. [Online]. Available:
https://who.rocq.inria.fr/Pierre-Alexandre.Bliman/publis.htm
C. Canudas-de-Wit, H. Olsson, K. J. Åström, and P. Lischinsky, “A
new model for control of systems with friction,” IEEE Transactions on
Automatic Control, vol. 40, no. 3, pp. 419–425, 1995.
F. H. Clarke, Optimization and nonsmooth analysis. Siam, 1990, vol. 5.
P. R. Dahl, “A solid friction model,” The Aerospace Corporation El
Segundo CA, Tech. Rep., 1968.
G. Ferretti, G. Magnani, and P. Rocco, “Single and multistate integral
friction models,” IEEE Transactions on Automatic Control, vol. 49,
no. 12, pp. 2292–2297, 2004.
A. F. Filippov, Differential equations with discontinuous righthand sides:
control systems. Kluwer Academic Publishers, 1988.
R. Goebel, R. G. Sanfelice, and A. R. Teel, Hybrid Dynamical Systems:
modeling, stability, and robustness. Princeton University Press, 2012.
J. W. Hagood and B. S. Thomson, “Recovering a function from a Dini
derivative,” The American Mathematical Monthly, vol. 113, no. 1, pp.
34–46, 2006.
H. K. Khalil, Nonlinear Systems, 3rd ed. Prentice Hall, 2002.
M. A. Krasnosel’skii and A. V. Pokrovskii, Systems with hysteresis.
Springer-Verlag, 1989.
R. I. Leine and N. Van de Wouw, Stability and convergence of mechanical systems with unilateral constraints. Springer Science & Business
Media, 2007, vol. 36.
H. Olsson, K. J. Åström, C. Canudas-de-Wit, M. Gäfvert, and P. Lischinsky, “Friction models and friction compensation,” European journal of
control, vol. 4, no. 3, pp. 176–195, 1998.
D. Putra, H. Nijmeijer, and N. van de Wouw, “Analysis of undercompensation and overcompensation of friction in 1 DOF mechanical systems,”
Automatica, vol. 43, no. 8, pp. 1387–1394, 2007.
E. Rabinowicz, “The nature of the static and kinetic coefficients of
friction,” Journal of applied physics, vol. 22, no. 11, pp. 1373–1379,
1951.
C. J. Radcliffe and S. C. Southward, “A property of stick-slip friction
models which promotes limit cycle generation,” in American Control
Conference, 1990. IEEE, 1990, pp. 1198–1205.
9
[22] R. T. Rockafellar and R. J.-B. Wets, Variational analysis. Springer
Science & Business Media, 2009, vol. 317.
[23] E. P. Ryan, “An integral invariance principle for differential inclusions
with applications in adaptive control,” SIAM Journal on Control and
Optimization, vol. 36, no. 3, pp. 960–980, 1998.
[24] E. D. Sontag, “Further facts about input to state stabilization,” IEEE
Transactions on Automatic Control, vol. 35, no. 4, pp. 473 – 476, 1990.
[25] J. Swevers, F. Al-Bender, C. G. Ganseman, and T. Projogo, “An
integrated friction model structure with improved presliding behavior
for accurate friction compensation,” IEEE Transactions on Automatic
Control, vol. 45, no. 4, pp. 675–686, 2000.
[26] N. Van de Wouw and R. I. Leine, “Attractivity of equilibrium sets of
systems with dry friction,” Nonlinear Dynamics, vol. 35, no. 1, pp. 19–
39, 2004.
[27] A. Visintin, Differential models of hysteresis. Springer-Verlag Berlin
Heidelberg, 1994, vol. 111.
A PPENDIX A
P ROOF OF C LAIM 1
Proof of Claim 1. The proof of the claim shows for each
possible initial conditions that for a suitable k the solution to
the affine system ξ˙ = fk (ξ) among (21a)-(21c) is also solution
to (15) on the interval [0, T ]. By Lemma 1, this is also the
unique solution to (15) on the same interval. The suitable k
for each initial condition is listed in the next table.
Case
Possible initial condition
(i)
(ii)
(iii)
(iv)
(v)
(vi)
(vii)
(viii)
(ix)
(x)
(xi)
v̄
v̄
v̄
v̄
v̄
v̄
v̄
v̄
v̄
v̄
v̄
>0
= 0, φ̄ > fc
= 0, φ̄ = fc , σ̄ > 0
= 0, φ̄ = fc , σ̄ = 0
= 0, φ̄ = fc , σ̄ < 0
= 0, |φ̄| < fc
= 0, φ̄ = −fc , σ̄ > 0
= 0, φ̄ = −fc , σ̄ = 0
= 0, φ̄ = −fc , σ̄ < 0
= 0, φ̄ < −fc
<0
k
1
1
1
0
0
0
0
0
−1
−1
−1
Define the components of the solution ξ as (ξσ , ξφ , ξv ). We
first prove item 1) and the left equation in (21d), addressing
only cases (i)-(v) as the remaining ones follow from parallel
arguments.
Case (i). We choose k = 1, that is, (21a). Because ξv (0) =
v̄ > 0, there exists a finite time T such that for all t ∈ [0, T ]
ξv (t) > 0. When the resulting solution ξ is substituted
into (15), −fc SGN(ξv (t)) = {−fc } for all t ∈ [0, T ] so that
˙ = Aξ(t) − b, and this is
for all t ∈ [0, T ] (15) becomes ξ(t)
satisfied precisely because ξ arises from (21a) (k = 1). Then
the solution ξ is also a solution to (15) for t ∈ [0, T ] because
˙
they have the same initial conditions and ξ(t)
∈ F (ξ(t)).
V (ξ(t)) = V1 (ξ(t)) for all t ∈ [0, T ] because f = fc is the
only possible selection in (18a) due to ξv (t) > 0 in [0, T ].
Case (ii). We prove that from (21a) (that is, k = 1) and
initial conditions ξ(0) = (σ̄, φ̄, 0) with φ̄ > fc it is implied
that there exist T > 0 such that for all t ∈ (0, T ], ξv (t) > 0.
Then, as in Case (i), we can conclude that ξ is also a solution
to (15). Indeed, the third state equation of (21a) reads ξ˙v =
ξφ −kv ξv −fc with ξv (0) = 0, ξφ (0) = φ̄ > fc , so that ξ˙v (0) =
ξφ (0)−fc > 0. Also here V (ξ(t)) = V1 (ξ(t)) because ξv (t) >
0 for all t ∈ (0, T ] and ξφ (0) > fc so that the minimizer
in (18a) at t = 0 is also f = fc .
Case (iii). We prove that from (21a) (k = 1) and initial
conditions ξ(0) = (σ̄, fc , 0) with σ̄ > 0 it is implied that
there exist T > 0 such that for all t ∈ (0, T ], ξv (t) > 0. Then,
as in Case (i), we can conclude that ξ is also a solution to (15).
Indeed, the second state equation of (21a) reads in this case
ξ˙φ = ξσ − kp ξv and ξ˙φ (0) = ξσ (0) > 0, so that there exists T
such that φ(t) > fc for all t ∈ (0, T ]. The third state equation
of (21a) reads ξ˙v = ξφ − kv ξv − fc and is ξ˙v (0) = 0 when
evaluated at time 0. Therefore, to show that ξv (t) > 0 for all
t ∈ (0, T ], we need to differentiate the third equation and get
ξ¨v (0) = σ̄ > 0 and by this we conclude that ξ˙v (t) > 0 in
(0, T ] and then also ξv (t) > 0 in (0, T ]. A similar reasoning
to the previous case shows that V (ξ(t)) = V1 (ξ(t)), for all
t ∈ [0, T ].
Case (iv). By choosing k = 0, we can verify that the
solution to (15) and (21b) is constant (specifically, an equilibrium). Then V (ξ(t)) = V0 (ξ(t)) for all t ∈ [0, +∞) because
V (ξ(0)) = V0 (ξ(0)).
Case (v). Here we have the explicit solution ξσ (t) =
σ̄ < 0, ξφ (t) = fc + σ̄t, ξv (t) = 0, for the interval
[0, T ] = [0, −2fc /σ̄]. This solution satisfies (15) because
an appropriate selection of the value in SGN(0) obtains
ξφ (t) − fc SGN(ξv (t)) = ξφ (t) − fc SGN(0) = {0} for all
t ∈ [0, −2fc /σ̄], as in this interval |ξφ (t)| ≤ fc . In this case
the minimizer in (18a) is f (t) = ξφ (t) for all t ∈ [0, T ], so
that V (ξ(t)) = V0 (ξ(t)) for all t ∈ [0, T ].
Finally, c = 2(kv kp − ki ) > 0 (by Assumption 1) is such
that the right inequality in (22) holds for each k ∈ {−1, 0, 1}
regardless of the initial condition
and
consideration.
case under
ξσ
T
ξσ
d
d
ξφ −fc
V1 (ξ(t)) = dt
P ξφ −fc
=
For k = 1 write dt
ξv
T ξv
ξσ
ξσ
(Aξ − b)T P ξφ −fc + ξφ −fc
P (Aξ − b) = −cξv2 , that
ξv
ξv
satisfies (22) in the interval [0, T ]. Parallel computations
hold for k = −1. For k = 0, V0 (ξ(t)) = kkvi ξv (t)2 so that
kv
d
2
˙
dt V0 (ξ(t)) = 2 ki ξσ ξσ = 0 ≤ 0 = cξv (t) .
| 3cs.SY
|
Universal Protocols for Information Dissemination
Using Emergent Signals
Bartłomiej Dudek1 and Adrian Kosowski∗2
arXiv:1705.09798v3 [cs.DS] 29 Nov 2017
1
University of Wrocław, Poland
2
Inria Paris, France
Abstract
We consider a population of n agents which communicate with each other in a decentralized
manner, through random pairwise interactions. One or more agents in the population may act as
authoritative sources of information, and the objective of the remaining agents is to obtain information from or about these source agents. We study two basic tasks: broadcasting, in which the agents
are to learn the bit-state of an authoritative source which is present in the population, and source
detection, in which the agents are required to decide if at least one source agent is present in the
population or not.
We focus on designing protocols which meet two natural conditions: (1) universality, i.e., independence of population size, and (2) rapid convergence to a correct global state after a reconfiguration, such as a change in the state of a source agent. Our main positive result is to show that both of
these constraints can be met. For both the broadcasting problem and the source detection problem,
we obtain solutions with a convergence time of O(log2 n) rounds, w.h.p., from any starting configuration. The solution to broadcasting is exact, which means that all agents reach the state broadcast
by the source, while the solution to source detection admits one-sided error on a ε-fraction of the
population (which is unavoidable for this problem). Both protocols are easy to implement in practice
and have a compact formulation.
Our protocols exploit the properties of self-organizing oscillatory dynamics. On the hardness
side, our main structural insight is to prove that any protocol which meets the constraints of universality and of rapid convergence after reconfiguration must display a form of non-stationary behavior
(of which oscillatory dynamics are an example). We also observe that the periodicity of the oscillatory behavior of the protocol, when present, must necessarily depend on the number #X of source
agents present in the population. For instance, our protocols inherently rely on the emergence of a
signal passing through the population, whose period is Θ(log #nX ) rounds for most starting configurations. The design of clocks with tunable frequency may be of independent interest, notably in
modeling biological networks.
Key words: Gossiping, Epidemic processes, Oscillatory dynamics, Emergent phenomena,
Population protocols, Broadcasting, Distributed clock synchronization.
∗
Corresponding Author. Email: [email protected]
1
Introduction
Information-spreading protocols, and more broadly epidemic processes, appear in nature, social interactions between humans, as well as in man-made technology, such as computer networks. For some
protocols we have a reasonable understanding of the extent to which the information has already spread,
i.e., we can identify where the information is located at a given step of the process: we can intuitively
say which nodes (or agents) are “informed” and which nodes are “uninformed”. This is the case for
usual protocols in which uninformed agents become informed upon meeting a previously informed
agent, cf. e.g. mechanisms of rumor spreading and opinion spreading models studied in the theory community [26, 29]). Arguably, most man-made networking protocols for information dissemination also
belong to this category.
By contrast, there exists a broad category of complex systems for which it is impossible to locate
which agents have acquired some knowledge, and which are as yet devoid of it. In fact, the question of
“where the information learned by the system is located” becomes somewhat fuzzy, as in the case of both
biological and synthetic neural networks. In such a perspective, information (or knowledge) becomes
a global property of the entire system, whereas the state of an individual agent represents in principle
its activation, rather than whether it is informed or not. As such, knowledge has to be treated as an
emergent property of the system, i.e., a global property not resulting directly from the local states of its
agents. The convergence from an uninformed population to an informed population over time is far from
monotonous. Even so, once some form of “signal” representing global knowledge has emerged, agents
may try to read and copy this signal into their local state, thus each of them eventually also becomes
informed. At a very informal conceptual level, we refer to this category of information-dissemination
protocols as protocols with emergent behavior. At a more technical level, emergent protocols essentially
need to rely on non-linear dynamical effects, which typically include oscillatory behavior, chaotic effects, or a combination of both. (This can be contrasted with simple epidemic protocols for informationspreading, in which nodes do not become deactivated.)
This work exhibits a simple yet fundamental information-spreading scenario which can only be addressed efficiently using emergent protocols. Both the efficient operation of the designed protocols,
and the need for non-stationary dynamical effects in any efficient protocol for the considered problems,
can be formalized through rigorous theoretical analysis. Our goal in doing this is twofold: to better
understand the need for emergent behavior in real-world information spreading, and to display the applicability of such protocols in man-made information spreading designs. For the latter, we describe
an interpretation of information as a (quasi-)periodic signal, which can be both decoded from states of
individual nodes, and encoded into them.
1.1
Problems and Model
We consider a population of n identical agents, each of which may be in a constant number of possible
states. Interactions between agents are pairwise and random. A fair scheduler picks a pair of interacting
agents independently and uniformly at random in each step. The protocol definition is provided through
a finite sequence of state transition rules, following the precise conventions of the randomized Population
Protocol model [6, 8] or (equivalently) of Chemical Reaction Networks [21].1
The input to the problem is given by fixing the state of some subset of agents, to some state of the
protocol, which is not available to any of the other agents. Intuitively, the agents whose state has been
fixed are to be interpreted as authoritative sources of information, which is to be detected and disseminated through the network (i.e., as the rumor source node, broadcasting station, etc.). For example, the
1
The activation model is thus asynchronous. The same protocols may be deployed in a synchronous setting, with scheduler
activations following, e.g., the independent random matching model (with only minor changes to the analysis) or the PULL
model [29] (at the cost of significantly complicating details of the protocol formulation).
2
problem of spreading a bit of information through the system is formally defined below.
Problem B IT B ROADCAST
Input States: X1 , X2 .
Promise: The population contains a non-zero number of agents in exactly one of the two input states
{X1 , X2 }.
Question: Decide if the input state present in the population is X1 or X2 .
We can, e.g., consider that the transmitting station (or stations) choose whether to be in state X1 or X2 in
a way external to the protocol, and thus transmit the “bit” value 1 or 2, respectively, through the network.
Broadcasting a bit is one of the most fundamental networking primitives.
The definition of the population protocol includes a partition of the set of states of the protocol
into those corresponding to the possible answers to the problem. When the protocol is executed on the
population, the output of each agent may be read at every step by checking, for each agent, whether its
state belongs to the subset of an output state with a given answer (in this case, the answer of the agent
will be the “bit” it has learned, i.e., 1 or 2). We will call a protocol exact if it eventually converges to
a configuration, such that starting from this configuration all agents always provide the correct answer.
We will say it operates with ε-error, for a given constant ε > 0, if starting from some step, at any given
step of the protocol, at most an ε-fraction of the population holds the incorrect answer, with probability
1 − O(1/n).
Time is measured in steps of the scheduler, with n time steps called a round, with the expected
number of activations of each agent per round being a constant. Our objective is to design protocols
which converge to the desired outcome rapidly. Specifically, a protocol is expected to converge in
O(poly log n) rounds (i.e., in O(npoly log n) steps), with probability 1 − O(1/n), starting from any
possible starting configuration of states in the population, conforming to the promise of the problem.2
Motivated by both applications and also a need for a better understanding of the broadcasting problem, we also consider a variant of the broadcasting problem in which no promise on the presence of the
source is given. This problem, called D ETECTION, is formally defined below.
Problem D ETECTION
Input State: X.
Question: Decide if at least one agent in state X is present in the population.
Detection of the presence of a source is a task which is not easier than broadcasting a bit. Indeed, any
detection protocol is readily converted into a broadcasting protocol for states {X1 , X2 } by identifying
X = X1 and treating X2 as a dummy state which does not enter into any interactions (i.e., is effectively
not visible in the network). Intuitively, the detection task in the considered setting is much harder: a
source X may disappear from the network at any time, forcing other agents to spontaneously “unlearn”
the outdated information about the presence of the source. This property is inherently linked to the
application of the D ETECTION problem in suppressing false rumors or outdated information in social
interactions. Specifically, it may happen that a certain part of a population find themselves in an informed
state before the original rumor source is identified as a source of false information, a false rumor may
be propagated accidentally because of an agent which previously changed state from “uninformed” to
“informed” due to a fault or miscommunication, or the rumor may contain information which is no
longer true. Similar challenges with outdated information and/or false-positive activations are faced
in Chemical Reaction Networks, e.g. in DNA strand displacement models [5]. In that context, the
2
We adhere to this strong requirement for self-stabilizing (or self-organizing) behavior from any initial configuration in the
design of our protocols. The presented impossibility results still hold under significantly weaker assumptions.
3
P ROBLEM :
B IT B ROADCAST
D ETECTION
Non-stationarity property:
(Applies to all fast O(1)-state protocols)
no fixed points while source
transmits random bits
no fixed points while source is
present
New protocols with emergent signal:
Convergence time:
– No error (exact output)
– One-sided ε-error
universal, 74 states
universal, 55 states
O(log2 n)
impossible
O(log2 n)
Other protocols with ω(1)-states:
Clock-Sync (in synchronized
round model) [11]
Time-to-Live [5]
detection problem has the intuitive interpretation of detecting if a given type of chemical or biological
agent (e.g., a contaminant, cancer cell, or hormonal signal) is present in the population, and spreading
this information among all agents.
1.2
Our Results
In Section 3, we show that both the B IT B ROADCAST, and the D ETECTION problem can be solved with
protocols which converge in O(log2 n) rounds to an outcome, with probability 1 − O(1/n), starting
from any configuration of the system. The solution to B IT B ROADCAST guarantees a correct output.
The solution to D ETECTION admits one-sided ε-error: in the absence of a source, all agents correctly
identify it as absent, whereas when the source is present, at any moment of time after convergence the
probability that at least (1 − ε)n agents correctly identify the source as present is at least 1 − O(1/n).3
Here, ε > 0 is a constant influencing the protocol design, which can be made arbitrarily small.
The designed protocols rely on the same basic building block, namely, a protocol realizing oscillatory dynamics at a rate controlled by the number of present source states in the population. Thus, these
protocols display non-stationary behavior. In Section 4, we show that such behavior is a necessary property in the following sense. We prove that in any protocol which solves D ETECTION in sub-polynomial
time in n and which uses a constant number of states, the number of agents occupying some state has to
undergo large changes: by a polynomially large factor in n during a time window of length proportional
to the convergence time of the protocol. For the B IT B ROADCAST problem, we show that similar volatile
behavior must appear in a synthetic setting in which a unique source is transmitting its bit as random
noise (i.e., selecting its input state {X1 , X2 } uniformly at random in subsequent activations).
We note that, informally speaking, our protocols rely on the emergence of a “signal” passing through
the population, whose period is Θ(log #nX ) rounds when the number of source agents in state X is #X. In
Section 5, we then discuss how the behavior of any oscillatory-type protocol controlled by the existence
of #X has to depend on both n and #X. We prove that for any such protocol with rapid convergence,
the cases of subpolynomial #X and #X = Θ(n) can be separated by looking at the portion of the
configuration space regularly visited by the protocol. This, in particular, suggests the nature of the
dependence of the oscillation period on the precise value of #X, and that the protocols we design with
period Θ(log #nX ) are among the most natural solutions to the considered problems.
The proofs of all theorems are deferred to the closing sections of the paper.
3
The existence of one-sided error is inherent to the D ETECTION problem in the asynchronous setting: indeed, if no agent
of the population has not made any communication with the source over an extended period of time, it is impossible to tell
for sure if the source has completely disappeared from the network, or if it is not being selected for interaction by the random
scheduler.
4
1.3
Comparison to the State-of-the-Art
Our work fits into lines of research on rumor spreading, opinion spreading, population protocols and
other interaction models, and emergent systems. We provide a more comprehensive literature overview
of some of these topics in Subsection 1.4.
Other work on the problems. The B IT B ROADCAST problem has been previously considered by
Boczkowski, Korman, and Natale in [11], in a self-stabilizing but synchronous (round-based) setting.
A protocol solving their problem was presented, giving a stabilized solution in Õ(log n) time, using a
number of states of agents which depends on population size, but exchanging messages of bit size O(1)
(this assumption can be modeled in the population protocol framework as a restriction on the permitted
rule set). In this sense, our result can be seen as providing improved results with respect to their approach, since it is applicable in an asynchronous setting and reduces the number of states to constant
(the latter question was open [11]). We remark that their protocol has a more general application to the
problem of deciding which of the sources X1 or X2 is represented by a larger number of agents, provided these two numbers are separated by a multiplicative constants. (Our approach could also be used
in such a setting, but the required separation of agent numbers to ensure a correct output would have to
be much larger: we can compare values if their logarithms are separated by a multiplicative constant.)
The protocol of [11] involves a routine which allows the population to create a synchronized modulo
clock, working in a synchronous setting. The period of this clock is independent of the input states of
the protocol, which should be contrasted with the oscillators we work on in this paper, which encode the
input into a signal (with a period depending on the number of agents in a given input state).
The D ETECT problem was introduced in a work complementary to this paper [5]. Therein we look
at applications of the confirmed rumor spreading problem in DNA computing, focusing on performance
on protocols based on a time-to-live principle and on issues of fault tolerance in a real-world model with
leaks. The protocols designed there require O(log n) states, and while self-stabilizing, do not display
emergent behavior (in particular, agents can be categorized as “informed” and “uninformed”, the number
of correctly informed agents tends to increase over time, while the corresponding continuous dynamical
system stabilizes to a fixed point attractor).
Originality of methods. The oscillatory dynamics we apply rely on an input-parameter-controlled
oscillator. The uncontrolled version of the oscillator which we consider here is the length-3 cyclic
oscillator of the cyclic type, known in population dynamics under the name of rock-paper-scissors (or
RPS). This has been studied intensively in the physics and evolutionary dynamics literature (cf.e.g. [46]
for a survey), while algorithmic studies are relatively scarce [15]. We remark that the uncontrolled
cyclic oscillator with a longer (but O(1) length cycle) has been applied for clock/phase synchronization
in self-stabilizing settings and very recently in the population protocol setting when resolving the leader
election problem [25]. (The connection to oscillatory dynamics is not made explicit, and the longer cycle
provides for a neater analysis, although it does not seem to be applicable to our parameter-controllable
setting.) Whereas we are not aware of any studies of parameter-controlled oscillators in a protocol design
setting (nor for that matter, of rigorous studies in other fields), we should note that such oscillators
have frequently appeared in models of biological systems, most notably in biological networks and
neuroendocrinology ( [27] for a survey). Indeed, some hormone release and control mechanisms (e.g.,
for controlling GnRH surges in vertebrates) appear to be following a similar pattern. To the best of our
knowledge, no computational (i.e., interaction-protocol-based) explanation for these mechanisms has
yet been proposed, and we hope that our work may provide, specifically on the D ETECT problem, may
provide some insights in this direction.
In terms of lower bounds, we rely on rather tedious coupling techniques for protocols allowing
randomization, and many of the details are significantly different from lower-bound techniques found
5
in the population protocols literature. We remark that a recent line of work in this area [3, 21] provides
a powerful set of tools for proving lower bounds on the number of states (typically Ω(log log n) states)
for fast (typically polylogarithmic) population protocols for different problems, especially for the case
of deterministic protocols. We were unable to leverage these results to prove our lower bound for the
randomized scenario studied here, and believe our coupling analysis is complementary to their results.
1.4
Other Related Work
Our work fits into the line of research on rumor spreading, population protocols, and related interaction
models. Our work also touches on the issue of how distributed systems may spontaneously achieve some
form of coordination with minimum agent capabilities. The basic work in this direction, starting with the
seminal paper [32], focuses on synchronizing timers through asynchronous interprocess communication
to allow processes to construct a total ordering of events. A separate interesting question concerns local
clocks which, on their own, have some drift, and which need to synchronize in a network environment
(cf. e.g. [35, 37], or [34] for a survey of open problems).
Rumor spreading. Rumor spreading protocols are frequently studied in a synchronous setting. In a
synchronous protocol, in each parallel round, each vertex independently at random activates a local rule,
which allows it either to spread the rumor (if it is already informed), or possibly also to receive it (if it has
not yet been informed, as is the case in the push-pull model). The standard push rumor spreading model
assumes that each informed neighbor calls exactly one uninformed neighbor. In the basic scenario, corresponding to the complete interaction network, the number of parallel rounds for a single rumor source
to inform all other nodes is given as log2 n + ln n + o(log n), with high probability [24, 42]. More general graph scenarios have been studied in [22] in the context of applications in broadcasting information
in a network. Graph classes studied for the graph model include hypercubes [22], expanders [45], and
other models of random graphs [23]. The push-pull model of rumor spreading is an important variation:
whereas for complete networks the speedup due to the pull process is in the order of a multiplicative
constant [29], the speed up turns out to be asymptotic, e.g., on preferential attachment graphs, where the
rumor spreading time is reduced from Θ(log n) rounds in the push model to Θ(log n/ log log n) rounds
in the push-pull model [18], as well as on other graphs with a non-uniform degree distribution. The
push-pull model often also proves more amenable to theoretical analysis. We note that asynchronous
rumor spreading on graphs, in models closer to our random scheduler, has also been considered in recent work [26, 40], with [26] pointing out the tight connections between the synchronous (particularly
push-pull) and asynchronous models in general networks.
Population protocols. Population protocols are a model which captures the way in which the complex behavior of systems (biological, sensor nets, etc.) emerges from the underlying local interactions
of agents. The original model of Angluin et al. [6, 7] was motivated by applications in sensor mobility.
Despite the limited computational capabilities of individual sensors, such protocols permit at least (depending on available extensions to the model) the computation of two important classes of functions:
threshold predicates, which decide if the weighted average of types appearing in the population exceeds
a certain value, and modulo remainders of similar weighted averages. The majority function, which belongs to the class of threshold functions, was shown to be stably computable for the complete interaction
graph [6]; further results in the area of majority computation can be found in [7, 9, 10, 38]. A survey of
applications and models of population protocols is provided in [9, 39]. An interesting line of research is
related to studies of the algorithmic properties of dynamics of chemical reaction networks [21]. These
are as powerful as population protocols, though some extensions of the chemical reaction model also
allow the population size to change in time. Two very recent results in the population protocol model are
6
worthy of special attention. Alistarh, Aspnes, and Gelashvili [4] have resolved the question of the number of states required to solve the Majority problem on a complete network in polylogarithmic time as
Θ(log n). For the equally notable task of Leader Election, the papers of Gasieniec and Stachowiak [25]
(for the upper bound) together with the work of Alistarh, Aspnes, Eisenstat, Gelashvili, and Rivest [3]
(for the lower bound) put the number of states required to resolve this question in polylogarithmic time
as Θ(log log n). Both of these results rely on a notion of a self-organizing phase clock.
Nonlinearity in interaction protocols. Linear dynamical systems, as well as many nonlinear protocols subjected to rigorous analytical study, have a relatively simple structure of point attractors and
repellers in the phase space. The underlying continuous dynamics (in the limit of n → +∞) of many
interaction protocols defined for complete graphs would fit into this category: basic models of randomized rumor spreading [42]; models of opinion propagation (e.g. [1, 14]); population protocols for
problems such as majority and thresholds [6, 7]; all reducible Markov chain processes, such as random
walks and randomized iterative load balancing schemes.
Nonlinear dynamics with non-trivial limit orbits are fundamental to many areas of systems science,
including the study of physical, chemical and biological systems, and to applications in control science.
In general, population dynamics with interactions between pairs of agents are non-linear (representable
as a set of quadratic difference equations) and have potentially complicated structure if the number of
states is 3 or more. For example, the simple continuous Lotka-Volterra dynamics [36] gives rise to a
number of discrete models, for example one representing interactions of the form A + B → A + A,
over some pairs A, B of states in a population (cf. [46] for further generalizations of the framework
or [15] for a rigorous analysis in the random scheduler model). The model describes transient stability
in a setting in which several species are in a cyclic predator-prey relation. Cyclic protocols of the type
have been consequently identified as a potential mechanism for describing and maintaining biodiversity,
e.g., in bacterial colonies [30, 31]. Cycles of length 3, in which type A2 attacks type A1 , type A3
attacks type A2 , and type A1 attacks type A3 , form the basis of the basic oscillator, also used as the
starting point for protocols in this work, which is referred to as the RPS (rock-paper-scissors) oscillator
or simply the 3-cycle oscillator, which we discuss further in Section 6.1. This protocol has been given
a lot of attention in the statistical physics literature. The original analytical estimation method applied
to RPS was based on approximation with the Fokker-Planck equation [44]. A subsequent analysis of
cyclic 3- and 4-species models using Khasminskii stochastic averaging can be found in [16], and a mean
field approximation-based analysis of RPS is performed in [41]. In [15], we have performed a study of
some algorithmic implications of RPS, showing that the protocol may be used to perform randomized
choice in a population, promoting minority opinions, in Õ(n2 ) steps. All of these results provide a good
qualitative understanding of the behavior of the basic cyclic protocols. We remark that the protocol
used in this paper is directly inspired by the properties of RPS, as we discuss further on, but has a more
complicated interaction structure (see Fig. 1).
For protocols with convergence to a single point in the configuration space in the limit of large
population size, a discussion of the limit behavior is provided in [12], who provide examples of protocols
converging to limit points at coordinates corresponding to any algebraic numbers.
We also remark that local interaction dynamics on arbitrary graphs (as opposed to the complete
interaction graph) exhibit a much more complex structure of their limit behavior, even if the graph has
periodic structure, e.g., that of a grid. Oscillatory behavior may be overlaid with spatial effects [46],
or the system may have an attractor at a critical point, leading to simple dynamic processes displaying
self-organized criticality (SOC, [43]).
7
2
2.1
Preliminaries: Building Blocks for Population Protocols
Protocol Definition
A randomized population protocol for a population of n agents is defined as a pair P = (Kn , Rn ),
where Kn is the set of states and Rn is the set of interaction rules. The interaction graph is complete.
We will simply write P = (K, R), when considering a protocol which is universal (i.e., defined in
the same way for each value of n) or if the value of n is clear from the context. All the protocols we
design are universal; our lower bounds also apply to some non-universal protocols. The set of rules
R ⊆ K 4 × [0, 1] is given so that each rule j ∈ R is of the form j = (i1 (j), i2 (j), o1 (j), o2 (j), qj ),
describing an interaction read as: “(i1 (j), i2 (j)) → (o1 (j), o2 (j))with probability qj ”. For all i1 , i2 ∈
{j ∈ R : (i1 (j), i2 (j)) = (i1 , i2 )} as the set of rules acting on the pair of states
K, we define Ri1 ,i2 = P
i1 , i2 , and impose that j∈Ri ,i qj ≤ 1.
1 2
For a state A ∈ K, we denote the number of agents in state AP
as #A, and the concentration of state
#
#
A as a = A/n, and likewise for a set of states A, we write A = A∈A #A.
In any configuration of the system, each of the n agents from the population is in one of states
from Kn . The protocol is executed by an asynchronous scheduler, which runs in steps. In every step
the scheduler uniformly at random chooses from the population a pair of distinct agents to interact:
the initiator and the receiver. If the initiator and receiver are in states i1 and i2 , respectively, then the
protocol executes at most one rule from set protocol Ri1 ,i2 , selecting rule j ∈ Ri1 ,i2 with probability qj .
If rule j is executed, the initiator then changes its state to o1 (j) and the receiver to o2 (j). The source
has a special state, denoted X in the D ETECT problem, or one of two special states, denoted {X1 , X2 }
in the B IT B ROADCAST problem, which is never modified by any rule.
All protocols are presented in the randomized framework, however, the universal protocols considered here are amenable to a form of conversion into deterministic rules discussed in [3], which simulates
randomness of rules by exploiting the inherent randomness of the scheduler in choosing interacting node
pairs to distribute weakly dependent random bits around the system.
All protocols designed in this work are initiator-preserving, which means that for any rule j ∈ R,
we have o1 (j) = i1 (j) (i.e., have all rules of the form A + B → A + C, also more compactly written as
A : B → C), which makes them relevant in a larger number of application. As an illustrative example,
we remark that the basic rumor spreading (epidemic) model is initiator-preserving and given simply as
1 : 0 → 1. All protocols can also obviously be rewritten to act on unordered pairs of agents picked by
the scheduler, rather than ordered pairs.
2.2
Protocol Composition Technique
Our protocols will be built from simpler elements. Our basic building block is the input-controlled
oscillatory protocol Po (see Fig. 1). We then use protocol Po as a component in the construction of
other, more complex protocols, without disrupting the operation of the original protocol.
Formally, we consider a protocol PB using state set B = {Bi : 1 ≤ i ≤ kb } and rule set RB , and
a protocol extension PBC using a state set B × C = B × {Ci : 1 ≤ i ≤ kc }, where C is disjoint from
B, and rule extension set RBC . Each rule extension defines for each pair of states from B × C (i.e., to
each element of (B × C) × (B × C)) a probability distribution over elements of C × C.
The composed protocol PB ◦ PBC is a population protocol with set of states B × C. Its rules are
defined so that, for a selected pair of agents in states (Bi , Cj ) and (Bi0 , Cj 0 ), we obtain a pair of agents
in states (Bi∗ , Cj ∗ ) and (Bi0∗ , Cj 0∗ ) according to a probability distribution defined so that:
• Each pair Bi0∗ , Bi0∗ appears in the output states of the two agents with the same probability as it
would in an execution of protocol PB on a pair of agents in states Bi and Bi0 .
8
• Each pair Ci0∗ , Ci0∗ appears in the output states of the two agents with the probability given by the
definition of PBC .
In the above, the pairs of agents activated by PB and PBC are not independent of each other. This is a
crucial property in the composition of protocol Po when composing it with further blocks to solve the
D ETECT problem.
We denote by 1B the identity protocol which preserves agent states on set of states B. For a protocol
P , we denote by P/2 a lazy version of a protocol P in which the rule activation of P occurs with
probability 1/2, and with probability 1/2 the corresponding rule of the identity protocol is activated.
Note that all asymptotic bounds on expected and w.h.p. convergence time obtained for any protocol
P also apply to protocol P/2, in the regime of at least a logarithmic number of time steps. We also
sometimes treat a protocol PBC extension as a protocol in itself, applied to the identity protocol 1B .
The independently composed protocol PB + PBC is defined as an implementation of the composed
protocol (PB ) ◦ (PC /2), realized with the additional constraint that in each step, either the rule of PB
is performed with an identity rule extension, or the rule extension of PBC is performed on top of the
identity protocol 1B . Such a definition is readily verified to be correct by a simple coupling argument,
and allows us to analyze protocols PB and PBC , observing that the pairs (identities) of agents activated
by the scheduler in the respective protocols are independent.
All the composed protocols (and protocol extensions) we design are also initiator-preserving, i.e.,
Ci∗ = Ci and Bi∗ = Bi , with probability 1. In notation, rules omitted from the description of protocol
extensions are implicit, occurring with probability 0 (where Cj ∗ 6= Cj ) or with the probability necessary
for the normalization of the distribution to 1, where the state is preserved (where Cj ∗ = Cj ).
As a matter of naming convention, we name the states in the separate state sets of the composed
protocols with distinct letters of the alphabet, together with their designated subscripts and superscripts.
The rumor source X is treated specially and uses a separate letter (and may be seen as a one state
protocol without any rules, on top of which all other protocols are composed; in particular, its state is
?
never modified). The six remaining states of protocol Po are named with the letters A? , as usual in its
definition. Subsequent protocols will use different letters, e.g., M? and L? .
3
3.1
Overview of Protocol Designs
Main Routine: Input-Controlled Oscillator Protocol Po
We first describe the main routine which allows us to convert local input parameters (the existence
of source into a form of global periodic signal on the population. This main building block is the
construction of a 7-state protocol Po following oscillator dynamics, whose design we believe to be of
independent interest.
The complete design of protocol Po is shown in Fig. 1. The source state is denoted by X. Addi++
tionally, there are six states, called A+
i and Ai , for i ∈ {1, 2, 3}. The naming of states in the protocol
is intended to maintain a direct connection with the RPS oscillator dynamics, which is defined by the
++
simple rule “Ai : Ai−1 7→ Ai , for i = 1, 2, 3”. In fact, we will retain the convention Ai = {A+
i , Ai }
++
+
++
and ai = a+
to be different flavors of the same species
i + ai , and consider the two states Ai and Ai
+
Ai , referring to the respective superscripts as either lazy ( ) or aggressive (++ ).
The protocol has the property that in the absence of X, it stops in a corner state of the phase space,
in which only one of three possible states appears in the population, and otherwise regularly (every
O(log n) steps) moves sufficiently far away from all corner states. An intuitive formalization of the
basic properties of the protocol is given by the theorem below.
Theorem 1. There exists a universal protocol Po with |K| = 7 states, including a distinguished source
state X, which has the following properties.
9
(1) Interaction with an initiator from the same species makes receiver aggressive:
?
Ai :
?
Ai
7→
A++
i
(2) Interaction with an initiator from a different species makes receiver lazy (case of no attack):
?
Ai :
?
Ai+1 7→
A+
i+1
(3) A lazy initiator has probability p of performing a successful attack on its prey:
A+
i :
?
Ai−1
(
A+
with probability p,
i ,
7→
+
Ai−1 , otherwise.
(4) An aggressive initiator has probability 2p of performing a successful attack on its prey:
A++
i :
?
Ai−1
(
A+
with probability 2p,
i ,
7→
+
Ai−1 , otherwise.
(5) The source converts any receiver into a lazy state of a uniformly random species:
X:
?
A?
+
A1 , with probability 1/3,
7→ A+
2 , with probability 1/3,
+
A3 , with probability 1/3.
Figure 1: Rules of the basic oscillator protocol Po . The adopted notation for initiator-preserving rules
is of the form A : B 7→ C, corresponding to transitions written as A + B → A + C in the notation of
chemical reaction networks or (A, B) → (A, C) in the notation of population dynamics. All rules apply
to i ∈ {1, 2, 3}, whereas a question mark ? in a superscript or subscript denotes a wildcard, matching
any permitted combination of characters, which may be set independently for each agent. Probability
p > 0 is any (sufficiently small) absolute constant.
1. For any starting configuration, in the absence of the source (#X = 0), the protocol always reaches
a configuration such that:
++
++
• all agents are in the same state: either A++
1 , or A2 , or A3 ;
• no further state transitions occur after this time.
Such a configuration is reached in O(log n) rounds, with constant probability (and in O(log2 n)
rounds with probability 1 − O(1/n)).
2. For any starting configuration, in the presence of the source (#X ≥ 1), we have with probability
1 − O(1/n):
• for each state i ∈ K, there exists a time step in the next O(log #nX ) rounds when at least a
constant fraction of all agents are in state i;
• during the next O(log #nX ) rounds, at least a constant fraction of all agents change their
state at least once.
10
The proof of the Theorem is provided in Section 6.
The RPS dynamics provides the basic oscillator mechanism which is still largely retained in our
scenario. Most of the difficulty lies in controlling its operation as a function of the presence or absence
of the rumor source. We do this by applying two separate mechanisms. The presence of rumor source X
shifts the oscillator towards an orbit closer to the central orbit (A1 , A2 , A3 ) = (1/3, 1/3, 1/3) through
rule (5), which increases the value of potential φ := ln(a1 a2 a3 ), where ai = #Ai /n. Conversely, independent of the existence of rumor source X, a second mechanism is intended to reduce the value
of potential φ. This mechanism exploits the difference between the aggressive and lazy flavors of the
species. Following rule (1), an agent belonging to a species becomes more aggressive if it meets another
from the same species, and subsequently attacks agents from its prey species with doubled probability
following rule (4). This behavior somehow favors larger species, since they are expected to have (proportionally) more aggressive agents than the smaller species (in which pairwise interactions between
agents of the same species are less frequent) — the fraction of agents in Ai which are aggressive would,
in an idealized static scenario, be proportional to ai . (This is, in fact, often far from true due to the
interactions between the different aspects of the dynamics). As a very loose intuition, the destabilizing
behavior of the considered rule on the oscillator is resemblant of the effect an eccentrically fitted weight
has on a rotating wheel, pulling the oscillator towards more external orbits (with smaller values of φ).
The intuition for which the proposed dynamics works, and which we will formalize and prove rigorously in Section 6, can now be stated as follows: in the presence of rumor source X, the dynamics will
converge to a form of orbit on which the two effects, the stabilizing and destabilizing one, eventually
compensate each other (in a time-averaged sense). The period of a single rotation of the oscillator around
such an orbit is between O(1) and O(log n), depending on the concentration of X. In the absence of X,
the destabilizing rule will prevail, and the oscillator will quickly crash into a side of the triangle.
For small values of #X > 0, the protocol can be very roughly (and non-rigorously) viewed as cyclic
composition of three dominant rumor spreading processes over three sets of states A1 , A2 , A3 , one
converting states A1 to A3 , the next from A3 to A2 , and the last from A2 to A1 , which spontaneously
take over at moments of time separated by O(log n) parallel rounds. For other starting configurations,
and especially for the case of #X = 0, the dynamics of the protocol, which has 5 free dimensions,
is more involved to describe and analyze (see Section 6.4). We provide some further insights into the
operation of the protocol in Section 7.1, notably formalizing the notion that an intuitively understood
oscillation (going from a small number of agents in some state Ai , to a large number of agents in state
Ai , and back again to a small number of agents in state Ai ) takes Θ(log #nX ) steps, with probability
1 − O(1/n). As such, protocol Po can be seen as converting local input #X into a global periodic signal
with period Θ(log #nX ). What remains is allowing nodes to extract information from this periodic signal.
Simulation timelines shown in Fig.9 in the Appendix illustrate the idea of operation of protocol Po
and its composition with other protocols.
3.2
Protocols for B IT B ROADCAST
A solution to B IT B ROADCAST is obtained starting with an independent composition of two copies of
oscillator Po , called Po[1] and Po[2] , with states in one protocol denoted by subscript [1] and in the other
by subscript [2]. The respective sources are thus written as X[1] and X[2] . In view of Theorem 1, in this
composition Po[1] + Po[2] , under the promise of the B IT B ROADCAST problem, one of the oscillators will
be running and the other will stop in a corner of its state space. Which of the oscillators is running can
be identified by the presence of states A+
i [z], which will only appear for z ∈ {1, 2} corresponding to
the operating oscillator. Moreover, by the same Theorem, every O(log n) rounds a constant number of
agents of this oscillator will be in such a state A+
i [z], for any choice of i ∈ {1, 2, 3}. We can thus design
the protocol extension Pb to detect this. This is given by the pair of additional output states {Y1 , Y2 } and
the rule extension consisting of the two rules shown in Fig. 2.
11
(6a) Interaction with agent state A+
?[1] sets agent output to Y1 :
++
(A+
?[1] , A?[2] , Y? ) :
(A??[1] , A??[2] , Y? )7→ Y1
(6a) Interaction with agent state A+
?[2] sets agent output to Y2 :
+
(A++
?[1] , A?[2] , Y? ) :
(A??[1] , A??[2] , Y? )7→ Y2
Figure 2: Protocol extension Pb of protocol (Po[1] + Po[2] ), with additional states {Y1 , Y2 }.
(5a-5b) Interaction with source X[1] alters corner configurations different from A1 :
X[1] :
A?3 7→ A+
1
X[1] :
A?2 7→ A+
3
(5c-5d) Interaction with source X[2] alters corner configurations different from A2 :
X[2] :
A?1 7→ A+
2
X[2] :
A?3 7→ A+
1
Figure 3: Protocol Po0 is obtained from the basic oscillator protocol Po by removing Rule (5) and replacing it by the two rules shown above.
Theorem 2 (Protocol for B IT B ROADCAST). Protocol (Po[1] + Po[2] ) + Pb , having |K| = 74 states,
including distinguished source states X[1] , X[2] converges to an exact solution of B IT B ROADCAST.
This occurs in O(log2 n) parallel rounds, with probability 1 − O(1/n). In the output encoding, agent
states of the form (·, ·, Yz ) represent answer “z”, for z ∈ {1, 2}.
The protocol (Po[1] + Po[2] ) + Pb is not “silent”, i.e., it undergoes perpetual transitions of state, even
once the output has been decided. As a side remark, we note that for the single-source broadcasting
problem, or more generally for the case when the number of sources is small, max{#X[1] , #X[2] } =
O(1), we can propose the following simpler silent protocol. We define protocol Po0 , by modifying
protocol Po as follows. We remove from it Rule (5), and replace it by to the four rules shown in
Fig. 3. The analysis of the modified protocol follows from the same arguments as those used to prove
Theorem 1(1). In the regime of max{#X[1] , #X[2] } = O(1), the effect of the source does not influence the
convergence of the process and each of the three possible corner configurations, with exclusively species
{A1 , A2 , A3 }, is reached in O(log n) steps with constant probability. However, rules (5a)−(5d) enforce
that the only stable configuration which will persist is the one in a corner corresponding to the identity
of the source, i.e., A1 for source X[1] and A2 for source X[2] ; the source will restart the oscillator in all
other cases. We thus obtain the following side result, for which we leave out the details of the proof.
Observation 1. Protocol Po0 , having |K| = 6 + 2 = 8 states, including distinguished source states X[1] ,
X[2] converges to an exact solution of B IT B ROADCAST, eventually stopping with all agents in state
12
(6) Meeting an agent of a different type resets majority setting:
((X or A?j ), M? ) :
(
M+1 , with probability 1/2
(A?i , M? ) 7→
,
M−1 , with probability 1/2
for j 6= i.
(7-10) Three-state majority protocol among agents of the same type:
(7) :
(A?i , M−1 ) :
(A?i , M+1 )7→M0 ,
with probability r.
(8) :
(A?i , M+1 ) :
(A?i , M+1 ) :
(A?i , M−1 ) :
(A?i , M−1 )7→M0 ,
(A?i , M0 ) 7→M+1 ,
(A?i , M0 ) 7→M−1 ,
with probability r.
(9) :
(10) :
with probability r.
with probability r.
Figure 4: Protocol extension Pm of protocol Po , with additional states {M−1 , M0 , M+1 }. This protocol
extension is applied through the composition (Po ◦ Pm ). Probability r > 0 is given by an absolute
constant, depending explicitly on s and p, whose value is sufficiently small.
++
A++
1 if source X[1] is present and stopping with all agents in state A2 if source X[2] is present, with no
subsequent state transition. The stabilization occurs within O(log2 n) parallel rounds, with probability
1 − O(1/n), if max{#X[1] , #X[2] } = O(1), i.e., the broadcast originates from a constant number of
sources.
3.3
Protocol for D ETECT
The solution to problem D ETECT is more involved. It relies on two auxiliary extensions added on top of
a single oscillator Po . The first, Pm , runs an instance of the 3-state majority protocol of Angluin et al. [7]
within each species Ai of the oscillator. For this reason, the composition between Po and Pm has to be of
the form Po ◦ Pm (i.e., it cannot be independent). The operation of this extension is shown in Fig. 5 and
analyzed in Section 7.2. It relies crucially on an interplay of two parameters: the time Θ(log #nX ) taken
by the oscillator to perform an orbit, and the time Ω(log #nX ) it takes for the majority protocol (which is
reset by the oscillator in its every oscillation) to converge to a solution. When parameters are tuned so
that the second time length is larger a constant number of times than the first, a constant proportion of
the agents of the population are involved in the majority computation, i.e., both of the clashing states in
the fight for dominance still include Ω(n) agents. In the absence of a source, shortly after the oscillator
stops, one of these states takes over, and the other disappears.
The above-described difference can be detected by the second, much simpler, extension Pl , designed
in Fig. 5 and analyzed in Section 7.3. The number of “lights” switched on during the operation of the
protocol will almost always be more than (1 − ε)n, where ε > 0 is a parameter controlled by the
probability of lights spontaneously disengaging, and may be set to and arbitrarily small constant.
Theorem 3 (Protocol for D ETECT). For any ε > 0, protocol (Po ◦ Pm ) + Pl , having |K| = 6 · 3 · 3 + 1 =
55 states, including a distinguished source state X, which solves the problem of spreading confirmed
rumors as follows:
1. For any starting configuration, in the presence of the source (#X ≥ 1), after an initialization
period of O(log n) rounds, at an arbitrary time step the number of agents in an output state
corresponding to a “yes” answer is (1 − ε)n, with probability 1 − O(1/n).
13
(11) Light switch progresses from L−1 to L+1 in the presence of initiator M−1 :
(A?? , M−1 , L? ) :
(A?? , M? , L−1 ) 7→ L+1
(12) Light switch progresses from L+1 to Lon in the presence of initiator M+1 :
(A?? , M+1 , L? ) :
(A?? , M? , L+1 ) 7→ Lon
(13) Light spontaneously turns off:
(A?? , M? , L? ) :
(A?? , M? , Lon ) 7→ L−1 ,
with probability q(ε).
Figure 5: Protocol extension Pl of protocol (Po ◦ Pm ), with additional states {L−1 , L+1 , Lon }. This
protocol extension is applied through the composition (Po ◦ Pm ) + Pl . Probability q(ε) > 0 is given by
an absolute constant, depending explicitly on s, p, r, and ε, whose value is sufficiently small.
2. For any starting configuration, in the absence of the source (#X = 0), the system always reaches
a configuration such that all agents are in output states corresponding to a “no” answer for all
subsequent time steps. Such a configuration is reached in O(log2 n) rounds, with probability
1 − O(1/n).
4
Impossibility Results for Protocols without Non-Stationary Effects
For convenience of notation, we identify a configuration of the population with a vector z = (z (1) , . . . , z (k) ) ∈
{0, 1, . . . , n}k = Z, where z (i) , for 1 ≤ i ≤ k, denotes the number of agents in the population having
state i, and kzk1 = n. Our main lower bound may now be stated as follows.
Theorem 4 (Fixed points preclude fast stabilization). Let ε1 > 0 be arbitrarily chosen, let P be any
k-state protocol, and let z0 be a configuration of the system with at most nε0 agents in state X, where
ε0 ∈ (0, ε1 ] is a constant depending only on k and ε1 . Let B be a subset of the state space around z0
such that the population of each state within B is within a factor of at most nε0 from that in z0 (for any
(i)
(i)
z ∈ B, for all states i ∈ {1, . . . , k}, we have z0 /nε0 < z (i) ≤ nε0 max{1, z0 }).
Suppose that in an execution of P starting from configuration z0 , with probability 1 − o(1), the
configurations of the system in the next n2ε1 parallel rounds are confined to B.
Then, an execution of P for n2ε0 parallel rounds, starting from a configuration in which state X
has been removed from z0 , reaches a configuration in a O(n6ε1 )-neighborhood of B, with probability
1 − o(1).
In the statement of the Theorem, for the sake of maintaining the size of the population, we interpret
“removing state X from z0 ” as replacing the state of all agents in state X by some other state, chosen
adversarially (in fact, this may be any state which has sufficiently many representatives in configuration
z0 ). The O(n6ε1 )-neighborhood of B is understood in the sense of the 1-norm or, asymptotically equivalently, the total variation distance, reflecting configurations which can be converted into a configuration
from B by flipping the states of O(n6ε1 ) agents.
The proof of Theorem 4 is provided in Section 8. It proceeds by a coupling argument between a
process starting from z0 and a perturbed process in which state X has been removed. The analysis
differently treats rules and states which are seldom encountered during the execution of the protocol
14
from those that are encountered with polynomially higher probability (such a clear separation is only
possible when k = O(poly log log n)). Eventually, the probability of success of the coupling reduces to
a two-dimensional biased random walk scenario, in which the coordinates represent differences between
the number of times particular rules have been executed in the two coupled processes.
We have the following direct corollaries for the problems we are considering. For D ETECT, if B
represents the set of configurations of the considered protocol, which are understood as the protocol
giving the answer “#X > 0”, then our theorem says that, with probability 1 − o(1), the vast majority
of agents will not “notice” that #X had been set to 0, even a polynomial number of steps after this has
occurred, and thus cannot yield a correct solution. An essential element of the analysis is that it works
only when state X is removed in the perturbed process. Thus, there is nothing to prevent the dynamics
from stabilizing even to a single point in the case of X = 0, which is indeed the case for our protocol
Pr . The argument for B IT B ROADCAST only applies to situations where the source agent is sending out
white noise (independently random bits in successive interactions). Such a source can be interpreted
as a pair of sources in states X1 and X2 in the population, each disclosing itself with probability 1/2
upon activation and staying silent otherwise. In the cases covered by the lower bound, the scenario in
which the source X1 is completely suppressed cannot be distinguished from the scenario in which both
X1 and X2 appear; likewise, the scenario in which the source X2 is completely suppressed cannot be
distinguished from the scenario in which both X1 and X2 appear. By coupling all three processes, this
would imply the indistinguishability of the all these configurations, including those with only source X1
and only source X2 , which would imply incorrect operation of the protocol.
Whereas we use the language of discrete dynamics for precise statements, we informally remark that
the protocols covered by the lower bound of Theorem 4 include those whose dynamics zt /n, described
in the continuous limit (n → +∞), has only point attractors, repellers, and fixed points. In this sense,
the use of oscillatory dynamics in our protocol seems inevitable.
The impossibility result is stated in reference to protocols with a constant number of states, however,
it may be extended to protocols with a non-constant number of states k, showing that such protocols
require nexp[−O(poly(k))] time to reach a desirable output. (This time is larger than polylogarithmic up to
some threshold value k = O(poly log log n).) The lower bound covers randomized protocols, including
those in which rule probabilities depend on n (i.e., non-universal ones).
5
Input-Controlled Behavior of Protocols for D ETECT
In this Section, we consider the periodicity of protocols for self-organizing oscillatory dynamics, in
order to understand how the period of a phase clock must depend on the input parameters. We focus on
the setting of the D ETECT problem, considering the value #X of the input parameter. In Section 3.1, we
noted informally that the designed oscillatory protocol performs a complete rotation around the triangle
in Θ(log n/#X) rounds. Here, we provide partial evidence that the periodicity of any oscillatory protocol
depends both on the value of #X and n. We do this by bounding the portions of the configuration space in
which a protocol solving D ETECT finds itself in most time steps, separating the cases of sub-polynomial
#X (i.e., #X < nε0 , where ε > 0 is a constant dependent on the specific protocol), and the case of
0
#X = Θ(n).
Any protocol on k states (not necessarily of oscillatory nature) can be viewed as a Markov chain
in its k-dimensional configuration space [0, n]k , and as in Section 4 we identify a configuration with a
vector z ∈ {0, 1, . . . , n}k = Z. The configuration at time step t is denoted z(t). In what follows, we
will look at the equivalent space of log-configurations, given by the bijection:
Z 3 z = (z (1) , . . . , z (k) ) 7→ (ln◦ z (1) , . . . , ln◦ z (k) ≡ ln◦ z ∈ {ln◦ 0, ln◦ 1, . . . , ln◦ n}k ),
where ln◦ a = ln a for a > 0 and ln◦ a = −1 for a = 0.
15
For z0 ∈ Z, we will refer to the d-log-neighborhood of z0 as the set of points {z ∈ Z : | ln◦ z −
ln◦ z0 | < d}.
Notice first that the notion of a box in the statement of Theorem 4 is closely related to the set of points
in the (ε0 ln n)-log-neighborhood of configuration z0 . It follows from the Theorem that any protocol for
solving D ETECT within a polylogarithmic number of rounds T with probability 1 − o(1) must, in the
case of 0 < #X < nε0 , starting from z0 at some time t0 , leave the (ε0 ln n)-log-neighborhood of z0
within T rounds with probability 1 − o(1). We obtain the following corollary.
Proposition 1. Fix a universal protocol P which solves the D ETECT problem with ε-error in T =
O(poly log n) rounds with probability 1 − o(1). Set 0 < #X < nε0 , where ε0 > 0 is a constant which
depends only on the definition of protocol #X. Let t0 be an arbitrarily chosen moment of time after at
least T rounds from the initialization of the protocol in any initial state. Then, within T rounds after
time t0 , there is a moment of time t such that z(t) is not in the (ε0 log n)-neighborhood of z(t0 ), with
probability 1 − o(1).
The above Proposition suggests that oscillatory or quasi-oscillatory behavior at low concentrations
of state X must be of length Ω(log n). By contrast, the following Proposition shows that in the case
#X = Θ(n), the protocol remains tied to a constant-size log-neighborhood of its configuration space.
Proposition 2. Fix a universal protocol P with set of states K which solves the D ETECT problem with
ε-error in T = O(poly log n) rounds with probability 1 − o(1). Then, there exists a constant δ0 > 0,
depending only on the design of protocol P , with the following property. Fix #X ∈ [cn, n/2], where 0 <
c < 1/2 is an arbitrarily chosen constant. Let t be an arbitrarily chosen moment of time, after at least T
rounds from the initialization of the protocol at an adversarially chosen initial configuration z(0), such
that each coordinate z (i) (0) satisfies z (i) (0) = 0 or z (i) (0) > 1/(2|K|), for all i ∈ {1, . . . , |K|}. Then,
Ω(1)
with probability 1 − e−n , z(t) is in the δ0 -neighborhood of z(0).
The proof of the Proposition is deferred to Section 9.
Note that, in the regime of a constant-size log-neighborhood of configuration z(0), the discrete
dynamics of the protocol adheres closely to the continuous-time version of its dynamics in the limit
n → +∞. (See Section 6, and in particular Lemma 1, for a further discussion of this property). Since
the latter is independent of n, any oscillatory behavior “inherited” from the continuous dynamic would
have a period of O(1) rounds. We leave as open the question whether some form of behavior of a
protocol with polylogarithmic (i.e., or more broadly, non-constant and subpolynomial) periodicity for
D ETECT can be designed in the regime of #X = Θ(n) despite this obstacle. In particular, the authors
believe that the existence of an input-controlled phase clock with a period of Θ(log n) for any #X > 0,
and the absence of operation for #X = 0, is unlikely in the class of discrete dynamical systems given by
the rules of population protocols.
The remaining sections of the paper provide proofs of the Theorems from Sections 3, 4, and 5.
6
Analysis of Oscillator Dynamics Po
This section is devoted to the proof of Theorem 1.
6.1
Preliminaries: Discrete vs. Continuous Dynamics
Notation. For a configuration of a population protocol, we write z = (z (1) , . . . , z (k) ) to describe the
number of agents in the k states of the protocol, and likewise use vector u = (u(1) , . . . , u(k) ) = z/n to
describe their concentrations. The concentration of a state called A which is the iA -th state in vector u
is equivalently written as a ≡ a(u) ≡ u(iA ) , depending on which notation is the easiest to use in a given
transformation.
16
If vector u represents the current configuration of the protocol and u0 := u0 |u is the random variable
describing the next configuration of the protocol after the execution of a single rule, we write ∆u :=
u0 − u. We also use the notation ∆ to functions of state u.
Next, we define the continuous dynamics associated with the protocol by the following vector differential equation:
du
u̇ ≡
:= nE(∆u)
dt
and likewise, for each coordinate, ȧ = nE(∆a) (we use the dot-notation and d/dt interchangeably for
time differentials). This continuous description serves for the analysis only, and reflects the behavior of
the protocol in the limit n → ∞.4
Warmup: the RPS oscillator. Our oscillatory dynamics may be seen as an extension of the rockpaper-scissors (RPS) protocol (see Related work). This is a protocol with three states A1 , A2 , A3 and
three rules:
Ai : Ai−1 7→ Ai with probability p,
where p > 0 is an arbitrarily fixed constant, and the indices of states Ai are always 1, 2, or 3, and any
other values should be treated as mod 3 in the given range. For i ∈ {1, 2, 3}, the change of concentration
of agents of state Ai in the population in the given step can be expressed for the RPS protocol as:
+1/n, with probability pai−1 ai ,
1
#
∆ai = · ∆ Ai = −1/n, with probability pai ai+1 ,
(1)
n
0,
otherwise,
Thus, the corresponding continuous dynamics for RPS is given as:
ȧi = nE(∆ai ) = pai−1 ai − pai ai+1 ,
for i = 1, 2, 3. The orbit of motion for this dynamics in R3 is given by two constants of motion. First,
a1 + a2 + a3 = 1 by normalization. Secondly, for any starting configuration with a strictly positive
number of agents in each of the three states, the following function φ of the configuration:
φ = ln(a1 a2 a3 )
(2)
is easily verified to be constant over time φ̇ = 0, hence φ = ln(a1 a2 a3 ) = const < 0 (or more simply,
a1 a2 a3 = const). Thus, for the continuous dynamics, the initial product of concentrations completely
determines its perpetual orbit, which is obtained by intersecting the appropriate curve a1 a2 a3 = const
with the plane a1 + a2 + a3 = 1. As a matter of convention, the plane a1 + a2 + a3 = 1 with conditions
ai ≥ 0 is drawn as an equilateral triangle (we adopt this convention throughout the paper, for subsequent
protocols). All of the orbits are concentric around the point (1/3, 1/3, 1/3), which is in itself a point
orbit maximizing the value of φ = − ln 27. The discrete dynamics follows a path of motion which
typically resembles random-walk-type perturbations around the path of motion, until eventually, after
Õ(n2 ) steps it crashes into one of the sides of the triangle. Subsequently, if ai = 0, for some i = 1, 2, 3,
then no rule can make ai increase. (If ai−1 > 0, in the next O(log n) steps, all remaining agents of Ai+1
will convert to Ai−1 , and there will be only agents from Ai−1 left.) Thus, the protocol will terminate in
a corner of the state space.
A further discussion of the RPS dynamics can be found in [15, 28].
4
We note that some of our results rely on the stochasticity of the random scheduler model, and do not immediately generalize
to the continuous case.
17
6.2
Proof Outline of Theorem 1
The rest of the section is devoted to the proof of Theorem 1. We start by noting some basic properties
in Subsection 6.3, then prove the properties of the protocol for the case of X = 0 (Subsection 6.4,
and finally analyze (the somewhat less involved) case of X > 0 (Subsection 6.5). For the case of
X = 0, the proof is based on a repeated application of concentration inequalities for several potential
functions (applicable in different portions of the 6-dimensional phase space). In two specific regions, in
√
the O(1/ n)-neighborhood of the center of the (A1 , A2 , A3 )-triangle and very close to its sides, we rely
on stochastic noise to “push” the trajectory away from the center of the triangle, and also to push it onto
one of its sides. Fortunately, each of these stages takes O(log n) parallel rounds, with strictly positive
probability. Overall, the O(log n) parallel rounds bound for the case of X = 0 is provided with constant
probability; this translates into O(log n) parallel rounds in expectation, since subsequent executions of
the process for O(log n) rounds have independently constant success probability, and the process has a
geometrically decreasing tail over intervals of length O(log n).
6.3
Properties of the Oscillator
In the following, we define s = a1 + a2 + a3 ∈ [0, 1]. Handling the case of s < 1 allows us not only
to take care of the fact that a fraction of the population may be taken up by rumor source X, but also
allows for easier composition of Po with other protocols (sharing the same population). We set p as a
constant value independent of n, which is sufficiently small (e.g., p = s2 /1012 is a valid choice; we
make no efforts in the proofs to optimize constants, but the protocol appears in simulations to work well
with much larger values of p).
We will occasionally omit an explanation of index i, which will then implicitly mean “for all i =
1, 2, 3”. We define amin := mini=1,2,3 ai and amax := maxi=1,2,3 ai .
From the definition of the protocol one obtains the distribution of changes of the sizes of states in a
step:
+
++
1
+1/n, with probability 3 x(s − ai ) + pai ai−1 + 2pai ai−1 ,
++
∆ai = −1/n, with probability 23 xai + pai a+
i+1 + 2pai+1 ai ,
0,
otherwise.
++
+1/n, with probability ai (ai − ai ),
∆a++
= −1/n, with probability xa++
+ (s − ai )a++
i
i
i ,
0,
otherwise.
++
Taking the expectations of the above random variables, and recalling that ai = a+
i +ai , we obtain:
++
+
++
ȧi = x(s/3 − ai ) + pa+
i ai−1 + 2pai ai−1 − pai ai+1 − 2pai+1 ai
++
= x(s/3 − ai ) + pai−1 (ai + a++
i ) − pai (ai+1 + ai+1 )
ȧ++
i
=
−xa++
i
+ ai (ai −
a++
i )
− (s −
ai )a++
i
=
−xa++
i
+
(3)
a2i
−
sa++
i .
Moreover, we have by a simple transformation:
X
X ȧi
x X s
++ ai−1
φ̇ =
=
−9 +p
ai
−1
ai
3
ai
ai
6.4
(4)
(5)
Stopping in O(n log n) Sequential Steps in the Absence of a Source
Throughout this subsection we assume that x = 0. We consider first the case where ai 6= 0, for
i = 1, 2, 3 (noting that as soon as ai = 0, we can easily predict the subsequent behavior of the oscillator,
as was the case for the RPS dynamics).
18
The dynamics of Po is defined in such a way that that when x = 0 and in the absence of the rules of
a2
the RPS oscillator, the value of a++
would be close to si . Consequently, we define κi , i = 1, 2, 3 as the
i
appropriate normalized corrective factor:
κi = s
a++
i
− ai
ai
thus
a++
=
i
ai
(ai + κi ).
s
Note that as 0 ≤ a++
≤ ai ≤ 1, thus −1 ≤ κi ≤ 1. Next, we introduce the following definitions:
i
δi = ai − ai−1
q
δ = δ12 + δ22 + δ32
q
κ = κ21 + κ22 + κ23
We also reuse potential φ from the original RPS oscillator. This time, it is no longer a constant of motion.
By (5) and the definition of κi , for x = 0 we upper-bound φ̇ as:
p X
ai−1
φ̇ =
− 1)
ai (ai + κi )(
s
ai
p X
(ai + κi )(ai−1 − ai )
=
s
X
p
1X
2
=
−
(ai − ai−1 ) +
κi (ai−1 − ai )
s
2
1 2
p
− δ + κδ
(6)
≤
s
2
The above change φ̇ of the potential is indeed negative when κ ≈ 0 (which is in accordance with our
intention in designing the destabilizing rules for the oscillator).
The functions δ, φ and κ are intricately dependent on each other. In general, we will try to show that
δ and φ increase over time, while κ stays close to 0. This requires that we first introduce a number of
auxiliary potentials based on these two functions.
First, for x = 0, we can rewrite (4) as:
˙ = a2 − sa++ = −a κ .
a++
i i
i
i
i
Next, introducing the definition of κi to (3), we obtain for x = 0:
++
++
++
ȧi = pai−1 (ai + a++
i ) − pai (ai+1 + ai+1 ) = pai (ai−1 − ai+1 ) + p(ai ai−1 − ai+1 ai ) =
1
1
= pai δi−1 + p
ai (ai + κi )ai−1 − ai+1 (ai+1 + κi+1 )ai =
s
s
p
= pai δi−1 + ai κi ai−1 − κi+1 ai+1 + (ai−1 ai − a2i+1 ) .
s
19
From the above, an upper bound on |ȧi | follows directly using elementary transformations:
p
|ȧi | ≤ pai |δi−1 | + ai |κi |ai−1 + |κi+1 |ai+1 + |ai−1 ai − a2i+1 | ≤
s
p
≤ pai δ + ai κ(ai−1 + ai+1 ) + |ai−1 ai − a2i+1 |) =
s
ai−1 + ai+1 p
= pai δ + pai κ
+ ai |(ai−1 − ai+1 )ai − (ai+1 − ai )ai+1 | ≤
s
s
p
≤ pai (δ + κ) + ai (|ai−1 − ai+1 |ai + |ai+1 − ai |ai+1 ) ≤
s
ai + ai+1
≤ pai (δ + κ) + pai δ
s
≤ pai (2δ + κ).
We are now ready to estimate κ̇i for x = 0, using the definition of κi and the previously obtained
formula for ȧ++
i :
++
ai
a++
ȧ++
i
i
= −sκi − s 2 + 1 ȧi
− 2 ȧi − ȧi
κ̇i = s
ai
ai
ai
++
a
ȧi
ȧi
= −sκi − s i − ai
− 2ȧi = −sκi − κi − 2ȧi .
ai
ai
ai
Next from the bound on |ȧi |:
ȧi
κ˙2i = 2κi κ̇i = 2 − sκ2i − κ2i − 2ȧi κi
ai
|
ȧ
|
i 2
κi + 2|ȧi ||κi |
≤ 2 − sκ2i +
ai
2
≤ 2 − sκi + p(2δ + κ)(κ2i + 2ai |κi |)
≤ 2 − sκ2i + p(2δ + κ)(κ + 2κ)
= −2sκ2i + p(12δκ + 6κ2 ).
Next:
X
1 X ˙2
1
κi ≤
−2s
κ2i + 3p(12δκ + 6κ2 ) ≤
2κ
2κ
1
2
≤
−2sκ + p(36δκ + 18κ2 ) =
2κ
= (−s + 9p)κ + 18pδ ≤
s
≤ − κ + 18pδ,
2
κ̇ =
where in the final transformation we took into account that p ≤ s/18.
Now, we define the potential η for any configuration with all ai > 0 as:
η=
s3
ln
−φ
27
1/2
=
−
3
X
i=1
ai
ln
s/3
!1/2
.
We remark that η is always well-defined when amin > 0, and that η ≥ 0.
20
(7)
A3
t3
t1
t0
δ~O(n -1/2)
t4
A1
t2
δ~s/12
t5
t7
t6
A2
Figure 6: Schematic illustration of order of phases in the proof of stabilization of protocol Po for X = 0.
Overview of the proof. The proof for the case of X = 0 proceeds by following the trajectory of the
discrete dynamics of Po , divided into a number of stages. We define a series of time steps t0 , t1 , . . . , t7
by conditions on the configuration met at time ti , and show that subject to these conditions holding,
we have tj+1 ≤ tj + O(n log n) (we recall that here time is measured in steps), with at least constant
probability. Overall, it follows that the configuration at time t7 , which corresponds to having reached a
corner state, is reached from t0 , which is any initial configuration with X = 0, in O(n log n) time steps,
with constant probability.
The intermediate time steps may be schematically described as follows (see Fig. 6). For configurations which start close to the center of the triangle (δ ≤ s/12), we define a pair of potentials ψ (1) , ψ (2) ,
based on a linear combination of modified versions of η and κ. The dynamics will eventually escape from
√
the area δ ≤ s/12; however, first it may potentially reach a very small area of radius O(1/ n) around
the center of the triangle with κ ≈ 0 (Lemma 3, time t1 , reached in O(n log n) steps by a multiplicative
drift analysis on potential ψ (2) < 0), pass through the vicinity of center of the triangle, escaping it with
κ ≈ 0 (Lemma 4, time t2 , reached in O(n log n) steps with constant probability by a protocol-specific
analysis of the scheduler noise, which with constant probability increases η without increasing κ too
much), and eventually escapes completely to the area of δ > s/12 (Lemma 5, exponentially increasing
value of potential ψ (1) > 0).
In the area of δ > s/12, we define a new potential ψ based on φ and κ. This increases (Lemma 8,
additive drift analysis on ψ with bounded variance) until a configuration at time t4 with a constant
number of agents of some species Ai is reached. This configuration then evolves towards a configuration
at time t5 at which some species has O(1) agents, and additionally its predator species is a constant part
of the population (Lemma 9, direct analysis of the process combined with analysis of potential ψ and a
geometric drift argument). Then, the species with O(1) agents is eliminated in O(n) steps with constant
probability (t6 , Lemma 10), and finally one more species is eliminated in another O(n log n) steps (at
time t7 , Lemma 11, straightforward analysis of the dynamics). At this point, the dynamics has reached
a corner.
Throughout the proof, we make sure to define boundary conditions on the analyzed cases to make
sure that the process does not fall back to a previously considered case with probability 1 − o(1).
ai
Phase with δ ≤ s/12. We then have ai ∈ [3s/12, 5s/12] and s/3
∈ [3/4, 5/4], for i = 1, 2, 3. In this
range, we have:
2
2
1 ai
ai
ai
3 ai
−1 <
− 1 − ln
<
−1 .
3 s/3
s/3
s/3
4 s/3
21
Summing the above inequalities for i = 1, 2, 3 and noting that
3
1X
3
i=1
P3
ai
i=1 ( s/3
− 1) = 0, we obtain:
2
2
3
ai
3 X ai
2
−1 <η <
−1 .
s/3
4
s/3
i=1
Next, we have:
2 2 X
3
3
X
ai
3
3δ 2
−1 =
(ai − s/3)2 = 2 .
s/3
s
s
i=1
i=1
Combining the two above expressions gives the sought bound between η and δ as:
δ
3δ
<η<
s
2s
and equivalently
δ ∈ ( 23 sη, sη).
We have directly from (6) and from the relations between η and δ:
p 1 2
p 2
p
ps
p
φ̇
≥
δ − κδ =
δ −
κδ ≥ η − κ,
η̇ = −
2η
2sη 2
4sη
2sη
9
2
(8)
and from (7):
s
s
κ̇ ≤ − κ + 18pδ ≤ − κ + 18ps η.
(9)
2
2
Moving to the discrete-time model, it is advantageous to eliminate the discontinuity of partial derivatives
of η and κ at points with η = 0 and κ = 0 respectively, which is a side-effect of the applied square root
transformation in the respective definitions of η and κ. We define the auxiliary functions η ∗ and κ∗ by
adding an appropriate corrective factor:
r
1
η∗ = η2 +
n
r
1
κ∗ = κ2 +
n
and derive accordingly from (8) and (9):
1
p
η
ps
ps
p
2ps
η−√
− κ ≥ η ∗ − κ∗ − √
η̇ ∗ = ∗ η̇ ≥
η
9
2
9
2
n
9 n
κ
s
1
s
s
∗
κ̇ = ∗ κ̇ ≤ −
κ− √
+ 18ps η ≤ − κ∗ + 18ps η ∗ + √ .
κ
2
2
n
n
(10)
(11)
Let u be the 5-dimensional vector representing the current configuration of the system: u :=
the last element a++
is determined as a++
=
3
3
folklore application of Taylor’s theorem.
+ ++ +
(1)
(5)
(a+
a++
1 ,P
1 , a2 , a2 , a3 ) ≡ (u , . . . , u ); note that
5
(i)
5
s − i=1 u . The following lemma is obtained by a
Lemma 1. Let f : R5 → R be a C 2 function in a sufficiently large neighborhood of u, with min1≤i≤5 u(i) ≥
˙
2/n. Then, |E∆f (u) − nf | ≤
2
n2
maxku∗ −uk∞ ≤1/n Df (u∗ ), where Df (u∗ ) := max1≤i,j≤5
∂f 2 (u∗ )
∂u(i) ∂u(j)
.
In principle it is also correct to represent u as a vector of dimension 6, i.e., including a++
in u as a free dimension.
3
2
However, such a representation would lead to second-order partial derivatives ∂u(i)∂∂u(j) η ∗ (u) which are too large for our
purposes.
5
22
Proof. Let u0 be the random variable representing the configuration of the system after its next transition
from configuration u. Observe that in every non-idle step of execution of the protocol, exactly one agent
changes its state, so ku0 − uk∞ ≤ 1/n.
Applying Taylor approximation we have:
X
X
E(∆f ) = E(f (u0 )|u) − f (u) =
(f (u0 ) − f (u)) Pr(u0 |u) =
∇f (u) · (u0 − u) + R2 (u, u0 ) Pr(u0 |u) =
u0
= ∇f (u) ·
u0
X
f˙
1
(u0 − u) Pr(u0 |u) + R2 (u) = ∇f (u) · (u̇(1) , . . . , u̇(5) )T + R2 (u) = + R2 (u),
n
n
0
u
(12)
where ∇f (u) is the gradient of f at u, R2 (u, u0 ) ∈ R denotes the second-order Taylor remainder for
function f expanded at point u along the vector towards point u0 , and R2 (u) ∈ R is subsequently an
appropriately chosen value, satisfying:
|R2 (u)| ≤
1
n2
max
ku∗ −uk∞ ≤1/n
Df (u∗ ).
The following lemma is obtained directly by computing and bounding all second order partial derivatives of functions η ∗ and κ∗ with respect to variables (u(1) , . . . , u(5) ).
Lemma 2. There exists a constant c1 > 1 depending only on s, p, such that, for any configuration u
with δ(u) ≤ s/12:
√
• maxku∗ −uk∞ ≤1/n Dη∗ (u∗ ) < c1 n,
√
• maxku∗ −uk∞ ≤1/n Dκ∗ (u∗ ) < c1 n.
In view of the above lemmas, we obtain from (8) and (9), for an appropriately chosen constant
c2 = 2c1 + s:
E∆η ∗ ≥ 1 ps η ∗ − p κ∗ − √c2 ,
when δ ≤ s/12,
n 18
2
n
E∆κ∗ ≤ 1 − s κ∗ + 18ps η ∗ + √c2 , when δ ≤ s/12.
n
3
n
For j = 1, 2, we now define two linear combinations of functions η ∗ and κ∗ :
ψ (j) = η ∗ −
3jp ∗
κ .
s
When δ ≤ s/12, we have:
1 ps ∗ p ∗
c2
3jp c2
1 ps ∗ jp ∗ 2c2
(j)
∗
2 ∗
√
E∆ψ ≥
η − κ − √ + jpκ − 54jp η −
≥
η + κ −√
≥
n 18
2
s
n 24
2
n
n
n
ps
3jp ∗
48c2
ps
c3
≥
η∗ +
κ − √
≥
|ψ (j) | − √
,
24n
s
24n
ps n
n
48c2
ps
s
72·54·2 .
(j)
We subsequently perform an analysis of ψt = ψ (j) (ut ), j = 1, 2,
(2)
(1)
(1)
(2)
∗
processes. We remark that ψt ≤ ψt , since ψt − ψt = 3p
s κ ≥ 0.
where we denoted c3 :=
and used the fact that p <
treating them as stochastic
Lemma 3. Let ut0 be an arbitrary starting configuration of the system. Then, with constant probability,
(1)
(2)
2c3
for some t1 = t0 + O(n log n), a configuration ut1 is reached such that ψt1 ≥ ψt1 ≥ − √
.
n
23
(2)
Proof. W.l.o.g. assume t0 = 0. We subsequently only analyze process ψt . Let t1 be the first time step
(2)
(2)
(2)
2c3
. If t1 6= 0, then ψ0 < 0. Note that then ψt < 0 for all t ≤ t1 , from which it
such that ψt1 > − √
n
s
follows by a straightforward calculation from the definition of ψ, κ, and η, that δt < 12
for all t ≤ t1 .
∗(2)
We now define the filtered stochastic process ψt
for t ≥ t1 . For all t ≥ 0, we then have:
∗(2)
E(∆ψt
∗(2)
Since 0 ≤ ψt
gives:
∗(2)
|ψt
∗(2)
as ψt
6= 0) ≤
∗(2)
(2)
:= |ψt | for t < t1 , and put ∆ψt
:= 0
ps ∗(2)
ψ .
48n t
< 9 for all time steps, a direct application of multiplicative drift analysis (cf. [19])
Et1 ≤
48n
ps
√
9 n
1 + ln
,
2c3
and the claim follows by Markov’s inequality.
(j)
2c3 √
Lemma 4. Let ut1 be an arbitrary starting configuration of the system such that ψt1 ∈ [− √
, 4cn3 ], for
n
j = 1, 2. Then, with constant probability, for some t2 = t1 + O(n), a configuration ut2 is reached such
(1)
4c3
that ψt2 ≥ √
.
n
(2)
Proof. W.l.o.g. assume that t1 = 0 and suppose that initially ψ0
Then, from the lower and upper bounds on
(1)
ψ0
2c3
3p ∗
(1)
(2)
κ0 = ψ0 − ψ0 ≤ √ +
s
n
4c
(1)
(2)
3
η0∗ = 2ψ0 − ψ0 ≤ 2 √ +
n
and
(2)
ψ0
(1)
≤ ψ0
<
4c3
√
n
(i.e., that t2 6= t1 ).
we obtain the following bounds on κ0 and δ0 :
4c
√ 3 =⇒ κ0 ≤
n
2c3
√ =⇒ η0 ≤
n
2c3 s
√ ,
p n
10c3
10c s
√ =⇒ δ0 < √ 3 .
n
n
√ 3 s , s + 10c
√ 3 s ] and a++ =
It follows that, for i = 1, 2, 3, ai,0 ∈ [ 3s − 10c
i,0
n 3
n
ai,0
s (ai,0 +κi,0 )
√ 3 )( s −
∈ [( 31 − 10c
3
n
√ 3 )( s + 10c
√ 3 s + 2c√3 s )]. For the sake of clarity of notation, we will simply write
+ 10c
3
n
n
p n
√
√
√
ai,0 = s/3 ± O(1/ n) and a++
=
s/9
± O(1/ n), hence also a+
i,0
i,0 = 2s/9 ± O(1/ n).
We will consider now the sequence of exactly n transitions of the protocol, between time steps
t = 0, 1, . . . , n.
Pt−1
(2)
c3 ps
For all t we have E∆ψt ≥ − 24n
3/2 . Consider the Doob submartingale Yt =
τ =0 Xt with
increments (Xt ) given as:
(
(2)
c3 ps
∆ψt + 24n
if Yt > − √c3n
3/2 ,
Xt =
0,
otherwise,
10c
√ 3s
n
−
2c√3 s
), ( 13
p n
Noting that |Xt | ≤ n9 , an application of the Azuma inequality for submartingales to (Yn ) gives: Pr[Yn ≤
− √c3n ] ≤ exp [−c23 /162] (cf. e.g. [13][Thm. 16]). From here it follows directly that:
Pr
ψn(2)
c3
c3 ps
(2)
> − √ + ψ0 − n
≥ 1 − exp [−c23 /162] > 1/2.
n
24n3/2
(2)
2c3
Noting that ψ0 ≥ − √
, we have:
n
Pr
ψn(2)
2c3
≥ −√
> 1/2.
n
(13)
We now describe the execution of transitions in the protocol for times t = 0, 1, . . . , n − 1 through the
following coupling. First, we select the sequence of pairs of agents chosen by the scheduler. Let V2+
24
+
(respectively, V1+ ) denote the subsets of the set of n agents, having initial states A+
2 (resp., A1 ) at time
0, respectively, which are involved in exactly one transition in the considered time interval, acting in it
as the initiator (resp., receiver). Let S ⊆ {0, 1, . . . , n − 1} denotes the subset of time steps at which the
scheduler activates a transition involving an element of V2+ as the initiator and an element of V1+ as the
receiver. The execution of the protocol is now given by:
• Phase PA : Selecting the sequence of pairs of elements activated by the scheduler in time steps
(0, 1, . . . , n − 1). This also defines set S. Executing the rules of the protocol in their usual order
for time steps from set {0, 1, . . . , n − 1} \ S.
• Phase PB : Executing the rules of the protocol for time steps from set S.
Observe that since elements of pairs activated in time steps from S are activated only once throughout
the n steps of the protocol, the above probabilistic coupling does not affect the distribution of outcomes.
Directly from (13), we obtain through a standard bound on conditional probabilities that at least a
(2)
2c3
constant fraction of choices made in phase PA leads to an outcome “ψn ≥ − √
” with at least constant
n
probability during phase PB :
2c3
(2)
(14)
Pr PA : Pr ψn ≥ − √ PA > 1/4 ≥ 1/3.
n
+
We now remark on the size of set S. The distribution of |S| depends only on a+
2,0 , a1,0 , and the choices
√
made by the random scheduler. We recall that a+
2,0 = 2s/9 ± O(1/ n). Since the expected number of
isolated edges in a random multigraph on n nodes (representing the set of agents) and n edges (representing the set of time steps) is (1 ± o(1))e−4 n, the number of such edges having the first endpoint in an
+
4e−4 s2
agent in state A+
2 and the second endpoint in an agent in state A1 is (1 ± o(1)) 81 n. A straightforward concentration analysis (using, e.g., the asymptotic correspondence between G(n, m) and G(n, p)
random graph models and an application of Azuma’s inequality for functions of independent random
−4 2
variables) shows that the bound |S| = (1 ± o(1)) 4e 81 s n holds with very high probability. In particular,
we have:
Pr[|S| > c4 n] = 1 − eΩ(−n) ,
(15)
for some choice of constant c4 which depends only on s.
Relations (14) and (15) provide all the necessary information about phase PA that we need. Subsequently, we will only analyse phase PB , conditioning on a fixed execution of phase PA such that the
following event FA holds:
2c3
(2)
Pr ψn ≥ − √ PA > 1/4 ∧ |S| > c4 n.
(16)
n
We remark that, by a union bound over (14) and (15), Pr[FA ] ≥ 1/3 − eΩ(−n) > 1/4.
In the remainder of our proof, our objective will be to show that:
4c3
Pr ψn(1) ≥ √ PA > c5 ,
n
for some constant c5 > 0 depending only on s, p, for any choice of PA for which event FA holds. When
this is shown, the claim of the lemma will follow directly, with a probability value given as at least
c5 Pr[FA ] > c5 /4 by the law of total probability.
We now proceed to analyze the random choices made during phase PB . Each of the considered |S|
+
interactions involves a pair of agents of the form (A+
2 , A1 ), and describes the following transition:
(
+
(A+
+
+
2 , A2 ), with probability p,
(A2 , A1 ) →
+
(A+
2 , A2 ), with probability 1 − p,
25
independently at random for each transition. The only state changes observed during this phase are from
+
A+
1 to A2 , and we denote by B the number of such state changes. The value of random variable B
completely describes the outcome of phase PB .
We have EB = p|S|, and by a standard additive Chernoff bound:
√
(17)
Pr |B − p|S|| ≤ 2 n PA ≥ 1 − 2e−4 > 7/8.
√
√
Let B ⊆ [p|S| − 2 n,p|S| + 2 n] be the subset of the considered interval containing values of B such
(1)
4c3
that ψn PA , B ∈ B ≥ √
. If Pr[B ∈ B|PA ] ≥ 1/8, then the claim follows directly.
n
√
√
Otherwise, it follows from (16) and (17) that there must exist a value b ∈ [p|S|−2 n, p|S|+2 n]\
B, such that:
2c3
ψn(2) PA , B = b ≥ − √ .
n
Given that:
4c
3
ψn(1) PA , B = b ≤ √ .
n
(2)
(1)
and recalling that ψn ≤ ψn , we obtain the following bound on ηn :
4c3
2c3
10c3
ηn∗ PA , B = b = 2ψn(1) − ψn(2) PA , B = b ≤ 2 √ + √ = √ .
n
n
n
(18)
(2)
We now consider lower bounds on the value of ψn , conditioned on PA , B = b+ (respectively,
√ 3 s , b + 21c
√ 3s ]
PA , B = b− ), where b+ (resp., b− ) is a value arbitrarily fixed in the range b+ ∈ [b + 20c
n
n
(resp., b− ∈ [b −
21c
√ 3s , b
n
−
20c
√ 3 s ]).
n
The executions of the protocol with B = b+ and B = b− differ
+
with respect to the execution with B = b in the number of executed transitions from a+
1 to a2 by at least
20c
√ 3 s . Recalling that δ2 = a2 − a1 , it follows that for some b0 ∈ {b+ , b− } we have after n steps:
n
20c3 s
δn PA , B = b0 ≥ |δ2,n | PA , B = b0 ≥ √ .
n
Subsequently, we will assume that b0 = b+ ; the case of b0 = b− is handled analogously. From the
relation η > δ/s and (18) we have:
20c3
10c3
ηn∗ PA , B = b+ ≥ √ ≥ ηn∗ PA , B = b + √ .
n
n
(19)
When comparing the value of κ∗n in the two cases, B = b+ and B = b, it is convenient to consider
√
κ∗ as the length of the vector (κ1 , κ2 , κ3 , 1/ n) in Euclidean space. For each of the coordinates κi ,
i = 1, 2, 3, we have:
40c3
κi,n PA , B = b+ − κi,n PA , B = b < √ ,
n
hence:
120c3
κ∗n PA , B = b+ < κ∗n PA , B = b + √ .
(20)
n
(1)
Introducing (19) and (20) into the definition of ψn , we obtain directly:
10c
3p 120c3
2c
10c3 3p 120c3
4c
3
√ ≥ −√3 +
√ > √3 ,
−
ψn(1) PA , B = b+ > ψn(1) PA , B = b + √ −
s
n
s
n
n
n
n
n
where we again used the fact that p is a sufficiently small constant w.r.t. s. We thus obtain:
4c
3
√ 3 s , b + 21c
√ 3s ] > √ ,
ψn(1) PA , B ∈ [b + 20c
n
n
n
26
where by the definition of random variable B as a sum of i.i.d. binary random variables and the choice
√ 3 s , b + 21c
√ 3 s ] holds with
of value b in the direct vicinity of the expectation of B, the event B ∈ [b + 20c
n
n
constant probability. The case of b0 = b− is handled analogously.
(1)
(2)
Lemma 5. Let ut2 be an arbitrary starting configuration of the system such that max{ψt2 , ψt2 } =
(1)
4c3
ψt2 ≥ √
. Then, with constant probability, for some t3 = t2 + O(n log n), a configuration ut3 is
n
reached such that δt3 > s/12.
(1)
Proof. We subsequently consider only the process ψt . We start by showing the following claim.
(1)
4c3
Claim. Suppose ψ0 = A ≥ √
. Then, with probability at least 1 − exp [−A2 psn/46656], for
n
some time step t ≤
72n
ps
(1)
≥ 2A, or δt > s/12.
P
Proof (of claim). Consider the Doob submartingale Yt = t−1
τ =0 Xt with increments (Xt ) given as:
(
(1)
A
∆ψt − psA
48n , if Yt > 2 or δt > s/12,
Xt =
0,
otherwise.
the process reaches a value ψt
Noting that |Xt | ≤ n9 , an application of the Azuma inequality for submartingales (cf. e.g. [13][Thm.
16]) to (YT ) with T = 72n
ps gives:
A
Pr[YT ≤ − ] ≤ exp [−A2 psn/46656],
2
Moreover, assuming the barrier δt > s/12 was not reached, we have:
A
psA
psA 72n A
(1)
(1)
ψT YT > −
= ψ0 +
T + YT > A +
− = 2A,
2
48n
48n ps
2
which completes the proof of the claim.
We now prove the lemma by iteratively applying the claim over successive intervals of time (τ0 , τ1 , . . .),
(1)
(1)
such that τ0 = t2 and τi+1 is the first time step not before τi such that ψτi+1 ≥ 2ψτi or δτi+1 ≥ s/12.
By the claim, we have:
72n
Pr τi+1 − τi ≤
≥ 1 − exp [−(ψτ(1)
)2 psn/46656].
i
ps
Noting that c3 > 48/(ps) by definition, and that before the barrier δ > s/12 is reached, we have
(1)
4c3 i
√ 2i , we obtain:
ψτi ≥ √
2 ≥ ps192
n
n
72n
Pr τi+1 − τi ≤
> 1 − exp [−4i+1 ].
ps
and further:
Pr τi+1
Y
i
i
X
72n
≤
(i + 1) >
1 − exp [−4j+1 ] > 1 −
exp [−4j+1 ] > 0.98.
ps
j=0
h
In particular, putting i = log2 n, Pr τi ≤
δτi ≥ s/12 (since otherwise we would
follows.
j=0
i
72n log2 n
> 0.98. Since for this value of i, we must have
ps
(1)
have ψτi = ω(1), which is impossible), the claim of the lemma
27
Phase with δ > s/12. The second phase of convergence corresponds to configurations of the system
which are sufficiently far from the center point (a1 , a2 , a3 ) = (s/3, s/3, s/3). Formally, we analyze a
variant of potential φ (with an additive corrective factor proportional to κ2 ) to show that, starting from
a configuration with δ > s/12, we will eliminate one of the three populations a1 , a2 , a3 in O(n log n)
sequential steps with constant probability, without approaching the center point too closely (a value of
δ = Ω(1) will be maintained throughout).
For this part of the analysis, we define the considered potential as:
ψ = η2 −
4p 2
s3
4p
κ = ln
− φ − 2 κ2 ,
2
s
27
s
(21)
for any configuration u with amin > 0.
We have directly from (6) and (7):
4p
p 1 2
p
2
ψ̇ = −φ̇ − 2 2κκ̇ ≥
δ − κδ + 4κ − 144 κδ =
s
s 2
s
2
δ
1p 2
p 1 2
p
δ +
− 2κ + 1 − 144 κδ ≥
δ ,
=
s 4
2
s
4s
(22)
where in the last transformation we took into account that p ≤ s/144.
For the sake of technical precision in formulating the subsequent lemmas, we also consider the
stochastic process ψt∗ , given as ψt∗ = ψ(ut ) for any t < td , where td is defined as the first time in the
evolution of the system such that a configuration with amin,td < c6 /n is reached, where c6 = 313600/s
∗
is a constant depending only on s. For all t ≥ td , we define ψt∗ := ψt−1
+ n1 .
Lemma 6. In any configuration ut with δ ≥ s/20 we have: E∆ψt∗ ≥
1 pδ 2
8 sn
>
ps
3600n .
Proof. We have:
4p
E∆(κ2 )
s2
Following the definition of φ in Eq. (2), we have by linearity of expectation:
X
X
X
∆ai
E∆φ = E
ln(ai + ∆ai ) −
ln ai =
E ln 1 +
.
ai
E∆ψ = −E∆φ −
(23)
Next, using the bound ln(1 + b) ≤ b which holds for b > −1, we have:
X
E ln(1 +
X E∆ai X ȧi /n
∆ai
)≤
=
= φ̇/n
ai
ai
ai
from which it follows that:
1
φ̇
(24)
n
To analyze E∆(κ2 ), we apply a variant of Lemma 1. A direct application of the lemma is not sufficient
due to the singularity related to the a−1
i term in the definition of κi ; however, this effect is compensated
when we take into account that any change of the value of κ2i occurs in the considered protocol with
probability at most proportional to ai . For the specific case of κ2i , for fixed i = 1, 2, 3, we consider
++
κ2i : R2 → R as a function of the restricted configuration ū = (a+
i , ai ), and we rewrite expression (12)
as:
X
E(∆κ2i ) =
∇f (u) · (u0 − u) + R2 (u, u0 ) Pr(u0 |u) ≤
E∆φ ≤
0
0
++
u0 =(a+
)6=u
i ,ai
≤
κ˙2i
1
+ 2
n
n
max
ku∗ −uk∞ ≤1/n
Dκ2 (u∗ )
X
Pr(u0 |u) ≤
i
0
0
++
u0 =(a+
)6=u
i ,ai
28
κ˙2i
1
+ 2
n
n
max
ku∗ −uk∞ ≤1/n
Dκ2 (u∗ )ai .
i
A straightforward computation from the definition of function κi shows that:
max
ku∗ −uk∞ ≤1/n
It follows that
E(∆κ2i )
Dκ2 (u∗ ) ≤
i
8s2
.
a2i
κ˙2i
ai 8s2
1 ˙2 8s2
≤
κi +
,
+ 2 2 =
n
n ai
n
ai n
and so:
E(∆κ2 ) ≤
1
n
24s2
κ˙2 +
amin n
.
(25)
Introducing (24) and (25) into (23), we obtain:
4p
1
1 4p
E∆ψ = −E∆φ − 2 E∆(κ2 ) ≥ − φ̇ −
s
n
n s2
392s
p 2
p
2
δ −
≥
≥
δ ,
4sn
amin n
8sn
1
24s2
96p
= ψ̇ − 2
≥
κ˙2 +
amin n
n
n amin
where in the second-to-last transformation we used (22), and in the last transformation we used the
2
relation a392s
≤ δ2 which holds when δ ≥ s/20 and amin ≥ c6 /n.
min n
∗
The claim thus follows when ψt∗ = ψt and ψt+1
= ψt+1 , i.e., for t < td . For larger values of t, the
∗
claim follows trivially from the definition of ψt .
The above Lemma is used to show that, starting from any configuration with δ > s/12, we quickly
reach a configuration in which some species has a constant number of agents.
Lemma 7. If δt ≥ s/20, we have:
(i) |∆ψt∗ | ≤ c7 ,
(ii) Var(∆ψt∗ ) ≤
c8
n.
where c7 > 0 and c8 > 0 are constants depending only on s. Moreover, in any configuration u with
amin ≥ 2/n, we have:
(iii) |∆ψ(u)| ≤
c7
namin .
(iv) Var(∆ψ(u)) ≤
c8
.
n2 amin
Proof. We first consider the case of a configuration with amin ≥ 2/n. Using the definition of ψ (and
within it, of φ and κ). Consider any transition from a configuration u to a subsequent configuration
u0 and let S ⊆ {1, 2, 3} be defined as the set of indices of configurations changing between u and u0
+ 0
++
++ 0
(S = {i : a+
i (u) 6= ai (u ) ∨ ai (u) 6= ai (u )}). We verify that there exists an absolute constant
c7 > 0 such that:
c7
|ψ(u0 ) − ψ(u)| ≤
≤ c7 .
n mini∈S ai
Moreover, by the definition of the protocol a transition from u to u0 occurs with probability Pr(u0 |u) ≤
mini∈S ai . Since there is only a constant number of possible successor configurations ut+1 for ut
(loosely bounding, not more than 36 ), it follows that:
( 36 c7 b
< n , for any b > 0
1
Pr |∆ψ(u)| >
7
b
= 0,
for b > nacmin
.
29
The bounds on the variance of Var(∆ψ(u)) and that of ∆ψt∗ = ∆ψ(u) (for t < td ) with amin,t ≥
(c + 6 + 1)/n follow directly. The analysis of ∆ψt∗ when amin,t = c6 /n and t < td is performed
analogously, noting that if the succeeding configuration u0 = ut+1 is such that amin (u0 ) < c6 /n, then
∆ψt∗ = n1 . Finally, for t ≥ td , the result holds trivially by the definition of ψt∗ .
Lemma 8. Let ut3 be an arbitrary starting configuration of the system such that δt3 > s/12. Then,
with probability 1 − O(1/n), for some t4 = t3 + O(n log n), a configuration ut4 is reached such that
amin,t4 = c6 .
Proof. W.l.o.g. assume that t3 = 0. First we remark that, by the relation between η and δ for δ ≤ s/12,
a process starting with δ0 > s/12 satisfies:
ψ0 = ψ0 = η02 −
4p 2 (s/12)2 12p
1
κ >
− 2 >
.
s2 0
s2
s
150
Moreover, for any configuration u0 with δ(u0 ) ≤ s/20 we have:
ψ(u0 ) = η 2 (u0 ) −
s 2
)
4p 2 ( 32 20
1
κ
≤
<
.
2
2
s
s
170
1
1
and as long as for all time steps t we have ψt ≥ 170
, the barrier condition
Thus, initially ψ0 > 150
1
1
δt ≥ s/20 has not been violated. Moreover, for ψt ∈ [ 170 , 150 ], we have by Lemma 6 that E∆ψt ≥ 0.
1
Moreover, by Lemma 7 (iii) and the fact that δt < 144
which implies amin,t > s/4, we have that
4c7
|∆ψt | ≤ sn .
It follows from a standard application of Azuma’s inequality for martingales (resembling the analysis
of the hitting time of the random walk with step size O(1/n), from one endpoint of a path of length Θ(1)
to the other) that:
1
Pr ∃t<n2 / ln n ψt <
= O(1/n),
170
hence also throughout the first n2 / ln n steps of the process we have δ > s/18, with probability 1 −
O(1/n).
We
Pare now ready to analyze the subsequent stages of the process, designing a Doob submartingale
Yt = t−1
τ =0 Xt with time increments (Xt ) defined as:
(
ps
∆ψt∗ − 3600
, if ψτ > 1/170 and amin,τ ≥ c6 for all τ ≤ t
Xt =
0,
otherwise.
Using Lemma 7 (i) and (ii) and applying the Azuma-McDiarmid inequality6 in the bounded variance
version (cf. e.g. [13][Thm. 18]) to Yt for tc = c3 n ln n, for some sufficiently large constant c > 0
depending only on s, we obtain:
#
"
"
#
4 ln2 n
c
c
ln
n
= exp −
= 1 − O(1/n).
Pr[Yt ≤ −c2 ln n] ≤ exp − c
2
2c8 + 23 cc7
2t n8 + 2c7 c3 ln n
ps
If the event Xt = ∆ψt∗ − 3600
were to hold for all t < tc with c = 2·3600
and if Ytc > −c2 ln n, then
ps
ps
we would have ψt∗c = ψ0∗ + Ytc + 3600 tc ≥ 0 − c2 ln n + 3c2 ln n = 2c2 ln n, which would mean that
ψt∗c 6= ψtc , since ψ ≤ 3 ln n + O(1) by definition. If ψt∗c 6= ψtc , then t4 < tc , and the proof is complete.
6
If our objective in the proof of the lemma were to show a bound on t4 which holds with constant probability (which would
be sufficient for our purposes later on), rather than a w.h.p. bound, then this specific step of the proof can also be performed
using Markov’s inequality. In any case, we would need to make use of the bounded variance of ψt∗ in the proof of the next
Lemma.
30
(Indeed, to reach a configuration with amin < c6 /n, the protocol has to pass through a configuration
with amin = c6 /n, since the size of each population changes by at most 1 in each transition.) Otherwise,
we must have that at least one of the following events holds: Ytc ≤ −c2 ln n, or ψτ ≤ 1/170 for some
τ < tc , or amin,τ < c6 for some τ < tc . We have established that each of the first two of these events
holds with probability O(1/n), whereas if the latter event holds, then t4 < tc . Thus, t4 < tc holds with
probability 1 − O(1/n) by a union bound.
Lemma 9. Let ut4 be a starting configuration of the system such that amin,t4 = c6 /n. Then, with
constant probability, for some t5 = t4 +O(n log n), a configuration ut5 is reached such that aj,t5 ≤ c6 /n
and aj+1,t5 > s/40, for some j ∈ {1, 2, 3}.
Proof. W.l.o.g. assume that arg mini=1,2,3 ai,t4 = 2. If a3,t4 > s/40, then the claim follows immediately, putting t5 = t4 and j = 2. Otherwise, we will show that with constant probability, the system will
evolve so that a2 will increase over time until within O(n log n) steps we will have a time step t5 with
j = 3 (i.e., a3,t5 ≤ c6 /n and a1,t5 > s/40).
In the considered case, w.l.o.g. assume t4 = 0. Next, let T = cn ln n for a sufficiently large constant
1
c; we choose as c := 2 log2 0.005ps
for convenience in later analysis. Intuitively, in view of Lemmas 6
∗
and 7, the potential ψT will be further increased in the next steps: the randompvariable (ψT∗√− ψ0∗ |u0 )
has an expected value of Θ(T /n) = Θ(log n), with a standard deviation of Θ( T /n) = Θ( log n).
By an application of the Azuma-McDiarmid inequality for martingales with bounded variance similar to that in the proof of Lemma 8, we obtain the following result:7
pst
T
∗
∗
Pr ∀t≤T ψt ≥ ψ0 +
−ε
= 1 − n−Ω(1) , for any constant ε > 0.
(26)
3600n
n
Observe that since a2,0 = c6 /n = O(1/n), we have ψ0∗ ≥ ln n − O(1). Taking this into account,
for our purposes, a slightly weaker and simpler form of expression (26) will be more convenient:
Pr ∀t∈[0.5n log2 n,T ] ψt∗ ≥ (1 + 10−4 ps) ln n = 1 − o(1).
(27)
The proof of the lemma is completed by a more fine-grained analysis of the considered protocol. In
the initial configuration t4 = 0, we have a2,0 = c6 /n (there are exactly c6 agents in state A2 ), and
since t4 6= t5 , we have a3,0 < s/40. Consequently, a1,0 = s − s/40 − O(1/n) > 0.9s. Informally,
since the prey of A2 (i.e., A1 ) is more than twice more numerous than its predator (i.e., A3 ), we should
++
observe the increase in the size of population of A2 , regardless of the activities (A+
i or Ai ) of the
agents in the population. We consider the evolution of the system, finishing at the earliest time te when
a2 (te ) > s/100. The following relations are readily shown (apply e.g. Lemma 14 with i = 2 and
u = 0):
0.05ps
a2
n
0.05ps
< 0.05s, t < te ) ≤ −
a3 .
n
E(∆a2,t |a3,t < 0.05s, t < te ) ≥
(28)
E(∆a3,t |a3,t
(29)
From (29), taking into account that |∆ai,t | ≤
for martingales shows that:
1
n
and a3,0 < s/40, an application of Azuma’s inequality
Pr[∀t≤min{te ,T } a3,t < 0.05s] = 1 − o(1).
(30)
Taking into account the above, by a straightforward geometric growth analysis (compare e.g. proof of
Lemma 5), we obtain from (28):
Pr[te < T ] = 1 − o(1).
(31)
7
Such an analysis can also be performed using Chebyshev’s inequality, obtaining a slightly weaker expression in the probability bound.
31
Moreover, since the speed of increase of a2 is bounded (even in the absence of predators) by that of a
standard push rumor spreading process (formally, E(∆a2,t ) ≤ a2,t ), we have (compare e.g. [20]):
Pr[te > 0.5n log2 n] = 1 − o(1).
(32)
Now, we observe that with constant probability, the size of population A2 does not decrease in the time
interval [0, te ] below the value a2,0 = c6 /n, attained at the beginning of this interval:
Pr ∀t∈[0,te ] a2,t ≥ c6 /n = Ω(1).
(33)
Indeed, with constant probability the value a2,t is initially non-decreasing: with constant probability,
in the first O(n) rounds each of the c6 = O(1) agents from A2 will be triggered by the scheduler O(1)
times in total, and each interaction involving an agent from A2 will have this agent as the initiator, and
an agent from the largest of the three populations, A1 , as the receiver (the prey). Thus, with constant
probability, the number of agents in population A2 is increased to an arbitrary large constant (e.g.,
1000c6 ). After this, we use the geometric growth property (28) to show that a2,t reaches the barrier
a2,t > s/100 (at time te ) before the event a2,t < c6 /n occurs (cf. e.g. proof of Lemma 5, or standard
analysis of variants of rumor-spreading processes in their initial phase [29]).
When the event from bound (33) holds, at least one of the following events must also hold:
(A) amin,t ≥ c6 /n, for all t ≤ te ,
(B) or there exists a time step t < te such that a1,t ≤ c6 /n,
(C) or there exists a time step t < te such that a3,t ≤ c6 /n.
To complete the proof, we will show that each of the events (A) and (B) holds with probability o(1).
Indeed, then in view of (33), event (C) will necessarily hold with probability Ω(1). This means that,
with probability Ω(1), there exists a time step t < te such that a3,t < c6 /n and a2,t < s/100 (since
t < te ), and so also a1,t > s − s/100 − c6 /n > 0.98s > s/40; thus, the claim of the lemma will hold
with t5 = t and j = 3.
To show that event (B) holds with probability o(1), notice that a2,t < s/100 by definition of te , and
moreover a3,t < 0.05s with probability 1 − o(1), hence the event a1,t < s − s/100 − 0.05s = 0.94s
holds with probability o(1).
To show that event (A) holds with probability o(1), notice that, substituting in (27) t = te , by a
union bound over (27), (31) and (32) we obtain:
Pr ψt∗e ≥ (1 + 10−4 ps) ln n = 1 − o(1).
This means that, with probability 1 − o(1), we have ψt∗e 6= ψte or ψte ≥ (1 + 10−4 ps) ln n. In the first
case, event (A) cannot hold. In the second case, observe that a2,te = s/100 + O(1/n)
by definition of
P
1
te , so a1,te > s − s/100 − O(1/n) − c6 /n > 0.98s, and it follows that ψte = 3i=1 ln ai,t
+ O(1) =
e
−4
ln n + O(1). Since the condition ψte ≥ (1 + 10 ps) ln n is not fulfilled, event (A) can only hold with
probability o(1).
Lemma 10. Let ut5 be a starting configuration of the system such that aj,t5 ≤ c6 /n and aj+1,t5 > s/40,
for some j ∈ {1, 2, 3}. Then, with constant probability, for some t6 = t5 + O(n), a configuration ut6 is
reached such that amin,t6 = 0.
Proof. We consider the pairs of interacting agents chosen by the scheduler in precisely the next n rounds
after time t5 . Given that set Aj,t5 has constant size, and set Aj+1,t5 has linear size in n, it is straightforward to verify that with constant probability, the set of randomly chosen n pairs of agents has all of the
following properties:
32
• Each agent from Aj,t5 belongs to exactly one pair picked by the scheduler, and is the receiver in
this pair.
• Each agent interacting in a pair with an agent from Aj,t5 belongs to exactly one pair.
• Each agent interacting in a pair with an agent from Aj,t5 belongs to set Aj+1,t5 .
Conditioned on such a choice of interacting pairs by the scheduler, the protocol changes the state of
all agents from set Aj,t5 to state j + 1 with probability at least p|Aj,t5 | ≥ pc6 = Ω(1). State j is then
effectively eliminated.
In the absence of species j, the interaction between species j − 1 and j + 1 collapses to a lazy
predator-prey process, with transitions of the form (Aj−1 , Aj+1 ) → (Aj−1 , Aj−1 ) associated with a
constant transition probability. A w.h.p. bound on the time of elimination of species j + 1 follows
immediately from the analysis of the push rumor spreading model, and we have the following Lemma.
Lemma 11. Let ut6 be a starting configuration of the system such that aj,t6 = 0, for some j ∈ {1, 2, 3}.
Then, with probability 1 − O(1/n), for some t7 = t6 + O(n log n), a configuration ut7 is reached such
that for all t ≥ t7 , aj,t = aj+1,t = 0 and aj−1,t = s.
After a further O(n log n) steps after time t7 , the final configuration of all agents in the oscillator’s
population will be a++
j−1,t = s.
6.5
Operation of the Oscillator in the Presence of a Source
In this section we prove properties of the oscillatory dynamics for the case #X > 0. It is possible to
provide a detailed analysis of the limit trajectories of the dynamics in this case, as a function of the
concentration of x. Here, for the sake of compactness we only show the minimal number of properties
of the oscillator required for the proof of Theorem 1. When the given configuration is such that amin is
sufficiently large, say amin > 0.02s, then both the subclaims of Theorem 1(2) hold for the considered
configuration. (The first subclaim hold directly; the second subclaim follows by a straightforward concentration analysis of the number of agents changing state in protocol Po over the next 0.01sn steps,
since we will always have amin ≥ 0.01s during the considered time interval.) Otherwise, the considered
configuration is close to one of the sides of the triangle. We will show that in the next O(n log n) steps,
with high probability, the protocol will either reach a configuration with amin > 0.02s, or will visit
successive areas around the triangle, as illustrated in Fig. 7. The following Lemmas show that within
each area, an exponential growth process occurs, which propagates the agent towards the next area.
Lemma 12. If ai−1 < 0.8s and ai+1 < 0.05s, then ȧi−1 ≤ xs/3 − 0.05psai−1 .
Proof. From the assumptions we have that ai > 0.15s. Starting from (3) we obtain:
++
ȧi−1 = x(s/3 − ai−1 ) + pai+1 (ai−1 + a++
i−1 ) − pai−1 (ai + ai )
≤ xs/3 + 2pai−1 ai+1 − pai−1 ai
≤ xs/3 + pai−1 (2ai+1 − ai )
≤ xs/3 + pai−1 (0.1s − 0.15s) = xs/3 − 0.05psai−1
33
A3
a2 = 0.05s
a3 = 0.8s
ta
a1 = 0.25s
A1
tb
A2
ta0
Figure 7: Configurations considered in Lemmas 12 and 13 (descending dashed slope) and Lemmas 14
and 15 (ascending dashed slope) for i = 1 (dark red) and i = 2 (green). Unless the process leaves
the shaded area towards the center o the triangle, the shape of the time trajectory is counter-clockwise
around the triangle, w.h.p.
From the above bound on expectation, the following Lemma follows directly by a standard concentration analysis. In what follows, we consider an execution in which the concentration x is strictly
positive and bounded by a sufficiently small absolute constant (i.e., #X is at most a given constant fraction of the entire population), with the required upper bounds on x used in the proofs of lemmas given
in their statements. This is a technical assumption, which allows us to simplify the proof structure. In
particular, the assumption #X ≤ c12 n can be omitted in the statement of Theorem 1, and the claim of the
theorem can even be proved for executions in which #X changes during the execution of the protocol,
as long as the invariant #X > 0 is preserved over the considered interval of time.
Lemma 13. Let uta be a starting configuration of the system such that ai−1,ta < 0.75s and ai+1,ta <
0.05s. Suppose x < 10−3 ps, starting from time ta . Then, for some tb ∈ [ta , ta + c9 n], where c9 is a
Ω(1)
constant depending only on p and s, with probability 1 − e−n ) , the system reaches a configuration
utb such that exactly one of the following two conditions is fulfilled:
• either amin,tb ≥ 0.02s,
• or ai+1,tb < 0.05s and ai−1,tb < 0.02s.
Proof. In the considered range of values of ai−1 , we have ai−1,ta < 0.75s and ai−1,t ≥ 0.02s, for all
t until we leave the considered area at time tb . Taking into account that x < 10−3 ps, it follows from
Lemma 12 that:
E∆ai−1 ≤
1
1
ps2
(xs/3 − 0.05psai−1 ) ≤ (x < 10−3 ps2 /3 − 10−3 ps2 ) < −0.0005
.
n
n
n
Taking into account that |∆ai−1 ≤ n1 |, it follows from a straightforward concentration analysis (cf. e.g.
proof of Lemma 5 for a typical analysis of this type of exponential growth process) that a boundary of
the considered area (either ai−1,t < 0.02s or amin,t > 0.02s) must be reached within O(n) steps with
very high probability, as stated in the claim of the lemma.
A similar analysis is performed for the next area.
Lemma 14. If ai+1 < 0.25s and ai−1 < 0.05s then ȧi+1 ≥ xs/12 + 0.6psai+1 and ȧi−1 ≤ xs/3 −
0.2psai−1 .
34
Proof. From assumptions we have that ai > 0.7s. Starting from (3) we obtain:
++
ȧi+1 = x(s/3 − ai+1 ) + pai (ai+1 + a++
i+1 ) − pai+1 (ai−1 + ai−1 )
≥ x(s/3 − ai+1 ) + pai+1 (ai − 2ai−1 )
≥ x(s/3 − 0.25s) + pai+1 (0.7s − 2 · 0.05s)
= xs/12 + 0.6psai+1
++
ȧi−1 = x(s/3 − ai−1 ) + pai+1 (ai−1 + a++
i−1 ) − pai−1 (ai+1 + ai+1 )
≤ xs/3 + pai−1 (2ai+1 − ai )
≤ xs/3 + pai−1 (2 · 0.25s − 0.7s) = xs/3 − 0.2psai−1 .
Again, a concentration result follows directly.
Lemma 15. Let utb be a starting configuration of the system such that ai−1,tb < 0.02s and ai+1,tb <
1
0.02s. Suppose x < 0.02ps, starting from time tb . Then, for some ta0 ∈ [tb , tb +c10 n ln max{1/n,a
]⊆
i+1,t }
b
[tb , tb + c10 n ln n], where c10 is a constant depending only on p and s, with probability 1 − O(1/n3 ),
the system reaches a configuration uta0 such that exactly one of the following two conditions is fulfilled:
• either amin,ta0 ≥ 0.02s,
• or ai−1,ta0 < 0.05s, ai+1,ta0 > 0.25s, and (consequently) ai,ta0 < 0.75s.
Proof. We first show that, starting from time tb onward, the process ai−1,t satisfies ai−1,t < 0.05s for
Ω(1)
all t ∈ [tb , t∗ ] with probability 1 − e−n , where t∗ is defined as the minimum of time ta + c10 n ln n
and the last time moment such that ai+1,t ≤ 0.25s holds for all t ∈ [tb , t∗ ]. By Lemma 14, we have for
all t ∈ [tb , t∗ ] such that ai−1,t > 0.01s:
E∆ai−1 =
1
1
1
0.01ps2
ȧi−1 ≤ (xs/3 − 0.2psai−1 ) < (0.02ps2 /3 − 0.02ps2 ) < −
,
n
n
n
n
where we took into account the assumption x < 0.02ps. The claim on ai−1,t < 0.05s follows from a
standard concentration analysis, noting that |∆ai−1,t | ≤ n1 .
In order to analyze the process ai+1,t , we apply a filter and consider the process a0i+1,t , starting at
time tb , defined as follows. For as long as ai−1,t < 0.05s, we put a0i+1,t := ai+1,t , and starting from
the first time t∗∗ when ai−1,t > 0.05s, we compute a0i+1,t+1 as the subsequent value of ai+1 after a
simulation of a single step of the process for some state u with concentrations of types: x(u) = x,
ai+1 (u) = a0i+1,t , ai−1 (u) = 0.05s, and ai (u) = s − ai−1 (u) − ai+1 (u).
For a given time step t, let Rt denote the event that ∆a0i+1,t 6= 0. By the construction of protocol Po ,
which always requires at least one agent of type X or type Ai+1 to be involved in an interaction which
creates or destroys an agent of type Ai+1 , we have:
Pr[Rt ] ≤ x + 2a0i+1,t .
Moreover, from Lemma 14 it follows that for a0i+1,t < 0.25s:
E∆a0i+1,t
ȧ0i+1,t
1
=
≥ (xs/12 + 0.6psa0i+1,t ).
n
n
35
Since ∆a0i+1,t |¬Rt = 0, we have:
E∆a0i+1,t |Rt ≥
+ 0.6psa0i+1,t )
≥
Pr[Rt ]
1
n (xs/12
+ 0.6psa0i+1,t )
0.3ps
≥
,
0
x + 2ai+1,t
n
1
n (xs/12
and moreover ∆a0i+1,t |Rt ∈ {− n1 , 0, n1 }. Analysis of this type of process is folklore (in the context of
epidemic models with infection and recovery) but somewhat tedious; we sketch the argument for the sake
of completeness. When considering only those steps for which event Rt holds, the considered process
can be dominated by a lazy random walk on the line {0, n1 , 2,n
, . . .}, with a constant bias towards its right
endpoint. To facilitate analysis, we define points Qc = cbα ln ncn, for c = 0, 1, . . ., where constant
α > 0 is subsequently suitably chosen, and for any point Qc to the right of the starting point of the walk
(i.e, c > cmin where cmin is the smallest integer such that Qcmin +1 > ai+1,tb ) define sc as the number of
steps of the walk until its first visit to sc . For a suitable choice of constants α and β > 0 sufficiently large,
we have that for any c, with probability at least 1 − O(1/n2 ), sc+1 − sc ≤ β ln n, and moreover between
its step sc and its step sc+1 , the walk is confined to the subpath (Qc−1 , Qc+1 ) of the considered path.
Considering the original time t of our process a0i+1,t (including moments with ¬Rt ), let tc be the moment
of time corresponding to the sc -th step of the walk. Conditioning on events which hold with probability
1 − O(1/n2 ), the value tc+1 − tc can be stochastically dominated by the sum of β ln n independent
n
geometrically distributed random variables, each with expected value O( max{1,(c−1)
ln n} ). Let cmax be
the largest positive integer such that Qcmax < 0.25s. Applying a union bound on the conditioning of
all intervals tc+1 − tc , for c ≥ cmin and a concentration bound on the considered geometric random
variables, we eventually obtain that with probability 1 − O(1/n3 ) the condition a0i+1,t is achieved for
time:
!
cX
cX
max
max
n
t<
(tc+1 − tc ) = O n + ln n
=
max{1, (c − 1) ln n}
c=cmin
c=cmin
1
O (n(ln n − ln max{1, ai+1,tb n} + 1)) = O n ln
.
max{1/n, ai+1,tb }
Recalling that a0i+1,t = ai+1,t holds throughout the considered time interval with very high probability,
the claim follows.
An iterated application of Lemmas 13 and Lemmas 15 moves the process along time moments ta , tb ,
where time moment t0a is again be fed to Lemma 13, considering the succeeding value of i. After
a threefold application of both Lemmas, the process has w.h.p. in O(n log n) steps either performed a
complete rotation, passing through three moments of time designated as “ta ”, rotated by one third of
a full circle, or has reached at some time t0 a point with amin,t0 ≥ 0.02s. In either case, the claim of
Theorem 1(2) follows directly.
t0a , . . .,
7
7.1
Analysis of Protocol for D ETECT
Further Properties of the Oscillator
We start by stating a slight generalization of Lemma 6, capturing the expected change of potential ψt∗
(given by (21)) for the case #X > 0, for configurations which are sufficiently far from both the center
and the sides of the triangle.
Lemma 16. In any configuration ut with 10−6 s2 ≤ amin ≤ 0.02s and x < c12 we have: E∆ψt ≥
where c12 > 0 is an absolute constant which depends only on s and p.
36
ps
7200n ,
Proof. We can condition the expectation of E∆ψt on the event Et , which holds if an agent in state
X participates in the current interaction. Conditioned on ¬Et , the analysis corresponds directly to the
computations performed for the case x = 0, where we remark that the assumptions of Lemma 6 are
satisfied due to the assumed upper bound on amin . Thus:
ps
E∆ψt |¬Et >
.
3600n
Next, taking into account the lower bound on amin , by exactly the same argument as in Lemma 7(iii),
we have |∆ψt | < c012 /n, for some choice of constant c012 > 0 which depends only on s and p. Obviously,
E∆ψt |Et > −c012 /n,
and since Pr[Et ] < 2x, by the law of total expectation:
E∆ψt > (1 − 2x)
ps
c0
ps
− 2x 12 >
,
3600n
n
7200n
where the last inequality holds for any x < c12 , where c12 :=
1
.
2(2c012 +1)
Lemma 17. Suppose amin,t0 < 10−6 s4 at some time t0 . Then, there exists an absolute constant c13 > 0,
Ω(1)
c
such that the following event holds with probability 1 − e−n ) : for all t ∈ [t0 , t0 + en 13 ], we have
amin,t < 0.01s2 .
3
s
Proof. Let ψ̂t ≡ ln 27
− ψt . Consider any t ≥ t0 such that amin,t < 10−6 s4 . Then, φt < ln amin,t <
−6
4
ln(10 s ) and consequently:
ψ̂t = φt +
4p
4p 2
κ ≤ ln(10−6 s4 ) + 4 ≤ ln(2 · 10−6 s4 ),
s4 t
s
where we recall that κ2t ≤ 1 and the last inequality follows for p chosen to be sufficiently small (4p/s4 <
ln 2).
Further, note that if for some time t we have ψ̂t < ln(8 · 10−6 s4 ), then:
φt = ψ̂t −
4p 2
κ < ln(8 · 10−6 s4 ),
s4 t
thus a1,t a2,t a3,t < 8 · 10−6 s4 , from which it follows that a2min,t < 16 · 10−6 s4 , and so amin,t < 0.01s2 .
Thus, for ψ̂t < ln(8 · 10−6 s4 ), at least one of the following holds:
• Either ψ̂t < ln(2 · 10−6 s4 ),
• Or ψ̂t ≥ ln(2 · 10−6 s4 ), thus amin,t ≥ 10−6 s4 . Then, taking into account that amin,t < 0.01s2 ,
ps
we have by Lemma 16: E∆ψ̂t < − 7200n
.
Taking into account the known properties of function ψt (Lemma 7), we have that starting from ψ̂t0 <
c
ln(2 · 10−6 s4 ), it takes time exponential in a polynomial of n (eΩ(n 13 ) , for some absolute constant
c13 > 0) to break the potential barrier for ψ̂, i.e., to reach the first moment of time t1 such that ψ̂t1 ≥
Ω(1)
ln(8 · 10−6 s4 ), with probability 1 − e−n ) , for some absolute constant c14 > 0. To complete the
proof, recall that for any t < t1 , we have ψ̂t < ln(8 · 10−6 s4 ), and so as previously established,
amin,t < 0.01s2 .
For any execution of the oscillator protocol Po , we can now divide the axis of time into maximal
time intervals of two types, which we call oscillatory and central. A central time interval continues
for as long as the condition amin,t ≥ 10−6 s4 is fulfilled, and turns into an oscillatory interval as soon
as this condition no longer holds. An oscillatory time interval continues for as long as the condition
amin,t < 0.01s2 is fulfilled, and turns into an oscillatory interval as soon as this condition no longer
holds. Lemma 17 implies that an oscillatory interval is of exponential length w.v.h.p.
37
Lemma 18. Suppose 0 < x < c12 . Let t0 be an arbitrary moment of time such that amin,t0 > 0.
1
Let T = Cn ln amin,t
, for an arbitrarily fixed constant positive integer C = O(1). With probability
0
Ω(1)
1 − e−n
, we have for all subsequent moments of time t ∈ [t0 , t0 + T ]:
ln
1
max{ n1 , amin,t }
≤ 100C ln
1
.
amin,t0
Proof. Without loss of generality assume t0 = 0. We can assume in the proof that amin,0 > n−0.01/C ,
otherwise the claim trivially holds. Thus, initially we have φ0 ≥ 3 ln amin,0 ≥ − 0.03
C ln n ≥ −0.03 ln n.
We proceed to show that potential φ does not decrease much during the considered motion. We have
at any time φ̇ ≥ −p, which follows directly from (5).
Suppose at some time t we have φt ≥ −0.2 ln n. We note that amin,t ≥ eφt ≥ n−0.2 . Applying
Lemma 1, we have under these assumptions:
!
p
1
1
E∆φt ≥ − − O
≥−
2
2
n
n
n amin,t
and moreover by the properties of the natural logarithm (cf. e.g. [15]):
|∆φt k <
4
namin,t
≤
4e−φt
≤ 4n−0.8 .
n
As usual, we apply a Doob martingale, with φ0t := φt until the first moment of time t such that φt <
−0.2 ln n, and subsequently φ0t+1 := φ0t for larger t. We have E∆φt ≥ −n−1 and |∆φt k < 4n−0.8 .
1
Considering T = Cn ln amin,0
< Cn ln n steps of the process starting from time 0, by a standard
Ω(1)
application of Azuma’s inequality, we obtain that with probability 1−e−n
φ0t − φ00 ≥ −
, for all t ∈ [0, T ] we have:
T
1
− n−0.2 ≥ −C ln
− 1 ≥ −0.1 ln n.
n
amin,0
From the last inequality it follows that φ0t ≥ φ00 − 0.1 ln n ≥ −0.2 ln n for all t ∈ T , with probability
Ω(1)
1 − e−n , and so φ0t = φt . We now rewrite the same bound for φt , using the relation φ ≥ 3 ln amin :
φt ≥ φ0 − C ln
1
amin,0
− 1 ≥ 3 ln amin,0 − C ln
1
amin,0
− 1 ≥ −(C + 4) ln
1
amin,0
≥ −100C ln
1
amin,0
.
1
Taking into account that − ln amin,t
≥ φt for amin,t 6= 0, we obtain the claim.
Lemma 19. Suppose 0 < x < c12 . Fix type i ∈ {1, 2, 3}. Let t0 be any time such that amin,t0 < 10−6 s4 .
Let t∗ > t be the first moment after t0 such that i is the most represented type, ai,t∗ = amax,t∗ . Let
t∗∗ > t∗ be the first moment after t∗ such that i is the least represented type, ai,t∗∗ = amin,t∗∗ .
1
Then, with probability 1−O(1/n3 ), t∗∗ ≤ t0 +c11 n ln max{1/n,a
, where c11 > 0 is an absolute
i+1,t0 }
constant depending only on s and p.
Proof. From Lemma 17, we have that w.v.h.p. the protocol is in an oscillatory interval which will
Ω(1)
c
last super polynomial time, i.e., with probability 1 − e−n ) : for all t ∈ [t0 , t0 + en 13 ], we have
amin,t < 0.02s. Acting as in the previous subsection, we iteratively apply Lemma 13 and Lemma 15.
After at most 6 applications of both Lemmas, the process has performed two complete rotations around
the triangle, w.h.p., passing in particular through time moments t∗ where the designated type i was a
maximal type and t∗∗ where type i was a minimal type. It remains to bound the time required to perform
these iterations. We consider as an example a single application of Lemma 15 starting at a time tb and
38
1
ending at a time ta0 , where ta0 < tb + c10 n ln max{1/n,a
i+1,t
1
Lemma 18 at time tb , we obtain ln amin,t
a0
b
}
with probability 1 − O(1/n3 ). Applying
1
≤ 100C ln amin,t
, w.v.h.p. We use this bound for the next
b
application of Lemma 13, and so on. After a total of at most 12 applications, we eventually obtain a
1
bound of the form c11 ln amin,t
on the length of the considered time interval, where the value of c11 is
0
computed as a function of C.
7.2
Protocol Extension Pm : Majority
(s)
The composition of the extension is specified in Fig. 4. In what follows, we denote Mi
(s)
(s)
and mi := Mi /n, for s ∈ {−1, 0, +1}.
:= #(A?i , Ms )
Lemma 20. Suppose 0 < x < c12 . Let t0 be an arbitrarily chosen moment of time and let T > 0. For
fixed i ∈ {1, 2, 3}, we have for all t ∈ [t0 , t0 + T ]:
(+1)
|mi,t
with probability 1 − O(e−n
1/6
(−1)
− mi,t | ≤ 2e4rT /n max{n−0.2 , ai,t0 },
).
(+1)
(−1)
Proof. W.l.o.g. assume t0 = 0 and i = 1. Denote A0 := #A1,0 , Dt := M1,t − M1,t , and Gt := Dt2 .
As usual, we denote ∆Dt = Dt+1 − Dt and ∆Gt = Gt+1 − Gt . For the subsequent analysis, we choose
to use the “squared potential” Gt to simplify considerations; this would be the usual potential of choice
to analyze an unbiased random walk with a fair coin toss.
First we remark that |D0 | ≤ A0 , so |G0 | ≤ A20 . Next, observe that since at most one agent changes
its state in a single time step, we have |∆Dt | ≤ 2, and so:
|∆Gt | ≤ (|Dt | + 2)2 − |Dt |2 ≤ 4(|Dt | + 1) ≤ 4n + 4.
(34)
We now upper bound the expectation E∆Gt . We condition this expectation on the disjoint set of events
R6 , R7 , R8 , R9 , R10 , R0 , where Rr , for 6 ≤ r ≤ 10, corresponds to Rule (r) being executed in the
current step, and R0 is the event that none of these rules is executed. We have the following:
• If event R0 or R6 holds, then at least one of the following three situations occurs: (1) the values
(+1)
(−1)
of M1
and M1
both remain unchanged at time t, (2) an agent changes state from type
A1 to another type, or (3) an agent turns from another type into type A1 . In case (1), we have
Dt+1 = Dt . In case (2), the probability that |Dt+1 | = |Dt | + 1 is not more than the probability
that |Dt+1 | = |Dt | − 1, since by the construction of the protocol, the choice of the agent leaving
the population is completely independent of its value Ms . In case (3), we have Pr[|Dt+1 | =
|Dt | − 1] = Pr[|Dt+1 | = |Dt | + 1] = 1/2 by construction. In all cases, Pr[|Dt+1 | = |Dt | + 1] ≤
Pr[|Dt+1 | = |Dt | − 1]. We therefore have:
E[∆Gt |R0 ∨ R6 ] ≤ ((|Dt | + 1)2 − |Dt |2 ) + ((|Dt | − 1)2 − |Dt |2 ) ≤ 2.
• For events R7 and R8 , we have Dt+1 |R7 = Dt − 1 and Dt+1 |R8 = Dt + 1. Since events R7 and
R8 hold with equal probability, it follows that:
E[∆Gt |R7 ∨ R8 ] ≤ ((|Dt | + 1)2 − |Dt |2 ) + ((|Dt | − 1)2 − |Dt |2 ) ≤ 2.
39
• Finally, for events R9 and R10 , we have Dt+1 |R9 = Dt + 1 and Dt+1 |R10 = Dt − 1. Since
(−1)
(+1)
Pr[R9 ] · M1
= Pr[R10 ] · M1
(+1)
E[∆Gt |R9 ∨ R10 ] ≤
max{M1
(−1)
, M1
}
A1
(+1)
min{M1
(−1)
, M1
A1
}
((|Dt | + 1)2 − |Dt |2 )+
((|Dt | − 1)2 − |Dt |2 ) ≤
4|Dt |2
4Gt
+2=
+ 2,
A1
A1
where we assume in notation that A1 > 0.
Applying the law of total expectation for ∆Gt over the set of events R0 ∨ R6 , R7 ∨ R8 , R9 ∨ R10 and
noting that Pr[R9 ∧ A10 ] ≤ r An1 , we eventually obtain:
Gt
+ 2.
(35)
n
Inequalities (34) and (35) are sufficient to lower-bound the evolution of random variable Gt , which
undergoes multiplicative drift with rate parameter 1 + 4r/n (up to lower order terms). Since known
multiplicative drift lower bounds (cf.e.g. [17, 33]) do not appear to cover this case explicitly, we sketch
the corresponding submartingale analysis (with slightly weaker parameters) for the sake of completeness.
Consider any moment of time t such that Gt ≥ n1.6 . Define target value Gmax = (1+8r)Gt ≥ n1.6 .
Consider the following filter, defining G0τ , τ ≥ 0 as the submartingale with G0τ = Gt+τ until the first
moment of time at which Gt+τ < Gmax , and G0τ = G0τ +1 + 5r
n Gmax for all subsequent moments of
5r
0
0
time. Note that E∆Gτ ≤ n Gmax by (35) and |∆Gτ | ≤ 4n + 4 ≤ 5n by (34) (where we conduct the
entire analysis for n sufficiently large with respect to absolute constants of the algorithm). By Azuma’s
inequality, we have for any τ > 0 and z > 0:
5r
−z 2
0
0
Pr[Gτ > G0 + Gmax τ + z] ≤ exp
.
n
2τ (5n)2
E∆Gt ≤ 4r
Next, choosing any τ ≤ n and z = rGmax ≥ rn1.6 and noting that:
5r
Gmax τ + z ≤ Gt + 5rGmax + rGmax ≤ Gmax ,
n
we rewrite the concentration equality as:
−(rn1.6 )2
0.19
0
Pr[Gτ > Gmax ] ≤ exp
< e−n .
2τ (5n)2
G00 +
Applying to the above a union bound over all τ ∈ [0, n], we obtain by another crude estimate:
Pr[∀τ ∈[0,n] G0τ ≤ Gmax ] ≥ 1 − e−n
0.18
,
from which it follows directly by the definition of G0τ that:
0.18
Pr[∀τ ∈[0,n] Gt+τ ≤ Gmax ] ≥ 1 − e−n
.
Thus, given that Gmax = (1 + 8r)Gt , the value of Gt increases by a factor of at most (1 + 8r) over
n steps, with very high probability. Iterating the argument at most a logarithmic number of times and
applying a union bound gives for arbitrary t:
Gt ≤ (1 + 8r)t/n+1 max{n1.6 , A20 } ≤ 2e8rt/n max{n1.6 , A20 },
1/6
with probability at least 1 − e−n , from which the claim of the lemma follows directly after taking the
square root and normalizing by a factor of n.
40
(−1)
(0)
(+1)
By considering the sizes of populations mi,t , mi,t , and mi,t (whose sum is ai,t ), we obtain the
1
following corollary of the above Lemma, applied for a suitably chosen value T = 0.001n
ln ai,t
.
r
0
Lemma 21. Suppose 0 < x < c12 . Let t0 be an arbitrarily chosen moment of time with ai,t0 ≤ 0.02s2 .
For fixed i ∈ {1, 2, 3}, we have for all t ∈ [t0 , t0 + 0.001n
ln max{ 11,a } ]:
r
i,t0
n
(+1)
|mi,t
with probability 1 − O(e−n
1/6
(−1)
− mi,t | ≤ 0.1ai,t + 0.05s,
).
(0)
The above Lemma provides a crucial lower bound on the size of population mi .
Lemma 22. Suppose 0 < x < c12 . Let t0 be an arbitrarily chosen moment of time with ai,t0 ≤ 0.02s2 .
For fixed i ∈ {1, 2, 3}, we have for all t ∈ [t0 , t0 + 0.005n
ln max{ 11,a } ] such that ai,t > 0.25s:
r
i,t0
n
(+1)
(−1)
min{mi,t , mi,t } ≥ 0.001s3 ,
Ω(1)
with probability 1 − e−n
.
Proof. Note first that by the conditions ai,t0 ≤ 0.02s2 < 0.1s < 0.25s < ai,t . Let t1 > t0 denote the
last moment of time before t2 such that ai,t1 −1 < 0.1s and let t2 ≥ t1 + 0.15sn denote the first moment
of time after t1 such that ai,t2 > 0.25s. We have t ≥ t2 + 0.05sn.
(+1)
(−1)
We now consider the process mi,τ (exactly the same arguments may be applied for process mi,τ ).
(+1)
We have ∆mi,τ
≤ 1/n. The analysis is divided into two phases:
(+1)
• Phase 1: τ ∈ [t1 , t2 ) (thus ai,τ ∈ [0.1s, 0.25s]). Initially, mi,t1 ≥ 0. Suppose for some step
(+1)
τ we have mi,τ
0.25s) ≥
most
< 0.02s. Rule (6) is executed with probability (ai,τ )(1 − ai,τ ) ≥ 0.1s(1 −
0.075s2 ,
(+1)
2rmi,τ
(+1)
whereas rule (8) or (9) which reduces mi,τ
< 0.04rs <
0.01s2 .
A computation of the expected value provides:
(+1)
(+1)
E[∆mi,τ |(mi,τ
< 0.02s)] >
(+1)
An application of Azuma’s inequality yields that mi,t2
probability 1 − e
−nΩ(1)
is executed with probability at
>
0.06s2
.
n
1 0.06s2
2 n (t2
− t1 ) > 0.004s3 , with
. In case of failure, we consider the process no further.
(+1)
• Phase 2: τ ∈ [t2 , t) (thus ai,τ ≥ 0.1s). From Phase 1, we have that initially mi,t2 > 0.004s3 .
(+1)
Suppose for some step τ we have 0.001s3 < mi,τ
< 0.01s. We consider two cases:
– If ai,τ ≥ 0.25s, then by Lemma 21 we have:
(+1)
|mi,t
(−1)
− mi,t | ≤ 0.1ai,t + 0.05s,
Ω(1)
with probability 1−e−n . In case of failure we interrupt the analysis (this is an implicit ap(+1)
plication of union bounds over successive steps τ ). Under the assumption mi,τ < 0.01sr,
we conclude:
(−1)
(+1)
mi,t ≤ 0.1ai,t + 0.05s + mi,t < 0.3ai,t ,
and hence:
(−1)
mi,t
1 (0)
≤ mi,t − 0.005s.
2
41
(36)
(+1)
Now, to compute the expected value ∆mi,τ , we remark that rule (6) does not decrease this
(0)
expected value since mi,τ . Moreover, in view of (36), the probability of executing rule (9)
(−1)
(which increases mi,t
by 1/n) exceeds the probability of executing one of the rules (7)
(−1)
or (8) (which decrease mi,t
(+1)
by 1/n) by 0.005smi,τ r > 5 · 10− 6s4 r by the assumption
(+1)
0.001s3 < mi,τ . We eventually obtain in this case:
(+1)
(+1)
E[∆mi,τ |(0.001s3 < mi,τ
(+1)
– If ai,τ ≥ 0.25s, then assuming mi,τ
the first phase to obtain:
(+1)
< 0.01s)] >
5 · 10− 6s4 r
.
n
(37)
< 0.01s, we can perform an analogous analysis as in
(+1)
E[∆mi,τ |(mi,τ
< 0.01s)] >
0.06s2
,
n
which, in particular, also implies (37).
(+1)
(+1)
We have thus shown that the expected change to mi,τ satisfies (37). Noting that initially mi,t2 >
0.004s3 = 0.001s3 + 0.003s3 , an application of Azuma’s inequality to an appropriate Doob
(+1)
martingale with (37) shows that the event mi,τ > 0.001s3 will hold for all remaining steps of
Ω(1)
the process τ , with probability 1 − e−n
.
Lemma 23. Let t be any moment of time with amin,t > 4r1/2 . For all i ∈ {1, 2, 3}, we have:
(+1)
min{mt
with probability 1 − O(e−n
Ω(1)
(−1)
, mt
} ≥ r3/2 s
).
Proof. Denote c = 4r1/2 . To show the claim, observe that necessarily for all τ ∈ [t − cn/2, t] we
(+1)
(−1)
have amin,τ > c/2. We consider the change of value mτ
over time (the argument for mτ
follows
(+1)
(+1)
symmetrically). Initially, we have mt−cn/2 ≥ 0, and at every step |∆mτ | ≤ 1/n. At any time τ such
(+1)
that mτ
< s/4 we have the following cases:
(+1)
• Rule (6) is executed, which happens with probability at least a2min,τ > c2 /4. Since mτ
conditioned on this event, the expected value of
(+1)
∆mτ
< s/4,
is at least 1/4.
• One of the rules (7)-(10) is executed, which occurs with probability at most r.
(+1)
• In all other cases, we have ∆mτ
= 0.
Noting that r = c2 /16, the claim follows from a standard application of Azuma’s inequality.
Lemma 24. Suppose 0 < x < c12 . Let t ≥ 2c11 n log n be an arbitrary moment of time. Then,
min{m+1,t , m−1,t } > c15 , with probability 1 − O(1/n), for some absolute constant c15 > 0 depending
only on s, p, and r.
42
Proof. Assume w.l.o.g. t = 2c11 n ln n. Instead of analyzing the evolution of the real system, we consider an execution of a system which is coupled with it over the first t steps as follows. First, starting
from time 0, we perform t steps of protocol Po (i.e., considering only rules (1)−(5) of its definition, and
without setting values of the second component M? ). Next, we once again activate the pairs of elements
which were activated in the first part of the coupling, in the same order, applying rules (6) − (10) of the
protocol with the same outcome which they would have received in the original execution. Clearly, at
time t the same configuration u(t) is reached by both the original and coupled execution.
Consider first the execution of Po from time 0. If amin,0 < 10−6 s4 , then the execution is in an
oscillatory interval at time 0, and will remain in it (amin < 0.02s2 ) until time t = c ln n with probability
Ω(1)
1 − e−n . Then, for all τ ∈ [0, t] we assume that the claim of Lemma 19 holds with t0 = τ for all
i ∈ {1, 2, 3}. By a crude union bound, this event holds with probability 1 − O(1/n); from now on we
assume this is true. (Formally, to allow us to proceed, in the analysis we can say with implicitly couple
the system with a different set of random choices to which the system switches in the low-probability
event that the claim of Lemma 19 were not to hold for some t0 = τ for the original system.) Given that
in the claim of Lemma 19 for t0 = 0 we have t∗∗ < t, and for t0 = t we have t∗ ≥ t, by the properties
of the time intervals [t∗ , t∗∗ ] we observe that there must exist a time τ ∈ [0, t] and a type i ∈ {1, 2, 3}
such that Ai is the least represented type at time τ and the most represented type at time t (amax,t = ai,t
and amin,τ = ai,τ ), and moreover t ∈ [t∗ , t∗∗ ] in the claim of Lemma 19 with choice of t0 = τ . Since
amax,t ≥ s/3, we now apply Lemma 22 with t0 = τ to obtain the claim, noting that amin,τ < 0.02s2
1
and moreover that t < τ + 0.005n
ln max{ 11,a } , given that we have t < τ + c11 n ln max{1/n,a
,
r
i+1,t }
n
i,t0
0
c1 1
where c11 is a constant depending only on s and p, and noting that we can choose r < 0.005
.
It remains to consider the cases when the execution starts at time 0 with amin,0 ≥ 10−6 s4 . Then, if
amin,t ≥ 10−6 s4 holds, the claim follows from Lemma 23 given that r is chosen so that 10−6 s4 > 4r1/2 .
Otherwise, there must exist some last time t0 ≤ t such that amin,t0 ≥ 10−6 s4 . We apply Lemma 19 with
t0 = t0 . If the obtained value t∗∗ satisfies t∗∗ < t, then we can apply an analogous analysis as in the
case amin,0 < 10−6 s4 to obtain the claim. Otherwise, we have that t ≤ t∗∗ ≤ t + cn, where the value of
constant c, depending only on s and p follows from Lemma 19. By an iterated application of Lemma 18,
we obtain that amin,t ≥ c0 , where the value of constant c0 > 0, depending only on s and p, follows from
the application of Lemma 18. Choosing r sufficiently small so that c0 > 4r1/2 , we complete the proof
using Lemma 23.
Finally, for the sake of completeness we state how the majority protocol stops in the case of #X = 0.
Lemma 25. Suppose x = 0. Then, there exists a moment of time ts such that either m+1,t = 0 or
m−1,t = 0 holds for all t > ts . Moreover, ts < c16 n log2 n with probability 1 − O(1/n), for some
absolute constant c16 > 0 depending only on s, p, and r.
Proof. By Theorem 1(1), there exists a moment of time t0 = O(n log2 n) such that the system reaches a
corner configuration (cf. Lemma 11). W.l.o.g., assume that a1 = s and a2 = a3 = 0. At this point, in the
majority protocol Rule (7) will never again be activated, whereas the execution of rules (8)-(11) follows
precisely the classical majority scenario of Angluin et al. [7]. By a standard concentration analysis (see
also [7]), one of the two species M+1 , M−1 will become extinct in O(n log2 n) steps with probability
1 − O(1/n).
As a side remark on Lemma 25 that it is possible to initialize the system entirely with a state (Ai , M0 )
so that m+1,t = m−1,t = 0 holds throughout the process (even if the designed protocol will never enter
such a configuration from most initial configurations).
43
7.3
Protocol Extension Pl : Detection with Lights
To complete the proof of Theorem 3, we design a protocol extension Pd , such that D ETECTION is solved
by the composition ((Po ◦ Pm ) + Pl ). Extension Pl , uses three states, {L−1 , L+1 , Lon }. We informally
refer to states L as lights. The composition is given in Fig. 5. Informally, state L−1 means that the agent
is “waiting for meeting M−1 ”, then after meeting M−1 it becomes L+1 , “waiting for M+1 ” and finally
it becomes Lon .
To analyze the operation of the protocol, consider first the case of x = 0. By Lemma 25, after
O(n log2 n) steps, agents in at least one of the states {M+1 , M−1 } are permanently eliminated from
the system. Thus, either rule (11) or rule (12) will never again be executed in the future. An agent
1
which is in state Lon will spontaneously move to another state following rule (13) within O( q(ε)
n log n)
2
steps, with probability 1 − O(1/n ), and will never reenter such a state, since this would require the
activation of both rule (11) and rule (12). By applying a union bound over all agents, we obtain that state
Lon never again appears in the population after O(n log n) steps from the termination of the majority
protocol, with probability 1 − O(1/n). Overall, all nodes reach a state having a different state than Lon
after O(n log2 n) steps from the start of the process, with probability 1 − O(1/n), and all leave such a
state eventually with certainty.
In the presence of the source X, the analysis of the process can be coupled with a Markov chain on
three states L−1 , L+1 , and Lon . In view of Lemma 24, transitions from state L−1 to L+1 and from state
L+1 to to L−1 occur with at least constant probability (except for an O(1/n)-fraction of all time steps),
this 3-state chain is readily shown to be rapidly mixing. For a choice of q(ε) > 0 depending only on
s, p, r, ε sufficiently small, we can lower-bound the number of agents occupying state Lon by (1 − ε)n,
with high probability.
Under the natural decoding of states as “informed” (having component Lon ) or “uninformed” (having component L−1 or L+1 ), the proof of Theorem 3 is complete. We remark that it is also possible to
design a related protocol in which exactly one state is recognized as “informed” and exactly one state is
recognized as “uninformed”; we omit the details of the construction.
8
Proof of Impossibility Result
This Section is devoted to the proof of Theorem 4. First, we restate some notation. We recall that
the vector z = (z (1) , . . . , z (k) ) ∈ {0, 1, . . . , n}k = Z describes the number of agents having particular states, and kzk1 = n. In this section we will identify the set of states with {1, . . . , k} = [1, k].
It is now also more convenient for us to work with a scheduler which selects unordered (rather than
ordered) pairs of interacting agents; we note that both models are completely equivalent in terms of
computing power under a fair random scheduler, since selecting an ordered pair of agents can be
seen as selecting an unordered pair, and then setting their orientation through a coin toss. Indexing with integers {1, 2, . . . , r} the set of all distinct rules of the protocol, where r ≤ k 4 , for a rule
j ≡ “{i1 (j), i2 (j)} → {o1 (j), o2 (j)}00 , 1 ≤ j ≤ r, i1 (j), i2 (j), o1 (j), o2 (j) ∈ {1, . . . , k}, we will
denote by qj , the probability (selected by the protocol designer) that rule j is executed as the next interaction rule once the scheduler has selected (i1 , i2 ) as the interacting pair, and by pj (z) the probability
(i (j)) (i (j))
that j is the next rule chosen in configuration z (we have pj (z) = qj z 1 nz2 2 (1−O(1/n)), where the
O(1/n) factor compensates the property of a scheduler which always selects a distinct pair of elements).
For any configuration z0 ∈ Z, we define the d-box Bd (z0 ) around z0 as the set of all states z ∈ Z
(i)
(i)
such that z0 /d ≤ z (i) ≤ d max{1, z0 }, for all 1 ≤ i ≤ k. We start the proof with the following
property of boxes.
Lemma 26. Fix k ∈ Z+ and let 0 < ε1 < 0.001 be arbitrarily fixed. There exists ε0 = ε0 (k, ε1 ),
0 < ε0 < ε1 , such that, for any interaction protocol P with k states and any configuration z0 ∈ Z, there
44
exists a value ε = ε(P, z0 ) ∈ [ε0 , ε1 ] such that, for any rule j of the protocol, 1 ≤ j ≤ r, exactly one of
the following bounds holds:
1. (i) for all z ∈ Bnε0 (z0 ), pj (z) ≤ nε−1 ,
2. (ii) for all z ∈ Bnε0 (z0 ), pj (z) ≥ n24ε−1 .
and for any state i, 1 ≤ i ≤ k, exactly one of the following bounds holds:
1. (iii) for all z ∈ Bnε0 (z0 ), z (i) ≤ nε ,
2. (iv) for all z ∈ Bnε0 (z0 ), z (i) ≥ n24ε .
4
Proof. Let k be fixed and let ε0 = 96−(k+k +f +1) ≤ 96−(k+r+f +1) , where f = log2 (1/ε1 ). Consider
(i)
the (multi)set M of real values M := {logn max{nε0 , z0 } : i ∈ {1, . . . , r}}∪{logn max{nε0 , npj (z0 )} :
j ∈ {1, . . . , r}} ⊆ [0, 1]. Since |M | = k + r, by the pigeonhole principle, there must exist an interval Il = [96−l , 96−l+1 ), for some l ∈ {f, . . . , k + r + f }, such that Il ∩ M = ∅. Now, we set
ε = 2 · 96−l > 96ε0 , we also have ε < 2 · 96−f < ε1 . We immediately obtain that for any state
(i)
(i)
i, 1 ≤ i ≤ k, we either have z0 ≤ nε/2 or z0 ≥ n48ε . Recalling that for any z ∈ Bnε0 (z0 ),
(i)
(i)
z0 /nε0 ≤ z (i) ≤ nε0 max{1, z0 }, claims (iii) and (iv) follow.
(i (j)) (i (j))
To show claims (i) and (ii), notice that if rule j, j ∈ {1, . . . , r} is such that min{z0 1 , z0 2 } ≤
nε , then for all z ∈ Bnε0 (z0 ) we have min{z (i1 (j)) , z (i2 (j)) } ≤ nε (by (iii) and (iv)), and so pj (z) ≤
(i (j)) (i (j))
nε−1 by the properties of the random scheduler. Otherwise, we have min{z0 1 , z0 2 } ≥ n24ε , and
1
so 2n2ε0 ≤ pj (z0 )/pj (z) ≤ 2n2ε0 , where we recall that ε > 96ε0 . Since we have pj (z0 ) ≤ nε/2 or
pj (z0 ) ≥ n48ε , claims (i) and (ii) follow.
Given any k-state protocol P , we will arbitrarily choose a value of ε for which the claim of the above
Lemma holds (e.g., the smallest possible such value of ε). Note that a similar analysis is also possible
for protocols using a super-constant number of states in n, however, then the value of ε0 is dependent on
4
n; retracing the arguments in the proof, we can choose appropriately ε0 ≥ nexp[−O(k )] . (We make no
effort to optimize the polynomial in k in the exponent.)
In what follows, let z0 be a fixed configuration of the protocol (admitting a certain property which
we will define later). We will then consider a rule j to be a low probability (LP) rule (writing j ∈ LP )
in box Bnε0 (z0 ) if it satisfies condition (i) of the Lemma, and a high probability (HP) rule in this box
(writing j ∈ HP ) if it satisfies condition (ii). Note that LP ∪ HP = {1, . . . , r}.
Likewise, for 1 ≤ i ≤ k, we will classify i as a low-representation (LR) state (writing i ∈ LR) in
box Bnε0 (z0 ) if i satisfies condition (iii) of the Lemma, and a high representation (HR) state (writing
i ∈ HR) in this box if i satisfies condition (iv). Note that LR ∪ HR = {1, . . . , k}. Moreover, we
define a set of very high representation (VHR) states, V HR ⊆ HR, as the set of all i such that for all
z 0 ∈ Bnε0 (z0 ), zi0 ≥ n1−8ε . Denoting HR0 = HR \ V HR, we have by the definition of a box that for
all i0 ∈ HR0 , for all z 0 ∈ Bnε0 (z0 ): zi0 ≤ n1−8ε /O(n2ε0 ) < n1−6ε .
From now on, we assume that configuration z0 admits the following property: for T = n1+2ε , an
execution of the protocol starting from configuration z0 passes through a sequence of configurations zt ,
t = 1, 2, . . . , T , such that the configuration does not leave the box around z0 in any step with sufficiently
large probability, lower-bounded by some absolute constant Π ∈ (0, 1]:
Pr[∀t<T zt ∈ B ⊆ Bnε0 (z0 )] ≥ Π,
(38)
where B is an arbitrarily fixed subset of Bnε0 (z0 ).
We now show that the above property has the following crucial implication: for an interacting pair
involving selected high and very high representation states, a rule creating a low representation state can
only be triggered with sufficiently small probability. Informally, it seldom happens that in the protocol
a low representation state is created out of any high representation state.
45
Lemma 27. For a protocol having the property given by Eq. (38), for i1 ∈ HR and i2 ∈ V HR, let
of the form {i1 , i2 } → {o1 , o2 }, taken over all o1 ∈ LR, o2 ∈ [1, r]. Then,
R
Pi1 ,i2 be the set of rules
−14ε ).
q
=
O(n
j
j∈Ri1 ,i2
P
Proof. Suppose, by contradiction, that j∈Ri ,i qj > 3n−14ε .
1 2
Associate with process zt a random variable Jt ∈ {0, 1}, defined as follows. For all t < te , where
te is the first moment of time such that zte ∈
/ Bnε0 (z0 ), we put Jt = 1 if a rule from Ri1 ,i2 is used for
the interaction made by the protocol in process zt at time t, and set Jt = 0 otherwise. For all t ≥ te , we
set Jt to 1. We have E(Jt |z1 , . . . , zt ) ≥ 2n2ε−1 ; indeed, for t < te , it holds that:
E(Jt |z1 , . . . , zt ) = Pr[Jt = 1|zt ] =
X
X
pj (zt ) =
j∈Ri1 ,i2
≥ 3n−14ε
j∈Ri1 ,i2
qj
z (i1 (j)) z (i2 (j))
(1 − O(1/n)) ≥
n2
n24ε n1−8ε
(1 − O(1/n)) > 2n2ε−1 .
n2
By a simple stochastic domination argument, (Jt ) can be lower-bounded by a sequence of independent
binomial trials with success probability n2ε−1 , hence by an application of a multiplicative Chernoff
bound for T = n1+2ε :
" T
#
" T
#
X
X
1
Pr
Jt > n4ε = Pr
Jt > 2n2ε−1 T = 1 − o(1),
2
t=1
t=1
where the o(1) factor is exponentially small in n.
We now show the following claim.
Claim. With probability Π − o(1), the following event holds: zt ∈ B for all t ∈ [0, T ) and the total
number of rule activations in the time interval [0, t) during which an agent changes state from a state in
LR to a different state is at most O(kn3ε ).
Proof (of claim). Acting similarly as before, we associate with process zt a random variable Lt ∈
{0, 1}, defined as follows. For all t < te , we put Lt = 1 if a rule acting on at least one agent in a state
from LR is made by the protocol in process zt at time t, and set Lt = 0 otherwise. For all t ≥ te , we set
Lt to a dummy variable set always to 0, i.a.r. We observe that:
nε
,
n
since |LR| ≤ k, and for t < te , zi < nε for any i ∈ LR, hence the scheduler selects an agent from a
ε
LR state into an interacting pair with probability at most 2k nn . Applying an analogous argument as in
the case of random variable Lt , this time for the upper tail, we obtain:
" T
#
" T
#
X
X
3ε
ε
Pr
Lt < 4kn
= Pr
Lt > 2 · 2kn T = 1 − o(1).
E(Lt |z1 , . . . , zt ) ≤ 2k
t=1
t=1
The claim follows directly.
Now, by a union bound we obtain:
" T
#
T
X
X
Pr
Jt > n4ε ∧
Lt < 4kn3ε = 1 − o(1).
t=1
t=1
Taking into account that te > T holds with probability Π = Ω(1) by (38), we have by a union bound
that with probability at least Π − o(1) = Ω(1), the
holds: zt ∈ B for all t ∈ [0, T ],
P
PTfollowingPevent
T
T
4ε , and L < 4kn3ε . However then
4ε
3ε > knε , so there
J
>
n
J
−
L
t
t=1 t
t=1 t
t=1 t > n − kn
(i)
must exist at time T a state i ∈ LR such that zT > nε . This is a contradiction with zT ∈ B ⊆ Bnε0 (z0 )
by Lemma 26(iii).
46
In the rest of the proof, we consider the evolution of a protocol starting from configuration z0 and
having property (38). We compare this evolution to the evolution of the same protocol, starting from a
perturbed configuration z0∗ , such that:
(C1) kz0 − z0∗ k1 ≤ nε .
∗(i)
(C2) for all low representation states i ∈ LR, we have z0
(i)
≤ z0 .
Intuitively, the perturbed state z0∗ may correspond to removing a small number of agents from z0 (and
replacing them by high representation states for the sake of normalization), e.g., as in the case of the
disappearance of a rumor source from a system which has already performed a rumor-spreading process.
Our objective will be to show that, with probability at least Π − o(1), after T = n1+2ε the process
∗
zT is still not far from z0 , being constrained to a box in a similar way as process zt . To achieve this, we
define a coupling between processes zt and zt∗ (knowing that process zt is constrained to a box around
z0 with probability Π). Informally, the analysis proceeds as follows. We run the processes together for
T = n1+2ε steps. In most steps, the 1-norm distance kzt − zt∗ k1 between the two processes remains
unchanged, without exceeding O(n3ε ). Otherwise, exactly one of the two processes executes a rule (and
the other pauses). With a frequency of roughly nε /n steps (i.e., roughly n3ε times in total during the
process), an LP rule is executed which increases the distance between these two states. We think of this
type of “error” as unfixable, contributing to the O(n3ε ) distance of the processes; however, such errors
are relatively uncommon. With a higher frequency of roughly n3ε /n steps (i.e., roughly once every
n1−3ε steps), a less serious “error” occurs, when some HP rule ι increases the distance between the two
states. The rate of such errors is too high to leave them unfixed, and we have a time window of about
n1−3ε steps to fix such an error (before the next such error occurs). We observe that since ι is an HP
rule, which is activated with probability at least n24ε−1 , rule ι will still be activated frequently during
this time window. The coupling of transitions of states zt and zt∗ is in this case performed so as to force
the two processes to execute rule ι lazily, never at the same time. The number of executions of rule
ι in the ensuing time window by each of the two processes follows the standard coupling pattern of a
pair of lazy random walks on a line, initially located at distance 1, until their next meeting (cf. e.g. [2]).
During this part of the coupling, we allow the distance kzt − zt∗ k1 to increase even up to n6ε (as a
result of executions of rule ι), but the entire contribution to the distance related to rule ι is reduced to 0
before the next HP rule “error” occurs, with sufficiently high probability (in this case, with probability
1 − O(n−6ε ). Overall, the coupling is successful with probability Π − O(n−ε ).
We remark that we use the bound on the number of states k to enforce a sufficiently large polynomial
separation between the frequencies of LR states and HR states, and likewise for LP rules and HP rules.
We also implicitly assume that k = no(1) , throughout the process. The analysis also works for a choice
of k = O(log log n), with a sufficiently small hidden constant. The separation between LR/HR states
and LP/HP rules is used in at least two places in the proof. First, it enforces that rules creating LR states
from VHR states may appear in the definition of the protocol only with polynomially small probability
∗(i)
(i)
(Lemma 27), which helps to maintain over time the invariant zt ≤ zt , for all LR states. Secondly,
we use the separation of LP/HP rules in the analysis of the coupling to show that a fixable “error” caused
by a HP rule can be sufficiently quickly repaired, before new errors occur.
In the formalization of the coupling, we make both processes zt and zt∗ lazy, i.e., add to each process
an additional independent coin-toss at each step, and enforce that with probability 1/2 no rule is executed
in a given step (i.e., the step is skipped by the protocol). We assume a random scheduler which picks
uniformly a random pair of nodes at each step. Thus, if the scheduler picks a pair of agents in states
{i1 , i2 }, and j is a rule acting on this pair of states, the probability that the interaction corresponding to
rule j will be qj /2. (The laziness of the process here is a purely technical assumption for the analysis,
and corresponds to using a measure of time which is scaled by a factor of 2 ± o(1) w.h.p.; this does not
affect the asymptotic statement of the theorem.)
47
We will also find it convenient to apply an auxiliary notation for representing the evolution of a state.
For process zt (resp., zt∗ , we define ρt (j) (resp. ρ∗t (j)), for all j ∈ [1, r], as the number of times rule j
has been executed since time 0. Observe that the pair (z0 , (ρt (j) : j ∈ [1, t]) completely describes the
evolution of a state (i.e., the order in which the rules were executed is irrelevant). Moreover, since each
execution of a rule changes the states of at most 4 agents, we have:
kzt −
zt∗ k1
≤4
r
X
|ρt (j) −
ρ∗t (j)|
+ kz0 −
j=1
z0∗ k1
≤ +4
r
X
|ρt (j) − ρ∗t (j)| + nε .
j=1
Definition of the coupling.
1. At each step t, we order the agents of configurations zt and zt∗ , so that al (t) denotes the type
of the l-th agent in zt and a∗l (t) is the type of the l-th agent in zt∗ . The orderings are such that
|{l : al (t) = a∗l (t)}| is maximized; in particular, for any state i such that z (i) (t) ≤ z ∗(i) (t)
(respectively, z ∗(i) (t) ≤ z (i) (t)) we have that if for some l, al (t) = i (resp., a∗l (t) = i), then
a∗l (t) = i (resp., al (t) = i).
2. The scheduler then picks a pair of distinct indices l1 , l2 ∈ {1, . . . , n} as the pair of interacting
agents.
2.1. If al1 (t) = a∗l1 (t) and al2 (t) = a∗l2 (t), then the same rule j = j ∗ acting on the pair of states
(al1 (t), al2 (t)) is chosen as the current interaction rule, with probability qj .
2.2. Otherwise, a pair of (clearly distinct) rules j and j ∗ are picked independently at random for
zt and zt∗ from among the rules available for state pairs (al1 (t), al2 (t)) and (a∗l1 (t), a∗l2 (t)),
with probabilities qj and qj ∗ , respectively.
3. The processes finally perform their coin tosses to decide which of the selected rules (j for zt and
j ∗ for zt∗ ) will be applied in the current step.
3.1. If j = j ∗ and rule j has been executed exactly the same number of times in the history of
the two processes (ρt (j) = ρ∗t (j)), then with probability 1/2 both of the processes execute
rule j, and with probability 1/2 neither execute their rule.
3.2. If j 6= j ∗ , or if j = j ∗ and rule j has been executed a different number of times in the history
of the two processes (ρt (j) 6= ρ∗t (j)), then exactly one of the two processes performs its
chosen rule and the other process waits, with the process performing the rule being chosen
as zt or zt∗ , with probability 1/2 each.
The correctness of the coupling (i.e., that the marginals zt and zt∗ each correspond to a valid execution
of the given protocol under a random scheduler) is immediate to verify.
Lemma 28. Let zt be a process satisfying property (38), and let z0∗ satisfy conditions (C1) and (C2).
Then, for T = n1+2ε , with probability Π − O(n−ε ) we have kzT∗ − zk1 = O(n6ε ), for some z ∈ B.
Proof. To prove the claim, it suffices to show that with probability Π − O(n−ε ) the provided coupling
succeeds, i.e., it maintains a sufficiently small difference zT (i) − zT∗ (i) for all states i, with zT ∈ B.
In the analysis of the provided coupling, we will assume that the box condition zt ∈ B holds always
throughout the process (otherwise, we assume the coupling does not succeed). To state this formally, we
work with auxiliary processes z̄t and z̄t∗ , given as z̄t = zt and z̄t∗ = zt∗ for all t < te , where te is the first
moment of time such that zt ∈
/ Bnε0 (z0 ), and set to the dummy value z̄t = z̄t∗ = z0 for all t ≥ te . At the
end of the process, we will thus have z̄T = zT and z̄T∗ = zT∗ with probability at least Π. In the following,
we silently assume that t < te − 1 (in particular, that zt ∈ B and zt+1 ∈ B), and we will simply show
48
that the coupling of z̄t and z̄t∗ is successful with probability 1 − n−ε . The condition of t ≥ te − 1 is
trivially handled.
In addition to the box condition (which is now enforced) we try to maintain, with sufficiently high
probability, throughout the first T steps of the process several invariants (all at a time), corresponding to
the following events holding:
∗(i)
• FD (t): for all states i ∈ LR, z̄t
∗(i)
• FLR (t): for all states i ∈ LR, z̄t
(i)
≤ z̄t . (LR domination condition)
(i)
≤ z̄t ≤ nε . (LR state condition)
• FLP (t): for all rules j ∈ LP , max{pj (z̄t ), pj (z̄t∗ )} ≤ 2nε−1 . (LP rule condition)
(i)
∗(i)
• FHR (t): for all states i ∈ HR, min{z̄t , z̄t
} ≥ n24ε /2. (HR state condition)
• FHP (t): for all rules j ∈ HP , min{pj (z̄t ), pj (z̄t∗ )} ≥ n24ε−1 /2. (HP rule condition)
(i)
∗(i)
• FHR0 (t): for all states i ∈ HR0 , max{z̄t , z̄t
} ≤ 2n1−6ε . (HR’ state condition)
• a family of possible events Sw,d (t), for some d ∈ {0, . . . , 4n3ε } and w ∈ {0, . . . , n6ε }, with
specific events defined as follows:
P
– S0,d (t) holds if for all rules j ∈ HP we have ρt (j) = ρ∗t (j), and j∈LP |ρt (j)−ρ∗t (j)| = d.
This implies, in particular, kz̄t∗ − z̄t k1 ≤ 4d + nε ≤ 5n3ε . (identical rate of HP execution)
– Sw,d (t) for w > 0 holds if there exists a rule ι ∈ HP suchP
that for all rules j ∈ HP \ {ι}
∗
∗
we have ρt (j) = ρt (j), |ρt (ι) − ρt (ι)| = w, and moreover i∈LP |ρt (i) − ρ∗t (i)| = d. This
implies, in particular, kz̄t∗ − z̄t k1 ≤ 4d + 4w + nε ≤ 5n6ε . (single HP execution difference)
We will call the coupling successful if for all t ≤ T , all events F· (t) and some event Sw,d (t) holds, and
we will say it is a failure otherwise. (We remark that condition FD (t) is implied by condition FLR (t),
but we retain both for convenience in discussion.)
The analysis of the coupled process is now the following. First, we remark that all of the given
events F· (t) and event S0,0 (t) hold for t = 0.
If the process meets condition S0,d at time t and all conditions F· (t), then we have the following:
• With probability at least 1−O(n3ε−1 ), the coupling will follow clauses 2.1 and 3.1 of its definition,
and the two processes z̄ and z̄ ∗ will execute the same rule j (or both pause). Hence, we continue
to step t + 1 satisfying condition S0,d and all of the conditions F· (t + 1), making use of the box
condition for process z̄t . (We note that, to show FLP (t), when considering the special case of a
rule involving a state from LR, we can make use of FLR (t) and note that the activation probability
of such a rule is bounded by 2nε−1 due to the nε bound on the population of a LR state).
• With probability at most O(n3ε−1 ), the coupling will, however, select distinct rules, j for z̄t and
j ∗ for z̄t∗ , and will select exactly one of them to execute, say j 0 ∈ {j, j ∗ }.
– If j 0 ∈ LP , which happens in the current step of the process with probability at most 2nε−1
by FLP , then the event S0,d+1 (t + 1) will hold in the next step (provided d + 1 ≤ 4n3ε ;
otherwise, if d + 1 > 4n3ε , we will say that the coupling has failed).
– If j 0 ∈ HP , which happens in the coupling with probability O(n3ε−1 ) (as bounded due to
clause 2.2), then the event S1,d (t + 1) will hold in the next step. The condition FD (t +
1) requires more careful consideration. Taking into account that FD (t) holds, we need to
consider two cases: either j 0 = j and the rule applied to z̄t changed at least one of the
two interacting states {i1 (j), i2 (j)}, say i1 (j) ∈ LR, so that z̄ i1 (j) (t) = z̄ ∗i1 (j) (t) and
49
failure
failure
failure
O(n−2.9ε−1 )
S0,0
S1,0
S1,1
S0,4n3ε
S0,d
S0,1
O(n3ε−1 )
1 − O(n3ε−1 )
failure
Pr ≤ 1
S1,d
failure
S2,d
failure
Sw,d
Π0t ≥
Π0t
n24ε−1
4
failure
O(n6ε−1 )
Sn6ε ,d
failure
Pr ≤ 1
Figure 8: Illustration of transitions between states Sw,d for the coupling in the lower bound proof.
z̄ i1 (j) (t + 1) ≤ z̄ ∗i1 (j) (t + 1) − 1, or j 0 = j ∗ and the rule applied to z̄t∗ created a pair of
states {o1 (j ∗ ), o2 (j ∗ )}, say o1 (j ∗ ) ∈ LR. In the first case, by the description of the ordering
given in clause 1 of the definition of the coupling, the problem occurs only if one of the
agents picked by the scheduler belongs to an LR state, and the other agent is at a position
in which the states of z̄ and z̄ ∗ differ in the ordering of the agents; hence, the probability
ε 3ε
that the coupling fails at this step is at most O( knn·n
) ≤ O(n5ε−2 ). In the second case,
2
we likewise analyze the ordering of the agents considered by the scheduler, and note that the
interacting agent, which belongs to the part of the ordering in which z̄t and z̄t∗ differ, must be
in a HR state, since the agents in a LR state in z̄ ∗ are matched by their counterparts in z̄ (as
noted in clause 1 of the discussion of the coupling). If the other interacting agent is in a state
3ε
1−6ε
from LR ∪ HR0 , then such an event occurs with probability O( n ·kn
) ≤ O(n−2.9ε−1 ),
n2
and we say that with this probability the coupling has failed. Finally, if the other interacting
agent is in a state from V HR, then by Lemma 27, we have that the probability of picking
a rule under which the coupling fails is at most O(n−14ε ), conditioned on the event j 6= j ∗
holding, hence overall the probability of failure is O(n−14ε n3ε−1 ) = O(n−11ε−1 ). Overall,
we obtain that FD (t + 1) holds with probability O(n−2.9ε−1 ). Given FD (t + 1), S1,d(t+1) ,
and the box condition, the remaining conditions F· (t + 1) follow directly.
Overall, we obtain that following a time t satisfying S0,d (t) and all conditions F· (t), we reach the
following successor state (see Fig. 8):
50
S0,d (t + 1) ∧ F· (t + 1),
S0,d+1 (t + 1) ∧ F· (t + 1),
S1,d (t + 1) ∧ F· (t + 1),
failure:
with probability 1 − O(n3ε−1 ),
with probability ≤ 2nε−1 , if d + 1 ≤ 4n3ε ,
3ε−1 ),
with
( probability O(n
with probability O(n−2.9ε−1 ), if d + 1 ≤ 4n3ε ,
with some probability ≤ 1,
otherwise.
At this point, before proceeding further, we can provide some intuition on the meaning of the respective
events S. The coupling process can be seen as a walk along the path (S0,d : d ≤ 4n3ε ), starting from
state S0,0 , and at each step, either staying in the current state S0,d , moving on to the next state S0,d+1 ,
branching to a side branch S1,d (which we will analyze later), or failing. The process also fails if it
reaches the endpoint of its path (d = 4n3ε ). Since the process is run for T = n1+2ε steps, the probability
that failure will occur before the end of the path is reached is O(n−0.9ε ), and the probability of reaching
the end of the path and failing is exponentially small in nε by a Chernoff bound (in expectation, the
process will progress halfway along the path). Hence, we have that the process succeeds with probability
1 − O(n−0.9ε ), or otherwise may fail in a side branch S·,d .
A side branch is entered with probability O(nε−1 ). To show that the coupling succeeds with the
required probability, it suffices to show that we return from any state S1,d to state S0,d with probability
at least 1 − O(n−6ε ); then, all (i.e., w.h.p. at most O(n1+2ε n3ε−1 ) = O(n5ε )) excursions into side
branches during the process will succeed with probability 1 − O(n−ε ).
Consider now an excursion into a side branch Sw,d (w ≥ 1) associated with a rule ι ∈ HP , which
has been executed a different number of times in z̄t and z̄t∗ . Now, if the process meets condition Sw,d at
time t and all conditions F· (t), then we have the following:
• With probability at least 1 − O(n6ε−1 ), the coupling will follow clause 2.1 of its definition, selecting a single rule j.
– If j 6= ι, then clause 3.1 will follow, and the two processes z̄ and z̄ ∗ will execute the same
rule j (or both pause). Hence, at time t + 1, all of the conditions F· (t + 1) and condition
Sw,d (t + 1) is satisfied.
– Else, the event j = ι occurs. The probability of such an event is denoted πt ∈ [pι (z̄t ) −
O(n6ε−1 ), pι (z̄t )] (due to the conditioning performed in the first clause of the coupling);
since pι (z̄t ) ≥ n24ε−1 by the box condition for HP rules, it follows that 2πt ≥ n24ε−1 −
O(n6ε−1 ) ≥ n24ε−1 /2. Now, following clause 3.2 of the coupling, depending on which of
the two processes z̄t , z̄t∗ is chosen to execute the rule, with probability πt /2 =: πt0 the system
moves to Sw−1,d (t + 1), and with probability πt0 the system moves to Sw+1,d (t + 1) (unless
w + 1 > n6ε , in which case the coupling has failed). As before, given there was no failure,
all conditions F· (t + 1) are readily verified to be satisfied in the new time step.
• With probability at most O(n6ε−1 ), for simplicity of analysis we assume the coupling has failed.
This time, for a time t satisfying Sw,d (t)
distribution of successor states:
Sw−1,d (t + 1) ∧ F· (t + 1),
Sw+1,d (t + 1) ∧ F· (t + 1),
failure:
Sw,d (t + 1) ∧ F· (t + 1),
for w ≥ 1 and all conditions F· (t), we obtain the following
with probability exactly πt0 ≥ n24ε−1 /4,
0
6ε
with
( probability exactly πt , if w + 1 ≤ n ,
with probability O(n6ε−1 ), if w + 1 ≤ n6ε ,
with some probability ≤ 1, otherwise,
otherwise.
51
The picture here corresponds to a lazy random walk along the side line Sw,d for w ∈ [0, n6ε ], with an
additional failure probability at each step. The walk starts at w = 1 and ends with a return to the primary
line S0,d if the endpoint w = 0 is reached, or ends with failure if the other endpoint w ≥ n6ε =: wmax is
reached. At each step, the walk is lazy (with probability of transition depending on the current step), but
unbiased with respect to transitions to the left or to the right. Assuming that failure does not occur sooner,
1
2 ) = O(n12ε )
with probability 1 − O( wmax
) = 1 − O(n−6ε ) the walk will reach point w = 0 in O(wmax
moves (transitions along the line), without reaching the other endpoint of the line sooner. Since a move
is made in each step t with probability πt0 ≥ n24ε−1 /4, by a straightforward Chernoff bound, the number
of steps spent on this line is given w.h.p. as at most O(n12ε /n24ε−1 ) = O(n1−12ε ). As the probability
of failure in each of these steps is O(n6ε−1 ), the probability that the process fails during these steps
is O(n−6ε ). Overall, by a union bound, we obtain that the process successfully returns to S0,d with
probability 1 − O(n−6ε ) (and within O(n1−12ε ) steps). In view of the previous observations, we have
that with probability 1 − O(n−ε ), all conditions F· and some condition Sw,d hold at time T . Thus, with
probability Π − n−ε , process zT∗ is sufficiently close to B, i.e., there exists a point z ∈ B such that
kzT∗ − zk1 = O(n6ε ).
9
Proof of Proposition 2
Proof. Fix protocol P with set of states K, in which the minimum positive probability of executing
some rule is p. Let K 0 ⊆ K, K 0 3 X be any minimal subset of the set of states such that no evolution
of protocol P starting in a configuration containing only states from set K 0 will ever contain an agent in
a state outside K 0 . Denote κ = |K 0 | − 1. Consider an initialization of protocol P at time t0 = 0, at a
configuration z(0) with x ∈ [c, 1/2] and with all other states from K 0 represented by the same number
of agents, i.e., for each Q ∈ K 0 , we have q(0) = (1 − x)/κ.
Let t ≥ kn be an arbitrarily chosen time step. Let t1 = t − (κ − 1)n. Fix Q1 ∈ K 0 \ {X}
as any state such that q1 (t1 ) ≥ 1/2κ (we can, e.g., fix q1 as the state from K 0 \ {X} having the
most agents at time t1 ). Observe that from the minimality of K 0 it follows that there must exist a
sequence of states (Q1 , . . . , Qκ ), with {Q1 , . . . , Qκ = K 0 \ {X}, such that in the definition of protocol
P , for all i ∈ {1, . . . , κ − 1}, some rule of protocol P creates at least one agent (i.e., either 1 or 2
agents) in state Qi+1 from an interaction of either the pair of agents in states (Qj , Qi ) or the pair of
agents in states (Qi , Qj ), for some j ∈ {1, . . . , i}. (Indeed, if for some i there was no possibility of
choosing Qi+1 in any way, then K 00 = {X, Q1 , . . . Qi } ⊆ K 0 would be closed under agent creation,
contradicting the minimality of the choice of K 0 .) Now, we consider intervals of time steps [ts , ts+1 ],
with ts = t1 + (s − 1)n for s > 1. We make the following claims:
(1) Fix i ∈ {1, 2, . . . κ}. If qi (ts ) = Ω(1), then qi (t) ≥ 0.11qi (ts ) for all t ∈ [ts , ts+1 , with probaΩ
bility 1 − e−n (1) . Indeed, in a sequence of n steps, the expected number of agents which do not
participate in any interaction in the time interval [ts , ts+1 ] following the asynchronous scheduler
n
is ( n−2
n ) n > 0.13n, and thus the number of non-interacting agents is at least 0.12n, with probaΩ
bility 1 − e−n (1) following standard concentration bounds for the number of isolated vertices in
a random graph on n nodes with n edges. Since the choice of agents by the scheduler is independent of their state, and the probability for a uniformly random agent to be in state Qi at time ts
is qi (ts ), a simple concentration bound shows that 0.11qi (ts )n having state Qi at time ts do not
participate in any interaction in the interval [ts , ts+1 ].
(2) Fix i ∈ {1, 2, . . . κ − 1}. Denote mi = minj≤i qj (ti ). If mi = Ω(1), then qi+1 (ti+1 ) ≥ 0.01pm2i ,
Ω
with probability 1 − e−n (1) . Indeed, consider the value j ≤ i such that the interaction (Qj , Qi )
or (Qi , Qj ) creates an agent in state Qi+1 . At any time t within the interval [ti , ti+1 ], we have by
Ω
Claim (1) that qj (t) ≥ 0.11mi and qi (t) ≥ 0.11mi , with probability 1 − e−n (1) . It follows that
52
an interaction creating a new agent in state Qi+1 is triggered with probability at least p(0.11mi )2
at each step. The number of agents in state Qi+1 at time step ti+1 may thus be dominated from
below by the number of successes in a sequence of n Bernoulli trials with success probability
p(0.11mi )2 , and the claim follows.
By applying Claim (2) iteratively for i = {1, 2, . . . κ − 1}, where we note that m1 ≥ 1/2κ, we
i
Ω
have qi+1 (ti+1 ) ≥ (0.01p/2κ)2 , with probability 1 − e−n (1) (through successive union bounds).
i
Then, applying Claim (1) up to time t = tκ−1 , we have qi+1 (tκ−1 ) ≥ 0.11κ−1−i−1 (0.01p/2κ)2 ≥
κ
Ω
(0.01p/2κ)2 , with probability 1 − e−n (1) . Applying once again a union bound, we have shown that
κ
Ω
for all q ∈ K 0 , we have q(t) ≥ (0.01p/2κ)2 ≡ C0 , with probability 1 − e−n (1) . The claim of the
lemma follows for a suitable choice of δ0 > 0.
Acknowledgment. We sincerely thank Dan Alistarh and Przemek Uznański for inspiring discussions,
and Lucas Boczkowski for many detailed comments which helped to improve this manuscript.
References
[1] M. A. Abdullah and M. Draief. Majority consensus on random graphs of a given degree sequence.
CoRR, abs/1209.5025, 2012.
[2] D. Aldous and J. A. Fill. Reversible markov chains and random walks on graphs, 2002. Unfinished
monograph, recompiled 2014, available at http://www.stat.berkeley.edu/˜aldous/
RWG/book.html.
[3] D. Alistarh, J. Aspnes, D. Eisenstat, R. Gelashvili, and R. L. Rivest. Time-space trade-offs in population protocols. In Proc. Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2017, Barcelona, Spain, pages 2560–2579, 2017.
[4] D. Alistarh, J. Aspnes, and R. Gelashvili. Space-optimal majority in population protocols. (To
appear, SODA 2018.) CoRR, abs/1704.04947, 2017.
[5] D. Alistarh, B. Dudek, A. Kosowski, D. Soloveichik, and P. Uznanski. Robust detection in leakprone population protocols. In DNA, volume 10467 of Lecture Notes in Computer Science, pages
155–171. Springer, 2017.
[6] D. Angluin, J. Aspnes, Z. Diamadi, M. J. Fischer, and R. Peralta. Computation in networks of
passively mobile finite-state sensors. Distributed Computing, 18(4):235–253, 2006.
[7] D. Angluin, J. Aspnes, and D. Eisenstat. A simple population protocol for fast robust approximate
majority. Distributed Computing, 21(2):87–102, 2008.
[8] D. Angluin, J. Aspnes, D. Eisenstat, and E. Ruppert. The computational power of population
protocols. Distributed Computing, 20(4):279–304, 2007.
[9] J. Aspnes and E. Ruppert. An introduction to population protocols. Bulletin of the EATCS, 93:98–
117, 2007.
[10] L. Becchetti, A. E. F. Clementi, E. Natale, F. Pasquale, R. Silvestri, and L. Trevisan. Simple
dynamics for majority consensus. CoRR, abs/1310.2858, 2013.
[11] L. Boczkowski, A. Korman, and E. Natale. Minimizing message size in stochastic communication
patterns: Fast self-stabilizing protocols with 3 bits. In Proc. Twenty-Eighth Annual ACM-SIAM
Symposium on Discrete Algorithms, SODA 2017, Barcelona, Spain, pages 2540–2559, 2017.
53
[12] O. Bournez, P. Fraigniaud, and X. Koegler. Computing with large populations using interactions.
In MFCS, volume 7464 of Lecture Notes in Computer Science, pages 234–246. Springer, 2012.
[13] F. R. K. Chung and L. Lu. Survey: Concentration inequalities and martingale inequalities: A
survey. Internet Mathematics, 3(1):79–127, 2006.
[14] C. Cooper, R. Elsässer, and T. Radzik. The power of two choices in distributed voting. In Proc. 41st
International Colloquium on Automata, Languages, and Programming, ICALP 2014, Copenhagen,
Denmark, Part II, pages 435–446, 2014.
[15] J. Czyzowicz, L. Gasieniec, A. Kosowski, E. Kranakis, P. G. Spirakis, and P. Uznanski. On convergence and threshold properties of discrete Lotka-Volterra population protocols. In Proc. 42nd International Colloquium on Automata, Languages, and Programming, ICALP 2015, Kyoto, Japan,
Part I, pages 393–405, 2015.
[16] A. Dobrinevski and E. Frey. Extinction in neutrally stable stochastic Lotka-Volterra models. Phys.
Rev. E, 85:051903, May 2012.
[17] B. Doerr, C. Doerr, and T. Kötzing. The right mutation strength for multi-valued decision variables.
In T. Friedrich, F. Neumann, and A. M. Sutton, editors, Proceedings of the 2016 on Genetic and
Evolutionary Computation Conference, Denver, CO, USA, July 20 - 24, 2016, pages 1115–1122.
ACM, 2016.
[18] B. Doerr, M. Fouz, and T. Friedrich. Social networks spread rumors in sublogarithmic time. Electronic Notes in Discrete Mathematics, 38:303–308, 2011.
[19] B. Doerr, D. Johannsen, and C. Winzen. Multiplicative drift analysis. CoRR, abs/1101.0776, 2011.
[20] B. Doerr and M. Künnemann. Tight analysis of randomized rumor spreading in complete graphs.
In 2014 Proc. Eleventh Workshop on Analytic Algorithmics and Combinatorics, ANALCO 2014,
Portland, Oregon, USA, January 6, 2014, pages 82–91, 2014.
[21] D. Doty. Timing in chemical reaction networks. In Proc. Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, Portland, Oregon, USA, January 5-7, 2014, pages
772–784, 2014.
[22] U. Feige, D. Peleg, P. Raghavan, and E. Upfal. Randomized broadcast in networks. In Algorithms,
International Symposium SIGAL ’90, Tokyo, Japan, 1990, pages 128–137, 1990.
[23] N. Fountoulakis and K. Panagiotou. Rumor spreading on random regular graphs and expanders.
Random Struct. Algorithms, 43(2):201–220, 2013.
[24] A. M. Frieze and G. R. Grimmett. The shortest-path problem for graphs with random arc-lengths.
Discrete Applied Mathematics, 10(1):57–77, 1985.
[25] L. Gasieniec and G. Stachowiak. Fast space optimal leader election in population protocols. (To
appear, SODA 2018.) CoRR, abs/1704.07649, 2017.
[26] G. Giakkoupis, Y. Nazari, and P. Woelfel. How asynchrony affects rumor spreading time. In Proc.
2016 ACM Symposium on Principles of Distributed Computing, PODC 2016, Chicago, IL, USA,
pages 185–194, 2016.
[27] A. Goldbeter. Computational approaches to cellular rhythms. Nature, 420(6912):238–245, Nov.
2002.
54
[28] J. Hofbauer and K. Sigmund. Evolutionary Games and Population Dynamics. Cambridge University Press, 1998.
[29] R. M. Karp, C. Schindelhauer, S. Shenker, and B. Vöcking. Randomized rumor spreading. In 41st
Annual Symposium on Foundations of Computer Science, FOCS 2000, 12-14 November 2000,
Redondo Beach, California, USA, pages 565–574, 2000.
[30] B. Kerr, M. A. Riley, M. W. Feldman, and B. J. M. Bohannan. Local dispersal promotes biodiversity in a real-life game of rock-paper-scissors. Nature, 418(6894):171–174, Jul 2002.
[31] B. C. Kirkup and M. A. Riley. Antibiotic-mediated antagonism leads to a bacterial game of rockpaper-scissors in vivo. Nature, 428(6981):412–414, Mar 2004.
[32] L. Lamport. Time, clocks, and the ordering of events in a distributed system. Commun. ACM,
21(7):558–565, 1978.
[33] P. K. Lehre and C. Witt. General drift analysis with tail bounds. CoRR, abs/1307.2559, 2013.
[34] C. Lenzen, T. Locher, P. Sommer, and R. Wattenhofer. Clock synchronization: Open problems
in theory and practice. In Proc. 36th Conference on Current Trends in Theory and Practice of
Computer Science, SOFSEM 2010, Spindleruv Mlýn, Czech Republic, pages 61–70, 2010.
[35] C. Lenzen, T. Locher, and R. Wattenhofer. Clock synchronization with bounded global and local
skew. In 49th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2008, October
25-28, 2008, Philadelphia, PA, USA, pages 509–518, 2008.
[36] A. J. Lotka. Contribution to the theory of periodic reactions. The Journal of Physical Chemistry,
14(3):271–274, 1909.
[37] J. Lundelius and N. A. Lynch. A new fault-tolerant algorithm for clock synchronization. In Proc.
Third Annual ACM Symposium on Principles of Distributed Computing, Vancouver, B. C., Canada,
pages 75–88, 1984.
[38] G. B. Mertzios, S. E. Nikoletseas, C. L. Raptopoulos, and P. G. Spirakis. Determining majority in
networks with local interactions and very small local memory. Distributed Computing, 30(1):1–16,
2017.
[39] O. Michail, I. Chatzigiannakis, and P. G. Spirakis. New Models for Population Protocols. Synthesis
Lectures on Distributed Computing Theory. Morgan & Claypool Publishers, 2011.
[40] K. Panagiotou and L. Speidel.
abs/1608.01766, 2016.
Asynchronous rumor spreading on random graphs.
CoRR,
[41] M. Parker and A. Kamenev. Extinction in the Lotka-Volterra model. Phys. Rev. E, 80:021129,
2009.
[42] B. Pittel. On spreading a rumor. SIAM J. Appl. Math., 47(1):213–223, Mar. 1987.
[43] V. B. Priezzhev, D. Dhar, A. Dhar, and S. Krishnamurthy. Eulerian walkers as a model of selforganized criticality. Phys. Rev. Lett., 77:5079–5082, Dec 1996.
[44] T. Reichenbach, M. Mobilia, and E. Frey. Coexistence versus extinction in the stochastic cyclic
Lotka-Volterra model. Phys. Rev. E, 74:051907, Nov 2006.
[45] T. Sauerwald. On mixing and edge expansion properties in randomized broadcasting. Algorithmica, 56(1):51–88, 2010.
[46] A. Szolnoki, M. Mobilia, L. Jiang, B. Szczesny, A. M. Rucklidge, and M. Perc. Cyclic dominance
in evolutionary games: A review. CoRR, abs/1408.6828, 2014.
55
56
Figure 9: Illustration of concentration of various features as a function of time steps for a simulation of the protocol Po for n = 106 , p = 6 · 10−2 , q =
10−2 , r = 10−1 , s = 1 in various scenarios. Left column shows concentration of species A1 , A2 , A3 , middle one the majority layer and right one light.
First row: initialization from a corner configuration (#X = 1, left panel), further dynamics of the protocol after rumor source is removed (#X = 0, middle
panel), further dynamics of the protocol after rumor source is reinserted (#X = 1, right panel). Middle row: initialization from a configuration with
A1 = A2 = A3 = 1/3 and #X = 0. Bottom row: as above, with #X = 1. The range of the horizontal time scale corresponds to 2500 parallel rounds of
the protocol.
APPENDIX: Simulation Examples for Oscillator Dynamics
| 8cs.DS
|
1
State measurement error-to-state stability results
based on approximate discrete-time models
arXiv:1803.08970v1 [cs.SY] 23 Mar 2018
A. J. Vallarella and H. Haimovich
Abstract—Digital controller design for nonlinear systems may be complicated by the fact that an exact discrete-time plant model is not known.
One existing approach employs approximate discrete-time models for
stability analysis and control design, and ensures different types of closedloop stability properties based on the approximate model and on specific
bounds on the mismatch between the exact and approximate models.
Although existing conditions for practical stability exist, some of which
consider the presence of process disturbances, input-to-state stability with
respect to state-measurement errors and based on approximate discretetime models has not been addressed. In this paper, we thus extend existing
results in two main directions: (a) we provide input-to-state stability
(ISS)-related results where the input is the state measurement error and
(b) our results allow for some specific varying-sampling-rate scenarios.
We provide conditions to ensure semiglobal practical ISS, even under
some specific forms of varying sampling rate. These conditions employ
Lyapunov-like functions. We illustrate the application of our results on
numerical examples, where we show that a bounded state-measurement
error can cause a semiglobal practically stable system to diverge.
Index Terms—Sampled-data; nonlinear systems; non-uniform sampling; input-to-state stability (ISS); approximate models; measurement
errors.
I. I NTRODUCTION
Digital controller design for nonlinear systems can be substantially
more complicated than for linear systems. One of the main obstacles
to the design of adequate digital controllers for nonlinear (continuoustime) plants is that even if a continuous-time plant model is known,
the corresponding exact discrete-time model, i.e. the model that
describes the state evolution at the sampling instants, can be difficult
or even impossible to obtain. This happens because the computation
of the exact discrete-time model requires knowledge of the solution
to a nonlinear differential equation.
Interesting existing results address controller design based on
an approximate discrete-time model of the plant and ensure the
stabilization of the original continuous-time plant in a practical sense
[1]–[4]. A general framework for stabilization of disturbance-free
sampled-data nonlinear systems via approximate discrete-time models
was developed in [1], [2] with further generalization for continuoustime plants with disturbances in [3]–[5] and for observer design in [6].
All of these approches are specifically suited to uniform-sampling,
i.e. the sampling rate is constant during operation.
In many sampled-data control systems, the use of a non-uniform
sampling period is necessary, usually imposed by performance requirements of the plant, hardware limitations or network communication constraints. The occurrence of non-uniform sampling is
typical of Networked Control Systems (NCSs). The main feature
of NCSs is that the different system components exchange data
over a communication network. Results for control design based
on approximate discrete-time models also exist for different NCS
scenarios. In [7], controller design for nonlinear NCSs with timevarying sampling periods, time-varying delays and packet dropouts
is considered, based on approximate discrete-time models constructed
The authors are with Centro Internacional Franco-Argentino
de Ciencias de la Información y de Sistemas (CIFASIS),
CONICET-UNR, Ocampo y Esmeralda, 2000 Rosario, Argentina.
{vallarella,haimovich}@cifasis-conicet.gov.ar
Work partially supported by ANPCyT grant PICT 2013-0852, Argentina.
for nominal sampling period and delays. All of the results mentioned so far ensure practical stability properties for sufficiently
small sampling periods, with ultimate bounds of decreasing size in
correspondence with decreasing maximum sampling periods.
Related but conceptually different results exist for dual- or multirate sampling, where some constraint exists on how fast measurements can be taken [8]–[10]. These results give closed-loop practical
stability warranties and employ approximate discrete-time models in
order to predict the state evolution at the control update instants, since
new measurements may not be available at each of these instants.
Specifically, [10] proposes a multi-rate sampled-data scheme to stabilize a NCS via output feedback using discrete-time approximations.
The results of [10] as well as those of [7] give practical stability and
robustness warranties for deviations about a nominal situation.
Non-uniform sampling is also a feature of event-triggered control,
where a triggering condition based on a continuous measurement of
system variables (e.g. the system state) determines when the control
action has to be updated [11], [12]. Strategies that also involve nonuniform sampling but require only sampled measurements are those
of self-triggered control [12]–[14]. Self-triggered control computes
both the control action and the next sampling instant at which the
control action should be recomputed. In both event-triggered and selftriggered control, we may thus say that sampling is controller-driven,
since the controller itself is in charge of computing the next sampling
instant.
The objective of this paper is to provide practical stability results
based on approximate discrete-time models, and allowing some
specific non-uniform sampling scenarios. Our results are, however,
independent of the mechanism employed to vary the sampling rate.
In this context, the main contribution of the current paper is to
provide conditions on the approximate discrete-time plant model and
the control law in order to ensure closed-loop semiglobal-practical
input-to-state stability (see Section II-B for the precise definition),
where the ‘input’ is the state measurement error.
Our results are novel even for the uniform sampling case. The fact
that state measurement errors be considered causes the analysis to
become substantially different from that of disturbances affecting the
plant dynamics (as in [3]). This is because if based on a perturbed
state measurement, the control action will itself have some error, and
this in turn will cause some intersample error additional to that introduced by the approximation in the discrete-time model. More specific
differences between the given and existing results will be explained
along the paper. The fact that some specific varying-sampling-rate
scenarios be covered increases the applicability of our results. Related
preliminary results dealing with closed-loop stability possibly under
controller-driven sampling (without measurement errors) have been
given in [15], [16].
The organization of this paper is as follows. This section ends
with a brief summary of the notation employed throughout the paper.
In Section II we state the problem and the required definitions and
properties. Our main results are given in Section III. An illustrative
example is provided in Section IV and concluding remarks are
presented in Section V. The appendix contains the proofs of some
intermediate technical points.
2
Notation: R, R≥0 , N and N0 denote the sets of real, nonnegative
real, natural and nonnegative integer numbers, respectively. A function α : R≥0 → R≥0 is of class-K (we write α ∈ K) if it is strictly
increasing, continuous and α(0) = 0. It is class-K∞ (α ∈ K∞ ) if in
addition α(s) → ∞ as s → ∞. A function β : R≥0 × R≥0 → R≥0
is of class-KL (β ∈ KL) if β(·, t) ∈ K for all t ≥ 0, and β(s, ·)
is strictly decreasing asymptotically to 0 for every s. We denote the
identity function by id. We denote the Euclidean norm of a vector
x ∈ Rn by |x|. Given a set X ⊂ Rn we denote its ǫ neighborhood
as N (X , ǫ) := {x : inf y∈X |x − y| ≤ ǫ}. We denote an infinite
sequence as {Ti } := {Ti }∞
i=0 . For any sequences {Ti } ⊂ R≥0 and
m
{e
i } ⊂ R , and any γ ∈ K, we take the following conventions:
P−1
i=0 Ti = 0 and γ(sup0≤i≤−1 |ei |) = 0. Given a real number
T > 0 we denote by Φ(T ) := {{Ti } : {Ti } ⊂ (0, T )} the set of all
sequences of real numbers in the open interval (0, T ). For a given
sequence we denote the norm k{xi }k := supi≥0 |xi |.
II. P RELIMINARIES
A. Problem statement
We consider the nonlinear continuous-time plant
ẋ = f (x, u),
x(0) = x0 ,
(1)
where x(t) ∈ Rn , u(t) ∈ Rm are the state and control vectors
respectively and f (0, 0) = 0. As in [1], the function f : Rn × Rm →
Rn is assumed to be such that for each initial condition and constant
control, there exists a unique solution on some interval [0, τ ) with
0 < τ ≤ ∞.
We consider that state measurements become available at time
instants tk , k ∈ N0 , that satisfy t0 = 0 and tk+1 = tk + Tk ,
where {Tk }∞
k=0 is the sequence of corresponding sampling periods.
As opposed to the uniform sampling case where Tk = T for all
k ∈ N0 , the sampling periods Tk may vary. We refer to this situation
as Varying Sampling Rate (VSR). The control signal is assumed
to be piecewise constant (i.e. zero-order hold is present) such that
u(t) = u(tk ) =: uk for all t ∈ [tk , tk+1 ). We denote the state
at the sampling instants by xk := x(tk ). The control action uk
may depend on the state measurement x̂k = xk + ek , with ek the
state measurement error, and also on the sampling period Tk , i.e.
we have uk = U (x̂k , Tk ). The dependence of the control action
at time tk , on the sampling period Tk = tk+1 − tk is possible in
the following situations: (a) uniform sampling; (b) controller-driven
sampling (such as self-triggered control [13], or others [17]). In the
situation (a), the constant sampling period employed can be known
before the controller is implemented. This is the setting in [2]–[4].
In (b), at each sampling instant tk , the controller may select the next
sampling instant tk+1 and, thus, the current sampling period Tk . As a
consequence, knowledge of Tk can be employed in order to compute
the current control action uk . However, our results will be still valid
in the particular case uk = Ū (x̂k ), where the control law does not
explicitly depend on the sampling period. In addition, our results are
independent of the way in which the sampling periods Tk may vary
over time. Fig. 1 depicts the complete closed-loop system.
The exact discrete-time model for a given nonlinear system is the
discrete-time system whose state matches the state of the continuoustime system at every sampling instant. From (1), and given that the
plant input u is held constant over each sampling interval, it is clear
that the state value at the next sampling instant, namely xk+1 , will
depend on the current state value xk , the current input value uk , and
the sampling period Tk . Hence, a discrete-time model for the case
of zero-order hold and VSR can be written as
xk+1 = F e (xk , uk , Tk ),
(2)
uk
u(t)
ZOH
x(t)
SENSOR
PLANT
Tk
x̂k=xk+ek
CONTROLLER
Fig. 1.
Schematic diagram of the controlled plant
where the superscript e stands for “exact”. Due to the fact that the
solution to (1) does not necessarily exist for all future times, then
the exact discrete-time model F e (·, ·, ·) may be not defined for every
(x, u, T ) ∈ Rn × Rm × R≥0 . However, we assume that f in (1) has
sufficient regularity so that for every pair of compact sets X ⊂ Rn
and U ⊂ Rm , there exists some T ∗ > 0 so that F e (x, u, T ) is defined
for every (x, u, T ) ∈ X × U × [0, T ∗ ). The exact model F e is in
general very hard or even impossible to obtain and thus some type
of approximation may be necessary for the design of a stabilizing
controller. A superscript a will be used to denote an approximate
model:
xk+1 = F a (xk , uk , Tk ).
(3)
We will use ‘F ’ with no superscript to denote any discrete-time
model (either exact or approximate). Under the feedback law uk =
U (x̂k , Tk ) = U (xk + ek , Tk ), the closed-loop discrete-time model
becomes
xk+1 = F (xk , U (xk + ek , Tk ), Tk ) =: F̄ (xk , ek , Tk ),
(4)
where the measurement error ek (disturbance) is regarded as an
input. Hence, given a control law/discrete-time model pair (U, F ),
the quantity F̄ will denote the composition of F and U as shown in
(4). Whenever required, we will write x(k, ξ, {ei }, {Ti }) to denote
the solution of (4) at time k with initial condition ξ at k = 0,
and corresponding to the measurement error and sampling period
sequences {ei } and {Ti }.
B. Stability properties for varying sampling rate
Our objective is to provide conditions that ensure stability properties of the exact discrete-time closed-loop system under the varyingsampling-rate case. The following definition constitute a natural
extension of the semiglobal practical ISS [3], [18], [19] property.
Definition 2.1: The system (4) is said to be Semiglobal Practical
ISS-VSR (SP-ISS-VSR) if there exist functions β ∈ KL and γ ∈ K∞
such that for every M > 0, E > 0 and R > 0 there exists T ⋆ > 0
such that the solutions of (4) satisfy
!
k−1
X
Ti + γ
|xk | ≤ β |x0 |,
sup |ei | + R,
(5)
i=0
0≤i≤k−1
for all1 k ∈ N0 , all {Ti } ∈ Φ(T ⋆ ), all |x0 | ≤ M and all k{ei }k ≤ E.
The function F̄ is said to be SP-ISS-VSR if the system (4) defined
by F̄ is SP-ISS-VSR.
Remark 2.2: Note that the sequences of constant sampling periods
{Ti } such that Ti = T > 0 for all i ∈ N0 are included in the set
Φ(T ⋆ ) whenever T ⋆ > T . This implies that if any of the above
properties holds for a certain system then the respective properties
for the uniform sampling case also hold.
1 As explained under “Notation” in Section I, for k = 0 we interpret
P−1
i=0 Ti = 0 and γ(sup0≤i≤−1 |ei |) = 0.
3
Remark 2.3: Setting k = 0 in (5), it follows that |x0 | ≤
β(|x0 |, 0) + R holds for every x0 ∈ Rn and every R > 0. As a
consequence, any function β ∈ KL characterizing SP-ISS-VSR has
the additional property that r ≤ β(r, 0) for all r ≥ 0. This property
will be repeatedly employed for the proof of our results.
Lemma 2.7: If (U, F a ) is MSEC with (U, F e ) then for each
compact sets X ⊂ Rn and E ⊂ Rn , and constants L > 0 and
η > 0, there exists T̃ > 0 such that, if {Ti }, {ei } and ξ satisfy
for all k for which
e
The next definition, as appears in [6, Definition 1], ensures that the
mismatch between the solutions of the exact and approximate systems
over one sampling period is bounded by a value that depends on the
sampling period, uniformly over states and inputs in compact sets.
This bound tends to zero in a specific way as the sampling period
becomes smaller.
Definition 2.4: The function F a is said to be consistent with F e
if for each compact set Ω ⊂ Rn × Rm , there exist ρ ∈ K and T0 > 0
such that, for all (x, u) ∈ Ω and all T ∈ [0, T0 ],
(6)
Definition 2.4 is similar to the definition of one-step consistency in
[1, Defintion 1]. However, the main difference lies in the fact that
Definition 2.4 is “open-loop” whereas Definition 1 of [1] involves
the feedback law. In the current setting of imperfect state knowledge,
consideration of the feedback law would be meaningless unless the
state measurement error be also considered (cf. Defintion 2.6 below).
The following lemma is a minor modification of Lemma 1 of [1]
that gives sufficient conditions for the consistency of F a with F e ,
based on f and F a , and hence without requiring knowledge of F e .
The required modification is due to the fact that in the current setting
the feedback law cannot assume perfect knowledge of the state. The
proof closely follows the proof of Lemma 1 of [1] but is given in
the Appendix for completeness.
Lemma 2.5: Suppose that
i) F a is consistent with F Euler , where F Euler (x, u, T ) := x +
T f (x, u), and
ii) for every pair of compact sets X ⊂ Rn and U ⊂ Rm , there exist
ρ ∈ K and M > 0 such that for all (x, u), (y, u) ∈ X × U,
a) |f (y, u)| ≤ M ,
b) |f (y, u) − f (x, u)| ≤ ρ(|y − x|).
a
Then, F is consistent with F .
The following definition extends the definition of multi-step consistency in [1] to the varying-sampling-rate scenario considered and to
the case of state measurement errors.
Definition 2.6: The pair (U, F a ) is said to be Multi-Step Error
Consistent (MSEC) with (U, F e ) if, for each L > 0, η > 0 and
compact sets X ⊂ Rn and E ⊂ Rn , there exist a function α :
R≥0 × R≥0 → R≥0 and T ∗ > 0 such that
(7)
for all xe , xa ∈ X satisfying |xe − xa | ≤ δ, all e ∈ E , and
P all T ∈
(0, T ∗ ), and for all {Ti } ∈ Φ(T ∗ ) and k ∈ N such that k−1
i=0 Ti ≤
L we have
k
z }| {
α (0, {Ti }) := α(. . . α(α( 0, T0 ), T1 ), . . . , Tk−1 ) ≤ η.
k
i=0
ek ∈ E ,
(9)
Ti ∈ [0, L], then, for these values of k
The next lemma gives a sufficient condition for MSEC. This lemma
can be regarded as the corresponding extension of [1, Lemma 3].
Lemma 2.8: If, for each compact sets X ⊂ Rn and E ⊂ Rn there
exist ρ0 ∈ K, a nondecreasing function σ : R≥0 → R≥0 and T ∗ > 0
such that for all T ∈ (0, T ∗ ) and all x, y ∈ X , e ∈ E we have
|F̄ e (x, e, T ) − F̄ a (y, e, T )| ≤ T ρ0 (T ) + (1 + T σ(T ))|x − y|
(11)
then, (U, F a ) is MSEC with (U, F e ).
Proof: Let L > 0 and η > 0, and let X ⊂ Rn and E ⊂ Rn be
compact sets. Define the compact set X˜ := N (X , η) and let X̃ and
E generate ρ0 ∈ K, σ and T ∗ > 0. Select T̂ > 0 sufficiently small
so that T̂ ≤ T ∗ and eσ(T̂ )L ρ0 (T̂ )L ≤ η. Define
α(δ, T ) := T ρ0 (T ) + (1 + T σ(T ))δ.
(12)
Let x, y ∈ X be such that |x − y| ≤ δ, let e ∈ E , and T ∈ (0, T̂ ).
From (11) and (12), then (7) is satisfied. Consider {Ti } ∈ Φ(T̂ ).
Using (12), we may recursively define η0 := 0 and, for k ≥ 1,
ηk := αk (0, {Ti }) = Tk−1 ρ0 (Tk−1 ) + (1 + Tk−1 σ(Tk−1 ))ηk−1
Recursively employing
the latter formula, we may write, for k ≥ 1
P
and such that k−1
i=0 Ti ≤ L,
ηk = Tk−1 ρ0 (Tk−1 ) +
k−2
X
Tj ρ0 (Tj )
j=0
≤ Tk−1 ρ0 (Tk−1 ) +
k−2
X
k−1
Y
(1 + Ti σ(Ti ))
i=j+1
Tj ρ0 (Tj )eσ(T̂ )
Pk−1
j+1 Ti
j=0
≤ Tk−1 ρ0 (Tk−1 ) + eσ(T̂ )L
k−2
X
Tj ρ0 (Tj )
j=0
e
|F̄ e (xe , e, T ) − F̄ a (xa , e, T )| ≤ α(δ, T )
Pk−1
|∆xk | := |x (k, ξ, {ei }, {Ti }) − xa (k, ξ, {ei }, {Ti })| ≤ η. (10)
C. Model consistency with non-ideal state measurements
|F e (x, u, T ) − F a (x, u, T )| ≤ T ρ(T ).
xa (k, ξ, {ei }, {Ti }) ∈ X ,
{Ti } ∈ Φ(T̃ ),
(8)
The next lemma constitutes the corresponding extension of [1,
Lemma 2] and shows that under MSEC, the error between the
approximate and exact solutions over a fixed time period is reduced
by making the maximum sampling period smaller. The proof is given
in the Appendix for the sake of completeness.
≤ eσ(T̂ )L
k−1
X
Tj ρ0 (Tj ) ≤ eσ(T̂ )L ρ0 (T̂ )
j=0
≤ eσ(T̂ )L ρ0 (T̂ )L ≤ η,
k−1
X
Tj
j=0
where we have used the facts that for every a > 0, 1 + a ≤ ea , and
Ti < T̂ for every i ∈ N0 . This shows that (U, F a ) is MSEC with
(U, F e ).
For the sake of completeness, we next state the definition of locally
uniformly bounded control law as in [3, Defintion II.4].
Definition 2.9: U (x, T ) is said to be locally uniformly bounded if
for every M > 0 there exist T ∗ = T ∗ (M ) > 0 and C = C(M ) > 0
such that |U (x, T )| ≤ C for all |x| ≤ M and T ∈ (0, T ∗ ).
Lemma 2.10 gives sufficient conditions for MSEC based on
consistency (as per Definition 2.4) and properties of the control law
and approximate model.
Lemma 2.10: Suppose that
i) F a is consistent with F e as per Definition 2.4.
ii) U (x, T ) is locally uniformly bounded as per Definition 2.9.
iii) For each compact sets X , E ⊂ Rn there exist a nondecreasing
function σ : R≥0 → R≥0 and T1 > 0 such that for all x, z ∈ X ,
e ∈ E and T ∈ (0, T1 ) we have
|F̄ a (x, e, T ) − F̄ a (z, e, T )| ≤ (1 + T σ(T ))|x − z|.
(13)
4
Then, (U, F a ) is MSEC with (U, F e ).
Proof: Let X ⊂ Rn and E ⊂ Rn be compact, and let M > 0 be
such that |x + e| ≤ M for all x ∈ X and e ∈ E . From ii), there exist
T ∗ (M ) > 0 and C(M ) > 0 such that |U (x + e, T )| ≤ C for all
T ∈ (0, T ∗ ), x ∈ X and e ∈ E . Define C := {u ∈ Rm : |u| ≤ C}
and Ω := X × C. From i), there exist ρ ∈ K and T0 > 0 such that
|F e (x, U (x + e, T ), T ) − F a (x, U (x + e, T ), T )| ≤ T ρ(T ) (14)
for all T ∈ (0, T2 ), x ∈ X and e ∈ E with T2 := min{T ∗ , T0 }. For
all x, z ∈ X and e ∈ E we have
|F e (x, U (x + e, T ), T ) − F a (z, U (z + e, T ), T )|
≤|F e (x, U (x + e, T ), T ) − F a (x, U (x + e, T ), T )|
+ |F a (x, U (x + e, T ), T ) − F a (z, U (z + e, T ), T )|.
|F e (x, U (x + e, T ), T )−F a (z, U (z + e, T ), T )|
(16)
a
hence (11) holds. By Lemma 2.8, then (U, F ) is MSEC with
(U, F e ).
In the next section, we give conditions on the approximate model
and control law in order to establish the semiglobal practical ISS for
the (exact) closed-loop system.
III. M AIN R ESULTS
In this section, we show that under MSEC, the SP-ISS-VSR
property for the approximate model carries over to the exact model.
Our first contribution is the following.
Theorem 3.1: Suppose that (U, F a ) is MSEC with (U, F e ), and
that the system xak+1 = F̄ a (xak , ek , Tk ) is SP-ISS-VSR. Then, the
system xek+1 = F̄ e (xek , ek , Tk ) is SP-ISS-VSR.
Proof: Let β ∈ KL and γ ∈ K∞ characterize the SP-ISS-VSR
property of xak+1 = F̄ a (xak , ek , Tk ). Define
γ̂(·) := β(2γ(·), 0) + γ(·).
(17)
Let M > 0, E > 0 and R > 0, and select η > 0 and Ra > 0 such
that
6η ≤ M,
β(6η, 0) + Ra < R/2,
Ra < η.
(18)
Since xak+1 = F̄ a (xak , ek , Tk ) is SP-ISS-VSR, in correspondence
with M̃ := M + γ(E), E and Ra , there exists T a > 0 such that for
all k ∈ N0 ,
!
k−1
X
a
a
|xk | ≤ β |x0 |,
sup |ei | + Ra
(19)
Ti + γ
i=0
0≤i≤k−1
|xa0 |
a
whenever
≤ M̃ and for all sequences {Ti } and {ei } such that
{Ti } ∈ Φ(T ) and k{ei }k ≤ E. Define
X̃ := {x ∈ Rn : |x| ≤ β(M̃ , 0) + γ(E) + Ra }
(20)
E := {e ∈ Rn : |e| ≤ E}.
(21)
Let L > 1 satisfy
β β M, 0 + γ(E) + Ra , L − 1 ≤ η.
|xe (k,ξ, {ei }, {Ti })| ≤ |xa (k, ξ, {ei }, {Ti })| + |∆x0k,ξ |
!
k−1
X
≤ β |ξ|,
Ti + γ
sup |ei | + Ra + η
≤β
|ξ|,
i=0
k−1
X
(22)
(23)
0≤i≤k−1
Ti
i=0
!
+ γ̂
sup
0≤i≤k−1
|ei | + R,
(24)
P
for all k for which k−1
i=0 Ti ≤ L, where we have used the facts that
γ ≤ γ̂ and η + Ra ≤ 6η + Ra ≤ β(6η, 0) + Ra < R/2 < R. For
every k ∈ N0 , define
)
(
r−1
X
Ti ≤ L .
(25)
s(k) := max r ∈ N0 : r ≥ k + 1,
i=k
Note that s(k) ≥ k + 1 for all k ∈ N0 because L > 1 and Ti < 1
Ps(k)−1
for all i ∈ N0 . Also, i=k Ti > L − T ⋆ > L − 1 holds for all
k ∈ N0 . We thus have
s(0)−1
X
Ti ≤ β(M, L − 1)
β |ξ|,
i=0
≤ β(β(M, 0) + γ(E) + Ra , L − 1) ≤ η, (26)
where we have used the fact M ≤ β(M, 0) and (22). Evaluating (23) at k = s(0), using (18) and (26), and defining
xek := xe (k, ξ, {ei }, {Ti }) it follows that
!
s(0)−1
X
e
|xs(0) | ≤ β |ξ|,
sup
|ei | + Ra + η
Ti + γ
0≤i≤s(0)−1
i=0
≤ γ (E) + 3η ≤ M̃ .
(27)
Claim 1: Suppose that |xer | ≤ γ(E) + 3η for some r ∈ N0 . Then,
≤ γ(E) + 3η and |xek | ≤ γ̂(E) + R for all r ≤ k ≤ s(r).
Proof of Claim 1: Set xa0 = xer and consider xak+1 =
F̄ a (xak , ek+r , Tk+r ) for all k ∈ N0 . For all r ≤ k ≤ s(r), we
have
|xes(r) |
|xek | = |xe (k, ξ, {ei }, {Ti })| = |xe (k − r, xer , {ei+r }, {Ti+r })|
≤ |xak−r | + |∆xrk,xer |.
(28)
Since |xa0 | ≤ γ(E) + 3η ≤ M̃ , {ei } ⊂ E , {Ti } ∈ Φ(T ⋆ ) ⊂ Φ(T a ),
and since F̄ a is SP-ISS-VSR, then for all k ∈ N0 we have
!
k−1
X
a
a
Ti+r + γ
|xk | ≤ β |x0 |,
sup |ei+r | + Ra
0≤i≤k−1
i=0
Take T̃ > 0 in correspondence with the tuple (X̃ , E , L, η), as per
Lemma 2.7. We will show that β ∈ KL and γ̂ ∈ K∞ characterize the
SP-ISS-VSR of xek+1 = F̄ e (xek , ek , Tk ), with T ⋆ := min{1, T a , T̃ }.
Consider sequences {ei } ⊂ E and {Ti } ⊂ Φ(T ⋆ ). For every k, ℓ ∈
N0 and ξ ∈ Rn , define
∆xℓk,ξ :=
From (19) and (20) it follows that if |xa0 | ≤ M̃ , then xak ∈ X̃ for all
k ∈ N0 . Consider an initial condition xe0 = xa0 = ξ such that |ξ| ≤
e
M ≤ M̃ . Since (U, F a ) is MSEC
2.7 it
Pk−1with (U, F ), by Lemma
follows that for all k such that i=0 Ti ≤ L, we have |∆x0k,ξ | ≤ η.
We thus have
(15)
Defining T̂ := min{T1 , T2 } and using (13) and (14) in (15), then
≤T ρ(T ) + (1 + T σ(T ))|x − z|,
|xe (k − ℓ, ξ, {ei+ℓ }, {Ti+ℓ }) − xa (k − ℓ, ξ, {ei+ℓ }, {Ti+ℓ })|.
≤β
γ(E) + 3η,
k−1
X
i=0
≤β
M̃ ,
k−1
X
Ti+r
i=0
!
Ti+r
!
+ γ(E) + Ra
(29)
+ γ(E) + Ra .
It thus follows that xak ∈ X̃ for all k ∈ N0 . By Lemma 2.7, we have
|∆xrk,xer | ≤ η
for all r ≤ k ≤ s(r).
(30)
5
Combining the bounds obtained so far, we reach, for all r ≤ k ≤
s(r),
!
k−1
X
|xek | ≤ β M̃ ,
(31)
Ti + γ (E) + 2η.
i=r
In particular at k = s(r), and taking into account that M̃ = M +
γ(E) ≤ β(M, 0) + γ(E) + Ra , we obtain
s(r)−1
X
e
Ti + γ (E) + 2η
|xs(r) | ≤ β β(M, 0) + γ(E) + Ra ,
i=r
≤ γ (E) + 3η
(32)
P
where we have used the fact that s(r)−1
Ti > L − 1 and (22).
i=r
From (28)–(30), we have, for all r ≤ k ≤ s(r),
|xek |
≤ β(γ(E) + 3η, 0) + γ(E) + Ra + η
≤ γ̂(E) + β(6η, 0) + Ra + η
(33)
where we have used the fact that η < 6η < β(6η, 0) + Ra < R/2.
This concludes the proof of the claim.
◦
Since from (27) we have |xes(0) | ≤ γ(E)+3η, iterative application
of Claim 1 and the fact that s(k) ≥ k + 1 for all k ∈ N0 show that
|xek | ≤ γ̂(E) + R
for all k ≥ s(0).
Combining the latter bound with (24), valid for 0 ≤ k ≤ s(0), it
follows that
!
k−1
X
e
Ti + γ̂(E) + R, ∀k ≥ 0.
|xk | ≤ β |ξ|,
(34)
i=0
The proof concludes by noting that, by causality, the trajectory xek
cannot depend on future values of ei , and hence E in (34) can be
replaced by sup0≤i≤k−1 |ei |.
Theorem 3.1 shows that the SP-ISS-VSR property for the exact
closed-loop discrete-time model can be ensured under MSEC if
the approximate closed-loop discrete-time model is SP-ISS-VSR.
It would thus be useful to have checkable conditions in order to
ensure that the approximate model is suitable. A set of conditions
is given by Theorem 3.2, which states Lyapunov-based sufficient
conditions for a discrete-time (closed-loop) model to be SP-ISSVSR. These conditions consist of specific boundedness and continuity
requirements and a Lyapunov-type condition on the closed-loop
model. Theorem 3.2 constitutes the main contribution of the current
paper and its proof, which is highly nontrivial, consitutes our main
technical result.
Theorem 3.2: Suppose that i)–iv) are satisfied.
i) There exists T̊ > 0 so that F̄ (0, 0, T ) = 0 for all T ∈ (0, T̊ ).
ii) There exists T̂ > 0 such that for every ǫ > 0 there exists
δ = δ(ǫ) > 0 such that |F̄ (x, e, T )| < ǫ whenever |x| ≤ δ,
|e| ≤ δ and T ∈ (0, T̂ ).
iii) For every M ≥ 0 and E ≥ 0, there exist C = C(M, E) > 0
and Ť = Ť (M, E) > 0, with C(·, ·) nondecreasing in each
variable and Ť (·, ·) nonincreasing in each variable, such that
|F̄ (x, e, T )| ≤ C for all |x| ≤ M , |e| ≤ E and T ∈ (0, Ť ).
iv) There exist α1 , α2 , α3 ∈ K∞ and ρ ∈ K such that for every
M ≥ R > 0 and E > 0 there exist T̃ = T̃ (M, E, R) > 0 and
V : Rn → R≥0 such that
α1 (|x|) ≤ V (x) ≤ α2 (|x|),
∀x ∈ Rn , and
V (F̄ (x, e, T )) − V (x) ≤ −T α3 (|x|),
ρ(|e|) + R ≤ |x| ≤ M,
|e| ≤ E,
(35)
∀(x, e, T ) satisfying
T ∈ (0, T̃ ).
i=0
0≤i≤k−1
Consider ρ ∈ K from iv). Define, ∀s ≥ 0, ∀r ≥ 0,
X1 (s, r) := {x ∈ Rn : |x| ≤ ρ(s) + r}
(38)
E (s) := {e ∈ Rn : |e| ≤ s}
n
o
T̄ (s, r) := min Ť (ρ(s) + s, s), Ť (ρ(r) + r, r), T̂ , T̊
S(s, r) := X1 (s, r) × E (s) × (0, T̄ (s, r))
(41)
σ(s, r) :=
(42)
sup
|F̄ (x, e, T )|.
(39)
(40)
(x,e,T )∈S(s,r)
≤ β(2γ(E), 0) + β(6η, 0) + γ(E) + Ra + η
≤ γ̂(E) + R,
Then, the system (4) is SP-ISS-VSR.
Proof: We aim to prove that there exist β ∈ KL and γ ∈ K∞
such that for all M0 > 0, E0 > 0 and R0 > 0 there exists T ⋆ > 0
such that for all {Ti } ∈ Φ(T ⋆ ), |x0 | ≤ M0 , k{ei }k ≤ E0 and
k ∈ N0 , the solutions of (4) satisfy
!
k−1
X
Ti + γ
|xk | ≤ β |x0 |,
sup |ei | + R0 .
(37)
(36)
From (38)–(39), we have X1 (0, 0) = {0} and E (0) = {0}. From
assumptions i)–ii), then T̄ (0, 0) > 0 and σ(0, 0) = 0. Given that
Ť (·, ·) is nonincreasing in each variable, note that Ť (ρ(s) + r, s) ≥
min{Ť (ρ(s) + s, s), Ť (ρ(r) + r, r)}, then T̄ (s, r) ≤ Ť (ρ(s) + r, s).
Defining ζ(s) := σ(s, s), from (42) we have
σ(s, r) ≤ ζ(s) + ζ(r).
(43)
Claim 2: There exists χ ∈ K∞ such that χ ≥ ζ.
Proof of Claim2: Let ǫ > 0 and take δ = δ(ǫ) according to ii).
Define δ̂ := min δ2 , ρ−1 ( δ2 ) (if 2δ ∈
/ dom ρ−1 , just take δ̂ = δ/2).
Then for all x ∈ X1 (δ̂, δ̂) and e ∈ E (δ̂) we have |x| ≤ δ and
|e| ≤ δ/2. Given that |F̄ (x, e, T )| < ǫ for every |x| ≤ δ, |e| ≤
δ and T ∈ (0, T̂ ) this shows that lims→0+ ζ(s) = σ(0, 0) = 0.
From iii), it follows that |F̄ (x, e, T )| ≤ C(s, s) for all |x| ≤ s,
|e| ≤ s and T ∈ (0, Ť (s, s)) for every s ≥ 0. From (40)–(42) and
the fact that C(·, ·) is nondecreasing in each variable, it follows that
ζ(s) ≤ C(s, s) ≤ C(s̄, s̄) for all s ∈ [0, s̄] with s̄ > 0. Thus,
ζ(s) < ∞ for every s ≥ 0. We have ζ : R≥0 → R≥0 , ζ(0) = 0, ζ
is right-continuous at zero and, by (38)–(42), nondecreasing. Then,
by Lemma 2.5 of [20], there exists χ ∈ K∞ such that χ ≥ ζ.
◦
Define η ∈ K∞ and η̃ ∈ K∞ via
η := max{χ, id},
(44)
η̃ := max{χ, ρ}.
(45)
σ(s, r) ≤ η̃(s) + η(r) ∀s ≥ 0, ∀r ≥ 0.
(46)
By (43) we have
Consider M0 > 0, E0 > 0 and R0 > 0 given and α1 , α2 , α3 ∈
1
K∞ from iv). Select E = E0 , R = η −1 ( 12 α−1
2 ( 3 α1 (R0 ))),
−1
and M = α1 ◦ α2 max{R, M0 , η̃(E0 ) + η(R0 )} . From (35), it
follows that M ≥ R. Also, R ≤ R0 follows because, from (35),
1 −1 1
α ( 3 α1 (r)) ≤ α−1
2 (α1 (r)) ≤ r for all r ≥ 0, and since η(r) ≥ r
2 2
from (44), then η −1 (r) ≤ r for all r ≥ 0. Let M, E, R with iv)
generate T̃ > 0 and V : Rn → R≥0 such that (35) and (36) hold.
Define T ⋆ = min{T̃ , T̄ (E, R)} and
X2 (s, r) := {x : V (x) ≤ α2 (η̃(s) + η(r))}.
(47)
Let x ∈ X1 (s, r). Then, α2 (|x|) ≤ α2 (ρ(s) + r) ≤ α2 (η̃(s) + η(r)),
and using (35), then V (x) ≤ α2 (η̃(s)+η(r)). Therefore, X1 (s, r) ⊂
X2 (s, r) for all s ≥ 0 and r ≥ 0. Let xk denote the solution to (4)
corresponding to |x0 | ≤ M0 , k{ei }k ≤ E0 and {Ti } ∈ Φ(T ⋆ ).
From (35) we have that α−1
2 (V (xk )) ≤ |xk |; using this in (36) then
V (xk+1 ) − V (xk ) ≤ −Tk α3 (|xk |) ≤ −Tk α(V (xk ))
if ρ(|ek |) + R ≤ |xk | ≤ M
(48)
6
where α := α3 ◦ α−1
2 .
Claim 3: If |x0 | ≤ M0 then |xk | ≤ M for all k ∈ N0 .
Proof of Claim 3: From (35), |x0 | ≤ M0 implies that V (x0 ) ≤
α2 (M0 ) and |x0 | ≤ α−1
1 ◦ α2 (M0 ), and by definition of M ,
then |x0 | ≤ M . By induction, we will prove that V (xk ) ≤
α2 (max{M0 , η̃(E0 )+η(R0 )}) for all k ∈ N0 . Note that the assertion
holds for k = 0. Suppose that V (xk ) ≤ α2 (max{M0 , η̃(E0 ) +
η(R0 )}). Then, |xk | ≤ M . If xk ∈
/ X1 (|ek |, R), then |xk | >
ρ(|ek |) + R and from (48), then V (xk+1 ) ≤ V (xk ). If xk ∈
X1 (|ek |, R), from (42)–(43) and the definition of η̃ and η we have
|xk+1 | ≤ η̃(|ek |) + η(R). Using (35), then V (xk+1 ) ≤ α2 (η̃(|ek |) +
η(R)) ≤ α2 (η̃(E0 ) + η(R0 )), and hence the induction assumption
holds for k + 1. Since V (xk ) ≤ α2 (max{M0 , η̃(E0 ) + η(R0 )})
implies that |xk | ≤ M , we have thus shown that |xk | ≤ M for all
k ∈ N0 .
◦
Claim 4: If xℓ ∈ X2 (k{ei }k, R) for some ℓ ∈ N0 then xk remains
in X2 (k{ei }k, R) for all k ≥ ℓ.
Proof of Claim 4: Let xℓ ∈ X2 (k{ei }k, R). If xℓ ∈
/
X1 (k{ei }k, R), then |xℓ | > ρ(k{ei }k) + R. Consequently, if
xℓ ∈ X2 (k{ei }k, R) \ X1 (k{ei }k, R), from (48) it follows that
and hence xℓ+1 ∈ X2 (k{ei }k, R). Next, consider that xℓ ∈
X1 (k{ei }k, R). From (42) and the definition of η̃ and η we have
|xℓ+1 | ≤ η̃(k{ei }k) + η(R). Using (35) and recalling (47), then
xℓ+1 ∈ X2 (k{ei }k, R). By induction, we have thus shown that if
xℓ ∈ X2 (k{ei }k, R) for some ℓ ∈ N0 , then xk ∈ X2 (k{ei }k, R) for
all k ≥ ℓ.
◦
P
Define tk = k−1
T
and
the
function
i
i=0
t − tk
(V (xk+1 ) − V (xk ))
Tk
if t ∈ [tk , tk+1 ) ,
which depends on the initial condition x0 , on the sampling period
sequence {Ti }, on the disturbance sequence {ei }, and on the given
constants M0 , E0 , R0 (through the fact that V depends on the latter
constants). Then,
ẏ(t) =
V (xk+1 ) − V (xk )
Tk
∀t ∈ (tk , tk+1 ) , ∀k ∈ N0 .
(49)
Note that
y(tk ) = V (xk ),
∀k ∈ N0 .
(51)
= inf{tk : xk ∈
Hence (51) holds up to t = t
where t
does not
X2 (k{ei }k, R)}. Note that the function α = α3 ◦ α−1
2
depend on any of the following quantities: x0 , {Ti }, {ei }, M0 , E0 ,
or R0 . Since α is positive definite, using Lemma 4.4 of [21], then
there exists β1 ∈ KL such that, for all t ∈ [0, tk∗ ) we have
k∗
V (x0 ),
Ti
i=0
!
+ α2 η̃(k{ei }k) + η(R) .
Using the fact that χ(a + b) ≤ χ(2a) + χ(2b) for every χ ∈ K and
(35) then
!
k−1
X
α1 (|xk |) ≤ β1 α2 (|x0 |),
Ti + α2 (2η̃(k{ei }k)) + α2 (2η(R)).
i=0
Define β ∈ KL via β(s, τ ) := α−1
1 (3β1 (α2 (s), τ )) and γ ∈ K∞
via γ(s) := α−1
1 (3α2 (2η̃(s))). Recalling the definition of R then
α−1
1 (3α2 (2η(R))) ≤ R0 , it follows that
!
k−1
X
|xk | ≤ β |x0 |,
(53)
Ti + γ(k{ei }k) + R0
i=0
for all k ∈ N0 , all {Ti } ∈ Φ(T ⋆ ), all |x0 | ≤ M0 and all k{ei }k ≤
E0 . We have thus established that (4) is SP-ISS-VSR.
IV. E XAMPLES
ẋ = x3 + u =: f (x, u)
whose Euler (approximate) discrete-time model is
xk+1 = xk + Tk x3k + uk =: F a (xk , uk , Tk ).
(54)
(55)
a
This open-loop Euler model F is consistent (as per Definition 2.4)
with the open-loop exact model F e . This can be established by
means of Lemma 2.5 as follows. Since F a coincides with F Euler , then
assumption i) of Lemma 2.5 holds; also, assumption ii) of Lemma 2.5
is easily shown to hold using f in (54).
We will consider two of the feedback laws considered in [1]:
U (xk , Tk ) = −xk − 3x3k ,
W (xk , Tk ) = −xk −
x3k .
(56)
(57)
In [1], both control laws were shown to achieve semiglobal-practical
stabilization under zero-order hold and uniform sampling. Our aim
here is to show that when state measurement errors are taken into
consideration, the control law (56) achieves SP-ISS-VSR whereas
(57) does not, not even under uniform sampling.
(50)
By Claim 3, for all |x0 | ≤ M0 and all tk such that xk ∈
/
X2 (k{ei }k, R), (48) holds. Combining (48) with (49) it follows that
for all t ∈ (tk , tk+1 )
ẏ(t) ≤ −α (y(tk )) ≤ −α (y(t)) .
V (xk ) ≤ β1
k−1
X
Consider the nonlinear continuous-time plant in Example 1 of [1]:
V (xℓ+1 ) ≤ V (xℓ ) − Tℓ α(V (xℓ )) ≤ V (xℓ )
y(t) := V (xk )+
(52), then, for all k ∈ N0 ,
k∗
y(t) ≤ β1 (y(0), t) .
By (50), for every k ∈ N0 such that xj ∈
/ X2 (k{ei }k, R) for all
0 ≤ j ≤ k, we have that
!
k−1
X
V (xk ) ≤ β1 V (x0 ),
Ti .
(52)
i=0
From Claim 4 and (47), if xj ∈ X2 (k{ei }k, R) then V (xj ) ≤
α2 (η̃(k{ei }k) + η(R)) for all k ≥ j. Combining the latter with
A. SP-ISS-VSR
Under the feedback law (56) and taking measurement errors into
account so that uk = U (xk + ek , Tk ), the closed-loop approximate
model is given by (4), with
F̄ a (x, e, T ) = x − T [2x3 + 9ex2 + (9e2 + 1)x + 3e3 + e]. (58)
We next prove that (U, F a ) is MSEC with (U, F e ) by means of
Lemma 2.10. Assumption i) has been already established, whereas
ii) is easily shown using (56). In order to establish iii), we use (58)
to evaluate
|F̄ a (x, e, T ) − F̄ a (z, e, T )| = |1 − T P (x, e, z)| |x − z|
≤ (1 + T |P (x, e, z)|)|x − z|,
where P (x, e, z) denotes a multinomial in the indeterminates x, e, z,
and we have used the fact that for every positive integer p, xp −
z p = q(x, z) · (x − z) for some multinomial q(x, z). Consider
compact sets X , E ⊂ R, and define the nonnegative constant
σ̄ := sup(x,e,z)∈(X ,E,X ) |P (x, e, z)|. Then, (13) holds for all T > 0
with σ : R≥0 → R≥0 defined as σ(T ) := σ̄, and assumption iii) of
Lemma 2.10 holds. By Lemma 2.10, (U, F a ) is MSEC with (U, F e ).
7
Next, we prove that xk+1 = F̄ a (xk , ek , Tk ) is SP-ISS-VSR.
The continuity and boundedness assumptions i), ii) and iii) of
Theorem 3.2, can be easily verified from (55), (56) and (58). To prove
assumption iv) of Theorem 3.2, let α1 , α2 , α3 , ρ ∈ K∞ be defined
via α1 (s) = α2 (s) = s2 , α3 (s) = 2.194s4 , and ρ(s) = 10s. Let
M ≥ R > 0 and E > 0 be given and define V (x) = x2 . Note that
V satisfies (35) and compute
V (F̄ a (x, e, T )) − V (x) = h(x, e)T + g(x, e)T 2 ,
(59)
h(x, e) = −2x[2x3 + 9ex2 + (9e2 + 1)x + (3e3 + e)],
(60)
g(x, e) = [2x3 + 9ex2 + (9e2 + 1)x + (3e3 + e)]2 .
(61)
2
Define G := sup|x|≤M,|e|≤E g(x, e) and let T̃ = 1.8R
. Expanding
G
h(x, e) and taking absolute values on sign indefinite terms,
h(x, e) ≤ −4x4 − (18e2 + 2)x2 + 18|e||x|3 + (6|e|3 + 2|e|)|x|
≤ −4x4 − 2x2 + 18|e||x|3 + (6|e|3 + 2|e|)|x|.
(62)
Whenever ρ(|e|) ≤ |x|, it follows that |e| ≤ 0.1|x| and
4
2
4
A PPENDIX
Proof of Lemma 2.5: Let Ω ⊂ Rn ×Rm be a given compact set.
Let X1 ⊂ Rn and U ⊂ Rm be compact sets such that Ω ⊂ X1 × U,
and define X := N (X1 , 1). Let ρ′ ∈ K and T0′ > 0 be given
by Definition 2.4 in correspondence with Ω by the fact that F a is
consistent with F Euler . Let hypothesis iia) generate ρ ∈ K and M > 0
in correspondence with X and U. Define T0 := min{T0′ , 1/M }. Let
(x0 , u) ∈ Ω and let φu (t, x0 ) denote the (unique) solution to
3
h(x, e) ≤ −4x − 2x + 1.8x + (0.006|x| + 0.2|x|)|x|
≤ −2.194x4 − 1.8x2 .
It thus follows that for ρ(|e|) + R ≤ |x| ≤ M and T ∈ (0, T̃ ),
V (F̄ a (x, e, T )) − V (x) ≤ (−2.194x4 − 1.8x2 )T + GT 2
≤ −T 2.194x4 + −1.8x2 + GT T
≤ −T α3 (|x|).
V. C ONCLUSIONS
We have given stability results for digital control design based
on discrete-time approximate models under varying sampling rate
(VSR) and in the presence of state measurement errors. We have
extended the concept of semiglobal practical ISS (SP-ISS) to the
VSR case (SP-ISS-VSR) and introduced the concept of multi-step
error consistency (MSEC). We have shown that if the approximate
closed-loop model is MSEC with the exact one, and if the control
law renders the approximate model SP-ISS-VSR, then the same
controller ensures SP-ISS-VSR of the exact discrete-time closedloop model. We have also given sufficient conditions for MSEC and
derived Lyapunov-based conditions that guarantee SP-ISS-VSR of a
discrete-time model. All of the given conditions are checkable without
assuming knowledge of the exact discrete-time model. We have also
illustrated application via numerical examples.
(63)
ẋ(t) = f (x(t), u),
where u is constant, so that
Therefore, assumption iv) of Theorem 3.2 also holds and the closedloop system xk+1 = F̄ a (xk , ek , Tk ) is SP-ISS-VSR. Theorem 3.1
then ensures that the exact closed-loop system is SP-ISS-VSR.
Consider next the feedback law (57), which was also shown
in [1] to achieve semiglobal practical stability under zero-order
hold and uniform sampling. We next show that, under bounded
state measurement errors and also uniform sampling, the true plant
state may diverge. Under the feedback law (57), the closed-loop
approximate model becomes
2
φu (t, x0 ) = x0 +
Z
t
f (φu (s, x0 ), u)ds.
0
From hypothesis iia), then |φu (t, x0 ) − x0 | ≤ M t ≤ 1, and hence
x(t) ∈ X , holds for all t ∈ [0, T0 ]. From hypothesis iib), it follows
that for all T ∈ [0, T0 ],
B. Practical stability but no SP-ISS-VSR
a
x(0) = x0 ,
2
3
F̄ (x, e, T ) = x − T [3ex + (1 + 3e )x + e + e].
(64)
Notice the absence of the cubic term in x within the square brackets
in (64) as compared with (58). Consider the constant error sequence
{ei } with ei = −1 for all i ∈ N0 . From (64), then
F̄ a (x, −1, T ) = x + T [3x2 − 4x + 2].
The polynomial between square brackets satisfies 3x2 − 4x + 2 ≥ x
for all x ≥ 1. Therefore, F̄ a (x, −1, T ) ≥ x(1 + T ) whenever x ≥
1 and T > 0. For every T ⋆ > 0, consider the constant sequence
{Ti } ∈ Φ(T ⋆ ) with Ti = T ⋆ /2 for all i ∈ N0 . It it thus clear that
for the selected constant sequences {ei } and {Ti }, and for x0 ≥ 1,
we have
lim xa (k, x0 , {ei }, {Ti }) = ∞
k→∞
and hence the approximate closed-loop system xk+1
=
F̄ a (xk , ek , Tk ) is not SP-ISS-VSR. However, the pair (W, F a ) is
indeed MSEC with (W, F e ), as can be shown following identical
steps to those in the previous example. Using Lemma 2.7, it can
be shown that the exact closed-loop system cannot be SP-ISS-VSR,
either.
Z
T
0
[f (φu (t, x0 ), u) − f (x0 , u)] dt
Z T
≤
ρ(|φu (t, x0 ) − x0 |)dt ≤ T ρ(M T ).
0
RT
Considering that F (x0 , u, T ) = x0 + 0 f (φu (t, x0 ), u)dt, that
R
T
F Euler (x0 , u, T ) = x0 + 0 f (x0 , u)dt, and that F a is consistent
with F Euler , then, for all T ∈ [0, T0 ], it follows that
e
|F e (x0 , u, T ) − F a (x0 , u, T )| ≤ |F Euler (x0 , u, T ) − F a (x0 , u, T )|
+ |F e (x0 , u, T ) − F Euler (x0 , u, T )|
Z T
[f (φu (t, x0 ), u) − f (x0 , u)] dt
≤ T ρ′ (T ) +
0
≤ T ρ′ (T ) + T ρ(M T ) =: T ρ1 (T )
where ρ1 ∈ K. Thus, F e is consistent with F a .
Proof of Lemma 2.7: Define X˜ := N (X , η). Since (U, F a ) is
MSEC with (U, F e ), in correspondence with the tuple (X̃ , E , L, η)
there exist α : R≥0 × R≥0 → R≥0 and T ∗ > 0 such that (7)
and (8) hold according to Definition 2.6. Let T̃ := T ∗ and consider
{Ti } ∈ Φ(T̃ ), {ei } ⊂ E , and ξ ∈ X . Note that (10) holds trivially
for k = 0 because
|∆x0 | = |ξ − ξ| = 0 ≤ η.
Pk
We proceed by induction
Pj on k. Let k ≥ 0 be such that i=0 Ti ∈
[0, L]. Note then that i=0 Ti ∈ [0, L] for every 0 ≤ j ≤ k. Suppose
also that xa (j, ξ, {ei }, {Ti }) ∈ X and that |∆xj | ≤ η, both for all
8
0 ≤ j ≤ k. Then, xe (j, ξ, {ei }, {Ti }) ∈ X̃ for all 0 ≤ j ≤ k.
Consider ∆xk+1 . From (10), (4) and (7), we have
|∆xk+1 | = F̄ e xek , ek , Tk − F̄ a xak , ek , Tk ≤ α(|∆xk |, Tk )
≤ αk+1 (|∆x0 |, {Ti }) = αk+1 (0, {Ti }) ≤ η.
We have thus established the result by induction.
R EFERENCES
[1] D. Nešić, A. Teel, and P. Kokotović, “Sufficient conditions for stabilization of sampled-data nonlinear systems via discrete-time approximations,” Systems & Control Letters, vol. 38, no. 4-5, pp. 259–270, 1999.
[2] D. Nešić and A. Teel, “A framework for stabilization of nonlinear
sampled-data systems based on their approximate discrete-time models,”
IEEE Trans. on Automatic Control, vol. 49, no. 7, pp. 1103–1122, 2004.
[3] D. Nešić and D. S. Laila, “A note on input-to-state stabilization for
nonlinear sampled-data systems,” IEEE Trans. on Automatic Control,
vol. 47, no. 7, pp. 1153–1158, 2002.
[4] D. Nešić and D. Angeli, “Integral versions of iss for sampled-data
nonlinear systems via their approximate discrete-time models,” IEEE
Trans. on Automatic Control, vol. 47, no. 12, pp. 2033–2037, 2002.
[5] D. Nešić and D. S. Laila, “Input-to-state stabilization for nonlinear
sampled-data systems via approximate discrete-time plant models,” in
Proc. 40th IEEE Conf. on Dec. and Control, vol. 1, 2001, pp. 887–892.
[6] M. Arcak and D. Nešić, “A framework for nonlinear sampled-data
observer design via approximate discrete-time models and emulation,”
Automatica, vol. 40, no. 11, pp. 1931–1938, 2004.
[7] N. van de Wouw, D. Nešić, and W. Heemels, “A discrete-time framework
for stability analysis of nonlinear networked control systems,” Automatica, vol. 48, no. 6, pp. 1144–1153, 2012.
[8] I. G. Polushin and H. J. Marquez, “Multirate versions of sampled-data
stabilization of nonlinear systems,” Automatica, vol. 40, no. 6, pp. 1035–
1041, 2004.
[9] X. Liu, H. J. Marquez, and Y. Lin, “Input-to-state stabilization for
nonlinear dual-rate sampled-data systems via approximate discrete-time
model,” Automatica, vol. 44, no. 12, pp. 3157–3161, 2008.
[10] H. Beikzadeh and H. J. Marquez, “Multirate output feedback control
of nonlinear networked control systems,” Automatic Control, IEEE
Transactions on, vol. 60, no. 7, pp. 1939–1944, 2015.
[11] P. Tabuada, “Event-triggered real-time scheduling of control tasks,” IEEE
Trans. on Automatic Control, vol. 52, no. 9, pp. 1680–1685, 2007.
[12] W. P. M. H. Heemels, K. H. Johansson, and P. Tabuada, “An introduction
to event-triggered and self-triggered control,” in 2012 IEEE 51st IEEE
Conference on Decision and Control (CDC), 2012, pp. 3270–3285.
[13] M. Velasco, J. Fuertes, and P. Marti, “The self-triggered task model
for real-time control systems,” in Proc. 24th IEEE Real-Time Systems
Symposium (RTSS), 2003, pp. 67–70.
[14] A. Anta and P. Tabuada, “To sample or not to sample: Self-triggered
control for nonlinear systems,” IEEE Transactions on Automatic Control,
vol. 55, no. 9, pp. 2030–2042, 2010.
[15] A. J. Vallarella, E. N. Osella, and H. Haimovich, “Stability results
for control design based on approximate discrete-time models and
under controller-driven sampling,” in 25 Congreso Argentino de Control
Automático (AADECA), Buenos Aires, Argentina, 2016.
[16] A. J. Vallarella, E. Osella, and H. Haimovich, “Equivalence between
uniform- and varying-sampling stability properties for discrete-time
exact and approximate models,” in XVII Workshop on Information
Processing and Control (RPIC), Mar del Plata, Argentina, 2017.
[17] H. Haimovich and E. Osella, “On controller-driven varying-samplingrate stabilization via Lie-algebraic solvability,” Nonlinear Analysis:
Hybrid Systems, vol. 7, no. 1, pp. 28–38, February 2013.
[18] E. D. Sontag, “Smooth stabilization implies coprime factorization,” IEEE
Trans. on Automatic Control, vol. 34, pp. 435–443, 1989.
[19] ——, “On the input-to-state stability property,” European Journal of
Control, vol. 1, no. 1, pp. 24–36, 1995.
[20] F. H. Clarke, Y. S. Ledyaev, and R. J. Stern, “Asymptotic stability and
smooth lyapunov functions,” Journal of differential Equations, vol. 149,
no. 1, pp. 69–114, 1998.
[21] Y. Lin, E. D. Sontag, and Y. Wang, “A smooth converse Lyapunov
theorem for robust stability,” SIAM J. Control Optim., vol. 34, no. 1,
p. 124, 160 1996.
| 3cs.SY
|
A Productivity Checker for Logic Programming
E. Komendantskaya1 and P. Johann2 and M.Schmidt3
1
Heriot-Watt University, Edinburgh, Scotland, UK
Appalachian State University, Boone, NC, USA
3
University of Osnabrück, Osnabrück, Germany
arXiv:1608.04415v3 [cs.PL] 19 Aug 2016
2
Abstract. Automated analysis of recursive derivations in logic programming is
known to be a hard problem. Both termination and non-termination are undecidable problems in Turing-complete languages. However, some declarative languages
offer a practical work-around for this problem, by making a clear distinction between whether a program is meant to be understood inductively or coinductively.
For programs meant to be understood inductively, termination must be guaranteed, whereas for programs meant to be understood coinductively, productive nontermination (or “productivity”) must be ensured. In practice, such classification
helps to better understand and implement some non-terminating computations.
Logic programming was one of the first declarative languages to make this distinction: in the 1980’s, Lloyd and van Emden’s “computations at infinity” captured
the big-step operational semantics of derivations that produce infinite terms as answers. In modern terms, computations at infinity describe “global productivity”
of computations in logic programming. Most programming languages featuring
coinduction also provide an observational, or small-step, notion of productivity
as a computational counterpart to global productivity. This kind of productivity
is ensured by checking that finite initial fragments of infinite computations can
always be observed to produce finite portions of their infinite answer terms.
In this paper we introduce a notion of observational productivity for logic programming as an algorithmic approximation of global productivity, give an effective
procedure for semi-deciding observational productivity, and offer an implemented
automated observational productivity checker for logic programs.
Keywords: Logic programming, corecursion, coinduction, termination, productivity.
1
Introduction
Induction is pervasive in programming and program verification. It arises in definitions
of finite data (e.g., lists, trees, and other algebraic data types), in program semantics
(e.g., of finite iteration and recursion), and proofs (e.g., of properties of finite data and
processes). Coinduction, too, is important in these arenas, arising in definitions of infinite
data (e.g., lazily defined infinite streams), in program semantics (e.g., of concurrency),
and in proofs (e.g., of observational equivalence, or bisimulation, of potentially infinite
processes). It is thus desirable to have good support for both induction and coinduction
in systems for reasoning about programs.
Given a logic program P and a term A, SLD-resolution provides a mechanism for
automatically (and inductively) inferring that P ⊢ A holds, i.e., that P logically entails
A. The “answer” for a program P and a query ? ← A is a substitution σ computed from
P and A by SLD-resolution. Soundness of SLD-resolution ensures that P ⊢ σ(A) holds,
so we also say that P computes σ(A).
Example 1 (Inductive logic program). The program P1 codes the Peano numbers:
0. nat(0) ←
1. nat(s(X)) ← nat(X)
To answer the question “Does P1 ⊢ nat(s(X)) hold?”, we represent it as the logic programming (LP) query ? ← nat(s(X)) and resolve it with P1 . It is standard in implementations
of traditional LP to use a topmost clause selection strategy, which resolves goals against
clauses in the order in which they appear in the program. Topmost clause selection gives
the derivation nat(s(X)) → nat(X) → true for P1 and ? ← nat(s(X)), which computes
the answer {X 7→ 0} in its last step. Since P1 computes nat(s(0)), one answer to our
question is “Yes, provided X is 0.”
While inductive properties of terminating computations are quite well understood [14],
non-terminating LP computations are notoriously difficult to reason about, and can arise
even for programs that are intended to be inductive:
Example 2 (Coinductive meaning of inductive logic program). If P1′ is obtained by reversing the order of the clauses in the program P1 from Example 1, then the SLD-derivation
for program P1′ and query ? ← nat(s(X)) does not terminate under standard topmost
clause selection. Instead, it results in an attempt to compute the “answer” {X 7→ s(s(...))}
by repeatedly resolving with Clause 1. Nevertheless, P1′ is still computationally meaningful, since it computes the first limit ordinal at infinity [14].
Some programs do not admit terminating computations under any selection strategy:
Example 3 (Coinductive logic program). No derivation for the query ? ← stream(X) and
the program P2 comprising the clause
0. stream(scons(0, Y)) ← stream(Y)
terminates with an answer, be it success or otherwise. Nevertheless, P2 has computational
meaning: it computes the infinite stream of 0s at infinity.
The importance of developing sufficient infrastructure to support coinduction in automated proving has been argued across several communities; see, e.g., [13,17,21]. In LP,
the ability to work with non-terminating and coinductive programs depends crucially
on understanding the structural properties of non-terminating SLD-derivations. To illustrate, consider the non-terminating programs P3 , P4 , and P5 :
Program Program definition
For query ? ← p(X), computes the answer:
P3
p(X) ← p(X)
id
P4
p(X) ← p(f(X))
id
P5
p(f(X)) ← p(X)
{X 7→ f(f...)}
Programs P3 and P4 each loop without producing any substitutions at all; only P5 computes an infinite term at infinity. It is of course not a coincidence that only P5 resembles
a (co)inductive data definition by pattern matching on a constructor, as is commonly
used in functional programming.
When an infinite SLD-derivation computes an infinite object, and this object can be
successively approximated by applying to the initial query the substitutions computed
at each step of the derivation, the derivation is said to be globally productive. The only
derivation for program P5 and the query ? ← p(X) is globally productive since it approximates, in the sense just described, the infinite term p(f(f...)). In terminology of [14],
it computes p(f(f...)) at infinity. Programs P2 and P1′ similarly give rise to globally
productive derivations. But no derivations for P3 or P4 are globally productive.
Since global productivity determines which non-terminating logic programs can be
seen as defining coinductive data structures, we would like to identify exactly when a
program is globally productive. But porting functional programming methods of ensuring productivity by static syntactic checks is hardly possible. Unlike pattern matching in
functional programming, SLD-resolution is based on unification, which has very different operational properties — including termination and productivity properties — from
pattern matching. For example, programs P1 , P1′ , P2 , and P5 are all terminating by termmatching SLD-resolution, i.e., resolution in which unifiers are restricted to matchers, as
in term rewriting. We thus call this kind of derivations rewriting derivations.
Example 4 (Coinductive program defining an irrational infinite term). The program P6
comprises the single clause
0. from(X, scons(X, Y)) ← from(s(X), Y)
For P6 and the query ? ← from(0, Y), SLD-resolution computes at infinity the answer substitution {Y 7→ [0, s(0), s(s(0)), . . .]}. Here [t1 , t2 , . . .] abbreviates scons(t1 , scons(t2 , . . .)),
and similarly in the remainder of this paper. This derivation depends crucially on unification since variables occurring in the two arguments to from in the clause head overlap.
If we restrict to rewriting, then there are no successful derivations (terminating or nonterminating) for this choice of program and query.
Example 4 shows that any analysis of global productivity must necessarily rely on
specific properties of the operational semantics of LP, rather than on program syntax
alone. It has been observed in [11,9] that one way to distinguish globally productive
programs operationally is to identify those that admit infinite SLD-derivations, but for
which rewriting derivations always terminate. We call this program property observational productivity. The programs P1 , P1′ , P2 , P5 , P6 are all observationally productive.
The key observation underlying observational productivity is that terminating rewriting derivations can be viewed as points of finite observation in infinite derivations. Consider again program P6 and query ? ← from(0, Y) from Example 4. Drawing rewriting
derivations vertically and unification-based resolution steps horizontally, we see that each
unification substitution applied to the original query effectively observes a further fragment of the stream computed at infinity:
{X7→[0,X′ ]}
→
from(0, X)
{X′ 7→[s(0),X′′ ]}
′
from(0, [0, X ])
′
from(s(0), X )
→
.
′′
→ ...
from(0, [0, s(0), X ])
from(s(0), [s(0), X′′ ])
from(s(s(0)), X′′ )
If we compute unifiers only when rewriting derivations terminate, then the resulting
derivations exhibit consumer-producer behaviour: rewriting steps consume structure (here,
the constructor scons), and unification steps produce more structure (here, new sconses)
for subsequent rewriting steps to consume. This style of interleaving matching and unification steps was called structural resolution (or S-resolution) in [9,12].
Model-theoretic properties of S-resolution relative to least and greatest Herbrand
models of programs were studied in [12]. In this paper, we provide a suitable algorithm
for semi-deciding observational productivity of logic programs, and present its implementation [19], see also Appendix B online. As exemplified above, observational productivity
of a program P is in fact a conjunction of two properties of P :
1. universal observability: termination of all rewriting derivations, and
2. existential liveness: existence of at least one non-terminating S-resolution or SLDresolution derivation.
While the former property is universal, the latter must be existential. For example, the
program P1 defining the Peano numbers can have both inductive and coinductive meaning. When determining that a program is observationally productive, we must certify
that the program actually does admit derivations that produce infinite data, i.e., that it
actually can be seen as a coinductive definition. Our algorithm for semi-deciding observational productivity therefore combines two checks:
1. guardedness checks that semi-decide universal observability: if a program is guarded,
then it is universally observable. (The converse is not true in general.)
2. liveness invariant checks ensuring that, if a program is guarded and exhibits an
invariant in its consumption-production of constructors, then it is existentially live.
This is the first work to develop productivity checks for LP. An alternative approach
to coinduction in LP, known as CoLP [7,21], detects loops in derivations and closes
them coinductively. However, loop detection was not intended as a tool for the study
of productivity and, indeed, is insufficient for that purpose: programs P3 , P4 and P5 ,
of which only the latter is productive, are all treated similarly by CoLP, and all give
coinductive proofs via its loop detection mechanism.
Our approach also differs from the usual termination checking algorithms in termrewriting systems (TRS) [22,1,8] and LP [3,16,18,20,15]. Indeed, these algorithms focus
on guaranteeing termination, rather than productivity, see Section 5. And although the
notion of productivity has been studied in TRS [4,5], the actual technical analysis of productivity is rather different there because it considers infinitary properties of rewriting,
whereas observational productivity relies on termination of rewriting.
The rest of this paper is organised as follows. In Section 2 we introduce a contraction ordering on terms that extends the more common lexicographic ordering, and argue
that this extension is needed for our productivity analysis. We also recall that static
guardedness checks do not work for LP. In Section 3 we employ contraction orderings in
dynamic guardedness checks and present a decidable property, called GC2, that characterises guardedness of a single rewriting derivation, and thus certifies existential observability. In Section 4 we employ GC2 to develop an algorithm, called GC3, that analyses
consumer-producer invariants of S-resolution derivations to certify universal observability. For universally observable programs, these invariants also serve as liveness invariant
checks. We also prove that GC3 indeed semi-decides observational productivity. In Section 5 we discuss related work and in Section 6 – implementation and applications of the
productivity checker. In Section 7 we conclude the paper.
2
Contraction Orderings on Terms
In this section, we will introduce the contraction ordering on first-order terms, on which
our productivity checks will rely. We work with the standard definition of first-order logic
programs. A signature Σ consists of a set F of function symbols f, g, . . . each equipped
with an arity. Nullary (0-ary) function symbols are constants. We also assume a countable
set Var of variables, and a set P of predicate symbols each equipped with an arity. We
have the following standard definition for terms, formulae and Horn clauses:
Definition 1 (Syntax of Horn clauses and programs).
Terms T erm ::= V ar | F (T erm, ..., T erm)
Atomic formulae (or atoms) At ::= P(T erm, ..., T erm)
(Horn) clauses CH ::= At ← At, ..., At
Logic programs P rog ::= CH, ..., CH
In what follows, we will use letters A, B with subscripts to refer to elements of At.
Given a program P , we assume all clauses are indexed by natural numbers starting from
0. When we need to refer to ith clause of program P , we will use notation P (i). To refer
to the head of clause P (i), we will use notation head (P (i)).
A substitution is a total function σ : Var → T erm. Substitutions are extended from
variables to terms as usual: if t ∈ T erm and σ is a substitution, then the application
σ(t) is a result of applying σ to all variables in t. A substitution σ is a unifier for t, u
if σ(t) = σ(u), and is a matcher for t against u if σ(t) = u. A substitution σ is a most
general unifier (mgu) for t and u if it is a unifier for t and u and is more general than any
other such unifier. A most general matcher (mgm) σ for t against u is defined analogously.
We can view every term and atom as a tree. Following standard definitions [2,14],
such trees can be indexed by elements of a suitably defined tree language. Let N∗ be the
set of all finite words (i.e., sequences) over the set N of natural numbers. A set L ⊆ N∗ is
a (finitely branching) tree language if the following two conditions hold: (i) for all w ∈ N∗
and all i, j ∈ N, if wj ∈ L then w ∈ L and, for all i < j, wi ∈ L, and (ii) for all w ∈ L, the
set of all i ∈ N such that wi ∈ L is finite. A tree language L is finite if it is a finite subset
of N∗ , and infinite otherwise. Term trees (for terms and atoms) are defined as mappings
from a tree language L to the given signature, see [2,14,9]. Informally speaking, every
symbol occurring in a term or an atom receives an index from L.
In what follows, we will work with term tree representation of all terms and atoms,
and for brevity we will refer to all term trees simply as terms. We will use notation t(w)
when we need to talk about the element of the term tree t indexed by a word w ∈ L.
Note that leaf nodes are always given by variables or constants.
Example 5. Given L = {ǫ, 0, 00, 01}, the atom stream(scons(0, Y)) can be seen as a term
tree t given by the map t(ǫ) = stream, t(0) = scons, t(00) = 0, t(01) = Y.
We can use such indexing to refer to subterms, and notation subterm(t, w) will refer to a
subterm of term t starting at node w. In the above example, where t = stream(scons(0, Y)),
subterm(t, 0) is scons(0, Y).
Two most popular tools for termination analysis of declarative programs are lexicographic ordering and (recursive) path ordering of terms. Informally, the idea can be
adopted to LP setting as follows. Suppose we have a clause A ← B1 , . . . , Bi , . . . , Bn . We
may want to check whether each Bi sharing the predicate with A is “smaller”’ than A,
since this guarantees that no infinite rewriting derivation is triggered by this clause. For
lexicographic ordering we will write Bi <l A and for path ordering we will write Bi <p A.
Using standard orderings to prove universal observability works well for program P2 ,
since stream(Y) <l stream(scons(0, Y)) and stream(Y) <p stream(scons(0, Y)), and
so any rewriting derivation for P2 terminates. But universal observability of P6 from
Example 4 cannot be shown by this method. Indeed, none of the four orderings
from(X, scons(X, Y)) <l from(s(X), Y), from(s(X), Y) <l from(X, scons(X, Y)),
from(X, scons(X, Y)) <p from(s(X), Y), and from(s(X), Y) <p from(X, scons(X, Y))
holds because the subterms pairwise disagree on the ordering. This situation is common
for LP, where some arguments hold input data and some hold output data, so that
some decrease while others increase in recursive calls. Nevertheless, P6 is universally
observable, and we want to be able to infer this. Studying the S-resolution derivation for
P6 in Section 1, we note that universal observability of P6 is guaranteed by contraction
of from’s second argument. It is therefore sufficient to establish that terms get smaller
in only one argument. This inspires our definition of a contraction ordering, which takes
advantage of the tree representation of terms.
Definition 2 (Contraction, recursive contraction). If t1 and t2 are terms, then t2
is a contraction of t1 (written t1 ⊲ t2 ) if there is a leaf node t2 (w) on a branch B in t2 ,
and there exists a branch B ′ in t1 that is identical to B up to node w, however, t1 (w) is
not a leaf. If, in addition, subterm(t1 , w) contains the symbol given by t2 (w), then t2 is
a recursive contraction of t1 .
We distinguish variable contractions and constant contractions according as t2 (w) is
a variable or constant, and call subterm(t1 , w) a reducing subterm for t1 ⊲ t2 at node w.
We call subterm(t1 , w) a recursive, variable or constant reducing subterm if t1 ⊲ t2 is a
recursive, variable or constant contraction, respectively.
Example 6 (Contraction orderings). We have from(X, scons(X, Y)) ⊲ from(s(X), Y), as the
leaf Y in the latter is “replaced” by the term scons(X, Y) in the former. Formally, scons(X, Y)
is a recursive and variable reducing subterm. It can be used to certify termination of all
rewriting derivations for P6 . Note that from(s(X), Y) ⊲ from(X, scons(X, Y)) also holds,
with (recursive and variable) reducing subterm s(X).
The fact that ⊲ is not well-founded makes reasoning about termination delicate. Nevertheless, contractions emerge as precisely the additional ingredient needed to formulate
our productivity check for a sufficiently general and interesting class of logic programs.
In general, static termination checking for LP suffers serious limitations; see, e.g., [3].
The following example illustrates this phenomenon.
Example 7 (Contraction ordering on clause terms is insufficient for termination checks).
The program P7 , that is not universally observable, is given by mutual recursion:
0. p(s(X1), X2, Y1, Y2) ← q(X2, X2, Y1, Y2)
1. q(X1, X2, s(Y1), Y2) ← p(X1, X2, Y2, Y2)
No two terms from the same clause of P7 can be related by any contraction ordering
because their head symbols differ. But recursion arises for P7 when a derivation calls its
two clauses alternately, so we would like to examine rewriting derivations for queries,
such as ? ← p(s(X1), X2, s(Y1), Y2) and ? ← p(s(X1), s(X2), s(Y1), s(Y2)), that exhibit its
recursive nature. Unfortunately, such queries are not given directly by P7 ’s syntax, and
so are not available for static program analysis.
As static checking for contraction ordering in clauses is not sufficient, we will define
dynamic checks in the next section. The idea is to build a rewriting tree for each clause,
and check whether term trees featured in that derivation tree obey contraction ordering.
3
Rewriting Trees: Guardedness Checks for Rewriting
Derivations
To properly reason about rewriting derivations in LP, we need to take into account that
i) in LP, unlike, e.g., in TRS, we have conjuncts of terms in the bodies of clauses, and ii)
a logic program can have overlapping clauses, i.e., clauses whose heads unify. These two
facts have been analysed in detail in the LP literature, usually using the notion of andor-trees and, where optimisation has been concerned, and-or-parallel trees. We carry on
this tradition and consider a variant of and-or trees for derivations. However, the trees
we consider are not formed by general SLD-resolution, but rather by term matching
resolution. Rewriting trees are so named because each of their edges represents a term
matching resolution step, i.e., a matching step as in term rewriting.
Definition 3 (Rewriting tree). Let P be a logic program with n clauses, and A be an
atomic formula. The rewriting tree for P and A is the possibly infinite tree T satisfying
the following properties.
–
–
–
–
–
A is the root of T
Each node in T is either an and-node or an or-node
Each or-node is given by P (i), for some i ∈ {0, . . . , n}
Each and-node is an atom seen as a term tree.
For every and-node A′ occurring in T , if there exist exactly k > 0 distinct clauses
P (j), . . . , P (m) in P (a clause P (i) has the form Bi ← B1i , . . . , Bni i for some ni ),
such that A′ = θj (Bj ) = . . . = θm (Bm ), for mgms θj , . . . , θm , then A′ has exactly
k children given by or-nodes P (j), . . . , P (m), such that, every or-node P (i) has ni
children given by and-nodes θi (B1i ), . . . , θi (Bni i ).
When constructing rewriting trees, we assume a suitable algorithm [9] for renaming free
variables in clause bodies apart. Figure 1 gives examples of rewriting trees. An andsubtree of a rewriting tree (a subtree in which a derivation always pursues only one
or-choice at a time) is a rewriting derivation, see [9] for a formal definition.
Because mgms are unique up to variable renaming, given a program P and an atom A,
rewriting tree T for P and A is unique. Following the same principle as with definition of
term trees, we use suitably defined finitely-branching tree languages for indexing rewriting
trees, see [9] for precise definitions. When we need to talk about a node of a rewriting
tree T indexed by a word w ∈ L, we will use notation T (w).
We can now formally define our notion of universal observability.
Definition 4 (Universal observability). A program P is universally observable if,
for every atom A, the rewriting tree for A and P is finite.
Programs P1 , P1′ , P2 , P5 , P6 are universally observable, whereas programs P3 , P4 and P7
are not. An exact analysis of why P7 is not universally observable is given in Example 9.
We can now apply the contraction ordering we defined in the previous section to
analyse termination properties of rewriting trees. A suitable notion of guardedness can
be defined by checking for loops in rewriting trees whose terms fail to decrease by any
contraction ordering. But note that our notion of a loop is more general than that used
in CoLP [7,21] since it does not require the looping terms to be unifiable.
Definition 5 (Loop in a rewriting tree). Given a program P and an atom A the
rewriting tree T for P and A contains a loop at nodes w and v, denoted loop(T, w, v), if
w properly precedes v on some branch of T , T (w) and T (v) are and-nodes whose atoms
have the same predicate, and parent or-nodes of T (w) and T (v) are given by the same
clause P (i).
Examples of loops in rewriting trees are given (underlined) in Figure 1.
If T has a loop at nodes w and v, and if t is a recursive reducing subterm for T (w) ⊲
T (v), then loop(T, w, v) is guarded by (P (i), t), where P (i) is the clause that was resolved
against to obtain T (w) and T (v). It is unguarded otherwise. A rewriting tree T is guarded
if all of its loops are guarded, and is unguarded otherwise. We write GC2(T ) when T is
guarded, and say that GC2(T ) holds.
Example 8. In Figure 1, we have (underlined) loops in the third rewriting tree (for
q(s(X′′ ), s(X′′ ), s(Y′ ), Y′′ ) and q(s(X′ ), s(X′′ ), Y′′ , Y′′ )) and the fourth rewriting tree (for
q(s(X′′ ), s(X′′ ), s(Y′ ), s(Y′′ )) and q(s(X′′ ), s(X′′ ), s(Y′′ ), s(Y′′ ))). Neither is guarded. In the
former case, there is a contraction on the third argument, but because s(Y′ ) and Y′′ do not
share a variable, it is not recursive contraction. In the latter loop, there is no contraction
at all.
By Definition 5, each repetition of a clause and predicate in a branch of a rewriting tree
triggers a check to see if the loop is guarded by some recursive reducing subterm.
Proposition 1 (GC2 is decidable). GC2 is a decidable property of rewriting trees.4
The proof of Proposition 1 also establishes that every guarded rewriting tree is finite.
The decidable guardedness property GC2 is a property of individual rewriting trees.
But our goal is to decide guardedness universally, i.e., for all of a program’s rewriting
trees. The next example shows that extrapolating from existential to universal guardedness is a difficult task.
Example 9 (Existential guardedness does not imply universal guardedness). For program
P7 , the rewriting trees constructed for the two clause heads p(s(X′ ), X′′ , Y′ , Y′′ ) and
q(s(X′ ), X′′ , s(Y′ ), Y′′ ) are both guarded since neither contains any loops at all. Nevertheless, there is a rewriting tree for P7 (the last tree in Figure 1) that is unguarded and
infinite. The third tree is not guarded (due to the unguarded loop), but it is finite.
The example above shows that our initial idea of checking rewriting trees generated by
clause heads is insufficient to detect all cases of nonterminating rewriting. Since a similar
4
All proofs are in an Appendix A supplied as supplementary material online. Corresponding
pseudocode algorithms are given in Appendix B.
Y ′ 7→s(Y ′ )
X ′′ 7→s(X ′′ )
→
′
′′
′
′′
′′
′
′′
p(s(X ), X , Y , Y )
p(s(X ), X , s(Y ), Y )
P7 (0)
P7 (0)
q(X′′ , X′′ , Y′ , Y′′ )
Y ′′ 7→s(Y ′′ )
→
′
→
′
′′
′′
′
′′
p(s(X ), s(X ), s(Y ), Y )
′′
′′
p(s(X1), s(X′′ ), s(Y′ ), s(Y′′ ))
P7 (0)
q(s(X ), s(X ), s(Y ), Y )
P7 (1)
′′
′′
P7 (0)
q(X′′ , X′′ , s(Y′ ), Y′′ )
′′
′
′′
q(s(X ), s(X′′ ), s(Y′ ), s(Y′′ ))
P7 (1)
′′
p(X , X , Y , Y )
′′
′′
P7 (1)
′′
′′
′′
′′
p(s(X ), s(X ), Y , Y )
′′
p(s(X ), s(X′′ ), s(Y′′ ), s(Y′′ ))
P7 (0)
′′
′′
P7 (0)
q(s(X ), s(X ), Y , Y )
′′
q(s(X ), s(X′′ ), s(Y′′ ), s(Y′′ ))
P7 (1)
...
Fig. 1. An initial fragment of the derivation tree (comprising four rewriting trees) for the program P7 of Example 7 and the atom p(s(X′ ), X′′ , Y′ , Y′′ ). Its third and fourth rewriting trees each
contain an unguarded loop (underlined), so both are unguarded. The fourth tree is infinite.
situation can obtain for any finite set of rewriting trees, universal observability, and
hence observational productivity, of programs cannot be determined by guardedness of
rewriting trees for program clauses alone. The next section addresses this problem.
4
Derivation Trees: Observational Productivity Checks
The key idea of this section is, given a program P , to identify a finite set S of rewriting
trees for P such that checking guardedness of all rewriting trees in S is sufficient for
guaranteeing guardedness of all rewriting trees for P . One way to identify such sets will
be to use the strategy of Example 9 and Figure 1: for every clause P (i) of P , to construct
a rewriting tree for the head of P (i), and, if that tree is guarded, explore what kind of
mgus the leaves of that tree generate, and see if applications of those mgus may give an
unguarded tree. As Figure 1 shows, we may need to apply this method iteratively until
we find a nonguarded rewriting tree. But we want the number of such iterations to be
finite. This section presents a solution to this problem.
We start with a formal definition of rewriting tree transitions, which we have seen
already in Figure 1, see also Figure 2.
Definition 6 (Rewriting tree transition). Let P be a program and T be a rewriting
tree for P and an atom A. If T (w) is a leaf node of T given by an atom B, and B
unifies with a clause P (i) via mgu σ, we define a tree Tw as follows: we apply σ to every
and-node of T , and extend the branches where required, according to Definition 3.
Computation of Tw from T is denoted T → Tw . The operation T → Tw is the tree
transition for T and w.
If a rewriting tree T is constructed for a program P and an atom A, a (finite or infinite)
sequence T → T ′ → T ′′ → . . . of tree transitions is an S-resolution derivation for P and
A. For a given rewriting tree T , several different S-resolution derivations are possible
from T . This gives rise to the notion of a derivation tree.
Definition 7 (Derivation tree, guarded derivation tree). Given a logic program
P and an atom A, the derivation tree D for P and A is defined as follows:
– The root of D is given by the rewriting tree for P and A.
– For a rewriting tree T occurring as a node of D, if there exists a transition T → Tw ,
for some leaf node w in T , then the node T has a child given by Tw .
A derivation tree is guarded if each of its nodes is a guarded rewriting tree, i.e., if GC2(T )
holds for each of its nodes T .
Figure 1 shows an initial fragment of the derivation tree for P7 and p(s(X′ ), X′′ , Y′ , Y′′ ).
Note that we now have three kinds of trees: term trees have signature symbols as
nodes, rewriting trees have atoms (term trees) as nodes, and derivation trees have rewriting trees as nodes. For a given P and A, the derivation tree for P and A is unique up to
renaming. We use our usual notation D(w) to refer to the node of D at index w ∈ L.
Definition 8 (Existential liveness, observational productivity). Let P be a universally observable program and let A be an atom. An S-resolution derivation for P and
A is live if it constitutes an infinite branch of the derivation tree for P and A. The program P is existentially live if there exists a live S-resolution derivation for P and some
atom A. P is observationally productive if it is universally observable and existentially
live.
To show that observational productivity is semi-decidable, we first show that universal
observability is semi-decidable by means of a finite (i.e., decidable) guardedness check.
We started this section by motivating the need to construct a finite set S of rewriting
trees checking guardedness of which will guarantee guardedness for any rewriting tree for
the given program. Our first logical step is to use derivation trees built for clause heads
as generators of such a set S. Due to the properties of mgu’s used in forming branches
of derivation trees, derivation trees constructed for clause heads generate the set of most
general rewriting trees. The next lemma exposes this fact:
Lemma 1 (Guardedness of derivation trees implies universal observability).
Given a program P , if derivation trees for P and each head (P (i)) are guarded, then P
is universally observable.
However, derivation trees are infinite, in general. So it still remains to define a method
that extracts representative finite subtrees from such derivation trees; we call such subtrees observation subtrees. For this, we need only be able to detect an invariant property
guaranteeing guardedness through tree transitions in the given derivation tree. To illustrate, let us check guardedness of the program P6 . As it consists of just one clause,
we take the head of that clause as the goal atom, and start constructing the infinite
derivation tree D for P6 and from(X, scons(X, Y)) as shown in Figure 2. The first rewriting tree in the derivation tree has no loops, so we cannot identify any invariants. We
make a transition to the second rewriting tree which has one loop (underlined) involving
the recursive reducing subterm [s(X), Y′ ]. This reducing subterm is our first candidate
invariant, it is the pattern that is consumed from the root of the second rewriting tree to
its leaf. We now need to check this pattern is added back, or produced, in the next tree
transition. The next mgu involves substitution Y′ 7→ [s(s(X)), Y′′ ]. Because this derivation
gradually computes an infinite irrational term (rational terms are terms that can be represented as trees that have a finite number of distinct subtrees), the two terms [s(X), Y′ ]
and [s(s(X)), Y′′ ] we have identified are not unifiable. We need to be able to abstract away
Y 7→[s(X),Y ′ ]}
Y ′ 7→[s(s(X)),Y ′′ ]
−→
Y ′′ 7→[s(s(s(X))),Y ′′ ]
−→
′
−→
fr(X, [X, Y])
fr(X, [X, s(X), Y ])
fr(X, [X, s(X), s(s(X)), Y ])
P6 (0)
P6 (0)
P6 (0)
fr(s(X), Y)
′
fr(s(X), [s(X), Y ])
fr(s(X), [s(X), s(s(X)), Y′′ ])
P6 (0)
P6 (0)
′
fr(s(s(X)), Y )
...
′′
fr(s(s(X)), [s(s(X)), Y′′ ])
P6 (0)
fr(s(s(s(X))), Y′′ )
Fig. 2. An initial fragment of the infinite derivation tree D for the program P6 from Example 4
and its clause head. It is also the observation subtree of D. We abbreviate scons by [, ], and
from by fr. The guarded loops in each of its rewriting trees are underlined.
from their current shape and identify a common pattern, which is [ , ]. By the properties
of mgu’s used in transitions, such most general pattern can always be extracted from the
clause head itself. Indeed, the subterm of the clause head from(X, scons(X, Y)) has the
subterm [X, Y] that is exactly the pattern we look for. Thus, our current (coinductive)
assumption is: given a rewriting tree T in the derivation tree D, [X, Y] will be consumed
by rewriting steps from its root to its leaves, and exactly [X, Y] will be produced (i.e.,
added back) in the next tree transition. X and Y are seen as placeholders for some terms.
Consumption is always finite (by the loop guardedness), and production is potentially
infinite.
We now need to check that this coinductive assumption will hold for the next rewriting
tree of D. The third rewriting tree indeed has guarded loops with recursive reducing
subterm [s(s(X)), Y′′ ], and the next mgu it gives rise to is Y′′ 7→ [s(s(s(X))), Y′′ ]. Again, to
abstract away the common pattern, we look for a subterm in the clause head of P6 (0) that
matches with both of these terms, it is the same subterm [X, Y]. Thus, our coinductive
assumption holds again, and we conclude by coinduction that the same pattern will hold
for any further rewriting tree in D. When implementing this reasoning, we take the
observation subtree of D up to the third tree shown in Figure 2 as a sufficient set of
rewriting trees to check guardedness of (otherwise infinite) D.
The rest of this section generalises and formalises this approach. In the next definition,
we introduce the notion of a clause projection to talk about the process of “abstracting
away” a pattern from an mgu σ by matching it with a subterm t of a clause head. When
t also matches with a recursive reducing subterm of a loop in a rewriting tree, we call t
a coinductive invariant.
Definition 9 (Clause projection and coinductive invariant). Let P be a program
and A be an atom, and let D be a derivation tree for P and A in which a tree transition
from T to T ′ is induced by an mgu σ of some P (k) and an atom B given by a leaf node
T (u).
The clause projection for T ′ , denoted π(T ′ ), is the set of all triples (P (k), t, v), where t
is a subterm of head (P (k)) at position v, such that the following conditions hold: σ(B)⊲B
with variable reducing subterm t′ , and t′ matches against t (i.e. t′ = σ ′ (t) for some σ ′ ).
Additionally, the coinductive invariant at T ′ , denoted ci(T ′ ), is a subset of the clause
projection for T ′ , satisfying the following condition. An element (P (k), t, v) ∈ π(T ′ ) is
also in ci(T ′ ), if T contains a loop in the branch leading from T ’s root to T (u) that is
guarded by (P (k), t′′ ) for some t′′ such that t′′ matches against t (t′′ = θ(t) for some θ).
Given a program P , an atom ASand a derivation tree D for P and A, the clause
projection set
S for D is cproj(D) = T π(T ) and the coinductive invariant set for D is
cinv(D) = T ci(T ), where these unions are taken over all rewriting trees T in D.
Example 10 (Clause projections and coinductive invariants). Coming back to Figure 2,
the mgu for the first transition is σ1 = {X′ 7→ s(X), Y 7→ scons(s(X), Y′)} (renaming
of variables in P6 (0) with primes), that for the second is σ2 = {X′′ 7→ s(s(X)), Y′ 7→
scons(s(s(X)), Y′′)} (renaming of variables in P6 (0) with double primes), etc. Clause
projections are given by π(T ) = {(P6 (0), scons(X, Y), 1)} for all trees T in this derivation,
and thus cproj(D) is the finite set. Moreover, for the first rewriting tree T , ci(T ) = ∅, and
ci(T ′ ) = {(P6 (0), scons(X, Y), 1)} for all trees T ′ except for the first one, so cinv(D) =
{(P6 (0), scons(X, Y), 1)} is the finite set too.
The clause projections for the derivation of Figure 1 are π(T ′ ) = π(T ′′′ ) = (P (1),
s(Y1), 2), and π(T ′′ ) = (P (0), s(X1), 0), where T ′ , T ′′ , T ′′′ refer to the second, third and
fourth rewriting tree of that derivation. All coinductive invariants for that derivation are
empty, since none of these rewriting trees contain guarded loops.
Generally, clause projection sets are finite, as the number of subterms in the clause
heads of P is finite. This property is crucial for termination of our method:
Proposition 2 (Finiteness of clause projection sets). Given a program P , an atom
A, and a derivation tree D for P and A, the clause projection set cproj(D) is finite.
In particular, this holds for derivation trees induced by clause heads.
We terminate the construction of each branch of a derivation tree when we notice
repeating coinductive invariant. A subtree we get as a result is an observation subtree.
Formally, given a derivation tree D for a program P and an atom A, with a branch in
which nodes D(w) and D(wv) are defined, if ci(D(w)) = ci(D(wv)) 6= ∅, then D has
a guarded transition from D(w) to D(wv) (denoted D(w) =⇒ D(wv)). Every guarded
transition thus identifies a repeated “consumer-producer” invariant in the derivation from
D(w) to D(wv). This tells us that observation of this branch of D can be concluded.
Imposing this condition on all branches of D gives us a general method to construct
finite observation subtrees of potentially infinite derivation trees:
Definition 10 (Observation subtree of a derivation tree). If D is a derivation
tree for a program P and an atom A, the tree D′ is the observation subtree of D if
1) the roots of D and D′ are given by the rewriting tree for P and A, and
2) if w is a node in both D and D′ , then the rewriting trees in D and D′ at node w are
the same and, for every child w′ of w in D, the rewriting tree of D′ at node w′ exists
and is the same as the rewriting tree of D at w′ , unless either
a) GC2 does not hold for D(w′ ), or
b) there exists a v such that D(v) =⇒ D(w).
In either case, D′ (w) is a leaf node. We say that D′ is unguarded if Condition 2a holds
for at least one of D’s nodes, and that D′ is guarded otherwise.
A branch in an observation subtree is thus truncated when it reaches an unguarded
rewriting tree or its coinductive invariant repeats. The observation subtree of any derivation tree is unique. The following proposition and lemma prove the two most crucial
properties of observation subtrees: that they are always finite, and that checking their
guardedness is sufficient for establishing guardedness of the whole derivation trees.
Proposition 3 (Finiteness of observation subtrees). If D is a derivation tree for
a program P and an atom A then the observation subtree of D is finite.
Lemma 2 (Guardedness of observation subtree implies guardedness of derivation tree). If the observation subtree for a derivation tree D is guarded, then D is
guarded.
Example 11 (Finite observation subtree of an infinite derivation tree). The initial fragment D′ of the infinite derivation tree D given by the three rewriting trees in Figure 2
is D’s observation subtree. The third rewriting tree T ′′ in D is the last node in the observation tree D′ because ci(T ′ ) = ci(T ′′ ) = {(P6 (0), scons(X, Y), 1)} =
6 ∅. Since D′ is
guarded, Lemma 2 above ensures that the whole infinite derivation tree D is guarded.
It now only remains to put the properties of the observation subtrees into practical
use, and, given a program P , construct finite observation subtrees for each of its clauses. If
none of these observation subtrees detects unguarded rewriting trees, we have guarantees
that this program will never give rise to infinite rewriting trees. The next definition,
lemmas and a theorem make this intuition precise.
Definition 11 (Guarded clause, guarded program). Given a program P , its clause
P (i) is guarded if the observation subtree for the derivation tree for P and the atom
head (P (i)) is guarded, and P (i) is unguarded otherwise. A program P is guarded if each
of its clauses P (i) is guarded, and unguarded otherwise. We write GC3(P (i)) to indicate
that P (i) is guarded, and similarly for P .
Lemma 3 uses Proposition 3 to show that GC3 is decidable.
Lemma 3 (GC3 is decidable). GC3 is a decidable property of logic programs.
Theorem 1 (Universal observability is semi-decidable). If GC3(P ) holds, then P
is universally observable.
Proof: If GC3(P ) holds, then the observation subtree for each P (i) is guarded. Thus,
by Lemma 2, the derivation tree for each P (i) is guarded. But then, by Lemma 1, P
is universally observable. Combining this with Lemma 3, we also obtain that universal
observability is semi-decidable.
The converse of Theorem 1 does not hold: the program comprising the clause p(a) ←
p(X) is universally observable but not guarded, hence the above semi-decidability result.
From our check for universal observability we obtain the desired check for existential
liveness, and thus for observational productivity:
Corollary 1 (Observational productivity is semi-decidable). Let P be a guarded
logic program. If there exists a clause P (i) such that the derivation tree D for P and P (i)
has an observation subtree D′ one of whose branches was truncated by Condition 2b of
Definition 10, then P is existentially live. In this case, since P is also guarded and hence
universally observable, P is observationally productive.
5
Related Work: Termination Checking in TRS and LP
Because observational productivity is a combination of universal observability and existential liveness, and the former property amounts to termination of all rewriting trees,
there is an intersection between this work and termination checking in TRS [22,1,8].
Termination checking via transformation of LP into TRS has been given in [20]. Here
we consider termination of restricted form of SLD-resolution (given by rewriting derivations), therefore a much simpler method of translation of LP into TRS can be used for
our purposes [6]: Given a logic program P and a clause P (i) = A ← B1 , . . . , Bn containing no existential variables, we define a rewrite rule A → fi (B1 , . . . , Bn ) for some fresh
function symbol fi . Performing this translation for all clauses, we get a translation from
P to a term-rewriting system TP . Rewriting derivations for P can be shown operationally
equivalent to term-rewriting reductions for TP ; see [6] for a proof. Therefore, for logic
programs containing no existential variables, any termination method from TRS may be
applied to check universal observability (but not existential liveness).
Algorithmically, our guardedness check compares directly with the method of dependency pairs due to Arts and Giesl [1,8]. Consider again TP obtained from a program
P . The set R of dependency pairs contains, for each rewrite rule A → fi (B1 , . . . , Bn )
in TP , a pair (A, Bj ), j = 1, . . . , n; see [6]. The method of dependency pairs consists of
checking whether there exists an infinite chain of dependency pairs (si , ti )i=1,2,3,... such
that σi (ti ) →∗ σi+1 (si+1 ). If there is no such infinite chain, then TP is terminating. Again
this translation from LP to dependency pairs in TRS is simpler than in [15], as rewriting
derivations are a restricted form of SLD-resolution. Due to the restricted syntax of TP
(compared to the general TRS syntax), generating the set of dependency pairs is equivalent to generating a set of rewriting trees for each clause of P and assuming σi = σi+1 (cf.
our GC2). To find infinite chains, a dependency graph is defined, in which dependency
pairs are nodes and arcs are defined whenever a substitution that allows a transition
from one pair to another can be found. Finding such substitutions is the hardest part
algorithmically. Note that every pair of neighboring and-nodes in a rewriting tree corresponds to a node in a dependency graph. Generating arcs in a dependency graph is
equivalent to using our GC3 to find a representative set of substitutions. However, the
way GC3 generates such substitutions via rewriting tree transitions differs completely
from the methods approximating dependency graphs [1,22], and relies on the properties
of S-resolution, rather than recursive path orderings. This is because GC3 additionally
generates coinductive invariants for checking existential liveness of programs.
Conceptually, observational productivity is a new property that does not amount
to either termination or nontermination in LP or TRS. E.g. programs P3 and P4 are
nonterminating (seen as LP or TRS), and P8 : p(X) ← q(Y ) is terminating (seen as LP
and TRS) but none of them is productive. This is why the existing powerful tools (such
as AProVE) and methods [1,8,15,20] that can check termination or nontermination in
TRS or LP are not sufficient to serve as productivity checks. To check termination of
rewriting trees, GC3 can be substituted by existing termination checkers for TRS, but
none of the previous approaches can semi-decide existential liveness as GC3 does.
6
Implementation and Applications
We implemented the observational productivity checker in parallel Go (golang.org) [19],
which allows to experiment with parallelisation of proof search [10]. Loading a logic
program P , one runs a command line to initialise the GC3 check. The algorithm then
certifies whether or not the program is guarded (and hence universally observable). If that
is the case, it also checks whether GC3 found valid coinductive invariants, i.e. whether
P is existentially live and hence admits coinductive interpretations for some predicates.
Appendix B (available in online version) gives further details.
In the context of S-resolution [11,9], observational productivity of a program is a
pre-condition for (coinductive) soundness of S-resolution derivations. This gives the first
application for the productivity checker. But the notion of global productivity (as related
to computations at infinity [14]) is a general property tracing its roots to the 1980s. A
program is productive, if it admits SLD- or S-resolution derivations that compute (or
produce) an infinite term at infinity. Thus the productivity checker has more general
practical significance for Prolog. In this paper we further exposed its generality by showing that productivity can be seen as a general property of logic programs, rather than
property of derivations in some special dialect of Prolog.
Based on this observation, we identify three applications for productivity checks encompassing the S-resolution framework. (1) In the context of CoLP [7,21] or any other
similar tool based on loop detection in SLD-derivations, one can run the observational
productivity checker for a given program prior to running the usual interpreter of CoLP.
If the program is certified as productive, all computations by CoLP for this program will
be sound relative to the computations at infinity [14]. It gives a way to characterise a
subset of theorems proven by CoLP that describe the process of production of infinite
data. I.e., as explained in Introduction, CoLP will return answers for programs P3 , P4 and
P5 . But if we know that only P5 is productive, we will know that only CoLP’s answers
for P5 will correspond to production of infinite terms at infinity.
(2) As our productivity checker also checks liveness of programs, it effectively identifies
which predicates may be given coinductive semantics. This knowledge can be used to
type predicates as inductive or coinductive. We can use these types to mark predicates
in CoLP or any other coinductive dialect of logic programming, cf. Appendix B.
(3) Observational productivity is also a guarantee that a sequence of mgus approximating the infinite answer can be constructed lazily even if the answer is irregular. E.g. our
running example of program P6 is irrational and hence cannot be handled by CoLP’s
loop detection. But even if we cannot form a closed-term answer for a query from(0, X),
the productivity checker gives us a weaker but more general certificate that lazy approximation of our infinite answer is possible.
These three groups of applications show that the presented productivity checker can
be implemented and applied in any dialect of logic programming, irrespective of the fact
that it initially arose from S-resolution research [11,9].
7
Conclusions
In this paper we have introduced an observational counterpart to the classical notion
of global productivity of logic programs. Using the recently introduced formalism of Sresolution, we have defined observational productivity as a combination of two program
properties, namely, universal observability and existential liveness. We have introduced
an algorithm for semi-deciding observational productivity for any logic program. We did
not impose any restrictions on the syntax of logic programs. In particular, our algorithm
handles both existential variables and non-linear recursion.
The algorithm relies on the observation that rewriting trees for productive and guarded
programs must show term reduction relative to a contraction ordering from their roots
to their leaves. But S-resolution derivations involving such trees can only proceed by
adding term structure back in transitioning to new rewriting trees via mgus. This “producer/consumer” interaction can be formally traced by observing a derivation’s coinductive invariants: these record exactly the term patterns that both reduce in the loops of
rewriting trees and are added back in transitions between these trees.
References
1. T. Arts and J. Giesl. Termination of term rewriting using dependency pairs. TCS,
236(12):133 – 178, 2000.
2. B. Courcelle. Fundamental properties of infinite trees. TCS, 25:95–169, 1983.
3. D de Schreye and S. Decorte. Termination of logic programs: the never-ending story. Journal
of Logic Programming, 19–20, Supplement 1:199–260, 1994.
4. J. Endrullis et al. Productivity of stream definitions. TCS, 411(4-5):765–782, 2010.
5. J. Endrullis et al. A coinductive framework for infinitary rewriting and equational reasoning.
In RTA, pages 143–159, 2015.
6. P. Fu and E. Komendantskaya. Operational semantics of resolution and productivity in
Horn clause logic. Formal Aspects of Computing, 2016.
7. G. Gupta et al. Coinductive logic programming and its applications. In ICLP, pages 27–44,
2007.
8. N.N. Hirokawa and A. Middeldorp. Dependency pairs revisited. In RTA, pages 249–268,
2004.
9. P. Johann et al. Structural resolution for logic programming. In Technical Communications
of ICLP, 2015.
10. E. Komendantskaya et al. Exploiting parallelism in coalgebraic logic programming. ENTCS,
(33):121–148, 2014.
11. E. Komendantskaya et al. Coalgebraic logic programming: from semantics to implementation. Journal of Logic and Computation, 26(2):745–783, 2016.
12. E. Komendantskaya and P. Johann. Structural resolution: a framework for coinductive proof
search and proof construction in Horn clause logic. Submitted, 2015.
13. K. R.M. Leino and M. Moskal. Co-induction simply - automatic co-inductive proofs in a
program verifier. In FM, pages 382–398, 2014.
14. J.W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 2nd edition, 1988.
15. M.T. Nguyen et al. Termination analysis of logic programs based on dependency graphs.
In LPOSTR 2007, pages 8–22, 2007.
16. F. Pfenning. Types in Logic Programming. The MIT Press, 1992.
17. A. Reynolds and J. Blanchette. A decision procedure for (co)datatypes in SMT solvers. In
CADE, pages 197–213, 2015.
18. E. Rohwedder and F. Pfenning. Model and termination checking for higher-order logic
programs. In ESOP, pages 296–310, 1996.
19. M. Schmidt. Productivity checker for LP, www.macs.hw.ac.uk/∼ek19/CoALP/, 2016.
20. P. Schneider-Kamp et al. Automated termination analysis for logic programs by term
rewriting. In LOPSTR, pages 177–193, 2006.
21. L. Simon et al. Co-logic programming: Extending logic programming with coinduction. In
ICALP, pages 472–483, 2007.
22. Terese. Term Rewriting Systems. Cambridge University Press, 2003.
Appendix A. Proofs
Proof of Proposition 1: GC2 is decidable
Any rewriting tree T is either finite or infinite. If T is finite, then its guardedness is
clearly decidable. So, we may, without loss of generality, assume T is infinite. If T is
infinite then it must have an infinite branch B. We now show that an infinite branch B
in T must necessarily contain an unguarded loop. Thus, whether T is finite or infinite,
its guardedness is decidable.
Assume B has only guarded loops, but it is infinite. Since the number of clauses and
the number of function symbols in Σ are finite, B must contain an infinite number of
loops. Consider one such infinite sequence q(t11 , ..., t1j ) → . . . → q(tk1 , ..., tkj ) → . . . →
q(tl1 , ..., tlj ) → . . ., where q(t11 , ..., t1j ), q(tk1 , ..., tkj ), q(tl1 , ..., tlj ), ... are all atoms with
the same predicate q obtained by rewriting using clause P (i). Because all loops in B
are guarded, we have q(t11 , ..., t1j ) ⊲ q(tk1 , ..., tkj ), q(t11 , ..., t1j ) ⊲ q(tl1 , ..., tlj ),... But since
q(t11 , ..., t1j ) is finite, there are only finitely many ways to construct a reducing subterm
on it. Thus, there will be a point when some terms q(tm1 , ..., tmj ) and q(tn1 , ..., tnj )
in the infinite sequence in B have the same recursive reducing subterm t∗ relative to
q(t11 , ..., t1j ).
Now, there are two cases: i) q(t11 , ..., t1j )⊲q(tm1 , ..., tmj ) and q(t11 , ..., t1j )⊲q(tn1 , ..., tnj )
hold, but q(tm1 , ..., tmj ) ⊲ q(tn1 , ..., tnj ) does not, and ii) the negation of this case.
– If q(t11 , ..., t1j )⊲q(tm1 ...tmj ) and q(t11 , ..., t1j )⊲q(tn1 , ..., tnj ) hold, but not q(tm1 , ..., tmj )⊲
q(tn1 , ..., tnj ), then q(tm1 , ..., tmj ) ⊲ q(tn1 , ..., tnj ) is an unguarded loop, which contradicts the assumption that all loops in B are guarded.
– If the negation holds — i.e., if q(tm1 , ..., tmj ) ⊲ q(tn1 , ..., tnj ) holds or q(t11 , ..., t1j ) ⊲
q(tm1 , ..., tmj ) does not hold or q(t11 , ..., t1j )⊲q(tn1 , ..., tnj ) does not hold — then there
are three cases. If either q(t11 , ..., t1j ) ⊲ q(tm1 , ..., tmj ) or q(t11 , ..., t1j ) ⊲ q(tn1 , ..., tnj )
does not hold, then the existence of this unguarded loop in B gives a contradiction.
So we need only consider the case when q(t11 , ..., t1j ) ⊲ q(tm1 , ..., tmj ), q(t11 , ..., t1j ) ⊲
q(tn1 , ..., tnj ), and q(tm1 , ..., tmj ) ⊲ q(tn1 , ..., tnj ) are all guarded loops in B. Let t∗∗
be the recursive reducing subterm for the loop q(tm1 , ..., tmj ) ⊲ q(tn1 , ..., tnj ). Since
the same recursive reducing subterm cannot be contracted twice along the same
path from q(t11 , ..., t1j ), we must have that t∗ 6= t∗∗ and, moreover, contracting t∗∗
must somehow “restore” t∗ to q(t11 , ..., t1j ). And this means that t∗ and t∗∗ must
be “independent”, in the sense of being on independent paths in q(t11 , ..., t1j ). But
then there will be cycles of terms in B in which one argument of q decreases in one
step and another independent one grows, and then the first argument grows while
the other one decreases. So q(t11 , ..., tij ) will appear in the infinite branch infinitely
many times (so definitely more than once!), and B will thus contain an unguarded
loop in this case as well. This is again a contradiction.
Proof of Lemma 1: Guardedness of derivation trees implies universal
observability
If P is not universally observable, then there exists an atom A such that the rewriting tree
T for P and A is infinite. Moreover, A must match some clause head (P (i)) via a mgm θ,
so in fact T is an infinite rewriting tree for P and θ(head (P (i))), with additional condition
that θ is also applied to all atoms of this tree. Then, as in the proof of Proposition 1,
there must exist an unguarded loop L on an infinite branch B of T . We claim that, if we
construct a derivation tree Di for the program P and head (P (i)), then some rewriting
tree in Di will contain an unguarded loop. Let us consider the construction of Di .
If the first rewriting tree of Di , i.e. the tree T ′ for P and head (P (i)) does not itself
contain an unguarded loop, then the branch in T ′ corresponding to B in T must have
a leaf node T ′ (w) given by an atom that unifies with a clause P (k1 ) via mgu σ1 , say.
Moreover, P (k1 ) is exactly the clause used to construct a node T (wi) of B in T via
its mgm with T (w). Now, consider the rewriting tree transition determined by the mgu
σ1 , i.e. consider T ′ → Tw′ . If the branch corresponding to B in Tw′ does not contain an
unguarded loop, then it too must have a leaf node Tw′ (u) that unifies with P (k2 ) via
mgu σ2 , say, and P (k2 ) is exactly the clause used to construct a node T (vj) of B in T
via its mgm with T (v). And so on. After some finite number n of tree transitions, of
“growing” the branch corresponding to B in T ′ by taking further mgu’s on its leaves,
we must come to a rewriting tree T ∗ for P and the root atom σ(head (P (i))) in Di that
contains an unguarded loop corresponding to L, and where σ = σn ◦ ... ◦ σ1 for the mgu’s
σ1 , ..., σn involved in the tree transitions in Di . Indeed, since branches B of T and T ∗
are constructed using mgm’s with exactly the same clauses at each step, and since the
mgu’s σ1 , ..., σn are all most general unifiers, we must have that σ is more general than
θ, and thus T ∗ is a more general version of T , and so contains an unguarded loop that is
a more general version of L.
Proof of Proposition 3: Finiteness of observation subtrees
Let D′ be the observation subtree of the derivation tree D for a program P and an atom
A. If D is finite, then D′ will necessarily be finite, so we may, without loss of generality,
suppose D is infinite.
If there exists a rewriting tree in D that is unguarded, then, by Condition 2a of
Definition 10, the branch of D on which that tree appears will end at that tree in D′
and will thus be finite. For D′ to be infinite, there must exist an infinite branch of D
containing only guarded rewriting trees such that coinductive invariants computed in
that branch never repeat. In fact, every infinite branch of D′ must satisfy these two
conditions.
Let T be any guarded rewriting tree on any infinite branch of D. We first note that
the coinductive invariant ci(T ) must be non-empty. In addition, T must itself be finite.
Indeed, if T were infinite then, by the completeness of breadth-first search, an unguarded
rewriting tree would have to exist at some finite depth on T ’s branch of D. Then, by the
argument of the preceding paragraph, T ’s branch of D would have to be finite. But this
is not the case.
So T must be a finite, guarded rewriting tree appearing on an infinite branch of D.
Now, although D itself is infinite, Proposition 2 ensures that D’s coinductive invariant
set still contains only finitely many clause projections, so any branch of D can add
only finitely many distinct elements to D’s coinductive invariant set. In particular, the
coinductive invariants for nodes on T ’s infinite branch of D must eventually all be equal.
Moreover, since ci(T ) 6= ∅, these coinductive invariants must eventually all be non-empty.
Thus Condition 2b of Definition 10 must eventually be satisfied and the branch of D′
corresponding to T ’s branch in D must thus be finite. Having argued that the branch of
D′ corresponding to any infinite branch of D is finite, we have that D′ is itself finite.
Proof of Lemma 2: Guardedness of observation subtree implies guardedness
of derivation tree
The proof proceeds by induction-coinduction. We assume the observation subtree D′ for
D is guarded and inductively examine every branch B ′ of D′ . This is possible because the
number of such branches and their lengths are all finite by Proposition 3. For any such
B ′ , either no parent of any leaf in the last coinductive tree of B ′ can be resolved with any
clause of P , or B ′ was terminated by Condition 2b of Definition 10. In the former case,
the entire branch B ′ will also appear in D, and each rewriting tree on the corresponding
branch B of D will be guarded. In the latter case, we can proceed coinductively.
If B ′ was terminated because it contains a guarded transition T =⇒ T ′ for T = D(w)
and T ′ = D(wv), then both T and T ′ were formed by resolving with some clause P (k).
In this case, we apply the following coinductive argument. Coinductive Hypothesis (CH):
The process of resolving with clause P (k) to produce a new guarded rewriting tree whose
coinductive invariant has first component P (k) can be repeated infinitely many times
in transition sequences originating from T . By computing that CH is again satisfied
for T ′ , we can make the following Coinductive Conclusion (CC): For any tree T in
any branch B containing B ′ , the process of resolving with clause P (k) to produce a
new guarded rewriting tree whose coinductive invariant has first component P (k) can be
repeated infinitely many times in transition sequences originating from T . So each of the
rewriting trees in B must be guarded.
Unfortunately, CC does not guarantee that no unguarded loop can possibly occur
in D by resolving with other clauses in the sequence of transitions from T ′ that occur in B but not in B ′ . But if it is possible to compute a sequence of rewriting tree
transitions in D from T ′ involving mgus θ1 , . . . , θn computed by resolving with clauses
P (k1 ), . . . , P (kn ) that lead to an unguarded rewriting tree in B, then, by completeness
of the breadth-first construction of the derivation tree D, there must be a a rewriting
tree T ∗ occurring in the sequence of rewriting trees in B ′ from T to T ′ that leads to
a sequence of rewriting tree transitions in another branch B ′′ of D involving exactly
the same sequence P (k1 ), . . . , P (kn ) of clauses and mgus θ1′ , . . . , θn′ such that, for each
i ∈ {1, ..., n}, θi′ = σi ◦ θi for some σi . This holds because rewriting tree transitions only
lead to further instantiations of variables, and the rewriting tree T ∗ appears earlier on
B than T ′ does, and hence is more general. But then an unguarded loop induced by the
mgus θ1′ , . . . , θn′ obtained by resolving with P (k1 ), . . . , P (kn ) will be found in one of the
branches of D′ to which T ∗ leads.
By inducting on all branches of the observation subtree D′ of D, and coinductively
terminating each, we conclude that if all branches of D′ are terminated by the above
coinductive argument with no unguarded rewriting tree being found, then no unguarded
loop can exist in any of the rewriting trees of D.
Proof of Lemma 3: GC3 is decidable
To decide guardedness of logic programs, we must let P be given and construct a set
of derivation trees, one derivation tree for each clause head of P , i.e. every such Di
is a derivation tree for P and head (P (i)). Moreover, we build these trees only until
we construct the observation subtree Di′ for each Di . We next check whether or not
each observation subtree Di′ is guarded. That is, we must check whether or not every
rewriting tree in Di′ is guarded and whether or not condition 2.a of Definition 10 was used
to construct Di′ . Since guardedness of rewriting trees is decidable by Proposition 1, and
there are only finitely many rewriting trees in any observation tree Di′ , guardedness of all
observation subtrees Di′ for this program is decidable. Since, by Lemma 2, guardedness
of observation subtrees implies guardedness of derivation trees, guardedness of P is also
decidable.
Appendix B. Implementation of Observational Productivity
Checks
Algorithmic overview of observational productivity checking
Definitions of contraction ordering, guarded rewriting trees, and observation subtrees
translate naturally into algorithmic forms that give rise to the implementation of our
observational productivity checks [19]. Below we give a high-level pseudocode representation of the formal definitions of this paper.
Algorithm 1 below captures the essence of our check that GC3(P ) holds for a logic
program P . It depends on the definition of the observation subtree (Definition 10), which
in turn depends on two conditions:
– finiteness of observation subtrees, as proven in Proposition 3, and
– guardedness of every rewriting tree in a program’s observation subtree.
These two conditions ensure termination of Algorithm 1, as expressed formally in Lemma 3.
In the main body of this paper we have written GC2(T ) to indicate that a rewriting tree
T is guarded. A pseudocode description of our check that GC2(T ) holds for a rewriting
tree T is given in Algorithm 2.
Termination of Algorithm 2 depends crucially on Proposition 1, i.e., on the fact that
it is impossible to construct an infinite rewriting tree without finding unguarded loops.
Algorithm 2 in turn relies on an algorithmic check that two terms are related via a contraction ordering, but we omit specifying this in pseudocode since it is entirely straightforward.
Implementation
Our observational productivity checker is implemented in Go (golang.org) as a command line program and is part of the general implementation of structural resolution
and coalgebraic logic programming (CoALP) [19]. Go was chosen as implementation language because it provides easy primitives for parallelization, which has been explored
Algorithm 1 Observational productivity check for a logic program
Require: P – a logic program over signature Σ
Require: LC – an empty list
n = number of clauses in P
for i = 0, . . . , n do
if observation subtree D′ of the derivation tree D for P and head (P (i)) is unguarded then
P (i) is not guarded.
else
P (i) is guarded.
if D′ contains transition D(v) =⇒ D(w) with coinductive invariant c then
LC := append(LC, c)
end if
end if
end for
if all P (i) are guarded then
Result1 := “P is guarded”
else
Result1 := “P is not guarded”
end if
if LC is not empty then
Result2 := “P is existentially live with LC”
else
Result2 := “P has finite derivations only”
end if
return (Result1, Result2)
Algorithm 2 Guardedness check in a rewriting tree
Require: T – the rewriting tree for a logic program P and an atom A
for i = 0, . . . , depth(T ) do
for nodes w1 , . . . wm at depth i do
if a node wj forms a loop with some node v above it then
if loop(T, v, wj ) is not guarded then
return “T is not guarded”
end if
end if
end for
end for
return “T is guarded”
to optimize proof search [10]. To compile and install the productivity checker follow the
instructions in the README file supplied in the program distribution available at [19].
CoALP can be used not only to check the productivity of logic programs, but to make
queries to guarded such programs as well. The checker takes Prolog-style programs saved
in text files as input. The format of programs corresponds exactly to that of Prolog. For
example, program P6 is represented as
from(X, scons(X, Y)) :- from(s(X), Y).
Unlike Prolog, our checker does not support built-in predicates or arithmetic functions.
To check a logic program for observational productivity, the path to the program file
has to be given as the first parameter:
guardcheck somefile.logic
The above command initialises the GC3(P ) check for a given logic program P in somefile.logic,
and, as GC3 involves computations of coinductive invariants, it simultaneously uses them
to detect existential liveness, as detailed in Algorithm 1. Many example files and tests as
well as the programs used in this paper can be found in the directory named “examples”
in [19].
The output for the observationally productive program P6 is:
Program is guarded.
Program is existentially live with coinductive invariants:
in clause 0 of "from": [{0 | scons(v3,v5) | [1]}]
Note that the first 0 in [0 | scons(v3,v5) | [1]] points to the clause P6 (0), and
would suggest that the predicate from in the head of this clause is a good candidate
to be given coinductive semantics and hence coinductive typing. We believe that this
general information can be used by CoLP or CoALP to determine typing for coinductive
predicates in their programs.
The output for the unguarded program P7 is:
Program is not guarded.
Goal q(s(v34),s(v34),s(v42),v36) results in unguarded loop
in path [(p:0), (q:0), (p:0)].
A more complex example
In this section, we consider a more challenging example of Sieve of Eratosthenes, known
for its difficulty in the literature on coinductive definitions [7]. The following program
P9 is an observationally productive reformulation of the original Sieve of Eratothsenes
program from [7]:
0. prime(X) ← inflist(I), sieve(I, L), member(X, L)
1. sieve(cons(H, T), cons(H, R)) ← filter(H, T, F), sieve(F, R)
2. filter(H, cons(K, T), cons(K, T1)) ← mod(X, K, H), less(0, X), filter(H, T, T1)
3. filter(H, cons(K, T), T1) ← mod(0, K, H), filter(H, T, T1)
4. int(X, cons(X, Y)) ← int(s(X), Y, Z1)
5. inflist(I) ← int(s(s(0)), I)
6. member(X, cons(X, L)) ←
7. member(X, cons(Y, L)) ← member(X, L)
8. less(0, s(X)) ←
9. less(s(X), s(Y)) ← less(X, Y)
The original program [7] does not use Prolog-style list notations and uses a structural
representation of numbers, which we avoid. We also assume a suitable implementation
of the modulo operator as the predicate mod above.
If we run the observational productivity check on this program, we obtain the following output:
Program is guarded.
Program is existentially live with coinductive invariants:
in clause 0 of "filter": [{0 | cons(v16,v17) | [1]} {0 | cons(v16,v18) | [2]}]
in clause 1 of "filter": [{1 | cons(v24,v25) | [1]}]
in clause 0 of "sieve": [{0 | cons(v10,v11) | [0]} {0 | cons(v10,v12) | [1]}]
in clause 1 of "member": [{1 | cons(v37,v38) | [1]}]
in clause 1 of "less": [{1 | s(v40) | [0]} {1 | s(v41) | [1]}]
Above, four predicates have been identified as potentially having a coinductive semantics: filter, sieve, member and less. Generally, most inductive definitions admit
coinductive interpretation, and predicates that we intuitively consider as inductive may
be identified as potentially coinductive. This situation was analysed in Example 2. Among
the four predicates, sieve, that admits only coinductive interpretation, was identified.
We note that the original formulation of [7] is not observationally productive, since it
does not possess universal observability property. If we run our checker on the formulation
in [7], failure of observational productivity is detected and reported as follows:
Program is not guarded.
Goal comember(v136,v140) results in unguarded loop
in path [(primes:0), (comember:0), (comember:0)].
As indicated by the checker output, the reason is that the following definition of comember
used in [7] is not universally observable and hence is not guarded:
0. comember(X, L) ← drop(X, L, L1), comember(X, L1)
1. drop(H, cons(H, T), T) ←
2. drop(H, cons(H, T), T) ← drop(H, T, T1)
Indeed, the definition of comember in Clause 0 above is not guarded by any constructors.
In our reformulation as program P9 above, we use a guarded definition of member
instead of the definitions of comember and drop used in [7]. The definition of member is
guarded by the constructor cons in Clause 7. Thus, in the case of the Sieve of Eratothsenes, the transition from an unproductive to a productive coinductive definition was a
simple matter of applying a program transformation that clearly preserves the intended
coinductive meaning of the coinductive definition of seive in Clause 1.
| 6cs.PL
|
arXiv:1801.03160v2 [cs.AI] 21 Mar 2018
A Formalization of Kant’s Second
Formulation of the Categorical Imperative
Felix LINDNER a,1 Martin Mose BENTZEN b
Group on the Foundations of Artificial Intelligence, Department of Computer
Science, Albert-Ludwigs-Universität Freiburg, Germany
b Management Engineering, Technical University of Denmark, Lyngby, Denmark
a Research
Abstract. We present a formalization and computational implementation of the
second formulation of Kant’s categorical imperative. This ethical principle requires
an agent to never treat someone merely as a means but always also as an end. Here
we interpret this principle in terms of how persons are causally affected by actions.
We introduce Kantian causal agency models in which moral patients, actions, goals,
and causal influence are represented, and we show how to formalize several readings of Kant’s categorical imperative that correspond to Kant’s concept of strict
and wide duties towards oneself and others. Stricter versions handle cases where an
action directly causally affects oneself or others, whereas the wide version maximizes the number of persons being treated as an end. We discuss limitations of our
formalization by pointing to one of Kant’s cases that the machinery cannot handle
in a satisfying way.
Keywords. Machine Ethics, Causal Networks, Categorical Imperative
1. Introduction
It has been suggested that artificial agents, such as social robots and software bots, must
be programmed in an ethical way in order to remain beneficial to human beings. One
prominent ethical theory was proposed by Immanuel Kant [1]. Here, we propose a formalization and implementation of Kant’s ethics with the purpose of guiding artificial
agents that are to function ethically. In particular, the system will be able to judge whether
actions are ethically permissible according to Kant’s ethics. In order to accomplish this
we focus on the second formulation of Kants categorical imperative. Kant proposed three
formulations of the categorical imperative. We formalize and implement the second formulation and do not take a stance on the interrelation of Kant’s three formulations. The
second formulation of Kant’s categorical imperative reads:
Act in such a way that you treat humanity, whether in your own person or in the
person of any other, never merely as a means to an end, but always at the same time
as an end. (Kant, 1785)
1 Corresponding Author: Department of Computer Science, University of Freiburg, Georges-Köhler-Allee
52, D-79110 Freiburg, Germany; E-mail: [email protected].
We take it to be the core of the second formulation of the categorical imperative that
all rational beings affected by our actions must be considered as part of the goal of the
action.
The paper is structured as follows: We first briefly review related work. Then, building upon our earlier work [2], we introduce an extension of Pearl-Halpern-style causal
networks which we call Kantian causal agency models. These models serve as a formal
apparatus to model the morally relevant aspects of situations. We then define an action’s
permissibility due to the categorical imperative, while considering two readings of being
treated as a means. To deal with Kant’s wider duties, we introduce an extra condition
according to which an agent should maximize the number of persons being treated as an
end. Finally, we briefly showcase the computational implementation of the categorical
imperative within the HERA software library2 .
2. Related work
In machine ethics, several ethical theories have been formalized and implemented, e.g.,
utilitarianism, see [3,4], the principle of double effect, see [5,6], pareto permissibility,
see [2], and Asimov’s laws of robotics, see [7].
It has been suggested for some time that Kant’s ethics could be formalized and implemented computationally, see [8,9]. Powers [8] suggests three possible ways of formalizing Kant’s first formulation of the categorical imperative, through deontic logic, nonmonotonic logic, or belief revision. The first formulation of the categorical imperative
states that you must be able to want that the subjective reasoning (or maxim) motivating
your action becomes a universal law and as Kant claims that this in some cases is a purely
formal matter, it should be possible to formalize it. However, Powers does not provide
details of a formalization or a computational implementation, so the formalization of the
first formulation in effect remains an open problem.
The work presented here differs in that we focus on the second formulation of the
categorical imperative and in that we present a precise formal representation and computational implementation of the formal theory. Rather than taking a starting point in one of
the paradigms Powers suggests, we use formal semantics and causal agency modelling as
this is fitting for the means-end reasoning central to the second formulation. Philosophically, our formalization is best seen as a rational reconstruction within this framework of
what we take to be the central ideas of Kant’s second formulation.
Ultimately, although we are sympathetic to Kant and sensitive to the original text,
the goal of our work is not to get close to a correct interpretation of Kant, but to show that
our interpretation of Kant’s ideas can contribute to the development of machine ethics.
To meet this goal, our interpretation has to be detailed and explicit enough to provide a
decision mechanism for the permissibility or not of specific actions in specific situations.
3. Kantian Causal Agency Models
In order to formalize the second formulation of the categorical imperative, we assume
some background theory. First, we assume that actions are performed by agents, and that
2 http://www.hera-project.com
actions and their consequences can affect a set of moral patients, i.e. persons who must
be considered ethically in a situation. The agent itself is also one of the moral patients.
The agent has available a set of actions which will have consequences given background
conditions. Some of the action’s consequences are the goals of the action. The actions
and consequences that together cause my goal are the means of the action. Patients,
who are affected by these means are treated as a means, and patients, who are affected
by my goal are treated as an end. For example, I (agent) have the option available to
press the light switch (action), and given that the light bulb is not broken (background
condition), the light will go on (consequence), which leads to me being able to read
my book (consequence). The last consequence was also my goal, and it affects me in a
positive way. The action thus treats me as an end.
Within this informally characterized framework, we can reformulate the second formulation of the categorical imperative as follows:
Act in such a way, that whoever is treated as a means through your action (positively
or negatively and including yourself), must also be treated as an end of your action.
The purpose of what follows is to formalize these intuitions. As a first step, we now
give the formal definition of the models we will be using in Definition 1. We call these
models Kantian causal agency models to set them apart from the causal agency models
we used in our earlier work [2], and which had no formal tools to consider moral patients
affected by one’s actions.
Definition 1 (Kantian Causal Agency Model)
A Kantian causal agency model M is a tuple (A, B,C, F, G, P, K,W ), where A is the set of
action variables, B is a set of background variables, C is a set of consequence variables,
F is a set of modifiable boolean structural equations, G = (Goal1 , . . . , Goaln ) is a list of
sets of variables (one for each action), P is a set of moral patients (includes a name for
the agent itself), K is the ternary affect relation K ⊆ (A ∪ C) × P × {+, −}, and W is a
set of interpretations (i.e., truth assignments) over A ∪ B.
A (actions), B (background variables) and C (consequences) are finite sets of boolean
variables with B and C possibly empty. W is a set of boolean interpretations of A ∪
B. Thus, the elements of W set the truth values of those variables that are determined
externally, and thus specify the concrete situation. We require that all interpretations in
W assign true to exactly one action a ∈ A. As a notational convention, by M, wa and
M, wb we dinstinguish two situations that only differ in that in the first situation, action a
is performed, and in the second situation, action b is performed.
Causal influence is determined by the set F of boolean-valued structural equations.
Each variable ci ∈ C is associated with the function fi ∈ F. This function will give ci its
value under an interpretation w ∈ W . An interpretation w is extended to the consequence
variables as follows: For a variable ci ∈ C, let {ci1 , . . . , cim−1 } be the variables of C \ {ci },
B = b1 , . . . , bk , and A = {a1 , . . . , an } the action variables. The assignment of truth values
to consequences is determined by:
w(ci ) = fi (w(a1 ), . . . , w(an ), w(b1 ), . . . , w(bk ), w(ci1 ), . . . , w(cim−1 ))
To improve readability, we will use the notation c := φ to express that c is true if
φ is true, where φ can be any boolean formula containing variables from A ∪ B ∪C and
its negations. For instance, the boolean structural equations for the light-switch example
will be written as F = {lightOn := press ∧ ¬bulbBroken, canReadBook := lightOn}.
In the general setting, it may be unfeasible to extend an interpretation from the action
variables to the rest of the variables, because it is possible that the value of some variable
depends on the value of another variable, and the value of the latter variable depends on
the value of the former. Dependence is defined in Definition 2.
Definition 2 (Dependence)
Let vi ∈ C, v j ∈ A ∪ B ∪ C be distinct variables. The variable vi depends on variable v j ,
if, for some vector of boolean values, fi (. . . , v j = 0, . . .) 6= fi (. . . , v j = 1, . . .).
Following Halpern [10], we restrict causal agency models to acyclic models, i.e.,
models in which no two variables are mutually dependent on each other. First, note that
the values of action variables in set A and the values of background variables in set B
are determined externally by the interpretations in W . Thus, the truth values of action
variables and background variables do not depend on any other variables. Additionally,
we require that the transitive closure, ≺, of the dependence relation is a partial order
on the set of variables: v1 ≺ v2 reads “v1 is causally modified by v2 ”. This enforces absence of cycles. In case of acyclic models, the values of all consequence variables can
be determined unambiguously: First, there will be consequence variables only causally
modified by action and/or background variables, and whose truth value can thus be determined by the values set by the interpretation. Call these consequence variables level one.
On level two, there will be consequence variables causally modified by action variable,
background variables, and level-one consequence variables, and so on [5,10].
Some of the definitions below will make use of causality. Thus, to take causation
into account, Definition 3 defines the relation of y being a but-for cause of φ , see [10].
Definition 3 makes use of external interventions on models. An external interventions
X consists of a set of literals (viz., action variables, consequence variables, background
variables, and negations thereof). Applying an external intervention to a causal agency
model results in a new causal agency model MX . The truth of a variable v ∈ A ∪ B ∪C in
MX is determined in the following way: If v ∈ X, then v is true in MX , if ¬v ∈ X, then v is
false in MX , and if neither v ∈ X nor ¬v ∈ X, then the truth of v is determined according
to its structural equation in M. External interventions thus override structural equations
of the variables occuring in X.
Definition 3 (Actual But-For Cause)
Let y be a literal and φ a formula. We say that y is an actual but-for cause of φ (notation:
y φ ) in the situation the agent choses option wa in model M, if and only if M, wa |= y∧ φ
and M{¬y} , wa |= ¬φ .
The first condition requires that both the cause and the effect must be actual.
The second condition requires that if y had not been the case, then φ would have
not occurred. Thus, in the chosen situation, y was necessary to bring about φ . Sometimes it is convenient to be able to talk about causality in the logical language used
to talk about the models. In accordance with Halpern’s [10] notation, we can define
that y is a but-for cause of φ in model M, wa iff M, wa |= y ∧ φ ∧ [y ← ⊥]¬φ holds.
This definition is equivalent to Definition 3. So, in the above light-switch example,
we might want to determine if ¬bulbBroken is a cause of canReadBook. And indeed,
if we consider the model M{bulbBroken}, w press , we can determine that canReadBook is
false, written M{bulbBroken}, w press 6|= canReadBook. Consequently, we obtain M, w press |=
[bulbBroken ← ⊤]¬canReadBook, i.e., under the intervention that sets bulbBroken true,
canReadBook is false.
The concept of but-for cause allows us to introduce the useful concept of direct
consequences via Definition 4.
Definition 4 (Direct Consequence)
A consequence c ∈ C is a direct consequence of v ∈ A ∪ B ∪ C in the situation M, wa iff
M, wa |= v c.
Persons can be affected by actions or consequences either in a positive or in a negative way. To represent that some action or consequence affects a person positively or
negatively, we introduce the notations ⊲+ and ⊲− , respectively. Thus, M, wa |= c ⊲+ p
holds iff (a, c, +) ∈ K, and M, wa |= c ⊲− p holds iff (a, c, −) ∈ K. We use ⊲ in case the
valence of affection is not relevant.
This finalizes the exposition of the background theory. We will now consider how to
make permissibility judgments about actions as defined in the context of Kantian causal
agency models using the categorical imperative.
4. Categorical Imperative Defined
The second formulation of the categorical imperative requires an agent to never treat
someone merely as a means but always also as an end. Thus, to formalize under which
conditions an action is permitted by the categorical imperative, we first define the concept
of someone being treated as an end (Definition 5). We then proceed to formalize two
possible readings of the concept of someone being treated as a means (Definition 6 and
Definition 7).
Definition 5 (Treated as an End)
A patient p ∈ P is treated as an end by action a, written M, wa |= End(p), iff the following
conditions hold:
1. Some goal g of a affects p positively.
W
M, wa |= g Goal(g) ∧ g ⊲+ p .
2. None of the goals of a affect p negatively.
V
M, wa |= g (Goal(g) → ¬(g ⊲− p))
Thus, being treated as an end by some action means that some goal of the action
affects one in a positive way. One could say that the agent of the action, by performing
that action, considers those who benefit from his goal. Things are less clear regarding
the concept ‘being treated as a means’. As a first step, we define two versions of the
concept which we refer to as Reading 1 and Reading 2. Both readings make use of the
causal consequences of actions. Reading 1 considers a person used as a means in case
she is affected by some event that causally brings about some goal of the action. As
a consequence, negative side effects are permitted. Consider, for instance, the classical
trolley dilemma, where the agent has the choice to either pull the lever to lead the tram
onto the second track killing one person, or refraining from pulling letting the tram kill
five persons on the first track. Under Reading 1, in case of pulling, the one agent is
not treated as a means. This reading is probably closest to what we informally mean
by ‘being treated as a means’. Reading 2 requires that everybody affected by any direct
consequence of the action is considered as a goal, hence, also the person on the second
track. As a consequence, everyone treated as a means according to Reading 1 is also
treated as a means according Reading 2, but Reading 2 may include additional patients.
This reading is further from the everyday understanding of means-end reasoning, but is
probably closer to what some people expect of a Kantian ethics. We consider it a feature
of a formal framework that we are able to formalize such distinctions, but we leave it for
the modeler to decide which one of the readings is more useful for a given application.
Definition 6 (Treated as a Means, Reading 1)
A patient p ∈ P is treated as a means by action a (according to Reading 1), written
M, wa |= Means1 (p), iff there is some v ∈ A ∪C, such that v affects p, and
v is a cause of
W
W
some goal g, i.e., M, wa |= v (a v ∧ v ⊲ p) ∧ g (v g ∧ Goal(g)) .
Definition 7 (Treated as a Means, Reading 2)
A patient p ∈ P is treated as a means by action a (according to Reading 2), written
M, wa |= Means2 (p), iff there is some direct
consequence v ∈ A ∪ C of a, such that v
W
affects p, i.e., M, wa |= v a v ∧ v ⊲ p .
Having defined both being treated as an end and being treated as a means, the permissibility of actions according to the second formulation of the categorical imperative
can now be defined in Definition 8. The formulation requires that noone is merely used
as a means, but always at the same time as an end.
Definition 8 (Categorical Imperative)
An action a is permitted according to the categorical imperative, iff for any p ∈ P, if p is
treated as a means (according to Reading N) then p is treated as an end
V
M, wa |= p∈P (MeansN (p) → End(p))
There are thus two main reasons why an action is not permitted. Either a patient is
treated as a means but is left out of consideration by the end of the action. Or, the action
is done for an end that affects someone negatively.
Note that the choice between Reading 1 and Reading 2 yields significant implications, as can be seen in the Trolley dilemma: According to Reading 1, pulling the lever
is permissible, because the one person on the other track is not treated as a means, and
therefore it is permissible to not consider her as an end. According to Reading 2, however, pulling the lever is impermissible, because the one person on the other track is
treated as a means while she does not benefit from the goal (and is thus not treated as an
end). Refraining from action is permitted by the categorical imperative according to both
the readings of ‘being treated as a means’ (more about this issue below in the Section on
wide duties).
5. Cases of Strict Duty
We will now provide examples that highlight aspects of the definition of the categorical
imperative. Although these do not prove it correct in any formal sense they can be used
to discuss its appeal as an ethical principle as an explication of Kant’s ideas. First, we
A = {suicide}
A = {amputate}
C = {dead}
C = {survives}
P = {Bob}
P = {Bob}
F = {dead := suicide}
F = {survives := amputate}
K = {(suicide, Bob, +)}
K = {(amputate, Bob, −),
G = (Goalsuicide = {dead})
(survives, Bob, +)}
G = (Goalamputate = {survives})
(a) Model M1
(b) Model M1∗
Figure 1. Kantian Causal Agency Models yielding the impermissibility of Suicide (M1 ) and the permissibility
of Amputation (M1∗ ).
rephrase three cases that contain what Kant calls strict duties (and two of which Kant
himself used to explain his ideas).
5.1. Example 1: Suicide
Bob wants to commit suicide, because he feels so much pain he wants to be relieved
from. This case can be modeled by a causal agency model M1 that contains one action
variable suicide and a consequence variable dead, see Figure 1a. Death is the goal of the
suicide action (as modeled by the set G), and the suicide affects Bob (as modeled by the
set K). In this case, it does not make a difference whether the suicide action affects Bob
positively or negatively.
We assume that the suicide affects noone other than Bob, because Kant’s argument
is not about the effect of suicide on other people but about the lack of respect of the
person commiting suicide. The reason why Bob’s suicide is not permitted is that the
person affected by the suicide, viz., Bob, does not benefit from the goal, because he is
destroyed and thus cannot be affected positively by it. He is thus treated as a means to his
own annihilation from which he receives no advancement. Therefore, the first condition
of the categorical imperative (Definition 8) is violated according to both readings (1 and
2), because M1 , wsuicide |= Means{1,2} (Bob) holds but M1 , wsuicide |= End(Bob) does not.
As noted above, we could say that the suicide affects Bob negatively, and the action
would also be impermissible. The reason for the impermissibility of suicide also in this
case is not due to the fact that Bob does something harmful towards himself. As Kant
also remarks, other harmful actions would be allowed, e.g., risking your life or amputating a leg to survive. To see this, consider Figure 1b, where M1 has been be slightly
modified to M1∗ : We rename suicide to amputate and dead to survives. Moreover, we
add (amputate, Bob, −) to K. In this case, Bob is positively affected by the goal, and thus
the act of amputation is permitted. The modified example also shows that in some cases,
the categorical imperative is more permissive than the principle of double effect, which
strictly speaking never allows negative means to an end (cf., [5]).
A = {give f lowers}
A = {give f lowers}
C = {alice happy, celia happy}
C = {alice happy, celia happy}
P = {Bob, Alice,Celia}
P = {Bob, Alice,Celia}
F = {alice happy := give f lowers
F = {alice happy := give f lowers
celia happy := alice happy}
celia happy := alice happy}
K = {(alice happy, Alice, +),
K = {(alice happy, Alice, +),
(celia happy,Celia, +)}
(celia happy,Celia, +)}
G = (Goalgive
f lowers
= {celia happy
})
(a) Model M2
G = (Goalgive
f lowers
= {celia happy,
alice happy})
(b) Model M2∗
Figure 2. Kantian Causal Agency Models yielding the impermissibility of giving flowers to Alice to make
Celia happy (M2 ) and the permissibility of doing so if making Celia happy is a goal as well (M2∗ ).
5.2. Example 2: Giving Flowers
We consider the fact that an action can be judged as impermissible by the categorical
imperative although noone is negatively affected a feature of the categorical imperative
that inheres in no other moral principles formalized so far. The following example showcases another situation to highlight this feature: Bob gives Alice flowers in order to make
Celia happy when she sees that Alice is thrilled about the flowers. Alice being happy
is not part of the goal of Bob’s action. We model this case by considering the Kantian
causal agency model M2 shown in Figure 2a. In the model M2 , the action give f lowers
is not permitted according to the categorical imperative, because Bob is using Alice as a
means to make Celia happy, but not considering her as part of the goal of the action. This
action is immoral, even though the action has positive consquences for all, and no bad
consequence are used to obtain a good one. Again, this example shows how the Kantian
principle differs from other ethical principles such as utilitarianism and the principle of
double effect, because these principles would permit the action.
The model M2 can be extended to model M2∗ shown in Figure 2b. In model M2∗ , Bob’s
action is permitted by the Kantian principle. The only thing in which M2∗ differs from
M2 is that the variable alice happy is added to the set Goalgive f lowers . In this case, Alice
is both treated as a means and treated as an end, which is permitted by the categorical
imperative.
The flower example demonstrates how demanding the categorical imperative is, because the principles requires that everybody affected by ones’ action must be treated as
a goal: This includes the taxi driver that drives you to your destination, as well as the
potential murderer you defend yourself against. In these examples, the ethical principle
requires one to, e.g., have the taxi driver’s earning money among one’s goals, and the
murderer’s not going to jail.
A = {borrow}
C = {bob keeps 100Dollar f orever}
P = {Alice, Bob}
F = {bob keeps 100Dollar f orever := borrow}
K = {(borrow, Bob, +), (borrow, Alice, −),
(bob keeps 100Dollar f orever, Bob, +),
(bob keeps 100Dollar f orever, Alice, −)}
G = (Goalborrow = {bob keeps 100Dollar f orever})
Figure 3. Model M3 for the case of Bob making a false promise to Alice.
5.3. Example 3: False Promise
We return to a case mentioned by Kant himself. Consider that Bob makes a false promise
to Alice. Bob borrows one 100 Dollars from Alice with the goal of keeping the money
forever. He knows that it is an inevitable consequence of borrowing the money that he
will never pay it back. Figure 3 shows the model of this situation, M3 .
The action is impermissible, because Alice is treated as a means (by both Reading
1, Definition 6, and Reading 2, Definition 7) while she is not treated as an end. In this
case, none of the two conditions for ‘being treated as an end’ (Definition 5) are met.
6. Cases of Wide Duty
Examples 1, 2 and 3 are instances of what Kant calls necessary, strict, narrower duties to
oneself and to others, and it seems obvious they involve using a person as a means. Kant
also presents two other examples to which we now turn. These involve what Kant calls
contingent, meritorious, or wider duties. His arguments for these appear more vague and
at least from our perspective harder to handle. We now turn to wide duties and discuss,
through an example, how actions that indirectly affect others by refraining from preventing harmful consequences could be handled, and we learn through another example
where the limitations of our formalization attempt are.
6.1. Example 4: Not Helping Others
Bob who has everything he needs, does not want to help Alice who is in need. Let us
assume she is drowning and Bob is refraining from saving her live. Formally, the situation
in the example can be represented with a causal agency model M4 that contains one
background variable accident representing the circumstances that led to Alice being in
dire straits, two action variables rescue and refrain and a consequence variable drown.
Moreover, ¬drown is the goal of rescue. See Figure 4 for the specification of the model.
A = {rescue, refrain}
B = {accident}
C = {drown}
P = {Alice, Bob}
F = {drown := accident ∧ ¬rescue}
K = {(drown, Alice, −), (¬drown, Alice, +)}
G = (Goalrescue = {¬drown}, Goalrefrain = 0)
/
Figure 4. Model M4 for the impermissibility of not helping others.
According to the categorical imperative using Readings 1 and 2 of ‘being treated as
a means’ both rescue and refrain are permitted. Bob is strictly speaking not using Alice
as a means by going about his business. Kant gives us a clue of how to formalize an
argument against refraining in that he says we have to make other people’s ends our own
as far as possible. Kant writes that ‘For a positive harmony with humanity as an end in
itself, what is required is that everyone positively tries to further the ends of others as far
as he can.’ One way of understanding this is as an additional requirement on top of the
categorical imperative of choosing an action whose goals affect most people positively.
This understanding is captured in Definition 9.
Definition 9 (Meritorious principle)
Among actions permitted by the categorical imperative, choose one whose goals affect
most patients positively.
The meritorious principle thus goes beyond simply avoiding to treat others as means
by acitvely helping them. As formulated here, the principle is compatible with the categorical imperative. In our example, it requires of the agent to choose saving Alice, because the goal advances her. There may be several actions advancing the same number
of agents, in which case the agent can choose freely (or randomly) amongst them. One
could also take Kant to imply a second condition to the meritorious principle, to prevent
as many people being negatively affected by circumstances as possible. In the current
example, both conditions would lead to the same result.
6.2. Unhandled Case: Not Using Your Talent
As a final example, consider the following situation: Bob has the talent to become a
great artist. However, he wonders whether it is permissible to just be lazy and enjoy life
instead of working hard to improve himself. Strictly speaking Bob is not working to
anyone’s disadvantage by being lazy and thus the definitions of ‘being treated as a means’
advanced above will not cover this example. As the goal of enjyoing life and the goal
of making art both benefit Bob, the meritorious principle also cannot be used to make
the distinction. What Kant says is that laziness could be consistent with the preservation
of humanity but does not harmonize with its advancement. He also writes that a rational
being necessarily wills that all his capacities are developed. However, it is not clear to
us what consistutes the advancement of humanity beyond the sheer feeling of happiness.
The example is further complicated by the fact that Kant says that this is a duty one has
towards oneself, not others. Therefore, it would be inappropriate to solve this case by
introducing others into the model that would benefit from Bob becoming an artist.
In the current formalization, we have no means to represent the relevant aspects that
render laziness impermissible and becoming an artist permissible for the right reasons.
We thus take this example to showcase a limitation of our treatment of Kant’s ethics, and
leave a formalization that could capture this last example for further research.
7. Implementation
The formalization of Kant’s second formulation of the categorical imperative has
been implemented within the Hybrid Ethical Reasoning Agent software library (short:
HERA).3 The general goal of the HERA project is to provide theoretically well-founded
and practically usable logic-based machine ethics tools for implementation in artificial
agents, such as companion robots with moral competence [11]. The core of HERA consists of a model checker for (Kantian) causal agency models. Thus, the situations the
agent can reason about are represented in terms of models, and ethical principles like the
categorical imperative are implemented as (sets of) logical formulae. To showcase the use
the categorical imperative from a Python program, Listing 1 reconsiders a representation
of the suicide case.
{
" actions " : [ " suicide " ] ,
" background " : [ ] ,
" consequences " : [ " dead " ] ,
" patients " : [ " Bob " ] ,
" mechanisms " : { " dead " : " suicide " } ,
" affects " : { " suicide " : [ [ " Bob " , "+" ] ] ,
" dead " : [ ] } ,
" goals " : { " suicide " : [ " dead " ] }
}
Listing 1: A sample JSON encoding of the suicide case.
The workflow for using HERA requires to first generate a causal agency model like
the one in Listing 1. Given such a model, arbitrary logical formulae can be checked for
being satisfied or not by this model. This way, the conditions of ethical principles like the
Kantian categorical imperative as defined in Definition 8 can be checked for satisfication.
To support the usage of the HERA library, the logical formulae to be checked for
ethical principles already included in HERA are encapsulated into prepared classes. Listing 2 shows a sample interaction. The first three commands load the implementations
of two syntactical entities of the logical language (the predicates Means and End), the
3 The HERA software is available from http://www.hera-project.com . It is fully implemented in
Python and can be installed via the PyPI repository (package name: ethics).
causal agency model from the semantics package, and the categorical imperative using Reading 1 of ‘being treated as a means’ from the principles package. The third
command loads the suicide example and sets the external variable suicide to the value
True. This way, the suicide action is chosen in the situation, and the truth values of the
consequence variables can be evaluated the way explained in Section 3. In the concrete
case, True will be assigned to the variable dead. The fourth command asks whether, in
the resulting situation, Bob is used as a means according to Reading 1 (see Definition
6). The answer is True, because Bob is affected by the action (suicide) and the action is
a but-for cause of Bob’s goal (dead). The fifth command asks if Bob is used as an end.
This query returns False, because Bob is not affected by the goal (see Section Example 1:
Suicide). All in all, the action is not permissible according to the categorical imperative,
and the output of the last command is accordingly.
from e t h i c s . l a n g u a g e import Means , End
from e t h i c s . s e m a n t i c s import C a u s a l M o d e l a s cm
from e t h i c s . p r i n c i p l e s import K a n t i a n H u m a n i t y P r i n c i p l e a s c i
m = cm ( ” s u i c i d e . j s o n ” , { ” s u i c i d e ” : T r u e } )
m. m o d e l s ( Means ( ” Reading −1” , ” Bob ” ) )
output: True
m. m o d e l s ( End ( ” Bob ” ) )
output: False
m. e v a l u a t e ( c i )
output: False
Listing 2: A sample interaction with the Python package ethics, which we develop and
maintain as the standard implementation of HERA.
8. Conclusion
We have shown proof of principle how Kant’s second formulation of the categorical imperative can be formalized and implemented computationally. The strict duties towards
yourself and others are defined, given goals, structural equations, and the affects relation. To define permissibility according the categorical imerative, we have defined ‘being
treated as an end’, and we formalized two readings of ‘being treated as a means’ that
meet different intuitions about this concept. The formalization deals well with Kant’s
own examples of strict duties. We were also able to partly deal with Kant’s wide duties by
defining an additional condition that requires agents to maximize the number of persons
being treated as an end.
We envision that the theory will be used as a tool for the comparison of morally
relevant aspects of different views on morally delicate cases, thus helping people to have
moral discussions. Moreover, we aim at allowing automatic moral judgments in line with
Kant in robots such as self-driving cars and care robots. Our future research will investigate whether and under which circumstances Kantian reasoning the way it is presented
here is perceived as appropriate for social robots as compared to other types of moral
reasoning.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
Kant, I. 1785. Grundlegung zur Metaphysik der Sitten.
Lindner, F., Bentzen, M., and Nebel, B. 2017. The HERA approach to morally competent robots. In
Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).
Horty, J. F. 2001. Agency and Deontic Logic. Oxford Uni- versity Press.
Arkoudas, K.; Bringsjord, S.; and Bello, P. 2005. Toward ethical robots via mechanized deontic logic.
Technical re- port, AAAI Fall Symposium on Machine Ethics, AAAI.
Bentzen, M. 2016. The principle of double effect applied to ethical dilemmas of social robots. In
Robophilosophy 2016/TRANSOR 2016: What Social Robots Can and Should Do. IOS Press. 268–279.
Govindarajuli, N. S., and Bringsjord, S. 2017. On automat- ing the doctrine of double effect. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial In- telligence (IJCAI). 4722–4730.
Winfield, A. F.; Blum, C.; and Liu, W. 2014. Towards an ethical robot: internal models, consequences
and ethical ac- tion selection. In Mistry, M.; Leonardis, A.; M.Witkowski; and Melhuish, C., eds., Advances in Autonomous Robotics Systems. Springer. 85–96.
Powers, T. M. 2006. Prospects for a kantian machine. IEEE Intelligent Systems 21(4):46–51.
Abney, K. 2012. Robotics, ethical theory, and metaethics: A guide for the perplexed. In Lina, P.; Abney,
K.; and Bekey, G. A., eds., Robot Ethics: The Ethical and Social Implications of Robotics. MIT Press.
35–52.
Halpern, J. Y. 2016. Actual Causality. The MIT press.
Lindner, F., and Bentzen, M. 2017. The hybrid ethical reasoning agent IMMANUEL. In Proceedings of
the Companion 2017 Conference on Human-Robot Interaction (HRI). 187–188.
| 2cs.AI
|
arXiv:1703.09611v1 [math.GR] 28 Mar 2017
Actions of automorphism groups of Lie groups
S.G. Dani
March 29, 2017
Introduction
The aim of this article is to discuss the actions on a connected Lie group by subgroups of its automorphism group. The automorphism groups are themselves Lie
groups (not necessarily connected) and the actions have, not surprisingly, played
an important role in the study of various topics, including geometry, dynamics,
ergodic theory, probability theory on Lie groups, etc.
We begin in § 1 with results on the structure of the automorphism groups
Aut(G), G a Lie group, generalities about their subgroups, connected components, etc.. The automorphism group of a connected Lie group can be realised
as a linear group via association with the corresponding automorphism of the Lie
algebra G of G, and § 2 is devoted to relating Aut(G) to algebraic subgroups,
and more generally “almost algebraic” subgroups, of GL(G); in particular the
connected component of the identity in Aut(G) turns out to be almost algebraic
and this has found considerable use in the study of various topics discussed in the
subsequent sections. In § 3 we discuss orbits of various subgroups of Aut(G) on
G, and especially conditions for them to be dense in G. In § 4 we discuss invariant measures and various ergodic-theoretic aspects of the actions of subgroups of
Aut(G) on G, and in § 5 some of the topics in topological dynamics in which various special features of Aut(G) play a role. The action of Aut(G) on G induces in
a natural way an action on the space of probability measures on G and properties
of this action have played an important role in various questions in probability
theory on Lie groups, in terms of extending certain aspects of classical probability
theory to the Lie group setting. These are discussed in § 6.
While our main focus will be on automorphism groups of connected Lie groups,
at various points along the way generalisations to the case of more general locally
compact groups are mentioned with references to the relevant literature.
1
Contents
.
.
.
.
3
3
4
5
6
.
.
.
.
.
.
.
.
7
8
8
9
10
10
11
12
13
.
.
.
.
.
.
.
13
14
15
16
17
18
19
20
.
.
.
.
.
.
21
21
22
23
24
24
25
5 Some aspects of dynamics of Aut(G)-actions
5.1 Stabilisers of continuous actions . . . . . . . . . . . . . . . . . . .
26
26
1 Structure of the automorphism groups
1.1 Preliminaries . . . . . . . . . . . . . .
1.2 Subgroups of Aut(G) . . . . . . . . . .
1.3 Connected components of Aut(G) . . .
1.4 Locally isomorphic Lie groups . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Almost algebraic automorphism groups
2.1 A decomposition . . . . . . . . . . . . . . . . .
2.2 Almost algebraic subgroups of Aut(G) . . . . .
2.3 Groups with Aut(G) almost algebraic . . . . . .
2.4 Automorphisms preserving additional structure
2.5 Linearisation of the Aut(G) action on G . . . .
2.6 Embedding of G in a projective space . . . . . .
2.7 Embedding in a vector space . . . . . . . . . . .
2.8 Algebraicity of stabilizers . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
3 Dense orbits
3.1 Aut(G)-actions with dense orbits . . . . . . . . .
3.2 Connected subgroups of Aut(G) with dense orbits
3.3 Automorphisms with dense orbits . . . . . . . . .
3.4 Zd -actions . . . . . . . . . . . . . . . . . . . . . .
3.5 Discrete groups with dense orbits . . . . . . . . .
3.6 Orbit structure of actions of some discrete groups
3.7 Aut(G)-actions with few orbits . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Ergodic theory of actions of automorphism groups
4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . .
4.2 Finite invariant measures . . . . . . . . . . . . . . . .
4.3 Convolution powers of invariant measures . . . . . . .
4.4 Infinite invariant measures . . . . . . . . . . . . . . .
4.5 Quasi-invariant measure and ergodicity . . . . . . . .
4.6 Stabilisers of actions of Lie groups . . . . . . . . . . .
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5.2 Anosov automorphisms . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Distal actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Expansive actions . . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Orbits on the space of probability measures
6.1 Preliminaries . . . . . . . . . . . . . . . . . . .
6.2 Factor compactness and concentration functions
6.3 Tortrat groups . . . . . . . . . . . . . . . . . . .
6.4 Convergence of types . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
References
1
.
.
.
.
.
.
.
.
.
.
.
.
27
28
29
29
29
30
31
32
33
Structure of the automorphism groups
In this section we discuss various structural aspects of the automorphisms of
connected Lie groups.
1.1
Preliminaries
Let G be a connected Lie group and G be the Lie algebra of G. We denote by
Aut(G) the group of all Lie automorphisms of G equipped with its usual topology,
corresponding to uniform convergence on compact subsets (see [55] Ch. IX for some
details). To each α ∈ Aut(G) there corresponds a Lie automorphism dα of G, the
derivative of α. We may view Aut(G) in a natural way as a closed subgroup of
GL(G), the group of nonsingular linear transformations of G, considered equipped
with its usual topology. Consider the map d : Aut(G) → Aut(G), given by
α 7→ dα, for all α ∈ Aut(G). As G is connected, dα uniquely determines α, and
thus the map is injective.
When G is simply connected the map is also surjective, by Ado’s theorem (cf.
[53], Theorem 7.4.1). When G is not simply connected the map is in general not
surjective; in this case G has the form G̃/Λ, where G̃ is the universal covering
group of G and Λ is a discrete subgroup contained in the center of G̃, and the
image of Aut(G) in Aut(G) consists of those elements for which the corresponding
automorphism of Aut(G̃) leaves the subgroup Λ invariant (note that G may be
viewed also, canonically, as the Lie algebra of G̃), and in particular it follows
that it is a closed subgroup. This implies in turn that d as above is a topological
isomorphism (see also [55], Ch. IX, Theorem 1.2). We shall view Aut(G) as a
subgroup of Aut(G), and in turn GL(G), via the correspondence.
3
As a closed subgroup of GL(G), Aut(G) is Lie group, and in particular a
locally compact group; it may be mentioned here that the question as to when
the automorphism group of a general locally compact group is locally compact is
analysed in [90].
We note that Aut(G) is a real algebraic subgroup of GL(G), namely the group
of R-points of an algebraic subgroup of GL(G⊗C), defined over R. While Aut(G)
is in general not a real algebraic subgroup (when viewed as a subgroup of GL(G) as
above), various subgroups of Aut(G) being algebraic subgroups plays an important
role in various results discussed in the sequel.
1.2
Subgroups of Aut(G)
We introduce here certain special classes of automorphisms which play an important role in the discussion in the following sections.
Let G be a connected Lie group. For any subgroup H of G let AutH (G)
denote the subgroup of Aut(G) consisting of automorphisms leaving H invariant;
viz. {τ ∈ Aut(G) | τ (H) = H}. A subgroup H for which AutH (G) = Aut(G) is
called a characteristic subgroup of G. Clearly, for any Lie group G the center of
G, the commutator subgroup [G, G], the (solvable) radical, the nilradical are some
of the characteristic subgroups of G. Similarly, together with any characteristic
subgroup, its closure, centraliser, normaliser etc. are characteristic subgroups.
For each g ∈ G we get an automorphism σg of G defined by σg (x) = gxg −1 for
all x ∈ G, called the inner automorphism corresponding to g. For a subgroup H
of G we shall denote by Inn(H) the subgroup of Aut(G) defined by {σh | h ∈ H}.
For any Lie subgroup H, Inn(H) is a Lie subgroup of Aut(G); it is a normal
subgroup of Aut(G) when H is a characteristic subgroup. In particular Inn(G)
is a normal Lie subgroup of Aut(G); the group Aut(G)/Inn(G) is known as the
group of outer automorphisms of G. We note that Inn(G) may in general not
be a closed subgroup of Aut(G); this is the case, for example, for the semidirect
product of R with C2 with respect to the action under which t ∈ R acts by
(z1 , z2 ) 7→ (eiαt z1 , eiβt z2 ) for all z1 , z2 ∈ C, with α, β fixed nonzero real numbers
such that α/β is irrational.
Let Z denote the center of G and ϕ : G → Z be a (continuous) homomorphism
of G into Z. Let τ : G → G be defined by τ (x) = xϕ(x) for all x ∈ G. It is easy
to see that τ ∈ Aut(G); we call it the shear automorphism, or more specifically
isotropic shear automorphism (as in [20]), associated with ϕ. We note that for
any continuous homomorphism ϕ as above the subgroup [G, G] is contained in the
kernel of ϕ, and hence the associated shear automorphism τ fixes [G, G] pointwise.
In particular, if [G, G] = G then there are no nontrivial shear automorphisms. The
shear automorphisms form a closed normal abelian subgroup of Aut(G), say S.
4
We note that if A = {τ ∈ Aut(G) | τ (z) = z for all z ∈ Z} then Aut(G) is the
semidirect product of A and S.
If H is the subgroup of G containing [G, G] and such that H/[G, G] is the
maximal compact subgroup of G/[G, G], then G/H is a vector group and the set
of continuous homomorphisms ϕ of G into Z such that H ⊂ ker ϕ has the natural
structure of a vector space, and in turn the same holds for the corresponding set of
shear automorphisms; the dimension of the vector space equals the product of the
dimensions of G/H and Z. When G/[G, G] is a vector group (viz. topologically
isomorphic to Rn for some n) S is a connected algebraic subgroup of Aut(G). On
the other hand, when G/[G, G] is compact then there are only countably many
distinct continuous homomorphisms of G into Z, and hence only countably many
shear automorphisms.
For a connected semisimple Lie group G, Inn(G) is a subgroup of finite index
in Aut(G); in the case of a simply connected Lie group this follows from the
corresponding statement for the associated Lie algebra (cf. [53], Theorem 5.5.14),
and the general case follows from the special case, since the inner automorphisms
of the simply connected covering group factor to the original group G. On the
other hand, for a connected nilpotent Lie group the group of outer automorphisms
is always of positive dimension (see [58], Theorem 4). An example of a 3-step
simply connected nilpotent Lie group G for which Aut(G) = Inn(G) · S, where S
is the group of all shear automorphisms of G, is given in [26]; in particular Aut(G)
is nilpotent in this case; a larger class of connected nilpotent Lie groups for which
Aut(G) is nilpotent is also described in [26].
We recall here that the center of a connected Lie group G is contained in a
connected abelian Lie subgroup of G (see [53], Theorem 14.2.8, or [55], Ch. 16,
Theorem 1.2). In particular the center is a compactly generated abelian group and
hence has a unique maximal compact subgroup; we shall denote it by C. Being a
compact abelian Lie subgroup, C is in fact the cartesian product of a torus with a
finite abelian group. As the unique maximal compact subgroup of G, C is Aut(G)invariant and hence we get a continuous homomorphism q : Aut(G) → Aut(C),
by restriction of the automorphisms to C. Since the automorphism group of a
compact abelian group is countable, it follows that Aut(G)/ ker q, where ker q is
the kernel of q, is a countable group.
1.3
Connected components of Aut(G)
As a closed subgroup of GL(G), Aut(G) is a Lie group with at most countably
many connected components (see [54] for an extensions of this to not necessarily
connected Lie groups). It is in general not connected; for Rn , n ≥ 1, the automorphism group, which is topologically isomorphic to GL(n, R), has two connected
5
components, while for the torus Tn (= Rn /Zn ), n ≥ 2, the automorphism group
is in fact an infinite (countable) discrete group.
We shall denote by Aut0 (G) the connected component of the identity in
Aut(G); it is an open (and hence also closed) subgroup of Aut(G). We denote by
c(G) the group of connected components, namely Aut(G)/Aut0 (G).
For a connected semisimple Lie group, since Inn(G) is of finite index in Aut0 (G)
(see § 1.2), Aut0 (G) = Inn(G). Also, for these groups the number of connected
components is finite; the number is greater than one in many cases (see [57]
and [71] for details on the group of connected components of Aut(G) for simply
connected semisimple groups G; see also the recent paper [50] where splitting of
Aut(G) into the connected component and the component group is discussed).
We note that Aut0 (G) acts trivially on the unique maximal compact subgroup C of the center of G, and hence when the image of the homomorphism
q : Aut(G) → Aut(C), as in § 1.2, is infinite, c(G) is infinite.
When C as above is the circle group, Aut(C) is of order two, and the image
of q has at most two elements, but nevertheless c(G) can be infinite. A natural
instance of this can be seen in the following: Suppose G has a closed normal subgroup H such that G/H is a torus of positive dimension; then there is a unique
minimal subgroup with the property and it is invariant under the action of Aut(G)
and hence, modifying notation, we may assume H to be Aut(G)-invariant. Then
we have (countably) infinitely many (continuous) homomorphisms ϕ : G/H → C
and for each such ϕ we have an isotropic shear automorphism of G (see § 1.2).
It can be seen that the shear automorphisms corresponding to distinct homomorphisms belong to distinct connected components of Aut(G), and hence Aut(G)
has infinitely many connected components. Thus c(G) is infinite in this case also.
The group c(G) is finite if and only if Aut(G) is “almost algebraic” as a
subgroup of GL(G); see §2.2 for a discussion on conditions for Aut(G) to be
almost algebraic.
1.4
Locally isomorphic Lie groups
Any connected Lie group is locally isomorphic to a unique (upto Lie isomorphism)
simply connected Lie group, namely its universal covering group. Now let G be
a simply connected Lie group and Z be the center of G. Then all connected
Lie groups locally isomorphic to G are of the form G/D, where D is a discrete
subgroup of Z. Moreover, for two discrete subgroups D1 and D2 of Z the Lie
groups G/D1 and G/D2 are Lie isomorphic if and only if there exists a τ ∈ Aut(G)
such that τ (D1 ) = D2 . Thus the class of Lie groups (viewed up to isomorphism
of Lie groups) locally isomorphic to a given connected Lie group G is in canonical
6
one-one correspondence with the orbits of the action of Aut(G) on the class of
discrete subgroups of it center Z, under the action induced by the Aut(G)-action
on Z, by restriction of the automorphisms to Z.
For G = Rn , n ≥ 1, all connected Lie groups isomorphic to G are of the form
Rm × Tn−m , with 0 ≤ m ≤ n, thus n + 1 of them altogether. For the group G of
upper triangular n × n unipotent matrices (n ≥ 2), which is a simply connected
nilpotent Lie group, the center is one-dimensional and all its nontrivial discrete
subgroups are infinite cyclic subgroups that are images of one another under automorphisms of G; thus in this case there are only two non-isomorphic Lie groups
locally isomorphic to G (including the simply connected one), independently of n.
On the other hand there are simply connected nilpotent Lie groups G for which
Aut(G) is a unipotent group (when viewed as a subgroup of Aut(G)) (see [26]);
hence in this case the Aut(G)-action on the center has uncountably many distinct orbits, and therefore there are uncountably many mutually non-isomorphic
connected Lie groups that are locally isomorphic to G.
Now let G be a connected Lie group with discrete center, say Z. Then Z
is finitely generated (see § 1.2) and hence Z has only countably many distinct
subgroups. Considering the indices of the subgroups in Z it can also be seen
that when Z is infinite there are infinitely many subgroups belonging to distinct
orbits of the Aut(G)-action on the class of subgroups; thus in this case there are
countably infinitely many connected Lie groups locally isomorphic to G. This
applies in particular when G is the universal covering group of SL(2, R). When
Z is finite the number of Lie groups locally isomorphic to G is finite, and at
least equal to the number of prime divisors of the order of Z. For simple simply
connected Lie groups the orbits of the Aut(G)-action on the class of subgroups of
the center have been classified completely in [48].
2
Almost algebraic automorphism groups
Let G be a connected Lie group and G be the Lie algebra of G. Let GL(G) be
realised as GL(n, R), where n is the dimension of G, via a (vector space) basis of
G. For g ∈ GL(G) let gij , 1 ≤ i, j ≤ n, denote the matrix entries of g and let det g
denote the determinant of g. A subgroup H of GL(G) being a real algebraic group
is equivalent to the condition that it can be expressed as the set of zeros (solutions)
of a set of polynomials in gij , 1 ≤ i, j ≤ n, and (det g)−1 as the variables; in the
present instance, the field being R, it suffices to consider a single polynomial in
place of the set of polynomials.
We call a subgroup of GL(G) almost algebraic if it is of finite index in a real
algebraic subgroup. A real algebraic subgroup is evidently a closed (Lie) subgroup
7
of GL(G), and since a connected Lie group admits no proper subgroups of finite
index it follows that any almost algebraic subgroup is an open subgroup of a real
algebraic subgroup, and in particular it is a closed subgroup of GL(G). It is known
that any real algebraic subgroup has only finitely many connected components (see
[10], Corollary 14.5, for a precise result on the number of components), and hence
the same holds for any almost algebraic subgroup.
2.1
A decomposition
Let G be a connected Lie group and G the Lie algebra of G. In what follows we
view Aut(G) as a subgroup of Aut(G) and of GL(G), via the identification introduced earlier (§ 1.1). We recall that any connected Lie group admits a maximal
torus (a subgroup topologically isomorphic to Tn = Rn /Zn for some n ≥ 0), and
any two maximal tori are conjugate to each other in G (see [53], Corollary 14.1.4).
Theorem 2.1. (cf. [19], [73]) Let G be a connected Lie group with Lie algebra
G. Let T be a maximal torus in G. Then there exists a closed connected normal
subgroup H of G such that the following conditions hold:
i) Inn(H) is an almost algebraic subgroup of Aut(G), and
ii) Aut(G) = Inn(H)AutT (G).
The subgroup H chosen in the proof is in fact invariant under all τ ∈ Aut(G)
and consequently Inn(H) is a normal subgroup of Aut(G). Since Inn(H)∩AutT (G)
is trivial this further implies that Aut(G) is a semidirect product of the two subgroups. We may also mention here that the subgroup H as chosen in [19] contains
[G, G], the commutator subgroup of G. When Inn(G) is an almost algebraic subgroup of Aut(G), H as in Theorem 2.1 can be chosen to be G itself; this applies
in particular when G is an almost algebraic subgroup of GL(n, R) for some n ≥ 1.
2.2
Almost algebraic subgroups of Aut(G)
In view of Theorem 2.1 Aut(G) is almost algebraic if and only if AutT (G) is almost
algebraic for a (and hence any) maximal torus T in G. In this respect we recall
the following specific results.
Theorem 2.2. ([19], [88], [73]) Let G be a connected Lie group and G be the Lie
algebra of G. Then the following statements hold.
i) If the center of G does not contain a compact subgroup of positive dimension,
then Aut(G) is an almost algebraic subgroup of GL(G).
ii) Aut0 (G) is an almost algebraic subgroup of GL(G);
8
iii) if R is the solvable radical of G and T is a maximal torus in G then
Aut(G) is almost algebraic if and only if the restrictions of all automorphisms from
AutT (G) to T ∩ R form a finite group of automorphisms of T ∩ R; in particular,
Aut(G) is almost algebraic if and only if Aut(R) is almost algebraic.
Assertions (i) and (ii) above were deduced from Theorem 2.1 in [19]; (ii) was
proved earlier by D. Wigner [88], and (iii) is due to W. H. Previts and S. T. Wu
[73], where improved proofs were also given for (i) and (ii).
We note in particular that, in the light of Theorem 2.2(ii), Inn(G) is contained
in an almost algebraic subgroup of Aut(G), namely Aut0 (G). This turns out to
be useful in various contexts on account of certain properties of actions of almost
algebraic subgroups (see for example §§4.6 and 5.1).
Let G be a connected Lie group. Let T be the unique maximal torus contained
in the center of G. Then G/T is a connected Lie group whose center contains no
nontrivial compact subgroup of positive dimension, and hence Aut(G/T ) is an
almost algebraic subgroup of Aut(G′ ), where G′ is the Lie algebra of G/T . Each
α ∈ Aut(G) induces an automorphism of Aut(G/T ), say ᾱ. If η : Aut(G) →
Aut(G′ ) is the canonical quotient homomorphism, defined by η((α) = ᾱ for all
α ∈ Aut(G), then the image of any almost algebraic subgroup of Aut(G) under
η is almost algebraic (as the map is a restriction of a homomorphism of algebraic
groups). In particular η(Aut0 (G)) is an almost algebraic subgroup.
2.3
Groups with Aut(G) almost algebraic
For a class of connected Lie groups the following characterisation, incorporating
a partial converse of Theorem 2.2(i) is proved in [11]; there the issue is considered
for all Lie groups with finitely many connected components, but we shall restrict
here to when G is connected; (the general case involves some technicalities in its
formulation).
Theorem 2.3. ([11]) Let G be a connected Lie group admitting a faithful finitedimensional representation. Then Aut(G) is almost algebraic if and only if the
maximal torus contained in the center of G is of dimension at most one, and it is
also the maximal torus in the radical of G.
Let G be as in Theorem 2.3 and C be the maximal torus of the center of G and
suppose that C is one-dimensional. In the context of the examples of Lie groups
with Aut(G) having infinitely many distinct connected components discussed in
§1.3 it may be noted that if C coincides with the maximal torus in the radical
of G, then there does not exist a closed normal subgroup H such that G/H is a
torus, as used in the argument there.
9
It could happen that Aut(G) may be almost algebraic (which in view of Theorem 2.2(ii) is equivalent to c(G) being finite), for a connected Lie group G (not
admitting a faithful finite-dimensional representation) even if its center contains
a torus of dimension exceeding 1 (in fact of any given dimension). An example
of this was indicated in [19] (page 451) and has been discussed in detail in [73].
The main idea involved in the example is that Lie groups G can be constructed
such that the central torus, though of higher dimension, is the product of onedimensional tori, each of which is invariant under a subgroup of finite index in
Aut(G). It is not clear whether there could be situations, with c(G) finite, for
which this may also fail, namely with no one-dimensional compact subgroups contained in the center and invariant under a subgroup of finite index in Aut(G). It
is proved in [73] (Proposition 3.1), however that for a connected Lie group G of
the form H × T, where H is a connected Lie group and T is the one-dimensional
torus, Aut(G) has infinitely many connected components, and hence is not almost
algebraic.
2.4
Automorphisms preserving additional structure
When the Lie group has additional structure, the group of automorphisms preserving the structure would also be of interest. In this subsection we briefly recall
some results in this respect.
By a result of Hochschild and Mostow [56] if G is a connected complex affine
algebraic group then the connected component of the identity in the group A of
rational automorphisms of G is algebraic (with respect to a canonical structure
arising from the associated Hoff algebra), and moreover A is itself algebraic if
either the center of G is virtually unipotent (namely if it admits a subgroup of
finite index consisting of unipotent elements) or the center of a (and hence any)
maximal reductive subgroup of G is of dimension at most 1; the result may be
compared with Theorem 2.2 for real Lie groups. It may be noted that the results
of [56] are in the framework of automorphism groups of affine algebraic groups
over an algebraically closed field of characteristic zero. Some further elaboration
on the theme is provided by Dong Hoon Lee [67].
For the connected component of the group of complex analytic automorphisms
of a faithfully representable complex analytic group, a result analogous to that of
Hochschild and Mostow recalled above was proved by Chen and Wu [12].
2.5
Linearisation of the Aut(G) action on G
Under certain conditions, a connected Lie group can be realised as a subset of Rn ,
or the projective space Pn−1 , for some n ≥ 2, in such a way that the automorphisms
10
of G are restrictions of linear or, respectively, projective transformations. We call
this linearisation of the Aut(G)-action on G. We next discuss various results in
this respect.
Let G be a connected Lie group. By an affine automorphism of G we mean a
transformation of the form Tg ◦ τ where τ ∈ Aut(G), and Tg is a left translation
by an element g in G, namely Tg (x) = gx for all x ∈ G. We denote by Aff(G) the
group of all affine automorphisms. We identify G canonically as a subgroup of
Aff(G), identifying g ∈ G with Tg as above. Then Aff(G) is a semidirect product of
Aut(G) with G, with G as the normal subgroup, and we shall consider it equipped
with the Cartesian product topology. We shall denote by Aff0 (G) the connected
component of the identity in Aff(G).
Let A and B be the Lie algebras of Aut(G) and Aff(G) respectively. Let
Ad : Aff(G) → GL(B) be the adjoint representation of Aff(G). Let a be the
dimensions of Aut(G) and V = ∧a B, the vector space of a-th exteriors over B.
Let ρ : Aff(G) → GL(V ) be the representation arising as the a-th exterior power of
Ad; we call ρ the linearising representation for G (for reasons that would become
clear below). Let L be the vector subspace ∧a A of V = ∧a B; since a is the
dimension of A, L is a one-dimensional subspace.
Consider the (linear) action of Aff(G) on V via the representation ρ. Let
P = P (V ) denote the corresponding projective space, consisting of all lines (onedimensional subspaces) in V , equipped with its usual topology. The Aff(G)-action
on V induces an action of Aff(G) on P . We have the corresponding actions
of Aut(G) and G by restriction of the action (G being viewed as the group of
translations as above).
2.6
Embedding of G in a projective space
Let p0 denote the point of P corresponding to the line L as above. Let S be the
stabiliser of p0 under the action of G. Consider any g ∈ S. Then L is invariant
under the action of g on V and this means that the subspace A is invariant under
the action of g on B. In turn we get that Aut0 (G) is normalised by g in Aff(G),
and since Aff0 (G) is a semidirect product of Aut0 (G) and G this implies that g is
contained in the center of Aff0 (G). Conversely it is easy to see that every element
of the center of Aff0 (G) fixes p0 . Thus S coincides with the center of Aff0 (G). We
note also that
S = {g ∈ G | τ (g) = g for all τ ∈ Aut0 (G)}.
The orbit map g 7→ gp0 of G into P induces a canonical continuous bijection
j of G/S onto its image in P , defined by j(gS) = gp0 for all g ∈ G. When S
is trivial j defines a continuous embedding of G into P . We note that for any
11
τ ∈ Aut(G) and g ∈ G we have τ (g)p0 = τ gτ −1 p0 = τ gp0 = τ (gp0 ). Thus the
map j is equivariant with respect to the actions of Aut(G) on G/S and P , viz.
j(τ (gS)) = τ j(gS) for all g ∈ G and τ ∈ Aut(G). Thus when S is trivial the
orbits of the Aut(G)-action on G/S are in one-one correspondence in a natural
way with orbits of the Aut(G)-action on P that are contained in the image of
G/S.
Recall that Aut0 (G) is an almost algebraic subgroup of GL(G) (see Theorem 2.2(ii)). From this it can be deduced that the restriction of the representation ρ as above to Aut0 (G) is an algebraic representation, viz. restriction of an
algebraic homomorphism of algebraic groups. Hence every orbit of Aut0 (G) on P
is locally closed, namely open in its closure (see [72], Lemma 1.22, for instance).
Thus the above argument shows the following.
Theorem 2.4. Let G be a connected Lie group and S be the center of Aff0 (G).
Then every orbit of the action of Aut0 (G) on G/S is locally closed (viz. open in
its closure).
The subgroup S, which is noted to be contained in the center of G, contains
the maximal compact subgroup of the center. In particular if S is trivial then by
Theorem 2.2(i) Aut0 (G) has finite index in Aut(G) and hence the above theorem
implies that all orbits of Aut(G) on G are locally closed.
We recall here that the condition of the orbits being locally closed is well
studied in a wider context and is equivalent to a variety of other conditions of
interest; see [44]; (see also [42] for a more general result). It represents in various
ways the opposite extreme of the action being ergodic.
2.7
Embedding in a vector space
In analogy with the embedding of G as a subset of P (modulo the subgroup S
as above) we can also get an embedding of G in the vector space V as defined
above, as follows. Let the notation be as above and let v0 be a nonzero point of
L. We note that S as above is also the stabilizer of v0 under the G-action on V ,
since if g fixes v0 it fixes p0 , and if it fixes p0 then we have τ (g) = τ gτ −1 = g
for all τ ∈ Aut0 (G) and hence it fixes v0 . As L is invariant under the action
of Aut(G) we get a continuous homomorphism s : Aut(G) → R∗ such that for
all τ ∈ Aut(G), τ (v0 ) = s(τ )v0 . For any g ∈ G and τ ∈ Aut(G) we have
τ (g)v0 = (τ gτ −1 )v0 = τ g(τ −1 (v0 )) = s(τ −1 )τ gv0 . Consider the (linear) action
of Aut0 (G) on V such that τ ∈ Aut0 (G) acts by v ∈ V 7→ s(τ −1 )τ v. Then
the Aut0 (G)-orbits on G/S are in canonical correspondence with the orbits on V
under this Aut0 (G)-action. Under this action also the orbits are locally closed, by
algebraic group considerations.
12
2.8
Algebraicity of stabilizers
Let G be a connected Lie group and ρ : Aff(G) → GL(V ) be the corresponding
linearising representation as in §2.5. We recall that Aut0 (G) is an almost algebraic
subgroup of GL(G), G being the Lie algebra of G, and that the restriction of ρ to
Aut0 (G) is the restriction of a homomorphism of algebraic groups. In particular,
for any v ∈ V the stabiliser {τ ∈ Aut0 (G) | τ (v) = v} is an almost algebraic
subgroup of Aut0 (G) (algebraic subgroup if Aut0 (G) is algebraic). Let v0 be as
in §2.7 and S, as before, be the center of Aff0 (G), which is the stabiliser of v0
under the action as in §2.7. Then by the preceding observation for all g ∈ G the
subgroup {τ ∈ Aut0 (G) | τ (gS) = gS} is almost algebraic. In particular for a
connected Lie group G such that the center of Aff0 (G) is trivial, for all g ∈ G the
stabiliser {τ ∈ Aut0 (G) | τ (g) = g} of g under the action of Aut0 (G) is an almost
algebraic subgroup; this holds in particular when the center of G is trivial.
Let G be a connected Lie group and for g ∈ G let S(g) denote the stabiliser
{τ ∈ Aut(G) | τ (g) = g}. It is shown in [41] that when G is a simply connected
solvable Lie group, S(g) is an algebraic subgroup for all g ∈ G. In the general
case it is shown that the connected component of the identity in S(g) is an almost
algebraic subgroup for all g which are of the form exp ξ for some ξ in the Lie
algebra of G, exp being the exponential map associated with G. It is noted that
S(g) itself need not be an algebraic subgroup, as may be seen in the case when G
is the universal covering group of SL(2, R) and g is one of the generators of the
center of G (the latter is an infinite cyclic subgroup); in this case S(g) = Inn(G),
and it is of index 2 in Aut(G), but not an algebraic subgroup.
For elements z contained in the center of G it is proved in [47] (see also [19])
that the connected component of the identity in S(z) is an almost algebraic subgroup, and if G has no compact central subgroup of positive dimension then S(z)
itself is almost algebraic. It is however not true that S(z) is almost algebraic
whenever Aut(G) is almost algebraic; a counterexample in this respect may be
found in [73] (p. 432).
We note in particular that if A is the subgroup of Aut(G) consisting of all
automorphisms which fix the center pointwise then the connected component of
the identity in A is an almost algebraic subgroup, and if G has no compact central
subgroup of positive dimension then A is almost algebraic.
3
Dense orbits
It is well-known that an automorphism α of the n-dimensional torus Tn , n ≥ 2,
admits dense orbits (and is ergodic with respect to the Haar measure; see §4
13
for a discussion on the ergodicity condition) when dα has no eigenvalue which
is a root of unity. These automorphisms constitute some of the basic examples
in ergodic theory and topological dynamics, and have been much studied for
detailed properties from the point of view of the topics mentioned (see [14], [87],
for instance; see also [79] and [22] for generalisations). In this section we discuss
actions of subgroups of Aut(G) with a dense orbit on G.
3.1
Aut(G)-actions with dense orbits
Let G be a connected abelian Lie group. Then it has the form Rm × Tn for some
m, n ≥ 0 and in this case it is easy to see that the Aut(G)-action on G has dense
orbits, except when (m, n) = (0, 1) (viz. when G is the circle group); if n = 0 the
complement of 0 in Rm is a single orbit. If m ≥ 1 then the action of Aut0 (G) also
has a dense orbit on G, as may be seen using the isotropic shear automorphisms
(see §1.2) associated with homomorphisms of Rm into Tn .
Next let G be a 2-step connected nilpotent Lie group, namely [G, G] is contained in the center of G. Let Z denote the center of G. Then G/Z is simply
connected, and hence is Lie isomorphic to Rn for some n. The Aut(G)-action
on G factors to an action on G/Z. Using shear automorphisms (see §1.2) it can
be seen that the Aut(G)-action on G has a dense orbit on G if and only if the
Aut(G)-action on G/Z has a dense orbit. The latter condition holds in particular
if G is a free 2-step simply connected Lie group; it may be recalled that G is called
a free 2-step nilpotent Lie group if its Lie algebra is of the form V ⊕ ∧2 V , with
V a finite-dimensional vector space over R, and the Lie product is generated by
the relations [u, v] = u ∧ v for all u, v ∈ V , and [u, v ∧ w] = 0 for all u, v, w ∈ V .
When G is a free 2-step simply connected nilpotent Lie group the quotient G/Z
as above corresponds to the vector space V and every nonsingular automorphism
of V is the factor of a τ ∈ Aut(G), which leads to the observation as above.
For a general simply connected 2-step nilpotent Lie group G the Lie algebra
G of G can be expressed as (V ⊕ ∧2 V )/W , where V is the vector space G/Z (Z
being the center of G) and W is a vector subspace of ∧2 V ; V ⊕ ∧2 V is the free
2-step nilpotent Lie algebra with the structure as above, and any vector subspace
of ∧2 V is a Lie ideal in the Lie algebra (see [5], where the example is discussed
in a different context). It is easy to see that in this case the image of Aut(G) in
GL(V ), under the map associating to each automorphism its factor on V = G/Z,
is the subgroup, say I(W ), consisting of g ∈ GL(V ) such that the corresponding
exterior transformation ∧2 (g) of ∧2 V leaves the subspace W invariant. It can be
seen that I(W ) has an open dense orbit on V for various choices of W , and in
these cases by the argument as above Aut(G) has open dense orbit on G. For
example, if e1 , . . . , en , n ≥ 2 is a linear basis of V , then this is readily seen to hold
14
if W is the subspace spanned by e1 ∧ e2 or, more generally, by sets of the form
{e1 ∧ e2 , e3 ∧ e4 , . . . , e2k−1 ∧ e2k }, for n ≥ 2k.
Along the lines of the above arguments it can be seen that for any k ≥ 2 if G
is a free k-step simply connected Lie group then the action of Aut(G) on G has
an open dense orbit. Also, given a simply connected k-step nilpotent Lie group
H there exists simply connected k + 1-step nilpotent Lie group G such that H is
Lie isomorphic to G/Z, where Z is the center of G, and every automorphism of H
is a factor of an automorphism of G. Therefore using the examples as above one
can also construct examples, for any k ≥ 2, of simply connected k-step nilpotent
Lie groups such that the Aut(G)-action on G has an open dense orbit.
In the converse direction we have the following.
Theorem 3.1. Let G be a connected Lie group. Suppose that there exists g ∈ G
such that the closure of the Aut(G)-orbit of g has positive Haar measure in G.
Then G is a nilpotent Lie group.
A weaker form of this, in which it was assumed that the orbit is dense in G,
was proved in [23] (Theorem 2.1), but the same argument is readily seen to yield
the stronger assertion as above. By a process of approximation, Theorem 2.1 in
[23] was extended to all finite-dimensional connected locally compact groups, and
in the same way one can also get that the assertion in Theorem 3.1 holds also for
all finite-dimensional connected locally compact groups.
A nilpotent Lie group G need not always have dense orbits under the action of
Aut(G). Examples of 2-step simply connected nilpotent Lie groups with no dense
orbits under the Aut(G)-action are exhibited in [26]. There are also examples in
[26] of 3-step simply connected nilpotent Lie groups for which Aut(G) is a unipotent group, namely when Aut(G) is viewed a subgroup of GL(G) all its elements
are unipotent linear transformation, and consequently all orbits of Aut(G) on G
are closed, and hence lower dimensional, submanifolds of G.
There has been a detailed study of groups in which Aut(G) has only finitely
many, or countably many, orbits, in the broader context of locally compact groups,
and also abstract groups. The results in this respect for connected Lie groups will
be discussed in §3.7.
3.2
Connected subgroups of Aut(G) with dense orbits
Theorem 3.1 implies in particular that there is no connected Lie group G for
which Aut(G) has a one-parameter subgroup whose action on G has a dense
orbit; by the theorem, a Lie group G with that property would be nilpotent and
since it has to be noncompact, going to a quotient we get that for some n ≥ 1,
15
GL(n, R) has a one-parameter subgroup acting on Rn with a dense orbit, but
simple considerations from linear algebra rule this out.
On R2n viewed as Cn , n ≥ 2, we have linear actions of Cn−1 × C∗ , which
have an open dense orbit: the action of (z1 , . . . , zn−1 , z), where z1 , . . . , zn−1 ∈ C
and z ∈ C∗ , is defined on the P
standard basis vectors {e1 , . . . , en } by ej 7→ zej
n−1
for j = 1, . . . , n − 1 and en 7→ j=1
zj ej + zen . The same also holds for (outer)
cartesian products of such actions, and in particular we have linear actions of C∗n
on R2n , with dense orbits. Since for n ≥ 2 we can realise Rn+1 as a dense subgroup
of C∗n , we get linear actions Rn+1 on R2n admitting dense (but not open) orbits;
we note that in analogy with the above we can get linear actions of R∗n on Rn
with open dense orbits, but they do not yield actions of Rn with dense orbits.
It can be seen using the Jordan canonical form that there is no linear action of
R2 admitting dense orbits, and hence there is no action of R2 on any Lie group
admitting dense orbits, by an argument as above, using Theorem 3.1.
On Rn , apart from GL(n, R), various proper subgroups act transitively on
the complement of {0}; e.g. S · O(n, R) where S is the subgroup consisting of
nonzero scalar matrices and O(n, R) is the orthogonal group, or the symplectic
group Sp(n, R) for even n. Also if O(p, q) is the orthogonal group of a quadratic
form of signature (p, q) on Rn , n = p + q then the action of S ·O(p, q), with S as
above, on Rn has an open dense orbit (the complement consists of the set of zeros
of the quadratic form, which is a proper algebraic subvariety of Rn ).
Subgroups of GL(n, R) acting with an open dense orbit have been a subject
of much interest in another context; Rn together with such a subgroup is called a
pre-homogeneous vector space; the reader is referred to [63] for details.
3.3
Automorphisms with dense orbits
Analysis of the issue of dense orbits was inspired by a question raised by P. R.
Halmos in his classic book on Ergodic Theory ([51], page 29) as to whether a noncompact locally compact group can admit a (continuous) automorphism which is
ergodic with respect to the Haar measure of the group, namely such that there
is no measurable set invariant under the automorphism such that both the set
and its complement have positive Haar measure (see § 4.1 for more on ergodicity).
An automorphism as in Halmos’ question would have a dense orbit (assuming
the group to be second countable), and in particular one may ask whether there
exists an automorphism of a noncompact locally compact group with a dense orbit. This question was answered in the negative in a paper of R. Kaufman and
M. Rajagopalan [62] and T. S. Wu [89] for connected locally compact groups, and
N. Aoki [4] for a general locally compact group (see also [74] for some clarifications
on the proof in [4]); it may also be mentioned here that the analogous question is
16
studied for affine automorphisms of locally compact groups (namely transformations which are composite of a continuous group automorphism with a translation
by a group element) in [17] for connected groups, and in [61] in the generality of
all locally compact groups.
Though there are nonabelian compact groups admitting automorphisms with
dense orbits, such as C Z where C is a compact nonabelian group, for which the
shift automorphism has dense orbits, a compact connected Lie group admits such
an automorphism only if it is a torus of dimension n ≥ 2. For a compact semisimple Lie group G, Inn(G) is a subgroup of finite index in Aut(G) and the orbits
of Aut(G) on G are closed submanifolds of dimension less than that of G, and in
particular not dense in G; a general compact connected Lie group has a simple
Lie group as a factor and hence the preceding conclusion holds in this generality
also.
3.4
Zd -actions
The analogue of Halmos’ question for Zd -actions, namely the multi-parameter case
with d ≥ 1, and more generally actions of abelian groups of automorphisms, was
considered in [24], where the following is proved.
Theorem 3.2. Let G be a connected Lie group. Suppose that there exists an
abelian subgroup H of Aut(G) such that the H-action on G has a dense orbit.
Then there exists a compact subgroup C contained in the center of G such that
G/C is topologically isomorphic to Rn for some n ≥ 0; in particular G is a
two-step nilpotent Lie group. If moreover the H-action leaves invariant the Haar
measure on G then G is a torus.
Recall that the n-dimensional torus Tn , where n ≥ 2, admits automorphisms
with a dense orbit, and one can find such an automorphism contained in a subgroup A of Aut(Tn ) ≈ GL(n, Z) which is isomorphic to Zd for d ≤ n − 1. This
gives examples of subgroups A of Aut(Tn ) isomorphic to Zd for d ≤ n − 1, acting
with a dense orbit. Conversely every subgroup A of Aut(Tn ) isomorphic to Zd
and acting with a dense orbit contains an (individual) automorphism which has
a dense orbit; see [7]; see also [76] for analogous results in a more general setting
of automorphisms of general compact abelian groups.
It is easy to see that we have a Z2 -action on R with a dense orbit; the automorphisms defined respectively by multiplication by eα and −eβ , α, β > 0, generate
such an action when α/β is irrational. More generally, G = Rn × Tm admits a
Zd -action with a dense orbit if any only if m 6= 1 and d ≥ (n + 2)/2; see [24]
for details. Examples of nonabelian two-step nilpotent Lie groups G admitting
Z2 -actions with a dense orbit are given in [24].
17
It may be mentioned here that Theorem 3.2 is extended in [38] to actions on
general locally compact groups G, where it is concluded that under the analogous
condition there exists a compact normal subgroup C such that the quotient G/C
is a (finite) product of locally compact fields of characteristic zero.
In line with the above it would be interesting to know about Lie groups admitting actions by nilpotent or, more generally, solvable groups of automorphisms
with a dense orbit.
3.5
Discrete groups with dense orbits
Consider a connected Lie group G with a Lie subgroup H of Aut(G) such that
the H-action on G has an open dense orbit; e.g. Rn , n ≥ 1, and H = SL(n, R)
– see §3.1 for more examples (recall also that such a G is nilpotent). Let g ∈ G
be such that the H-orbit is open and dense in G. Let L be the stabiliser of g
under the H-action, viz. L = {τ ∈ Aut(G) | τ (g) = g}. Then it is easy to see
that for a subgroup Γ of H the Γ-action on G has a dense in G if and only if the
L-action on H/Γ has a dense orbit; this phenomenon is known as “duality” - see
for instance [6], [22] for some details). In certain situations, such as when Γ is a
lattice in H (viz. H/Γ admits a finite measure invariant under the action of H
on the left), the question of whether the action of a subgroup has dense orbits on
H/Γ is amenable via techniques of ergodic theory.
For G = Rn , n ≥ 2, we have Aut(G) ≈ GL(n, R), and there exist many discrete
subgroups of the latter whose action on Rn admits dense orbits. Let H = SL(n, R)
and Γ be a lattice in SL(n, R). We choose g as e1 where {e1 , . . . , en } is the
coordinate basis of Rn and let L be its stabiliser. It is known that this subgroup
acts ergodically on H/Γ and hence as noted above the Γ-action on Rn has a dense
orbit. This applies in particular to the subgroup SL(n, Z), consisting of integral
unimodular matrices, which is indeed a lattice in SL(n, R) (see [70], Ch. 7 or [75],
Ch. 10). We shall discuss more about the orbits of these in the next section.
There are also natural examples of discrete subgroups of SL(n, R) other than
lattices which have dense orbits under the action on Rn . For example if n is even
and Γ is a lattice in the symplectic group Sp(n, R) then it has dense orbits on
Rn ; stronger statements analogous to those for lattices in SL(n, R) are possible
but we shall not go into the details. There are also other examples, arising from
hyperbolic geometry. Let Γ be the fundamental group of a surface of constant
negative curvature whose associated geodesic flow is ergodic. Then Γ may be
viewed canonically as a subgroup of PSL(2, R) = SL(2, R)/{±I}, where I denotes
the identity matrix, and if Γ̃ is the lift of Γ in SL(2, R), then the action of Γ̃ on R2
has dense orbits; the action is ergodic with respect to the Lebesgue measure (see
[6] for some details; see §4 for a discussion on ergodicity). Analogous examples
18
can also be constructed in higher dimensions.
In place of Rn one may consider other nilpotent connected Lie groups G which
admit an open dense orbit under the action of Aut(G). It would be interesting to
know analogous results for discrete groups of automorphisms of other connected
nilpotent Lie groups, which however does not seem to be considered in the literature.
3.6
Orbit structure of actions of some discrete groups
In most of the cases considered in the earlier subsections where we conclude existence of a dense orbit, not all orbits may be dense, and in general there is no
good description possible of the ones that are not dense. However in certain cases
a more complete description is possible.
Expressing Tn as Rn /Zn , Aut(Tn ) can be realised as GL(n, Z), the group of
all n × n matrices with integer entries and determinant ±1. For v ∈ Rn whose
coordinates with respect to the standard basis (generating Zn ) are rational the
orbits under the Aut(Tn )-action are easily seen to be finite. It turns out, and not
too hard to prove (see [30], for instance) that conversely for any v at least one of
whose coordinates is irrational the Aut(Tn )-orbit is dense.
For the case of Rn , n ≥ 1 we noted above that if Γ is a lattice in SL(n, R)
then there exist dense orbits, by duality and ergodicity considerations. In fact
in this case it is possible to describe the dense orbits precisely. If SL(n, R)/Γ
is compact (viz. if Γ is a “uniform” lattice) then the orbit of every non-zero
point in Rn is dense in Rn . When SL(n, R)/Γ is noncompact (but a lattice) the
set of points whose orbits are not dense is contained in a union of countably
many lines (one-dimensional vector subspaces) in Rn . In the case of SL(n, Z) the
exceptional lines involved are precisely those passing through points in Rn with
rational coordinates (see [35] for more general results in this direction). These
results are consequences of the study of flows on homogeneous spaces which has
been a much studied topic in the recent decades, thanks to the work of Marina
Ratner on invariant measures of unipotent flows. We shall not go into details on
the topic here; the interested reader is referred to the expository works [22] and
[64] and other references there, for exploration of the topic. The result in the
special case of SL(n, Z) recalled above was first proved in [15] (see also [34] for a
strengthening in another direction).
As in the case of the issues in the previous section, it would be interesting
to know results analogous to the above for discrete groups of automorphisms of
other nilpotent connected Lie groups.
19
3.7
Aut(G)-actions with few orbits
Observe that when G = Rn , n ∈ N, Aut(G) is GL(n, R) and the action is transitive
on the complement of the zero element. Thus the action of Aut(G) on G has only
two orbits. The question as to when there can be only finitely many, or countably
many, orbits has attracted attention, not only for Lie groups, but in the general
context of locally compact groups; we shall indicate some of the results in that
generality, giving references, but our focus shall be on Lie groups.
Theorem 3.3. Let G be a locally compact group. Suppose that the action of
Aut(G) on G has only countably many orbits. Then the connected component
G0 of the identity in G is a simply connected nilpotent Lie group. Moreover, the
number of Aut(G)-orbits in G0 is finite and one of them is an open orbit.
The first statement in the theorem was proved in [84]. We note that for Lie
groups it can be deduced from Theorem 3.1 (which is a generalised version of
a theorem from [23]); since under the condition in the hypothesis at least one
of the orbits has to be of positive measure Theorem 3.1 yields that the group
is nilpotent, but on the other hand the condition also implies that there is no
compact subgroup of positive dimension contained in the center, so G must in
fact be a simply connected nilpotent Lie group. It is proved in [85] (Theorem 6.3)
that under the condition in the hypothesis one of the orbits is open. Moreover, as
G is a simply connected nilpotent Lie group, the Aut(G)-orbits are in canonical
one-one correspondence with Aut(G)-orbits on G via the exponential map, where
G is the Lie algebra of G, and the latter being an action of a real algebraic group,
the cardinality of the orbits can be countable only if it is finite and one of the
orbits is open.
Now let G be a connected Lie group such that the Aut(G)-action on G has
only finitely many orbits. Since the identity element is fixed, there are at least
two orbits; the group is said to be homogeneous if there are only two orbits. The
groups Rn , n ≥ 1, are homogeneous, and they are also readily seen to be the
only connected abelian Lie groups for which the Aut(G)-action has only finitely
many orbits. It turns out that Rn , n ≥ 1, are in fact the only connected locally
compact groups that are homogeneous (see [82], Theorem 6.4); all (not necessarily
connected) homogeneous locally compact groups have also been determined in [82];
K n , where K is the field of p-adic numbers, with p is a prime, or Q (with the
discrete topology), and n ∈ N, are some of the other examples of homogeneous
groups.
A locally compact group is said to be almost homogeneous if the Aut(G)-action
on G has 3 orbits. The class of groups with the property has been studied in [84]
and [85]. An almost homogeneous Lie group is a Heisenberg group, namely a
20
group defined on V ⊕ Z, where V and Z are vector spaces, with the product
defined by (v, x) · (w, y) = (v + w, x + y + 12 hv, wi), where h·, ·i is an alternating
non-degenerate bilinear form over V with values in Z; (we note that here the
group structure is viewed via identification with the corresponding Lie algebra,
and the factor 12 is introduced so that hv, wi is the Lie bracket of v and w).
Moreover, the pair of dimensions of the vector spaces V and Z are either of the
form (2n, 1), (4n, 2), (4n, 3), with n ∈ N, or one of (3, 3), (6, 6), (7, 7), (8, 5), (8, 6)
or (8, 7), and each of these pairs determines an almost homogeneous Heisenberg
group. Automorphism groups of these Heisenberg groups have been discussed
in [83].
It is easy to see that if G is a k-step simply connected nilpotent Lie group then
the number of Aut(G)-orbits on G is at least k + 1. If G is one of the Heisenberg
groups as above then for any r ≥ 1, the action of Aut(Gr ) on Gr has 2r + 1 orbits
(cf. [85], Proposition 6.8). Thus there exist nilpotent Lie groups with arbitrarily
large finite number of orbits under the action of the respective automorphism
group. We refer the reader to [85], and other references there, for further details
and also open problems on this theme, for general (not necessarily connected)
locally compact groups.
4
Ergodic theory of actions of automorphism
groups
In this section we discuss various aspects of ergodic theory of actions on Lie groups
by groups of automorphisms.
4.1
Preliminaries
We begin by briefly recalling some definitions and conventions which will be followed throughout. By a measure we shall always mean a σ-finite measure. Given
a measure space (X, M), an automorphism τ : X → X is said measurable if
τ −1 (E) ∈ M for all E ∈ M; a measure µ on (X, M) is said to be invariant under
a measurable automorphism τ if µ(τ −1 (E)) = µ(E) for all E ∈ M, and it is said
to be quasi-invariant if, for E ∈ M, µ(τ −1 (E)) = 0 if and only if µ(E) = 0; a
measure is said to be invariant or, respectively, quasi-invariant, under a group
of measurable automorphisms if it has the property with respect to the action
of each of the automorphisms from the group. A measure which is invariant or
quasi-invariant with respect to an action is said to be ergodic with respect to the
action if there are no two disjoint measurable subsets invariant under the action,
each with positive measure.
21
Two measures on a measurable space X are said to be equivalent if they have
the same sets of measure 0. It is easy to see that given a measure which is quasiinvariant under an action there exists a finite measure equivalent to it, which
is also quasi-invariant; in particular, given an infinite invariant measure there
exists a finite quasi-invariant measure equivalent to it (which may however not be
invariant under the action).
A measure µ on G is called a probability measure if µ(G) = 1. We denote by
P (G) the space of probability measures on G. For actions of a locally compact
second countable group every quasi-invariant probability measure can be “decomposed as a continuous sum” of ergodic quasi-invariant probability measures in a
canonical way, and hence it suffices in many respects to understand the ergodic
quasi-invariant measures (see, for instance, [70], Theorem 14.4.3).
In our context the measure space structure will always be with respect to the
Borel σ-algebra of the topological space in question.
4.2
Finite invariant measures
When G is the torus Tn , n ≥ 2, the Haar measure is invariant under the action of
Aut(G), which is an infinite discrete group. More generally, if G is a connected Lie
group such that the center contains a torus of positive dimension, then the Haar
measure of the maximal torus of the center, viewed canonically as a measure on G,
is invariant under Aut(G), which can be a group with infinitely many connected
components (see §2.2). It turns out however that when we restrict to almost
algebraic subgroups of Aut(G) the situation is quite in contrast, as will be seen
in Theorem 4.1 below.
Let G be a connected Lie group. For µ ∈ P (G) we denote by supp µ the
support of µ, namely the smallest closed subset of G whose complement is of
µ-measure 0. Let
I(µ) = {τ | µ is invariant under the action of τ },
and
J(µ) = {τ ∈ Aut(G) | τ (g) = g for all g ∈ supp µ}.
Then it can be seen that I(µ) and J(µ) are both closed subgroups of Aut(G) and
J(µ) is a normal subgroup of I(µ). It turns out that when G has no compact
subgroup of positive dimension contained in the center, the quotient I(µ)/J(µ) is
compact. In fact we have the following.
Theorem 4.1. (cf. [20]) Let G be a connected Lie group. Let A be an almost
algebraic subgroup of Aut(G). Then for any µ ∈ P (G), (I(µ) ∩ A)/(J(µ) ∩ A) is
compact.
22
Consider a connected Lie group G such that Aut(G) is almost algebraic (see
Theorem 3.1) then I(µ)/J(µ) is compact and in turn for any x ∈ supp µ the I(µ)orbit of x in G is compact; thus supp µ can be expressed as a disjoint union of
compact orbits of I(µ).
We recall that if G is a compact Lie group then the Haar measure of G is
a finite measure invariant under all automorphisms. On the other hand when
G = Rn for some n ≥ 1, the only finite invariant measure is the point mass at 0
(see Corollary 4.2 below). The general situation is, in a sense, a mix of the two
kinds of situations.
From Theorem 4.1 one can deduce the following, by going modulo the maximal compact subgroup contained in the center and applying Theorem 3.1 on the
quotient.
Corollary 4.2. Let G be a connected Lie group and H be a subgroup of Aut(G).
Let µ be a finite H-invariant measure on G. Then for any g ∈ supp µ the H-orbit
of g is contained in a compact subset of G. In particular if H does not have an
orbit other than that of the identity which is contained in a compact subset, then
the point mass at the identity is the only H-invariant probability measure on G.
For G = Rn , n ≥ 1, Therorem 4.1 implies in particular the following result,
which was proved earlier in [18].
Theorem 4.3. For µ ∈ P (Rn ), I(µ) is an algebraic subgroup of Aut(Rn ) =
GL(n, R).
Analogous results are also proved in [18] for projective transformations. The
approach involves the study of non-wandering points of the transformations, which
has been discussed in further detail in [28].
4.3
Convolution powers of invariant measures
We recall here the following result concerning convolution powers of probability
measures invariant under the action of a compact subgroup of positive dimension.
For a probability measure µ and k ∈ N we denote by µk the k-fold convolution
power of µ.
Theorem 4.4. Let µ be a probability measure on Rn , n ≥ 1. Suppose that µ
is invariant under the action of a compact connected subgroup K of Aut(G) =
GL(n, R) of positive dimension. Then one of the following holds:
i) µn is absolutely continuous with respect to the Lebesgue measure on Rn ;
ii) there exists an affine subspace W of Rn such that µ(W ) > 0.
23
This is a variation of Theorem 3.2 in [29] whose proof can be read off from
that of the latter; it can be seen that Condition (ii) above fails to hold if the Kaction has no nonzero fixed point and there is no proper vector subspace U with
µ(U) > 0 (the subspace U can also be stipulated to be invariant), so under these
assumptions µn is absolutely continuous with respect to the Lebesgue measure;
this is the formulation of Theorem 3.2 in [29].
It would be interesting to have a suitable analogue of the above theorem for
actions on a general connected Lie group G by compact subgroups of positive
dimension in Aut(G), (with the Lebesgue measure replaced by the Haar measure
of G).
4.4
Infinite invariant measures
The Lebesgue measure on Rn is invariant under the action of a large subgroup
of Aut(G) = GL(n, R), namely SL(n, R), the special linear group, but not the
whole of Aut(G). On the other hand if G is a connected semisimple Lie group
then Inn(G) is of finite index in Aut(G), and as G is unimodular it follows that
the Haar measure is invariant under the action of Aut(G). In many Lie groups of
common occurrence it is readily possible to determine whether the Haar measure is
invariant under all automorphisms, but there does not seem to be any convenient
criterion in the literature to test it.
Let G be a connected Lie group and let S denote the subgroup which is the
center of Aff0 (G), when G is viewed canonically as a subgroup of Aff(G) (see § 2.5).
Recall that by Theorem 2.4 all orbits of Aut0 (G), and more generally of any
almost algebraic subgroup A of Aut(G), on G/S are locally closed. By a theorem
of Effros [42] this implies that for every measure on G which is quasi-invariant and
ergodic under the action of an almost algebraic subgroup A of Aut(G), there exists
an A-orbit O such that the complement of O has measure 0 (that is, the measure
is “supported” on O, except that the latter is a locally closed subset which may
not be closed); in particular this applies to any infinite measure invariant under
the action of an almost algebraic subgroup A of Aut(G) as above.
4.5
Quasi-invariant measure and ergodicity
We recall that for an action of a group on a locally compact second countable
space, given an ergodic quasi-invariant measure µ, for almost all points in supp µ,
the orbit is dense in supp µ. Together with the results on existence of dense orbits
in § 3 this yields the following:
Theorem 4.5. Let G be a connected Lie group and let λ be a Haar measure of
24
G. Then λ is quasi-invariant under the action of Aut(G). Moreover the following
holds:
i) if λ is ergodic with respect to the Aut(G)-action on G then G is a nilpotent
Lie group.
ii) if λ is ergodic under the action of an abelian subgroup of Aut(G) then G is
a two-step nilpotent Lie group, with [G, G] compact.
iii) if λ is ergodic under the action of some α ∈ Aut(G) then G is a torus.
iv) if λ is ergodic under the action of an almost algebraic subgroup A of Aut(G)
then the A-action on G/C, where C is the maximal compact subgroup of G, has
an open dense orbit.
The first three statements are straightforward consequences of results from § 3
on the existence of a dense orbit and the observation above. The last assertion
follows from the fact that {ᾱ | α ∈ A} is an almost algebraic subgroup of Aut(G′ ),
where G′ is the Lie algebra of G/C, and hence its orbits on G′ are locally closed;
as G/C is a simply connected nilpotent Lie group the exponential map is a homeomorphism and hence the preceding conclusion implies that the A-orbits on G/C
are locally closed, and in particular an orbit which is dense is also open in G/C.
4.6
Stabilisers of actions of Lie groups
Let G be a connected Lie group and consider a (Borel-measurable) action of G on a
standard Borel space X. By the stabiliser of a point x ∈ X we mean the subgroup
{g ∈ G | gx = x}, and we denote it by Gx . Each Gx is a closed subgroup of G (see
[86], Corollary 8.8). When the action is transitive, namely when the whole of X
is a single orbit, then the stabilisers of any two points in X are conjugate to each
other. One may wonder to what extent this generalises to a general ergodic action,
with respect to a measure which is invariant or quasi-invariant under the action of
G. In this respect the following is known from [25]; the results in [25] strengthen
those from [45] in the case of Lie groups, while in [45] the issue of conjugacy of
the stabilisers is considered in the wider framework of all locally compact groups.
Theorem 4.6. Let G be a connected Lie group acting on a standard Borel space
X. Let µ be a measure on X which is quasi-invariant and ergodic with respect
to the action of G. Suppose also that for all x ∈ X the stabiliser Gx of x has
only finitely many connected components. Then there exists a subset N of X
with µ(N) = 0 such that for x, y ∈ X\N there exists α ∈ Aut0 (G) such that
α(Gx ) = Gy ; in particular Gx and Gy are topologically isomorphic to each other.
This is a variation of Corollary 5.2 in [25], where the ergodicity condition
involved is formulated in terms of a σ-ideal of null sets, rather than a quasi25
invariant measure. Without the assumption of Gx having only finitely many
connected components it is proved that for x, y in the complement of a null set
N, the connected components of the identity in Gx and Gy are topologically
isomorphic to each other ([25], Corollary 5.7). The proofs of these results are
based on consideration of the action of Aut(G) on the space of all closed subgroups
of G, the latter being equipped with the Fell topology, and the orbits or stabilisers
Gx , x ∈ X, under the action. The argument depends on the fact Aut0 (G) is an
almost algebraic subgroup of GL(G) containing Inn(G), G being the Lie algebra
of G (see Theorem 2.2(ii)); the proof shows that automorphisms α as in the
conclusion of the theorem may in fact be chosen to be from the smallest almost
algebraic subgroup containing Inn(G), and in particular if Inn(G) is an almost
algebraic subgroup of GL(G) then stabilisers of almost all points are conjugates
to each other. Other conditions under which the stabilisers of almost all points
may be concluded to be actually conjugate in G, and also examples for which it
does not hold, are discussed in [25]. We shall not go into the details of these here.
5
Some aspects of dynamics of Aut(G)-actions
In this section we discuss certain results on dynamics in which some features of
the actions of Aut(G) on G described in earlier sections play a role.
5.1
Stabilisers of continuous actions
Let G be a connected Lie group and consider a continuous action of G on a
compact Hausdorff space X. Here we recall a topological analogue of Theorem 4.6.
As before we denote by Gx the stabiliser of x under the action in question and we
denote by G0x the connected component of the identity in Gx . We recall that an
action is said to be minimal if there is no proper nonempty closed subset invariant
under the action. The following is proved in [27], Proposition 3.1.
Theorem 5.1. Let G be a connected Lie group acting continuously on a locally
compact Hausdorff space X. Suppose that there exists x ∈ X such that the orbit of
x is dense in X. Then there exists an open dense subset Y of X such that for all
y ∈ Y there exists α ∈ Aut0 (G) such that α(G0x ) = G0y . If the action is minimal
then G0x , x ∈ X, are Lie isomorphic to each other.
Further conditions which ensure G0x , x ∈ X, being conjugate in G are discussed
in [27]. The results of [27] generalise earlier results of C. C. Moore and G. Stuck
proved for special classes of Lie groups (see [27] for details).
26
5.2
Anosov automorphisms
Anosov diffeomorphisms have played an important role in the study of differentiable dynamical systems; we shall not go into much detail here on the issue –
the reader is referred to Smale’s expository article [81] for a perspective on the
topic. Hyperbolic automorphisms of tori Tn , n ≥ 2, namely automorphisms α
such that dα has no eigenvalue (including complex) of absolute value 1, serve
as the simplest examples of Anosov automorphisms. In [81] Smale described an
example, due to A. Borel, of a non-toral compact nilmanifold admitting Anosov
automorphisms; a nilmanifold is a homogeneous space of the form N/Γ where
N is a simply connected nilpotent Lie group and Γ is discrete subgroup, and an
Anosov automorphism of N/Γ is the quotient on N/Γ of an automorphism α of
N such that α(Γ) = Γ and dα has no eigenvalue of absolute value 1. Such a system can have a nontrivial finite group of symmetries and factoring through them
leads to some further examples on what are called infra-nilmanifolds, known as
Anosov automorphisms of infra-nilmanifolds. Smale conjectured in [81] that all
Anosov diffeomorphisms are topologically equivalent to Anosov automorphisms
on infra-nilmanifolds.
A broader class of Anosov automorphisms of nilmanifolds, which includes also
the example of Borel referred to above, was introduced by L. Auslander and
J. Scheuneman [5]. Their approach involves analysing Aut(N) for certain simply
connected nilpotent Lie groups N to produce examples of Anosov automorphisms
of compact homogeneous spaces of N/Γ for certain discrete subgroups Γ. In [16]
the approach in [5] was extended using some results from the theory of algebraic
groups and arithmetic subgroups, and some new examples of Anosov automorphisms were constructed. A large class of examples of Anosov automorphisms
were constructed in [31] by associating nilpotent Lie groups to graphs and studying their automorphism group in relation to the graph. Study of the automorphism groups of nilpotent Lie groups has also been applied to construct examples
of nilmanifolds that can not admit Anosov automorphisms (see [16] and [26]).
Subsequently Anosov automorphisms of nilmanifolds have been constructed
via other approaches; the reader is referred to [46] and [40] and various references
cited there for further details. There has however been no characterisation of
nilmanifolds admitting Anosov automorphisms, and it may be hoped that further
study of the automorphism groups of nilpotent Lie groups may throw more light
on the issue.
27
5.3
Distal actions
An action of a group H on a topological space X is said to be distal if for any pair
of distinct points x, y in X the closure of the H-orbit {(gx, gy) | g ∈ G} of (x, y)
under the componentwise action of G on the cartesian product space X × X does
not contain a point on the “diagonal”, namely of the form (z, z) for any z ∈ X.
The action on a locally compact group G by a group H of automorphisms of G is
distal if and only if under the H-action on G the closure the orbit of any nontrivial
element g in G does not contain the identity element.
Distality of actions is a classical topic, initiated by D. Hilbert, but the early
studies were limited to actions on compact spaces. The question of distality
of actions on Rn , n ≥ 1, by groups of linear transformations was initiated by
C. C. Moore [69] and was strengthened by Conze and Guivarc’h [13]; see also
H. Abels [1], where the results are extended to actions by affine transformations.
It is proved that the action of a group H of linear automorphisms of Rn is distal
if and only if the action of each h ∈ H (viz. of the cyclic subgroup generated by
it) is distal, and that it holds if and only if all (possibly complex) eigenvalues of
h are of absolute value 1 (see [13], [1]).
It is proved in [2] that the action of a group H of automorphisms of a connected
Lie group G is distal if and only if the associated action of H on the Lie algebra
G of G is distal (to which the above characterisations would apply).
For a subgroup H of GL(n, R), the H-action on Rn is distal if and only if
the action of its algebraic hull (Zariski closure) in GL(n, R) is distal, and the
action of an algebraic subgroup H of GL(n, R) on Rn is distal if and only if the
unipotent elements in H form a closed subgroup U (which would necessarily be
the unipotent radical) such that H/U is compact. (see [1], Corollaries 2.3 and 2.5).
Via the above correspondence these results can be applied also to actions of groups
of automorphisms of a general connected Lie group G.
An action of a group H on a space X is called MOC (short for “minimal orbit
closure”) if the closures of all orbits are minimal sets (viz. contain no proper
nonempty invariant closed subsets). When X is compact, the MOC condition is
equivalent to distality. For the action on a topological group G by a group of
automorphisms MOC implies distality. The converse is known in various cases,
including for actions on connected Lie groups G by groups of automorphisms (see
[2]). The reader is also referred to [3], [80] and [78] for some generalisations of this
as well as some of the other properties discussed above to more general locally
compact groups).
A connected Lie group G is said to be of type R if for all g ∈ G all (possibly
complex) eigenvalues of Ad g are of absolute value 1. Thus in the light of the
results noted above G is of type R if and only if the action of Inn(G) on G is
28
distal. This condition is also equivalent to G having polynomial growth, viz. for
any compact neighbourhood V of the identity the Haar measure of V n grows at
most polynomially in n; see [60] and also [49] and [68] for related general results.
5.4
Expansive actions
A homeomorphism ϕ of a compact metric space (X, d) is said to be expansive if
there exists ǫ > 0 such that for any pair of distinct points x, y there exists an
integer n such that d(ϕn (x), ϕn (y)) > ǫ; the notion may also be defined in terms
of a uniformity in place of a metric. An automorphism τ of a topological group
G is expansive if there exists a neighbourhood V of the identity such that for any
nontrivial element g in G there exists an integer n such that τ n (g) ∈
/ V . More
generally the action of a group Γ of automorphisms of a topological group G is
said to be expansive if there exists a neighbourhood V of the identity in G such
that for any nontrivial g in G there exists a γ ∈ Γ such that γ(g) ∈
/ V.
A compact connected topological group admits a group of automorphisms acting expansively only if it is abelian (cf. [65]) and finite dimensional (cf. [66]). For
compact abelian groups the expansiveness condition on actions of automorphism
groups has been extensively studied using techniques of commutative algebra (see
[79] for details).
If G is a connected Lie group with Lie algebra G, the action of a subgroup Γ
of Aut(G) on G is expansive if and only if for the induced action of Γ on G, for
any nonzero ξ ∈ G the Γ-orbit of ξ is unbounded in G (cf. [9], where the issue is
also considered for semigroups of endomorphisms). It is also deduced in [9] that
if Γ is a virtually nilpotent Lie group (viz. with a nilpotent subgroup of finite
index) of Aut(G) whose action on G is expansive, then Γ contains an element
acting expansively on G.
6
Orbits on the space of probability measures
The action of Aut(G) on G, where G is a Lie group, induces an action of Aut(G)
on the space of probability measures on G (see below for details). This action
plays an important role in many contexts. This section is devoted to recalling
various results about the action and their applications.
6.1
Preliminaries
Let G be a connected Lie group and as before let P (G) denote the space of
probability measures on G. We consider P (G) equipped with the weak∗ topology
29
with respect to the space of bounded continuous functions; we note that the
topology is Rmetrizable, Rand a sequence {µj } in P (G) converges to µ ∈ P (G) if
and only if G f dµj → G f dµ, as j → ∞, for all bounded continuous functions
f on G. We recall that the action of Aut(G) on G induces an action on P (G),
defined, for τ ∈ Aut(G) and µ ∈ P (G), by τ (µ)(E) = µ(τ −1 (E)) for all Borel
subsets E of G.
For λ, µ ∈ P (G), λ ∗ µ denotes the convolution product of λ and µ, and for any
µ ∈ P (G) and k ∈ N we denote by µk the k-fold convolution product µ ∗ · · · ∗ µ.
Behaviour of orbits of probability measures under actions of various subgroups
is of considerable interest in various contexts. One of the issues is to understand
conditions under which orbits of automorphism groups A ⊂ Aut(G) are locally
closed, namely open in their closures. We recall here that by a result of Effros [42]
this condition is equivalent to a variety of other “smoothness” conditions for the
action, including that the orbits map being an open (quotient) map onto its image
(the latter being considered with respect to the induced topology from P (G)).
As before for a connected Lie group G we view Aut(G) as a subgroup of GL(G),
where G is the Lie algebra of G, and a subgroup A of Aut(G) is said to be almost
algebraic if it is an almost algebraic subgroup of GL(G) (see § 2).
Theorem 6.1. (cf. [36], Theorem 3.3) Let G be a connected Lie group and A be
an almost algebraic subgroup of Aut(G). Let C be the maximal compact subgroup
contained in the center of G. Suppose that for any g ∈ G, {g −1 τ (g) | τ ∈ A} ∩ C
is finite. Then for any µ ∈ P (G) the A-orbit {τ (µ) | τ ∈ A} is open in its closure
in P (G). Moreover, if A consists of unipotent elements in GL(G) then the A-orbit
is closed in P (G).
The theorem implies in particular that when G has no compact subgroup of
positive dimension contained in its center, for the action of any almost algebraic
subgroup A of Aut(G) (which includes also the whole of Aut(G) in this case) the
orbits of A are locally closed.
6.2
Factor compactness and concentration functions
Let µ ∈ P (G). We denote by G(µ) the smallest closed subgroup of G containing
supp µ, the support of µ, by N(µ) the normaliser of G(µ) in G and by Z(µ) the
centraliser of supp µ, namely {g ∈ G | gx = xg for all x ∈ supp µ}.
A λ ∈ P (G) is called a factor of µ if there exists ν ∈ P (G) such that µ =
λ ∗ ν = ν ∗ λ. It is known that any factor of µ has its support contained in N(µ),
more specifically in a coset of G(µ) contained in N(µ) (see [32], Proposition 1.1
for the first assertion; the second is an easy consequence of the first).
30
A question of interest, which is not yet understood in full generality, is whether
given µ ∈ P (G) and a sequence {λj } of its factors there exists a sequence {zj }
in Z(µ) such that {λj zj } is relatively compact in P (G) (in place of Z(µ) one
may also allow in this respect the subgroup {g ∈ G | gµg −1 = µ}, which can
be bigger, but the distinction turns out to be a rather technical issue, and we
shall not concern ourselves with it here). It is known that given a sequence of
−1
factors {λj } there exists a sequence {xj } in N(µ) such that {xj µx−1
j }, {xj µxj }
and {xj λj } are relatively compact [32], [33], and it suffices to show that for such
a sequence {xj } the sequence of cosets {xj Z(µ)} is relatively compact in G/Z(µ).
It is known that this is true when G is an almost algebraic Lie group (and also
under some weaker, somewhat technical, conditions) (see [32], [36]); the proofs
however fall back on reducing the question to vector space situation and a more
direct approach connecting the question to Theorem 6.1 would be desirable.
A similar issue, or rather a more general one, arises in the study of the decay
of concentration functions of µn as n → ∞, where one would like to know under
what conditions a sequence of the form {xj µx−1
j } is relatively compact. We shall
however not go into the details of the concept or the results about it here; a
treatment of the topic in this perspective may be found in [21] and [39]; for a
complete resolution of the problem itself, which however involves a somewhat
different approach, the reader is referred to [59] and other references cited there.
6.3
Tortrat groups
A locally compact group is called a Tortrat group if for any µ ∈ P (G) which is
not idempotent (viz. such that µ2 6= µ) the closure of {gµg −1 | g ∈ G} in P (G)
does not contain any idempotent measure. We note that the issue concerns the
closure of the orbits of µ in P (G) under the action of the subgroup Inn(G); the
latter being contained in Aut0 (G) which is an almost algebraic subgroup is useful
in this respect.
Recall (see § 5.3) that a connected Lie group is said to be type R if for all
g ∈ G all eigenvalues of Adg are of absolute value 1. It was shown in [37] that for
a Lie group the two properties are equivalent:
Theorem 6.2. (cf. [37]) A connected Lie group is a Tortrat group if and only if
it is of type R.
The result implies also that a connected locally compact group is a Tortrat
group if and only if it is of polynomial growth (see [37] for details and the related
references).
31
6.4
Convergence of types
Let {λj } and {µj } be two sequences in P (G) converging to λ and µ respectively.
Suppose further that there exists a sequence {τj } in Aut(G) such that τj (λj ) = µj
for all j. A question, arising in various contexts in the theory of probability
measures on groups, is under what (further) conditions on λ and µ can we conclude
that there exists a τ ∈ Aut(G) such that τ (λ) = µ; the reader can readily convince
herself/himself that further conditions are indeed called for. When there exists a
τ as above we say that convergence of types holds. Though the term “type of a
measure” does not seem to make an appearance in literature freely, implicit in the
terminology above is the idea that {τ (µ) | τ ∈ Aut(G)} constitutes the “type of
µ”, or that µ and τ (µ) are of the same type for any τ ∈ Aut(G), and the question
is if you have pair of sequences of measures with the corresponding measures of
the same type, converging to a pair of measures, under what conditions can we
conclude the limits to be of the same type.
We now recall some results in this respect; for reasons of simplicity of exposition we shall not strive for full generality (see [20] for more details). As in § 2 let
Aff(G) be the group of affine automorphisms, and let ρ : Aff(G) → GL(V ) be the
linearising representation (see §2.5); recall that ρ is defined over V = ∧a B, where
B is the Lie algebra of Aff(G) and a is the dimension of Aut(G).
We say that µ ∈ P (G) is ρ-full if there does not exist any proper subspace U
of V that is invariant under ρ(g) for all g ∈ supp µ; though this condition is rather
technical, as it involves the linearising representation, it is shown in [20] that for
various classes of groups it holds under simpler conditions; for instance if G is
an almost algebraic subgroup of GL(n, R) then the condition holds for any µ for
which supp µ is not contained in a proper almost algebraic subgroup of G. The
following is a result which is midway between the rather technical Theorem 1.5
and the specialised result in Theorem 1.6 from [20], whose proof can be read off
from the proof of Theorem 1.6.
Theorem 6.3. Let G be a connected Lie group. Let {λj } and {µj } be sequences
in P (G) converging to λ and µ respectively. Suppose that there exists a sequence
{τj } in Aut0 (G) such that µj = τj (λj ) for all j and that λ and µ are ρ-full. Then
there exist sequences {θj } and {σj } in Aut(G) such that {θj } is contained in a
compact subset of Aut(G), {σj } are isotropic shear automorphisms, and τj = θj σj
for all j. If moreover supp λ is not contained in any proper closed normal subgroup
M of G such that G/M is a vector space of positive dimension, then {σj } is also
relatively compact.
We note that when {τj } as in the theorem is concluded to be relatively compact, convergence of types holds (for the given sequences); with the notation as
32
in the theorem, if τ is an accumulation point of {τj } then τ (λ) = µ. In the light
of the conclusion of the theorem it remains mainly to understand the asymptotic
behaviour under sequences of shear automorphisms.
A special case of interest is when {λj } are all equal, say λ. For this case we recall also the following result, for shear automorphisms, proved in [29] (Theorem 4.3
there); the result played an important role in the proof of the main theorem there
concerning embeddability in a continuous one-parameter semigroup, for a class of
infinitely divisible probability measures; we shall not go into the details of these
concepts here. It may be noted that one of the conditions in the hypothesis is as
in statement (i) of Theorem 4.4.
Theorem 6.4. Let G be a connected Lie group and T be a torus contained in
the center of G. Let H be a closed normal subgroup of G such that G/H is
topologically isomorphic to Rn for some n ≥ 1. Let W = G/H and θ : G → W
be the canonical quotient homomorphism. Let λ ∈ P (G) be such that θ(µ)n is
absolutely continuous with respect to the Lebesgue measure on W . Let {ϕj } be
a sequence of homomorphisms of W into T and for all j let τj ∈ Aut(G) be the
shear automorphism of G corresponding to ϕj . Suppose that the sequence {τj (λ)}
converges to a measure of the form τ (λ) for some τ ∈ Aut(G) fixing T pointwise.
Then there exists a sequence {σj } in Aut(G) such that σj (µ) = µ and {τj σj } is
relatively compact.
References
[1] Herbert Abels, Distal affine transformation groups, J. Reine Angew. Math. 299/300
(1978), 294 - 300.
[2] Herbert Abels, Distal automorphism groups of Lie groups, J. Reine Angew. Math.
329 (1981), 82 - 87.
[3] Herbert Abels, Which groups act distally? Ergodic Theory Dynam. Systems 3
(1983), no. 2, 167 - 185.
[4] Nobuo Aoki, Dense orbits of automorphisms and compactness of groups, Topology
Appl. 20 (1985), no. 1, 1 - 15.
[5] Louis Auslander and John Scheuneman, On certain automorphisms of nilpotent
Lie groups, 1970 Global Analysis (Proc. Sympos. Pure Math., Vol. XIV, Berkeley,
Calif., 1968) pp. 9 - 15, Amer. Math. Soc., Providence, R.I., USA.
[6] M. Bachir Bekka and Matthias Mayer, Ergodic theory and topological dynamics of
group actions on homogeneous spaces, London Mathematical Society Lecture Note
Series, 269. Cambridge University Press, Cambridge, 2000. x+200 pp.
33
[7] Daniel Berend, Ergodic semigroups of epimorphisms, Trans. Amer. Math. Soc. 289
(1985), no. 1, 393 - 407.
[8] Vitaly Bergelson and Alexander Gorodnik, Ergodicity and mixing of noncommuting epimorphisms, Proc. Lond. Math. Soc. (3) 95 (2007), no. 2, 329 - 359.
[9] Siddhartha Bhattacharya, Expansiveness of algebraic actions on connected groups,
Trans. Amer. Math. Soc. 356 (2004), no. 12, 4687 - 4700
[10] A. Borel and J. Tits, Groupes réductifs, Inst. Hautes Études Sci. Publ. Math.
No. 27 (1965), 55 - 150.
[11] P. B. Chen and T.S. Wu, On the component group of the automorphism group of
a Lie group, J. Austral. Math. Soc. Ser. A 57 (1994), no. 3, 365 - 385.
[12] P. B. Chen and T.S. Wu, Automorphism groups of complex analytic groups,
Tamkang J. Math. 33 (2002), no. 3, 259 - 264.
[13] J.-P. Conze and Yves Guivarc’h, Remarques sur la distalité dans les espaces vectoriels, C. R. Acad. Sci. Paris Sér. A 278 (1974), 1083 - 1086.
[14] I. P. Cornfeld, S. V. Fomin and Ya. G. Sinai, Ergodic theory, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
245, Springer-Verlag, New York, 1982, x+486 pp.
[15] J. S. Dani, Density properties of orbits under discrete groups, J. Indian Math. Soc.
(N.S.) 39 (1975), 189 - 217 (1976).
[16] S. G. Dani, Nilmanifolds with Anosov automorphism, J. London Math. Soc. (2) 18
(1978), 553 - 559.
[17] S. G. Dani, Dense orbits of affine automorphisms and compactness of groups, J.
London Math. Soc. (2) 25 (1982), 241 - 245.
[18] S. G. Dani, On ergodic quasi-invariant measures of group automorphism, Israel J.
Math. 43 (1982), 62 - 74.
[19] S. G. Dani, On automorphism groups of connected Lie groups, Manuscripta Math.
74 (1992), 445 - 452.
[20] S. G. Dani, Invariance groups and convergence of types of measures on Lie groups,
Math. Proc. Cambridge Philos. Soc. 112 (1992), 91 - 108.
[21] S. G. Dani, Scattering of products of random matrices, In: Analysis, geometry and
probability, 33 - 53, Texts Read. Math., 10, Hindustan Book Agency, Delhi, 1996.
[22] S. G. Dani, Dynamical systems on homogeneous spaces, In: Dynamical systems,
ergodic theory and applications, Edited and with a preface by Ya. G. Sinai, Encyclopaedia of Mathematical Sciences, 100, Mathematical Physics, I. Springer-Verlag,
Berlin, 2000. xii+459 pp.
34
[23] S. G. Dani, On automorphism groups acting ergodically on connected locally compact groups, Ergodic theory and harmonic analysis (Mumbai, 1999), Sankhya,
Ser. A 62 (2000), 360 - 366.
[24] S. G. Dani, On ergodic Zd -actions on Lie groups by automorphisms, Israel J. Math.
126 (2001), 327 - 344.
[25] S. G. Dani, On conjugacy classes of closed subgroups and stabilizers of Borel actions
of Lie groups, Ergodic Theory Dynam. Systems 22 (2002), 1697 - 1714.
[26] S. G. Dani, Some two-step and three-step nilpotent Lie groups with small automorphism groups. Trans. Amer. Math. Soc. 355 (2003), 1491 - 1503.
[27] S. G. Dani, On stabilizers of continuous actions of Lie groups, In: Algebraic groups
and arithmetic, 555 - 567, Tata Institute of Fundamental Research, Mumbai, 2004.
[28] S. G. Dani, Dynamical properties of linear and projective transformations and their
applications, Indian J. Pure Appl. Math. 35 (2004), 1365 - 1394.
[29] S. G. Dani, Yves Guivarc’h and Riddhi Shah, On the embeddability of certain
infinitely divisible probability measures on Lie groups), Math. Z. 272 (2012), no.
1-2, 361 - 379.
[30] S. G. Dani and M. Keane, Ergodic invariant measures for actions of SL(2, Z), Ann.
Inst. H. Poincaré Sect. B (N.S.) 15 (1979), no. 1, 79 - 84.
[31] S. G. Dani, and Meera G. Mainkar, Anosov automorphisms on compact nilmanifolds associated with graphs, Trans. Amer. Math. Soc. 357 (2005), no. 6, 2235 2251.
[32] S. G. Dani and M. McCrudden, Factors, roots and embeddability of measures on
Lie groups, Math. Z. 199 (1988), no. 3, 369 - 385.
[33] S. G. Dani and M. McCrudden, Convolution roots and embeddings of probability
measures on Lie groups, Adv. Math. 209 (2007), 198 - 211.
[34] S. G. Dani and Arnaldo Nogueira, On SL(n, Z)+ -orbits on Rn and positive integral
solutions of linear inequalities, J. Number Theory 129 (2009), no. 10, 2526 - 2529.
[35] S. G. Dani, and S. Raghavan, Orbits of Euclidean frames under discrete linear
groups. Israel J. Math. 36 (1980), no. 3-4, 300 - 320.
[36] S. G. Dani, and C.R.E. Raja, Asymptotics of measures under group automorphisms
and an application to factor sets, In: Lie groups and ergodic theory (Mumbai,
1996), 59 - 73, Tata Inst. Fund. Res. Stud. Math. 14, Tata Institute of Fundamental
Research, Bombay, 1998.
[37] S. G. Dani and C.R.E. Raja, A note on Tortrat groups, J. Theoret. Probab., 11
(1998), 571 - 576.
35
[38] S. G. Dani, Nimish A. Shah, and George A. Willis, Locally compact groups with
dense orbits under Zd -actions by automorphisms, Ergodic Theory Dynam. Systems
26 (2006), 1443 - 1465.
[39] S. G. Dani, and Riddhi Shah, Collapsible probability measures and concentration
functions on Lie groups, Math. Proc. Cambridge Philos. Soc. 122 (1997), no. 1, 105
- 113.
[40] Karel Dekimpe and Kelly Verheyen, Constructing infra-nilmanifolds admitting an
Anosov diffeomorphism, Adv. Math. 228 (2011), no. 6, 3300 - 3319.
[41] Dragomir Z. Djoković, Exponential map and automorphism group of a connected
Lie group, Amer. J. Math. 99 (1977), no. 5, 973 - 984.
[42] Edward G. Effros, Transformation groups and C ∗ -algebras. Ann. of Math. (2) 81
1965, 38 - 55.
[43] Manfred Einsiedler and Thomas Ward, Ergodic theory with a view towards number theory, Graduate Texts in Mathematics, 259. Springer-Verlag, London, 2011.
xviii+481 pp.
[44] J. Glimm, Locally compact transformation groups, Trans. Amer. Math. Soc. 101
(1961), 124 - 138.
[45] Valentin Ya. Golodets and Sergey D. Sinel’shchikov, On the conjugacy and isomorphism problems for stabilizers of Lie group actions, Ergodic Theory Dynam.
Systems 19 (1999), no. 2, 391 - 411.
[46] V. V. Gorbatsevich, On algebraic Anosov diffeomorphisms on nilmanifolds (Russian), Sibirsk. Mat. Zh. 45 (2004), 995 - 1021; translation in Siberian Math. J. 45
(2004), 821 - 839.
[47] Morikuni Goto, Standard Cartan subalgebras and the automorphism group of a
Lie group, Bull. Inst. Math. Acad. Sinica 8 (1980), no. 2-3, part 1, 251 - 261.
[48] Morikuni Goto and Edward T. Kobayashi, On the subgroups of the centers of
simply connected simple Lie groups – Classification of simple Lie groups in the
large, Osaka J. Math. 6 (1969), 251 - 281.
[49] Yves Guivarc’h, Croissance polynomiale et priodes des fonctions harmoniques, Bull.
Soc. Math. France 101 (1973), 333 - 379.
[50] Hasan Gündogan, The component group of the automorphism group of a simple Lie
algebra and the splitting of the corresponding short exact sequence, J. Lie Theory
20 (2010), no. 4, 709 - 737.
[51] P. R. Halmos, Lectures on ergodic theory, Chelsea Publishing Co., New York 1960,
vii+101 pp.
36
[52] Sigurdur Helgason, Differential geometry, Lie groups, and symmetric spaces, Corrected reprint of the 1978 original, Graduate Studies in Mathematics 34, American
Mathematical Society, Providence, RI, 2001. xxvi+641 pp.
[53] Joachim Hilgert and Karl-Hermann Neeb, Structure and geometry of Lie groups.
Springer Monographs in Mathematics. Springer, New York, 2012. x+744 pp.
[54] G. Hochschild, The automorphism group of a Lie group, Trans. Amer. Math. Soc.
72, (1952), 209 - 216.
[55] G. Hochschild, The structure of Lie groups. Holden-Day, Inc., San FranciscoLondon-Amsterdam, 1965, ix+230 pp.
[56] G. Hochschild and G.D. Mostow, Automorphisms of affine algebraic groups, J.
Algebra 13 (1969), 535 - 543.
[57] N. Jacobson, Structure and automorphisms of semisimple Lie groups in the large,
Ann. of Math. (2) 40, (1939), 755 - 763.
[58] N. Jacobson, A note on automorphisms and derivations of Lie algebras, Proc. Amer.
Math. Soc. 6, (1955), 281 - 283.
[59] Wojciech Jaworski, On shifted convolution powers and concentration functions in
locally compact groups, In: Probability on algebraic structures (Gainesville, FL,
1999), 23 - 41, Contemp. Math., 261, Amer. Math. Soc., Providence, RI, 2000.
[60] J. W. Jenkins, Growth of connected locally compact groups. J. Functional Analysis
12 (1973), 113 - 127.
[61] Tatsuro Kasuga, Affine endomorphisms with a dense orbit, Proc. Amer. Math. Soc.
95 (1985), no. 4, 659 - 661.
[62] R. Kaufman and M. Rajagopalan, On automorphisms of a locally compact group,
Michigan Math. J. 13 (1966), 373 - 374.
[63] Tatsuo Kimura, Introduction to prehomogeneous vector spaces, Translated from
the 1998 Japanese original by Makoto Nagura and Tsuyoshi Niitani and revised by
the author, Translations of Mathematical Monographs, 215. American Mathematical Society, Providence, RI, 2003. xxii+288 pp.
[64] Dmitry Kleinbock, Nimish Shah and Alexander Starkov, Dynamics of subgroup
actions on homogeneous spaces of Lie groups and applications to number theory,
Handbook of dynamical systems, Vol. 1A, 813 - 930, North-Holland, Amsterdam,
2002.
[65] Ping-fun Lam, On expansive transformation groups, Trans. Amer. Math. Soc. 150
(1970), 131 - 138.
37
[66] Wayne Lawton, The structure of compact connected groups which admit an expansive automorphism, In: Recent advances in topological dynamics (Proc. Conf.,
Yale Univ., New Haven, Conn., 1972; in honor of Gustav Arnold Hedlund), pp. 182
- 196. Lecture Notes in Math., Vol. 318, Springer, Berlin, 1973.
[67] Dong Hoon Lee, On the group of automorphisms of affine algebraic groups, Trans.
Amer. Math. Soc. 237 (1978), 145 - 152.
[68] Viktor Losert, On the structure of groups with polynomial growth. Math. Z. 195
(1987), no. 1, 109 - 117.
[69] Calvin C. Moore, Distal affine transformation groups, Amer. J. Math. 90 (1968),
733 - 751.
[70] Dave Witte Morris, Introduction to arithmetic groups, Deductive Press, 2015.
xii+475 pp.
[71] Shingo Murakami, On the automorphisms of a real semi-simple Lie algebra, J.
Math. Soc. Japan 4, (1952), 103 - 133; Supplements and corrections, J. Math. Soc.
Japan 5, (1953). 105 - 112.
[72] Gopal Prasad, Lattices in semisimple groups over local fields, Studies in algebra
and number theory, pp. 285 - 356, Adv. in Math. Suppl. Stud., 6, Academic Press,
New York-London, 1979.
[73] W. H. Previts and S.T. Wu, On the group of automorphisms of an analytic group,
Bull. Austral. Math. Soc. 64 (2001), 423 - 433.
[74] W. H. Previts and T.S. Wu, Dense orbits and compactness of groups, Bull. Austral.
Math. Soc. 68 (2003), no. 1, 155 - 159.
[75] M. S. Raghunathan, Discrete subgroups of Lie groups. Ergebnisse der Mathematik
und ihrer Grenzgebiete, Band 68. Springer-Verlag, New York-Heidelberg, 1972.
ix+227 pp.
[76] C. R. E. Raja, On the existence of ergodic automorphisms in ergodic Zd -actions on
compact groups, Ergodic Theory Dynam. Systems 30 (2010), 1803 - 1816.
[77] C. R. E. Raja and Riddhi Shah, Distal actions and shifted convolution property,
Israel J. Math. 177 (2010), 391 - 412.
[78] C. R. E. Raja and Riddhi Shah, Some properties of distal actions of locally compact
groups, preprint.
[79] Klaus Schmidt, Dynamical systems of algebraic origin, [2011 reprint of the 1995
original], Modern Birkhuser Classics. Birkhuser/Springer Basel AG, Basel, 1995,
xviii+310 pp.
38
[80] Riddhi Shah, Orbits of distal actions on locally compact groups, J. Lie Theory 22
(2012), 587 - 599.
[81] S. Smale, Differentiable dynamical systems, Bull. Amer. Math. Soc. 73 (1967), 747
- 817.
[82] Markus Stroppel, Homogeneous locally compact groups, J. Algebra 199 (1998), 528
- 543.
[83] Markus Stroppel, Homogeneous symplectic maps and almost homogeneous Heisenberg groups, Forum Math. 11 (1999), 659 - 672.
[84] Markus Stroppel, Locally compact groups with many automorphisms, J. Group
Theory 4 (2001), 427 - 455.
[85] Markus Stroppel, Locally compact groups with few orbits under automorphisms,
Proceedings of the 16th Summer Conference on General Topology and its Applications (New York), Topology Proc. 26 (2001/02), 819 - 842.
[86] V. S. Varadarajan, Geometry of quantum theory. Second edition. Springer-Verlag,
New York, 1985. xviii+412 pp.
[87] Peter Walters, An introduction to ergodic theory, Graduate Texts in Mathematics,
79. Springer-Verlag, New York-Berlin, 1982, ix+250 pp.
[88] D. Wigner, On the automorphism group of a Lie group. Proc. Amer. Math. Soc.
45 (1974), 140 - 143; Erratum: Proc. Amer. Math. Soc. 60 (1976), 376.
[89] T. S. Wu, Continuous automorphisms on locally compact groups, Math. Z. 96
(1967), 256 - 258.
[90] Ta Sun Wu, On the automorphism group of a connected locally compact topological
group. Proc. Edinburgh Math. Soc. (2) 35 (1992), no. 2, 285 - 294.
S.G. Dani
Department of Mathematics
Indian Institute of Technology Bombay
Powai, Mumbai 400005
India
E-mail: [email protected]
39
| 4math.GR
|
March 23, 2018
GAUSSIAN ELIMINATION IN SYMPLECTIC AND ORTHOGONAL GROUPS
SUSHIL BHUNIA, AYAN MAHALANOBIS, PRALHAD SHINDE AND ANUPAM SINGH
IISER Pune, Dr. Homi Bhabha Road, Pashan, Pune 411008, INDIA.
A BSTRACT. This paper studies algorithms similar to the Gaussian elimination algorithm in symplectic and
orthogonal groups. We discuss two applications of this algorithm. One computes the spinor norm and the other
computes the double coset decomposition with respect to Siegel maximal parabolic subgroup.
arXiv:1504.03794v3 [math.GR] 15 Sep 2017
1. I NTRODUCTION
Gaussian elimination is a very old theme in Mathematics. It appeared in print as chapter eight in a
Chinese mathematical text called, “The nine chapters of the mathematical art”. It is believed, a part of that
book was written as early as 150 BCE. For a historical perspective on Gaussian elimination, we refer to a
nice work by Grcar [9].
In this paper, we work with Chevalley generators [4, §11.3]. Chevalley generators for the special linear
group are elementary transvections. Transvections of the special linear groups have many interesting applications. Hildebrand [12] showed that random walk based on transvections become close to the uniform
distribution fast.
Chevalley generators for other classical groups are known for a very long time. However, its use in rowcolumn operations in symplectic and orthogonal groups is new. We develop row-column operations, very
similar to the Gaussian elimination algorithm for special linear groups. We call our algorithms Gaussian
elimination in symplectic and orthogonal groups respectively. Similar algorithm for unitary groups [13] is
available.
From our algorithm, one can compute the spinor norm easily, see Section 6. Murray and RoneyDougal [15] studied computing spinor norm earlier. Our algorithm can also be used to compute the double
coset decomposition corresponding to the Siegel maximal parabolic subgroup, see Section 7.
2. E XISTING WORK
We report on some existing work that are relevant to this paper. In computational group theory, Gaussian
elimination algorithms are seen as a subprocess of the constructive group recognition project. In this case,
a group G is defined by a set of generators hXi = G, the problem is to write g ∈ G as a word in X.
Brooksbank [3, Section 5] presents an idea similar to that in our algorithm. His main interest lies in
constructive group recognition, writing g as a word in X. He is not particularly interested in developing
a Gaussian elimination algorithm. In a Gaussian elimination algorithm two important components are:
elementary matrices and elementary operations. Brooksbank’s elementary matrices are the output of a
probabilistic Las Vegas algorithm. So, it is difficult to judge, if he has the same elementary matrices as
ours. He does not define elementary operations. He uses a low dimensional oracle to solve the word
problem. Our algorithm is more straightforward and works directly with elementary matrices. It seems
that his methods could be modified to produce a Gaussian elimination algorithm in all classical groups in
finite fields of all characteristics. However, his treatment depends on the primitive element ρ of F×
q and
on expressing Fq as a finite dimensional vector space over F p – the prime subfield. This suggests that his
algorithm would only work for finite fields.
E-mail address: [email protected], [email protected], [email protected],
[email protected].
2010 Mathematics Subject Classification. 11E57, 15A21.
Key words and phrases. Symplectic group, orthogonal groups, Gaussian elimination, spinor norm.
This work is supported by a SERB research grant.
1
Costi [8] develops an algorithm similar to ours using standard generators. These standard generators are
defined using a primitive element ω of the finite field. When one uses the primitive element of the finite
field, then to work with an arbitrary field element, one needs to solve the discrete logarithm problem in ω.
Discrete logarithm problem in finite fields are hard to solve. There is no known polynomial time algorithm
to solve them. This makes programs using the standard generators slower. In the Magma implementation
that we talk about later, for small fields, the field elements are represented by Zech logarithm. So, there is
no need to compute the discrete logarithm. However, one needs to compute various powers of ω and ω −1
in Costi’s algorithm. This is the main reason behind Costi’s algorithm to be slower than ours. Furthermore,
Costi’s algorithms are recursive. Moreover, Costi’s algorithm cannot be extended to infinite fields because
of its use of a primitive element.
Cohen, Murray and Taylor [7] proposed a generalized algorithm using the row-column operations, using
a representation of Chevalley groups. The key idea there was to bring down an element to a maximal
parabolic subgroup and repeat the process inductively. Here we use the natural matrix representation of
these groups. Thus our algorithm is more direct and works with matrices explicitly and effectively. A
novelty of our algorithm is that we do not need to assume that the Chevalley generators generate the group
under consideration. Thus our algorithm proves independently the fact that these groups are generated by
elementary matrices.
3. M AIN R ESULT
Algorithms that we develop in this paper work only for a given bilinear form (or quadratic form) β (see
Equations 4.1- 4.5). Though in our algorithm, we work with only one bilinear form (or quadratic form)
β , given by a fixed basis, with a suitable change of basis matrix our algorithm works on any equivalent
bilinear forms (or quadratic forms). Our algorithm work well on fields of all characteristics for symplectic
and orthogonal groups.
Another way to look at this paper, we have an algorithmic proof of this well-known theorem. For definitions of elementary matrices, one can look ahead to Section 5.
Theorem A. Let k be a field. For d ≥ 4 or l ≥ 2 following holds:
(A ) Every element of the orthogonal group O(d, k) can be written as a product of elementary matrices
and a diagonal matrix. Furthermore, the diagonal matrix is of the form
diag(1, . . . , 1, λ , 1, . . . , 1, λ −1 )
λ ∈ k×
for O+ (2l, k)
diag(ϑ , 1, . . . , 1, λ , 1, . . . , 1, λ −1 ) λ ∈ k× and ϑ = ±1 for O(2l + 1, k)
diag(1, 1, 1, . . . , λ , 1, . . . , λ −1 )
| {z } | {z }
(l−1)
λ ∈ k×
for O− (2l, k).
(l−1)
(B) Every element of the symplectic group Sp(2l, k) can be written as a product of elementary matrices.
This theorem has a surprising corollary. It follows that:
Corollary B. Let char(k) be odd. In a split orthogonal group O+ (d, k), the image of λ in k× /k×2 is the
spinor norm.
Furthermore, the spinor norm can also be computed using our algorithm for the twisted orthogonal group
in odd characteristics, see Section 6.1.
Corollary C. Let k be a perfect field of characteristics 2. Every element of the orthogonal group O(2l +1, k)
can be written as the product of elementary matrices.
We have an efficient algorithm to compute the spinor norm. Since the commutator subgroup of the
orthogonal group is the kernel of the spinor norm restricted to special orthogonal group, the above corollary
is a membership test for the commutator subgroup in the orthogonal group. In other words, an element g
in the special orthogonal group belongs to the commutator subgroup if and only if the λ it produces in the
Gaussian elimination algorithm is a square in the field, see Equation 4.6.
2
The bilinear form that we use and the generators that we define have its roots in the abstract root system
of a semisimple Lie algebra and Chevalley groups defined by Chevalley and Steinberg [6, 16]. However we
assume no knowledge of Lie theory or Chevalley groups in this paper.
4. O RTHOGONAL AND S YMPLECTIC G ROUPS
We begin with a brief introduction to orthogonal and symplectic groups. We follow Carter [4], Taylor [17]
and Grove [10] in our introduction. In this section, we fix some notations which will be used throughout
this paper. We denote the transpose of a matrix X by TX. As usual, tei, j denote the matrix unit with t in the
(i, j) place and 0 everywhere else.
Let V be a vector space of dimension d over a field k. We write the dimension of V as d where d = 2l + 1
or d = 2l and l ≥ 1. Let β : V × V → k be a bilinear form. By fixing a basis of V we can associate a
matrix to β . We shall abuse the notation slightly and denote the matrix of the bilinear form by β itself.
Thus β (x, y) = Txβ y where x, y are column vectors. We will work with non-degenerate bilinear forms,
which implies, det β 6= 0. A symmetric or skew-symmetric bilinear form β satisfies β = Tβ or β = −Tβ
respectively. In even order, for characteristic 2, the symmetric and skew-symmetric forms are the same. By
fixing a basis for V , we identify GL(V ) with GL(d, k) and treat orthogonal groups and symplectic groups
as subgroups of the matrix group GL(d, k).
Definition 4.1 (Symplectic Group). A square matrix X of size d is called symplectic if TXβ X = β where β
is skew-symmetric. The set of symplectic matrices form the symplectic group.
In this paper, we deal with the symplectic group defined by the bilinear form defined by Equation 4.1. So
any mention of symplectic group means this one particular symplectic group, unless stated otherwise.
Up to equivalence, there is an unique non-degenerate skew-symmetric (or alternating) bilinear form over
a field k. Furthermore a non-degenerate skew-symmetric (or alternating) bilinear form exists only in even
dimension. Fix a basis of V as {e1 , . . . , el , e−1 , . . . , e−l } so that the matrix β is:
0 Il
(4.1)
β=
.
−Il 0
The symplectic group with this β is denoted by Sp(2l, k).
Definition 4.2. Let Q is a quadratic form. The orthogonal group associated with Q is defined as:
O(d, Q) := {X ∈ GL(V ) | Q(X(v)) = Q(v) for all v ∈ V }.
As the quadratic form is defined in a slightly different way in case of even characteristics we describe the
orthogonal groups for odd and even characteristics separately.
4.1. Orthogonal groups for char(k)6= 2. Recall that one can easily recover the bilinear form from the
quadratic form Q by the formula:
1
B(x, y) = {Q(x + y) − Q(x) − Q(y)}
2
and it is easy to see that a matrix X satisfies TXβ X = β if and only if Q(X(x)) = Q(x) for all x ∈ V . We
work with a particular non-degenerate quadratic form Q, however when the characteristics of k is not 2 this
corresponds to β being non-degenerate.
Let k be a field of odd characteristic. We work with the following non-degenerate bilinear forms: Fix
a basis {e0 , e1 , . . . , el , e−1 , . . . , e−l } for odd dimension and {e1 , . . . , el , e−1 , . . . , e−l } for even dimension so
that the matrix β is:
2
0
0
0 0 Il when d = 2l + 1
(4.2)
β=
0 Il 0
0 Il
when d = 2l.
Il 0
3
For the twisted form, we fix a basis {e1 , e−1 , e2 , . . . , el , e−2 , . . . , e−l } so that the matrix β is:
β0 0
0
0 Il−1 ,
(4.3)
β =0
0 Il−1 0
1 0
where β0 =
and ε is a fixed non-square in k.
0 ε
Note that if k(= Fq ) is a finite field of odd characteristic. If d is odd then there is only one orthogonal
group up to conjugation [10, Page 79] and thus we can fix β as above. However, up to conjugation there are
two different orthogonal groups [10, Page 79] in even dimension d = 2l. For orthogonal group with even
dimension we fix β as above. Thus, we cover all finite orthogonal groups over a field of odd characteristics.
4.2. Orthogonal groups for char(k) = 2. Assume that char(k)=2, in this case the quadratic form is defined
in a slightly different way. A quadratic form Q is defined as follows:
Q(λ x + µy) = λ 2 Q(x) + µ 2 Q(y) + λ µB(x, y)
for all x, y ∈ V , λ , µ ∈ k, and B a symmetric bilinear form on V which is called the associated bilinear form
of Q.
Let k be a field of even characteristics. We work with following non-degenerate quadratic forms [4,
Page 10]. Fix a basis {e0 , e1 , . . . , el , e−1 , . . . , e−l } for odd dimension and {e1 , . . . , el , e−1 , . . . , e−l } for even
dimension so that the quadratic forms and their associated bilinear forms are as follows:
0
0
0
x02 + x1 x−1 + . . . + xl x−l , with associated form 0 0 Il , when d = 2l + 1
(4.4)
Q(x) =
0 Il 0
0 Il
with associated form
, when d = 2l.
x1 x−1 + . . . + xl x−l ,
Il 0
For the twisted form, we fix a basis {e1 , e−1 , e2 , . . . , el , e−2 , . . . , e−l } so that the quadratic form and its
associated bilinear form is as follows:
β0 0
0
2
0 Il−1 ,
(4.5)
Q(x) = α(x12 + x−1
) + x1 x−1 + . . . + xl x−l , with the associated form 0
0 Il−1 0
0 1
2
where αt + t + α is irreducible in k[x] and β0 =
.
1 0
Note that if k is a perfect field of even characteristics. It is well known that if dim(V ) = 2l + 1 then there
is only one non-degenerate quadratic form upto equivalence. However, in case of dim(V ) = 2l there are
two quadratic forms upto equivalence. Thus, we cover all orthogonal groups over a perfect field of even
characteristics.
Let Q be a non-degenerate quadratic form defined as above. For a fixed basis, we note that any isometry
g satisfying Q(g(v)) = Q(v) for all v ∈ V also satisfies Tgβ g = β . However the converse is not true. We
denote othogonal groups associated with Q by O(2l + 1, k), O+ (2l, k) and O− (2l, k) respectively.
Let k be a field of odd characteristics. We denote by Ω(d, k) the commutator subgroup of the orthogonal
group O(d, k) which is equal to the commutator subgroup of SO(d, k). There is a well known exact sequence
(4.6)
2
1 −→ Ω(d, k) −→ SO(d, k) −→ k× /k× −→ 1
Θ
m
where Θ is the spinor norm. The spinor norm is defined as Θ(g) = ∏ Q(vi ) where g = ρv1 · · · ρvm is written
i=1
as a product of reflections.
4
5. S OLVING THE WORD PROBLEM IN G
In computational group theory, one is always looking for algorithms that solve the word problem. When
G is a special linear group, one has a well-known algorithm to solve the word problem – the Gaussian
elimination. One observes that the effect of multiplying an element of the special linear group by an
elementary matrix (also known as elementary transvection) from left or right is either a row or a column
operation respectively. Using this algorithm one can start with any matrix g ∈ SL(l + 1, k) and get to the
identity matrix thus writing g as a product of elementary matrices [1, Proposition 6.2]. One of the objective
of this paper is to discuss a similar algorithm for orthogonal and symplectic groups, with a set of generators
that we will call elementary matrices in their respective groups.
We first describe the elementary matrices and the row-column operations for the respective groups. These
row-column operations are nothing but multiplication by elementary matrices from left and right respectively. Here elementary matrices used are nothing but Chevalley generators which follows from the theory
of Chevalley groups.
The basic idea of the algorithm is to use the fact that multiplying any orthogonal matrix by any one
of the generators enables us to perform row or column operations. The relation Tgβ g = β gives us some
compact relations among the blocks of g which can be used to make the algorithm more faster. To make the
algorithm simple we will write the algorithm for O(2l + 1, k), O+ (2l, k) and O− (2l, k) separately.
5.1. Groups in which Gaussian elimination works.
• Symplectic groups: Since all non-degenerate skew-symmetric bilinear forms are equivalent [10,
Corollary 2.12], we have a Gaussian elimination algorithm for all symplectic groups over an arbitrary field.
• Orthogonal groups:
– Since non-degenerate symmetric bilinear forms over a finite field of odd characteristics are
classified [10, Page 79] according to the β ( see Equations 4.2 & 4.3), we have a Gaussian
elimination algorithm for all orthogonal groups over a finite field of odd characteristics.
– Since non-degenerate quadratic forms over a perfect field of even characteristics can be classified [4, Page 10] according to quadratic forms Q(x) defined in Equations 4.4 & 4.5, we have a
Gaussian elimination algorithm for all orthogonal groups over a perfect field of even characteristics.
– Furthermore, we have Gaussian elimination algorithm for orthogonal groups that are given by
the above bilinear forms or quadratic forms over arbitrary fields. This algorithm also works for
bilinear or quadratic forms that are equivalent to the above forms.
5.2. Gaussian elimination for matrices of even size – orthogonal group O+ (d, k) and symplectic groups.
Recall that the bilinear forms β are the following:
0 Il
• For symplectic group, Sp(d, k), d = 2l and β =
.
−I
0
l
0 Il
+
• For orthogonal group, O (d, k), d = 2l and β =
.
Il 0
T
Note that any isometry g satisfying the quadratic form Q also
satisfy
gβ g = β . The main reason our
A B
algorithm works is the following: Recall that a matrix g =
, where A, B,C & D are matrices of
C D
size l, is orthogonal or symplectic if Tgβ g = β for the respective β . After some usual calculations, for
orthogonal group it becomes
T
CA + TAC TCB + TAD
0 Il
(5.1)
TDA + TBC TDB + TBD = I
l 0
The above equation implies among other things, TCA + TAC = 0. This implies that TAC is skew-symmetric.
In an almost identical way one can show, if g is symplectic, TAC is symmetric. The working principle of
our algorithm is simple – use the symmetry of TAC. The problem is, for arbitrary A and C, it is not easy
5
to use this symmetry. In our case we were able to reduce A to a diagonal matrix and then it is relatively
straightforward to use this symmetry. We will explain the algorithm in details later. First of all, let us
describe the elementary matrices and the row-column operations for orthogonal and symplectic groups.
The genesis of these elementary matrices lie in the Chevalley basis of simple Lie algebras. We won’t
go into details of Chevalley’s theory in this paper. Furthermore, we don’t need to, the algorithm that we
produce will show that these elementary matrices are generators for the respective groups.
Next we present the elementary matrices for the respective groups and then the row-column operations
in a tabular form.
5.2.1. Elementary matrices (Chevalley generators) for orthogonal group O+ (d, k) of even size. Following
the theory of root system in a simple Lie algebra, we index rows by 1, 2, . . . , l, −1, −2, . . . , −l. For t ∈ k, the
elementary matrices are defined as follows:
TABLE 1. Elementary matrices for O+ (2l, k)
char(k)
both
Elementary matrices
xi, j (t)
I + t(ei, j − e− j,−i ),
i 6= j
xi,− j (t)
I + t(ei,− j − e j,−i ),
i< j
x−i, j (t)
I + t(e−i, j − e− j,i ),
i< j
wi
I − ei,i − e−i,−i + ei,−i + e−i,i 1 ≤ i ≤ l
Let us note the effect of multiplying g by elementary matrices. We write g ∈
O+ (2l, k)
A B
as g =
,
C D
where A, B,C & D are l × l matrices.
TABLE 2. The row-column operations for O+ (2l, k)
Row operations
ER1
ith 7→ ith + t jth row and
− jth 7→ − jth − t(−i)th row
ER2
ith 7→ ith + t(− j)th row and
jth 7→ jth − t(−i)th row
ER3
−ith 7→ −ith − t jth row and
− jth 7→ − jth + tith row
wi Interchange ith and (−i)th row
Column operations
EC1
jth 7→ jth + tith column and
−ith 7→ −ith − t(− j)th column
EC2
−ith 7→ −ith − t jth column and
− jth 7→ − jth + tith column
th
EC3
j 7→ jth + t(−i)th column and
ith 7→ ith − t(− j)th column
Interchange ith and (−i)th column
5.2.2. Elementary matrices (Chevalley generators) for symplectic groups. For t ∈ k, the elementary matrices are defined as follows:
TABLE 3. Elementary matrices for Sp(2l, k)
char(k)
both
xi, j (t)
xi,− j (t)
x−i, j (t)
xi,−i (t)
x−i,i (t)
Elementary matrices
I + t(ei, j − e− j,−i ),
i 6= j
I + t(ei,− j − e j,−i ),
i< j
I + t(e−i, j − e− j,i ),
i< j
I + tei,−i ,
1≤i≤l
I + te−i,i ,
1≤i≤l
6
A B
Let us note the effect of multiplying g by elementary matrices. We write g ∈ Sp(2l, k) as g =
,
C D
where A, B,C & D are l × l matrices.
TABLE 4. The row-column operations for symplectic groups
Row operations
Column operations
th
th
th
ER1
7→ i + t j row and
EC1
j 7→ jth + tith column and
th
th
th
− j 7→ − j + t(−i) row
−ith 7→ −ith + t(− j)th column
ER2
ith 7→ ith + t(− j)th row and
EC2
−ith 7→ −ith + t jth column and
jth 7→ jth + t(−i)th row
− jth 7→ − jth + tith column
ER3
−ith 7→ −ith + t jth row and
EC3
jth 7→ jth + t(−i)th column and
th
th
th
− j 7→ − j + ti row
ith 7→ ith + t(− j)th column
ER1a
ith 7→ ith + t(−i)th row
EC1a
−ith 7→ −ith + tith column
ER2a
−ith 7→ −ith + tith row
EC2a
ith 7→ ith + t(−i)th column
wi
Interchange ith and (−i)th rows
Interchange ith and (−i)th columns
th
with a sign change in the i row.
with a sign change in the ith column.
ith
5.2.3. Gaussian elimination for Sp(2l, k) and O+ (2l, k).
Step 1: Use ER1 and EC1 to make A into a diagonal matrix. This makes A into a diagonal matrix and
changes other matrices A, B, C and D. For sake of notational convenience we keep calling these
changed matrix as A, B, C and D as well.
Step 2: There are two possibilities. One, the diagonal matrix A is of full rank and two, the diagonal
matrix A is of rank r less than l. This is clearly identifiable by looking for zeros in the diagonal of
A.
Step 3: Make r rows of C, corresponding to the non-zero entries in the diagonal of A zero by using
ER3. If r = l, we have C a zero matrix. If not let us assume that ith row is zero in A. Then we
interchange the ith row with the −ith row in g. We do this for all zero rows in A. The new C is a zero
matrix. We claim that the new A must have a full rank. This follows from Equation 5.1; in particular
TCB + TAD = I . If C is zero matrix then A is invertible. Now make A a diagonal matrix by using
l
1 0 ··· 0
0 1 · · · 0
Step 1. Then one can make A a matrix of the form
... ... . . . ... where λ is a non-zero scalar
0 0 ··· λ
using ER1 [1, Proposition 6.2]. Once A is diagonal and C a zero matrix the equation TCB + TAD = Il
makes D a diagonal matrix of full rank.
Step 4: Use ER2 to make B a zero matrix. The matrix g becomes a diagonal matrix of the form
1 0 ··· 0
0 1 · · · 0
A 0
. . .
where
A
is
of
the
form
.. .. . . ... .
0 A−1
0 0 ··· λ
Step 5: (only for symplectic groups) Reduce the λ to 1 using Lemma 5.1.
Lemma 5.1. For Sp(2l, k), the element diag(1, . . . , 1, λ , 1, . . . , 1, λ −1 ) is a product of elementary matrices.
Proof. Observe that (I + tel,−l )(I − t −1 e−l,l )(I + tel,−l ) = I − el,l − e−l,−l + tel,−l − t −1 e−l,l and denote it
by wl (λ ) and then the diagonal element is wl (λ )wl (−1).
•
Remark 5.2. As we saw in the above algorithm, we will have to interchange ith and −ith rows for i =
1, 2, . . . , l. This can be done by premultiplying with a suitable matrix.
7
Let I be the 2l × 2l identity matrix over k. To swap ith and −ith row in O+ (2l, k), swap ith and −ith rows
in the matrix I. We will call this matrix wi . It is easy to see that this matrix wi is in O+ (2l, k) and is of
determinant −1. Premultiplying with wi does the row interchange we are looking for.
th
In the case of symplectic group Sp(2l, k), we again swap two rows i and −ith rows in I. However we
do a sign change in the ith row and call it wi . Simple computation with our chosen β shows that the above
matrices are in O+ (2l, k) and Sp(2l, k) respectively.
However there is one difference between orthogonal and symplectic groups. In symplectic group, wi
can be generated by elementary matrices because wi = xi,−i (1)x−i,i (−1)xi,−i (1). In the case of orthogonal
groups that is not the case. This is clear, the elementary matrices come from the Chevalley generators and
those generates Ω the commutator of the orthogonal group. All matrices in Ω have determinant 1. However
wi has determinant −1. So we must add wi as a elementary matrix for O+ (2l, k).
Remark 5.3. This algorithm proves, every element in the symplectic group is of determinant 1. Note, the
elementary matrices for the symplectic group is of determinant 1, and we have an algorithm to write any
element as product of elementary matrices. So this proves that the determinant is 1.
Remark 5.4. This algorithm proves, if X is an element of an symplectic group then so is TX. The argument
is similar to above, here we note that the transpose of an elementary matrix in symplectic groups is an
elementary matrix.
5.3. Gaussian elimination for matrices of odd size – the odd-orthogonal group. In this case, matrices
are of odd size and there is only one family of group to consider, it is the odd orthogonal group O(2l + 1, k).
This group will be referred to as the odd-orthogonal group.
5.4. Elementary Matrices (Chevalley generators) for O(2l + 1, k). Following the theory of Lie algebra,
we index rows by 0, 1, . . . , l, −1, . . . , −l. These elementary matrices are listed in Table 5.
TABLE 5. Elementary matrices for O(2l + 1, k)
char(k)
both
odd
even
Elementary matrices
xi, j (t)
I + t(ei, j − e− j,−i ),
xi,− j (t)
I + t(ei,− j − e j,−i ),
x−i, j (t)
I + t(e−i, j − e− j,i ),
xi,0 (t)
I + t(2ei,0 − e0,−i ) − t 2 ei,−i
x0,i (t) I + t(−2e−i,0 + e0,i ) − t 2 e−i,i
xi,0 (t)
I + te0,−i + t 2 ei,−i
x0,i (t)
I + te0,i + t 2 e−i,i
i 6= j
i< j
i< j
1≤i≤l
1≤i≤l
1≤i≤l
1≤i≤l
Elementary matrices for the odd-orthogonal group in even characteristics differs from that of odd characteristics. In above table we made that distinction and listed them separately in different rows according
to the characteristics of k. If char(k) even, we can construct the elements wi , which interchanges the ith row
with −ith row as follows
wi = (I + e0,i + e−i,i )(I + e0,−i + ei,−i )(I + e0,i + e−i,i ) = I + ei,i + e−i,−i + ei,−i + e−i,i .
Otherwise, we can construct wi , which interchanges the ith row with −ith with a sign change in ith , −ith and 0th
row in odd-orthogonal group as follows:
wi = x0,i (−1)xi,0 (1)x0,i (−1) = I − 2e0,0 − ei,i − e−i,−i − ei,−i − e−i,i .
The Gaussian elimination algorithm for O(2l + 1, k) follows the earlier algorithm for symplectic and evenorthogonal group closely, except that
we need to
take care of the zero row and the zero column. We write
α X Y
an element g ∈ O(2l + 1, k) as g = E A B where A, B,C & D are l × l matrices, X and Y are 1 × l
F C D
8
2 0 0
matrices, E and F are l × 1 matrices, α ∈ k and β = 0 0 Il . Then from the condition Tgβ g = β we
0 Il 0
get the following equations.
2TXX + TAC + TCA = 0
2α TX + TAF + TCE = 0
2αY + TED + TFB = 0
2TXY + TAD + TCB = Il
(5.2)
(5.3)
(5.4)
(5.5)
Let us note the effect of multiplying g by elementary matrices.
TABLE 6. The row-column operations for O(2l + 1, k)
ER1
(both)
ER2
(both)
ER3
(both)
ER4
(odd)
ER5
(odd)
ER6
(even)
ER7
(even)
wi
(odd)
Row operations
7→ ith + t jth row and
− jth 7→ − jth − t(−i)th row
ith 7→ ith + t(− j)th row and
jth 7→ jth − t(−i)th row
−ith 7→ −ith − t jth row and
− jth 7→ − jth + tith row
0th 7→ 0th − t(−i)th row and
ith 7→ ith + 2t0th − t 2 (−i)th row
0th 7→ 0th + tith row and
th
(−i) 7→ (−i)th − 2t0th − t 2 ith row
0th 7→ 0th + t(−i)th row and
ith 7→ ith + t 2 (−i)th row
0th 7→ 0th + tith row and
(−i)th 7→ (−i)th + t 2 ith row
Interchange ith and (−i)th rows
with a sign change in ith , −ith and 0th rows
ith
EC1
(both)
EC2
(both)
EC3
(both)
EC4
(odd)
EC5
(odd)
EC6
Column operations
7→ jth + tith column and
−ith 7→ −ith − t(− j)th column
−ith 7→ −ith − t jth column and
− jth 7→ − jth + tith column
jth 7→ jth + t(−i)th column and
ith 7→ ith − t(− j)th column
0th 7→ 0th + 2tith column and
(−i)th 7→ (−i)th − t0th − t 2 ith column
0th 7→ 0th − 2t(−i)th column and
ith 7→ ith + t0th − t 2 (−i)th column
(−i)th 7→ (−i)th + t0th + t 2 ith column
jth
(even)
EC7
ith 7→ ith + t0th + t 2 (−i)th column
(even)
wi
(odd)
wi
Interchange ith and (−i)th column
with a sign change in ith , −ith and 0th columns
wi
(even)
Interchange
ith
and
(−i)th
row
(even)
Interchange ith and (−i)th column
5.5. Gaussian elimination for O(2l + 1, k).
Step 1: Use ER1 and EC1 to make A into a diagonal matrix but in the process it changes other matrices
A, B,C, D, E, F, X, and Y . For the sake of notational convenience, we keep calling these changed
matrices as A, B,C, D, E, F, X, and Y as well.
Step 2: Now there will be two cases depending on the rank r of matrix A. The rank of A can be easily
determined using the number of non-zero diagonal entries. Use ER3 and non-zero diagonal entries
of A to make corresponding r rows of C zero.
(1) If r = l then C becomes zero matrix.
(2) If r < l then interchange all zero rows of A with corresponding rows of C using wi so that the
new C becomes a zero matrix.
Once C becomes zero, note that Relation 5.2 if char(k) is odd or Relation Q(g(v)) = Q(v) if char(k)
is even guarantees that X becomes zero. Relation 5.5 guarantees that A has full rank l which also
9
makes D a diagonal with full rank l. Thus Relation 5.3 shows that F becomes zero as well. Then
1 0 ··· 0
0 1 · · · 0
use Step 1 to reduce A =
... ... . . . ... .
0 0 ··· λ
Step 3: Now if char(k) is even then Relation 5.4 guarantees that E becomes zero as well. If char(k) is
odd then use ER4 to make E a zero matrix.
Step 4: Use ER2 to make B a zero matrix. For char(k) even the relation Q(g(v)) = Q(v) guarantees
that Y is a zero matrix and for char(k) odd Relation 5.4 implies that Y becomes zero.
±1 0 0
Thus the matrix g reduces to 0 A 0 , where A = diag(1, · · · , 1, λ ) and D = diag(1, · · · , 1, λ −1 ).
0 0 D
Proof of Corollary C. Let k be a perfect filed of characteristics 2. Note that we can write the diagonal matrix
diag(1, · · · , 1, λ , 1, · · · , 1, λ −1 ) as a product of elementary matrices as follows:
diag(1, · · · , 1, λ , 1, · · · , 1, λ −1 ) = xl,−l (t)x−l,l (−t −1 )xl,−l (t), where t2 = λ .
•
and hence we can reduce the matrix g to identity.
5.6. Elementary matrices (Chevalley generators) for twisted orthogonal groups O− (2l, k). In this section, we describe row-column operations for twisted Chevalley groups.
Thesegroups are also known as
A0 X Y
the Steinberg groups. An element g ∈ O− (2l, k) is denoted g = E A B , where A, B,C & D are
F C D
(l − 1) × (l − 1) matrices, X and Y are 2 × (l − 1) matrices, E and F are (l − 1) × 2 matrices and A0 is a 2 × 2
matrix. In the Gaussian elimination algorithm that we discuss, we reduce X,Y, E&F to zero and A and D to
diagonal matrices. However, unlike the previous cases we were unable to reduce A0 to an identity matrix.
However, for odd characteristics we were able to reduce A0 to a two-parameter subgroup. Furthermore, our
algorithm provides for the spinor norm as before.
Let us go ahead and talk about the output of the algorithm.
Inthe output we will have a 2 × 2 block (also
1 0
called A0 ) which will satisfy TA0 β0 A0 = β0 , where β0 =
for odd characteristics, as defined earlier.
0 ε
a b
Then A0 is a orthogonal group given by the bilinear form β0 . Now if we write A0 =
, then we get
c d
the following equation:
a2 + c2 ε = 1
ab + cdε = 1
b2 + d 2 ε = ε
Considering the fact that det(A0 ) = ±1, one more equation ac − bd = ±1 and this leads to two cases either
a = d and b = cε or a = −d and b = −cε. Recall
that, since
ε is not a square, d 6= 0. Then if c = 0, then
±1 0
there are four choices for A0 and these are A0 =
.
0 ±1
To summarize, in the output of the algorithm A0 will have either of the six forms
t −sε
t sε
(5.6)
or
where t 2 + s2 ε = 1
s t
s −t
±1 0
(5.7)
or
.
0 ±1
There are now two ways to describe the algorithm, one is to leave A0 as it is in the output of the algorithm
and the other is to include these matrices as generators. For the purpose of uniform exposition we choose
10
the later and included the following two generators
x1 (t, s) = I + (t − 1)e1,1 − (t + 1)e−1,−1 + s(e−1,1 + εe1,−1 ); t 2 + εs2 = 1
x2 = I − 2e−1,−1
in the list of elementary generatorsin Table
7. In the case of even characteristics no such reduction is
t p
possible and we included the matrix
in the list of generators with the condition that the determinant
r s
is 1.
The elementary matrices for O− (2l, k) depends on characteristics of k. We describe them separately in
the following table. Let α be an Arf-invariant, 2 ≤ i, j ≤ l and t ∈ K, ξ ∈ k∗ .
TABLE 7. Elementary matrices for O− (2l, k)
char(k)
both
odd
even
Elementary matrices
xi, j (t)
I + t(ei, j − e− j,−i ),
i 6= j
xi,− j (t)
I + t(ei,− j − e j,−i ),
i< j
x−i, j (t)
I + t(e−i, j − e− j,i ),
i< j
wi
I − ei,i − e−i,−i + ei,−i + e−i,i
2≤i≤l
xi,1 (t)
I + t(e1,i − 2e−i,1 ) − t 2 e−i,i
2≤i≤l
2
x1,i (t)
I + t(−e1,−i + 2ei,1 ) − t ei,−i
2≤i≤l
xi,−1 (t)
I + t(e−1,i − 2εe−i,−1 ) − εt 2 e−i,i
2≤i≤l
2
x−1,i (t)
I + t(−e−1,−i + 2εei,−1 ) − εt ei,−i
2≤i≤l
2
x1 (t, s) I + (t − 1)e1,1 − (t + 1)e−1,−1 + s(e−1,1 + εe1,−1 ), t + εs2 = 1
x2
I − 2e−1,−1
x1,−i (t)
I + te1,−i + tei,−1 + αt 2 ei,−i
2≤i≤l
2
x−1,−i (t)
I + te−1,−i + tei,1 + αt ei,−i
2≤i≤l
xA0
I + (t − 1)e1,1 + (s − 1)e−1,−1 + pe1,−1 + re−1,1 , ts + pr = 1.
Let us note the effect of multiplying g by elementary matrices. Elementary matrices for the twisted
orthogonal group in even characteristics differs from that of odd characteristics so in the following table
(Table 8 & 9) we made that distinction and listed them separately in different row according to the characteristics of k.
TABLE 8. The row operations for O− (2l, k)
ER1(both)
ER2 (both)
ER3(both)
ER4(odd)
ER5(odd)
ER6(odd)
ER7(odd)
ER8 (even)
ER9(even)
wi (both)
Row operations
row and and − jth 7→ − jth − t(−i)th row
th
th
i 7→ i + t(− j)th row and jth 7→ jth − t(−i)th row
−ith 7→ −ith − t jth row and − jth 7→ − jth + tith row
1st 7→ 1st − t(−i)th row and ith 7→ ith + 2t1st − t 2 (−i)th row
1st 7→ 1st + tith row and (−i)th 7→ (−i)th − 2t1st − t 2 ith row
(−1)th 7→ (−1)th − t(−i)th row and ith 7→ ith + 2εt(−1)th − εt 2 (−i)th row
(−1)th 7→ (−1)th + tith row and (−i)th 7→ (−i)th − 2εt(−1)th − εt 2 ith row
1st 7→ 1st + t(−i)th row and ith 7→ ith + t(−1)th + αt 2 (−i)th row
(−1)th 7→ (−1)th + t(−i)th row and ith 7→ ith + t1st + αt 2 (−i)th row
Interchange ith and (−i)th row
ith
7 ith + t jth
→
11
TABLE 9. The column operations for O− (2l, k)
EC1(both)
EC2(both)
EC3(both)
EC4(odd)
EC5(odd)
EC6 (odd)
EC7 (odd)
EC8 (even)
EC9(even)
wi (both)
Column operations
7→
column and −ith 7→ −ith − t(− j)th column
th
th
−i 7→ −i − t jth column and − jth 7→ − jth + tith column
jth 7→ jth + t(−i)th column and ith 7→ ith − t(− j)th column
1st 7→ 1st + 2tith column and (−i)th 7→ (−i)th − t1st − t 2 ith column
1st 7→ 1st − 2t(−i)th column and ith 7→ ith + t1st − t 2 (−i)th column
(−1)th 7→ (−1)th + (2εt)ith column and (−i)th 7→ (−i)th − t(−1)th − εt 2 ith column
(−1)th 7→ (−1)th − 2εt(−i)th column and ith 7→ ith + t(−1)th − εt 2 (−i)th column
(−1)th 7→ (−1)th + tith column and (−i)th 7→ (−i)th + t1st + αt 2 ith column
1st 7→ 1st + tith column and (−i)th 7→ (−i)th + t(−1)th + αt 2 ith column
Interchange ith and (−i)th column
jth
jth + tith
Note that any isometry g satisfying the quadratic form Q also satisfy Tgβ g = β . The main reason the
following algorithm works is the closed condition Tgβ g = β which gives the following relations:
T
(5.8)
A0 β0 A0 + TFE + TEF = β0 ,
(5.9)
T
(5.10)
T
A0 β0 X + TFA + TEC = 0,
A0 β0Y + TFB + TED = 0,
(5.11)
T
(5.12)
T
Xβ0 X + TCA + TAC = 0,
Xβ0Y + TCB + TAD = Il−1 .
5.7. The Gaussian elimination algorithm for O− (2l, k).
Step 1: Use ER1 and EC1 to make A into a diagonal matrix but in the process it changes other matrices
A0 , A, B,C, D, E, F, X, and Y . For sake of notational convenience, we keep calling these changed
matrices as A0 , A, B,C, D, E, F, X, and Y as well.
Step 2: Now there will be two cases depending on the rank r of the matrix A. The rank of A can be
easily determined by the number of non-zero diagonal entries.
Step 3: Use ER3 and non-zero diagonal entries of A to make corresponding r rows of C zero.
If r = l − 1 then C becomes zero matrix.
If r < l − 1 then interchange all zero rows of A with corresponding rows of C using wi so that
the new C becomes a zero matrix.
Once C becomes zero one can note that the relation TXβ0 X + TCA + TAC = 0 if char(k) is odd or the
relation Q(g(v)) = Q(v) and the fact that αt 2 + t + α is irreducible when char(k) is even guarantees
that X becomes zero (see Lemma 5.7). Then the relation TA0 β0 X + TFA + TEC = 0 shows that F
becomes zero as well and the relation TXβ0Y + TCB + TAD = Il−1 guarantees that A has full rank
l − 1 which also makes D a diagonal with full rank l − 1. Now we diagonalize A again to the form
1 0 ··· 0
0 1 · · · 0
. . .
.. .. . . ... as in Step 1.
0 0 ··· λ
Step 4: Use EC4 and EC6 when char(k) is odd or use EC8 and EC9 when char(k) is even to make E
zero.
Note that the relation TA0 β0 A0 + TFE + TEF = β0 shows that A0 is invertible. Thus the relation
T A β Y + T FB + T ED = 0 guarantees that Y becomes zero.
0 0
12
A0 0 0
Step 5: Use ER2 to make B a zero matrix. Thus the matrix g reduces to: g = 0 A 0 , where
0 0 D
−1
A = diag(1, 1, · · · , 1, λ ) and D = diag(1, 1, · · · , 1, λ ). Now if char(k) is odd then go to step 6
otherwise go to step 7.
t −εs
T
Step 6: Using the relation A0 β0 A0 = β0 it is easy to check that A0 has the form
or
s
t
t εs
. If the determinant of A0 is −1, multiply g by x2 to get new g of the above form such that
s −t
I2 0 0
A0 has determinant 1. Now using the elementary matrix x1 (t, s) we can reduce g to 0 A 0 ,
0 0 D
where A and D are as above.
I2 0 0
Step 7: Using elementary matrix xA0 we can reduce g to 0 A 0 , where A = diag(1, 1, · · · , 1, λ )
0 0 D
−1
and D = diag(1, 1, · · · , 1, λ ).
A0 X Y
Lemma 5.5. Let k be a field of characteristics 2 and let g = E A B , where A = diag(1, 1, · · · , 1, λ ),
F 0 D
−
be an element of O (2l, k) then X = 0.
Proof. Let {e1 , e−1 , e2 , · · · , el , e−2 , · · · , e−l } be the standard basis of the vector space V . Recall that for a
column vector x = (x1 , x−1 , x2 , · · · , xl , x−2 , · · · , x−l )t the action of the quadratic form Q is given by Q(x) =
2 )+x x
2
α(x12 + x−1
over k[t]. By definition, for any g ∈
1 −1 + . . . + xl x−l , where αt + t + α is irreducible
x
·
·
·
x
11
1(l−1)
−
O (2l, k) we have Q(g(x)) = Q(x) for all x ∈ V . Let X =
be a 2 × (l − 1) matrix. Comx21 · · · x2(l−1)
2 + x2 ) + x x = 0. If x = 0 then we
puting Q(g(ei )) = Q(ei ) for all 2 ≤ i ≤ l, we can see that α(x1i
1i 2i
2i
2i
can see that x1i = 0. Suppose x2i 6= 0 for some i then we rewrite the equation by dividing it by x2i as
α( xx1i2i )2 + xx2i1i + α = 0, which is a contradiction to the fact that αt 2 + t + α is irreducible over k[t]. Thus,
x2i = 0 for all 2 ≤ i ≤ l and hence X = 0.
•
5.8. Time-complexity of the above algorithm. We establish that the worst case time-complexity of the
above algorithm is O(l 3 ). We mostly count the number of field multiplications.
Step 1 We make A a diagonal matrix by row-column operations. That has complexity O(l 3 ).
Step 2 In making both C and B zero-matrix we multiply two rows by a field element and additions. In the
worst case, it has to be done O(l) times and done O(l 2 ) many times. So the complexity is O(l 3 ).
Step 3 In odd-orthogonal group and twisted orthogonal group we clear X,Y, E, F, this clearly has complexity O(l 2 )
Step 4 has only a few steps that is independent of l.
Then clearly, the time-complexity of our algorithm is O(l 3 ).
6. C OMPUTING SPINOR NORM FOR ORTHOGONAL GROUPS
In this section, we show how we can use our Gaussian elimination algorithm to compute the spinor norm
for orthogonal groups. Throughout this section we assume that the field k is of odd characteristics. The
classical way to define spinor norm is via Clifford algebras [10, Chapters 8 & 9]. Spinor norm is a group
2
homomorphism Θ : O(d, k) → k× /k× , restriction of which to SO(d, k) gives Ω(d, k) as kernel. However,
in practice, it is difficult to use that definition to compute the spinor norm. Wall [18], Zassenhaus [19] and
Hahn [11] developed a theory to compute the spinor norm. For our exposition, we follow [17, Chapter 11].
13
Let g be an element of the orthogonal group. Let g̃ = I − g and Vg = g̃(V ) and V g = ker(g̃). Using β we
define Wall’s bilinear form [ , ]g on Vg as follows:
[u, v]g = β (u, y), where, v = g̃(y).
This bilinear form satisfies following properties:
(1) [u, v]g + [v, u]g = β (u, v) and [u, u]g = Q(u) for all u, v ∈ Vg .
(2) g is an isometry on Vg with respect to [ , ]g .
(3) [v, u]g = −[u, gv] for all u, v ∈ Vg .
(4) [ , ]g is non-degenerate.
Then the spinor norm is
Θ(g) = disc(Vg , [ , ]g ) if g 6= I
extended to I by defining Θ(I) = 1. An element g is called regular if Vg is non-degenerate subspace of V
with respect to the form β . Hahn [11, Proposition 2.1] proved that for a regular element g the spinor norm
is Θ(g) = det(g̃|Vg )disc(Vg ). This gives,
Proposition 6.1.
(1) For a reflection ρv , Θ(ρv ) = Q(v).
(2) Θ(−1) = disc(V, β ).
(3) For a unipotent element g the spinor norm is trivial, i.e., Θ(g) = 1.
Murray and Roney-Dougal [15] used the formula of Hahn to compute spinor norm. However, we show
(Corollary B) that the Gaussian elimination algorithm we developed in Section 5 ouptuts the spinor norm.
We compute the spinor norm for elements in twisted orthogonal group separately. First we observe the
following:
Lemma 6.2. For the group O+ (d, k), d ≥ 4,
(1) Θ(xi, j (t)) = Θ(x−i, j (t)) = Θ(xi,− j (t)) = 1. Furthermore, in odd case we also have Θ(xi,0 (t)) = 1 =
Θ(x0,i (t)).
(2) Θ(wl ) = 1.
(3) Θ(diag(1, . . . , 1, λ , 1, . . . , 1, λ −1 )) = λ .
Proof. We use Proposition 6.1. The first claim follows from the fact that all elementary matrices are unipotent. The element wl = ρ(el +e−l ) is a reflection thus Θ(wl ) = Q(el + e−l ) = 1.
For the third part we note that diag(1, . . . , 1, λ , 1, . . . , 1, λ −1 ) = ρ(el +e−l ) ρ(el +λ e−l ) and hence the spinor
•
norm Θ(diag(1, . . . , 1, λ , 1, . . . , 1, λ −1 )) = Θ(ρ(el +λ e−l ) ) = Q(el + λ e−l ) = λ .
Lemma 6.3. The spinor norm of elementary matrices in O− (d, K) are:
(1) Θ(xi, j (t)) = Θ(x−i, j (t)) = Θ(xi,− j (t)) = 1̄.
(2) Θ(wi ) = 1̄.
(3) Θ(diag(1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = λ̄ ,
Θ(diag(1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = ελ ,
Θ(diag(−1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = λ̄ ,
Θ(diag(−1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = ελ .
(4) Θ(x1 (t, s)) = 2(1 − t) whenever t 6= 1.
(5) Θ(x2 ) = ε
Proof: The first one follows from previous proposition as all elementary matrices are unipotent. The
element wi = ρei +e−i is a reflection thus Θ(wi ) = Q(ei + e−i ) = 1̄.
For the third part we note that diag(1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 ) = ρel +e−l ρel +λ e−l and hence,
Θ(diag(1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = Θ(ρel +e−l )Θ(ρel +λ e−l )
Θ(diag(1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = Q(el + λ e−l )
Θ(diag(1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = λ̄ .
14
Observe that
diag(1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 ) = diag(1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )ρe−1
diag(−1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 ) = diag(1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )ρe1
implies that
Θ(diag(1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = ελ ,
Θ(diag(−1, 1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = λ .
Similarly we can see that
diag(−1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = diag(1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )ρe1
implies that Θ(diag(−1, −1, 1, ..., 1, λ , 1, ..., 1, λ −1 )) = ελ .
For the fourth part we observe that x1 (t, s) = ρ(t−1)e1 +se−1 and hence
Θ(x1 (t, s)) = Q(ρ(t−1)e1 +se−1 )
= (t − 1)2 + εs2
= 2(1 − t)
as t 2 + εs2 = 1.
Note that x2 = ρe−1 and Θ(ρe−1 ) = ε̄ implies that Θ(x2 ) = ε.
Proof of Corollary B. Let g ∈ O+ (d, k). From Theorem A, we write g as a product of elementary matrices
and a diagonal matrix diag(1, . . . , 1, λ , 1, . . . , 1, λ −1 ) and hence we can find the spinor norm of g.
•
6.1. Computing the spinor norm inO− (d, k). Before moving to Step 6 of the algorithm, look
0 . If
at A
t −sε
1 0
det(A0 ) = 1, it is of the form
where s 6= 0. Furthermore, if s = 0 then it is of the form
or
s
t
0
1
−1 0
the spinor norm then can be computed from the lemma above. If the determinant is −1, then A0
0 −1
is x2 times one of the previous matrices. So the spinor norm can be computed because it is multiplicative.
7. D OUBLE COSET DECOMPOSITION FOR S IEGEL MAXIMAL PARABOLIC
In this section, we compute the double coset decomposition with respect to Siegel maximal parabolic
subgroup using our algorithm. Let P be the Siegel maximal parabolic of G where G is either split orthogonal
group O(d, k) or Sp(2l, k), where char(k) is odd. In Lie theory, a parabolic is obtained by fixing a subset of
simple roots [4, Section 8.3]. Siegel maximal parabolic corresponds to the subset consisting of all but the
last simple root. Geometrically, a parabolic subgroup is obtained as fixed subgroup of a totally isotropic
flag [14, Proposition 12.13]. The Siegel maximal parabolic is the fixed subgroup of following isotropic flag
(with the basis in Section 4):
{0} ⊂ {e1 , . . . , el } ⊂ V.
α 0 Y
A
B
Thus P is of the form E A B in O(2l + 1, k) and
in Sp(2l, k) and O(2l, k).
0 D
F 0 D
F
The problem is to get the double coset decomposition P\G/P. That is, we want to write G =
PωP as
b
ω∈W
b is a finite subset of G. Equivalently, given g ∈ G we need an algorithm to determine
disjoint union where W
b
the unique ω ∈ W such that g ∈ PωP. If G is connected with Weyl group W and suppose WP is the Weyl
group corresponding to P then [5, Proposition 2.8.1]
P\G/P ←→ WP \W /WP .
b as follows:
We need a slight variation of this as the orthogonal group is not connected. We define W
b = {ω0 = I, ωi = w1 · · · wi | 1 ≤ i ≤ l}
W
15
where wi were defined earlier for each class of groups.
Theorem 7.1. Let P be the Siegel maximal parabolic subgroup in G, where G is either O(d, k) or Sp(d, k).
b the set
Let g ∈ G. Then there is an efficient algorithm to determine ω such that g ∈ PωP. Furthermore, W
of all ω is a finite set of l + 1 elements where d = 2l or 2l + 1.
Proof. In this proof we proceed with a similar but slightly different Gaussian
elimination
algorithm. Recall
α
X
Y
A B whenever g belongs to Sp(2l, k) or O(2l, k) or g = E A B whenever g belongs to
that g = C
D
F C D
O(2l + 1, k). In our algorithm, we made A into a diagonal matrix. Instead of that, we can use elementary
matrices ER1 and EC1 to make C into a diagonal matrix and then do the row interchange to make A into a
diagonal matrix and C a zero matrix. If we do that, we note that elementary matrices E1 and E2 are in P.
The proof is just keeping track of elements of P in this Gaussian elimination
algorithm. The step 1 in the
A
B
algorithm says that there are elements p1 , p2 ∈ P such that p1 gp2 = C1 D1 where C1 is a diagonal matrix
1
1
with m non-zero entries. Clearly m = 0 if and only if g ∈ P. In that case g is in the double coset Pω0 P = P.
Now suppose m ≥ 1. Then inStep 2 we
multiply by E2 to make the first m rows of A1 zero, i.e., there is a
Ã
B̃
1 where first m rows of Ã1 are zero. After this we interchange rows i
p3 ∈ P such that p3 p1 gp2 = 1
C1 D1
B2 ∈ P.
with −i for 1 ≤ i ≤ m which makes C1 zero, i.e., multiplying by ωm we get ωm p3 p1 gp2 = A02 D
2
Thus g ∈ Pωm P.
For O(2l + 1, k) we note that the elementary matrices E1, E2 and E4a are in P. Rest of the proof is similar
to the earlier case and follows by carefully keeping track of elementary matrices used in our algorithm in
Section 5.
•
8. C ONCLUSIONS
Now some implementation results, we implemented our algorithm in magma [2]. We found our implementation to be fast and stable. In magma, Costi and C. Schneider installed a function ClassicalRewriteNatural. It is the row-column operation developed by Costi [8] in natural representation. We tested the time
taken by our algorithm and the one taken by the Magma function. To do this test, we followed Costi [8, Table 6.1] as closely as possible. Two kind of simulations were done. In one case, we fixed the size of the
field at 710 and varied the size of the matrix from 20 to 60. To time both these algorithms for any particular
input, we took one thousand random samples from the group and run the algorithm for each one of them.
Then the final time was the average of this one thousand random repetitions. The times were tabulated and
presented below.
In the other case, we kept the size of the matrix fixed at 20 and we varied the size of the field, keeping the
characteristic fixed at 7. In many cases the magma computation for the function ClassicalRewriteNatural
will not stop in a reasonable amount of time or will give an error and not finish computing. In those cases,
though our algorithm worked perfectly, we were unable to get adequate data to plot and are represented by
gaps in the graph drawn. Here also the times are the average of one thousand random repetitions. It seems
F IGURE 1. Some simulations comparing our algorithm with the one inbuilt in Magma for
even-order orthogonal groups
that our algorithms perform better than that of Costi’s on all fronts.
16
F IGURE 2. Some simulations comparing our algorithm with the one inbuilt in Magma for
symplectic groups
R EFERENCES
1. J.L. Alperin and Rowen B. Bell, Groups and representations, Springer, 1995.
2. Wieb Bosma, John Cannon, and Catherine Playoust, The Magma algebra system. I. The user language, J. Symbolic Comput.
24 (1997), no. 3-4, 235–265, Computational algebra and number theory (London, 1993).
3. Peter Brooksbank, Constructive recognition of classical groups in their natural representation, Journal of Symbolic Computation 35 (2003), 195–239.
4. Roger Carter, Simple groups of Lie type, Pure and Applied Mathematics, vol. 28, John Wiley & Sons, 1972.
, Finite groups of Lie type, John Wiley & Sons, 1993.
5.
6. C. Chevalley, Sur certains groupes simples, Tohoku Math. J. 7 (1955), no. 2, 14–66.
7. Arjeh M. Cohen, Scott H. Murray, and D. E. Taylor, Computing in groups of Lie type, Mathematics of computation 73
(2003), no. 247, 1477–1498.
8. Elliot Costi, Constructive membership testing in classical groups, Ph.D. thesis, Queen Mary, Univ. of London, 2009.
9. Joseph F. Grcar, Mathematicians of Gaussian elimination, Notices of the AMS 58 (2011), no. 6, 782–792.
10. Larry C. Grove, Classical groups and geometric algebra, vol. 39, American Mathematical Society, Graduate Studies in
Mathematics, 2002.
11. Alexander J. Hahn, Unipotent elements and the spinor norms of Wall and Zassenhaus, Arch. Math. (Basel) 32 (1979), no. 2,
114–122.
12. Martin Hildebrand, Generating random elements in SLn (Fq ) by random transvections, Journal of algebraic combinatorics 1
(1992), 133–150.
13. Ayan Mahalanobis and Anupam Singh, Gaussian elimination in unitary groups with an application to cryptography, Journal
of Algebra Combinatorics Discrete Structures and Applications 4 (2017), no. 3, 247–260.
14. Gunter Malle and Donna Testerman, Linear algebraic groups and finite groups of lie type, Cambridge University Press,
2011.
15. Scott H. Murray and Colva M. Roney-Dougal, Constructive homomorphisms for classical groups, Journal of Symbolic
Computation 46 (2011), 371–384.
16. Robert Steinberg, Lectures on Chevalley groups. notes prepared by John Faulkner and Robert Wilson, Yale University, 1968.
17. Donald E. Taylor, The geometry of the classical groups, Heldermann Verlag, 1992.
18. G. E. Wall, The structure of a unitary factor group, Inst. Hautes Études Sci. Publ. Math (1959), no. 1, 23 pp.
19. Hans Zassenhaus, On the spinor norm, Arch. Math. (1962), no. 13, 434–451.
17
| 4math.GR
|
arXiv:1606.03256v3 [math.CO] 23 Feb 2017
Combinatorial Results Implied by Many Zero Divisors
in a Group Ring
Fedor Petrov
February 24, 2017
Abstract
It has been recently proved [2, 3] that for a group G = Gn0 , where G0 6= {1, −1}m is
a fixed finite Abelian group and n is large, any subset A without 3-terms progressions
(triples x, y, z of different elements with xy = z 2 ) contains at most |G|1−c elements,
where c > 0 is a constant depending only on G0 . This is known to be false when
G is, say, large cyclic group, see current records in [7]. The aim of this note is to
show that algebraic property which corresponds to this difference is the following:
in the first case a group algebra F[G] contains a subspace X with low codimension
such that X 3 = 0. We discuss which bounds are obtained for finite Abelian p-groups
and matrix p-groups: Heisenberg group over Fp and unitriangular group over Fp . All
bounds are of the form |G|1−c with some c > 0. Also we show how the method works
for further generalizations by Kleinberg–Sawin–Speyer and Ellenberg.
Let CN denote a cyclic group of order N in multiplicative notation. Denote κN =
minx>0 x−(N −1)/3 (1 + x + · · · + xN −1 ). The number of points in {0, 1, . . . , N − 1}n with sum
of coordinates at most n(N − 1)/3 grows as (cN + o(1))n for large n.
In a recent paper [2] Ernie Croot, Vsevolod Lev, and Peter Pach proved by a clever
combination of polynomial method (in spirit of Alon’s Combinatorial Nullstellensatz [1]),
linear algebraic dimension reasoning and law of large numbers that any subset A of a
group G = C4n containing more than κn4 elements, has three distinct elements a, b, c for
which ab = c2 (they used another formula for the same constant κ4 = 3.61 . . . , but I
prefer this unified formula, which appears also below in Theorem 4). It have been soon
observed that their method, being slightly modified (and even simplified! They started
from combinatorially harder, “ramified” problem), works for other groups, like Cpn for
prime p [3], where we may get the same type bound κnp for 3-progression-free sets. It
also works for other combinatorial problems of the same spirit, like Sunflower problems of
Erdös–Szemeredi and Erdös – Rado [4]. Robert Kleinberg, Will Sawin and David Speyer
[9] observed that [3] actually contains the bound for tri-colored sum-free sets (see the
definition below). Important feature of this generalization is that the constant κp in this
question is proved to be sharp (the proof was finished independently by Sergey Norin and
1
Luce Pebody [10, 11]). Next, Jordan Ellenberg proposed further generalization proving
that the sumset of any two sets may be covered by sumsets of small subsets.
It is natural that polynomial method works well for the groups which are contained
either in additive or multiplicative group of a field. For more general groups sometimes it
may be successfully replaced by considering generating functions in group rings, historically
this approach appeared even earlier than Combinatorial Nullstellensatz: John E. Olson [6]
used it for computing Davenport constant of finite Abelian p-groups.
1
Example: Cauchy–Davenport Theorem
For illustrating a parallelism between Combinatorial Nullstellensatz method and group
rings method, we give two proofs of the Cauchy–Davenport theorem, which is convenient
to formulate as follows:
If A, B, C are non-empty subsets of a cyclic group G = Cp of prime order p and |A| +
|B| + |C| = p + 2, then ABC := {abc|a ∈ A, b ∈ B, c ∈ C} = G.
1) Polynomial proof (see [1], where this proof is expressed bit differently). We identify
Cp with an additive group of Fp . Assume that some element h does not belong to A+B +C.
Then a polynomial ϕh (x, y, z) = (x + y + z − h)p−1 − 1 takes zero values on A × B × C. Let
u, v, w be Fp -valued functions on A, B, C to be specified later. Consider the following sum
X
u(x)v(y)h(z)ϕh (x, y, z) = 0.
x∈A,y∈B,z∈C
Now change order of summation: sum up by monomials of ϕh (x, y, z). For each monomial
xα y β z γ we have
X
X
X
X
u(x)v(y)h(z)xα y β z γ =
u(x)xα ·
v(y)y β ·
w(z)z γ .
(1)
x∈A,y∈B,z∈C
x∈A
y∈B
z∈C
Now we are ready to say what we require from functions u, v, w. Assume that
(
X
0, if α 6 |A| − 2
u(x)xα =
1, if α = |A| − 1,
x∈A
analogous conditions for β and γ. Such functions do exist: linear systems to their values
have Vandermonde matrices, thus non-degenerated. Then for any monomial xα y β z γ in ϕh
either α = |A| − 1, β = |B| − 1, γ = |C| − 1 or there is zero multiple in right hand side
(p−1)!
of (1). But ϕh has this monomial with non-zero coefficient (|A|−1)!(|B|−1)!(|C|−1)!
. Thus total
sum is not equal to 0, a contradiction.
2) Group ring proof. Now we use a multiplicative notation for G = Cp , fix a generator
g0 in Cp and denote τ = g0 − 1, we have τ p = 0. The ring Fp [G] is filtrated by the powers
of its augmentation ideal Aug (G) = τ · Fp [G].
uct
Let u, v, w be Fp -valued functions on A, B, C respectively. Consider the following prod!
!
!
X
X
X
u(a)a
v(b)b
w(c)c
a∈A
b∈B
c∈C
in Fp [G]. Our goal is to find functions u, v, w so that this product has all non-zero coefficients, this clearly implies ABC = G.
Assume that the first multiple, being expressed in powers of τ , has coefficient 1 of τ |A|−1
and coefficients 0 of τ i for i = 0, . . . , |A| − 2. Analogous conditions are imposed on two
other multiples. Then our product equals
τ p−1 =
p−1
X
(−1)i g0p−1−i ,
i=0
all coefficients are non-zero as desired. It remains to understand why we may choose
function, say, u satisfying these conditions. Some conceptual argument should exist, but in
any case we may denote A = {(τ + 1)α1 , . . . , (τ + 1)α|A| }, where 0 6 α1 < · · · < α|A| 6 p −1,
then finding appropriate
coefficients u(a) is solving a linear system. A matrix of this system
αi
has entries j , 1 6 i 6 |A|, 0 6 j 6 |A| − 1. It is non-degenerated as a generalized
Vandermonde type matrix.
These two proofs look like the same thing said on different languages. But if we change
the problem a bit, say, consider restricted product sets {abc|a ∈ A, b ∈ B, c ∈ C, a 6=
b} (Erdös–Heilbronn problem), then the first proof generalizes easily (we should replace
polynomial to (x − y)((x + y + z − h)p−1 − 1)). And what about the second? On the other
hand, what are “polynomial” analogues of group rings proofs for groups not embeddable
to fields? In some important cases the modular properties of binomial polynomials nx do
work (see [13]), but what in general? This parallelism is still unclear for me.
Now we explain a group ring version of Croot–Lev–Pach ideas. For making the exposition self-contained and comparing two technologies easier for a reader we also include the
polynomial proof for the bounds in the groups Fnp .
2
Linear algebraic lemma and polynomial approach
We start with a simple linear algebraic
Lemma 1. Let F be a field, A be a finite set of size |A| = d. Denote by FA the ddimensional space of functions f : A → F. Let X1 , . . . .Xk be linear subspaces of FA .
Denote by ti = codim Xi = d − dim Xi their codimensions. Consider the space Y ⊂ FA
spanned by functions of the form f1 . . . fk , fi ∈ Xi . Then
1) codim Y 6 t1 + . . . tk ;
2) if X1 = · · · = Xk , then codim Y 6 t1 ;
3) in general,P
if Z1 , . . . , Zr are the all pairwise distinct subspaces in the family {X1 , . . . , Xk },
then codim Y 6 ri=1 codim Zi ;
P
4) further, in notations of p.3), if a∈A f (a) = 0 for all f ∈ Y , then we have
(P
codim Zi , if r > 2
d6
2t1 ,
if r = 1.
Proof. (Idea of this proof is suggested by the answer given by Ilya Bogdanov on a similar
question on Mathoverflow [5]).
By Gauss elimination process we may find a basis {v1 , . . . , vd−t1 } of the space X1 and
elements p1 , . . . , pd−t1 in A so that vi (pj ) = δi,j is Kronecker delta. The claim 2) already
k
follows: functions v1k , . . . , vd−t
are linearly independent.
1 +1
Take f1 = v1 + · · · + vd−t1 . For a linear operator L: v → f1 v on FA we have
dim L(X2 ) = dim X2 − dim Ker L|X2 > dim X2 − dim Ker L > d − t2 − t1 ,
this proves claim 1) for k = 2. General k case follows by straightforward induction.
Claim 3) directly follows from the claims 1) and 2).
Now we prove Claim 4). Without loss of generality X1 = . . . Xℓ and Xj 6= X1 for j > ℓ.
Denote ℓ̃ = min(k − 1, ℓ). Then by Claim 2) the functions of the form f1 . . . fℓ̃ span a
subspace of codimension at most t1 and by
PClaim 3) the functions of the form fℓ̃+1 . . . fk
span a subspace of codimension at most
codim Zi − δ · t1 , where δ = 1 if r < k and
A
δ = 0 if r = k. These two
subspaces
of
F
are
mutually orthogonal with respect to a usual
P
bilinear form hf, gi = a∈A f (a)g(a), thus sum of their codimensions is at least d.
Before formulating the group rings approach we present the polynomial argument which
allows to get exponential bounds for progression-free subsets in the additive group of Fnp
for odd prime power p. It is essentially the same as Ellenberg–Gijswijt proof [3] and is
based on the ideas of Croot–Lev–Pach [2].
n
n
Let TP
ti ∈ Lnp \ Ti
1 , . . . , Tk be subsets of Lp := {0, 1, . . . , p − 1} such that whenever
P
we have
ti ∈
/ Lnp , i.e., there exists a coordinate j P
∈ {1, . . . , n} such that i tij > p. A
natural example is provided by Ti = {(x1 , . . . , xn ) : xi 6 (p − 1)n/k}.
Now assume that A ⊂ Fnp is a subset such that if n1P
g1 + · · · + nk gk = 0 for gi ∈ A, then
g1 = · · · = gk , where ni are integers coprime to p and
ni = 0.
Consider the subspaces X1 , . . . , Xk of FA
(the
space
of functions on A): Xi = {f :
p
P
Q λi
λ
λ
ai for a = (a1 , . . . , an ),
a∈A f (a)a = 0} for all λ ∈ Ti . Here we naturally denote a =
λ = (λ1 , . . . , λn ). Clearly codim Xi 6 |Ti |. Assume that fi ∈ Xi . Let F (x1 , x2 , . . . , xk ) be
any polynomial of kn variables (xi has n coordinates xi1 , . . . , xin ) such that total degree of
F in variables x1j , x2j , . . . , xkj does not exceed p − 1 for all j = 1, . . . , n. Then
X
f1 (a1 ) . . . fk (ak )F (a1 , . . . , ak ) = 0
ai ∈A
Q
whenever fi ∈ Xi . This is clear for any monomial aλi i of F : the sum factorizes and there
exists a zero factor , since λj ∈ Tj for some j: this is exactly what we require from Ti ’s
and F . Apply this to the polynomial
F (x1 , . . . , xk ) =
n
Y
1−(
k
X
!
ni xij )p−1 .
i=1
j=1
We get
0=
X
f1 (a1 ) . . . fk (ak )F (a1 , . . . , ak ) =
ai ∈A
X
f1 (a) . . . fk (a),
a∈A
theP
last identity follows from the assumption that P
A does not contain non-trivial solutions
of
ni ai = 0. Now Lemma implies that |A| 6
ti , and even better estimates hold if
some Ti ’s coincide. In
P particular, for k = 3 and arithmetic progressions taking T1 = T2 =
T3 = {(x1 , . . . , xn ) :
xi 6 (p − 1)n/3} we get |A| 6 2|T1 |.
3
Group rings approach
Now let G be a finite group (not
Pnecessary Abelian), n1 , . . . , nk (where k > 2) be non-zero
integers which sum up to 0 =
ni . Assume that all ni are coprime to |G|. Let a subset
A ⊂ G be so that the equation g1n1 . . . gknk = 1, gi ∈ A, holds only for g1 = · · · = gk .
Arithmetic progressions of length 3 correspond to the case k = 3, n1 = n2 = 1, n3 = −2. It
is probably more natural to call a solution of g1 g2−1g3 g2−1 = 1 “an arithmetic progression”
in non-Abelian setting, but, alas, I do not know how to modify the argument for such
equations.
Assume that for some field F we managed to find subspaces X1 , . . . , Xk of a group
algebra F[G] satisfying X1 . . . Xk = 0, i.e., u1 . . . uk = 0 for ui ∈ Xi . Denote ti = codim Xi .
General fact is the following
Theorem 1. Call two indices i, j ∈ {1, . . . , k} equivalent if Xi = Xj and ni = nj . If I is
a maximal system of mutually non-equivalent indices, then
(P
if |I| ≥ 2
i∈I ti
|A| 6
2t1
if |I| = 1.
P
In particular, we always have |A| 6
ti .
Proof. For g ∈ G, u ∈ F[G], let [g]u denote a coefficient of g in u. Denote Ar = {ar , a ∈ A}
for r = 1, 2, . . . . Then |Ar | = |A| if r and |G| are coprime. Let further Wi denote the
span of Ani in F[G], we have dim Wi = |A|. Finally denote Li = Xi ∩ Wi , we have
dim
Xi + dim Wi − |G| = |A| − ti . Consider functions f1 , . . . , fk on A such that
P Li > dim
ni
f
(a)a
∈ Li . Then
a∈A i
!
!
!
X
X
X
X
fk (a)ank = 0,
f1 (a) . . . fk (a) = [1]
f2 (a)an2 · . . . ·
f1 (a)an1 ·
a∈A
a∈A
a∈A
a∈A
first equality follows from our assumption on A. It remains to use Claim 4 of Lemma.
Remark. If k = 2, then even A = G satisfies the condition of the Theorem. So, what we
actually get in this case is that if X1 X2 = 0 for two subspaces of F[G], then codim X1 +
codim X2 > |G| (that is, the result is about group rings, not about combinatorics). But
already for 3 multiples for some groups there exist subspaces X1 , X2 , X3 of low codimension
(low means o(|G|) or sometimes even O(|G|c), c < 1) such that X1 X2 X3 = 0.
4
Kleinberg–Sawin–Speyer refinement
The set A ⊂ G without arithmetic progressions of length 3 produces a set of triples
{(xa , ya , za ) := (a, a, a−2 ), a ∈ A} ⊂ G3 such that xa yb zc = 1 if and only if a = b =
c. Following [9] we call such a set of ordered triples a “tri-colored product-free sets”.
Analogously we may defined k-colored product-free sets of k-tuples. Note that the above
proof for the sets without arithmetic progressions allows to show the following
Theorem 2. Assume that the subspaces X1 , X2 , . . . , Xk of a group P
algebra F[G] satisfy
X1 . . . Xk = 0. A k-colored product free sets may contain at most
ti k-tuples, where
ti = codim Xi .
We remark that for k = 3 we get the estimate 3codim X (where X 3 = 0, X ⊂ F[G]), for
the tri-colored product-free sets, but 2codim X for the sets without non-trivial solutions of
xy = z 2 (provided that |G| is odd.) This improvement comes from the coincidence of two
spaces of functions in the latter case.
What is remarkable is that the estimate we get on this way for tri-colored sum-free
subsets of Fnp is asymptotically tight (in logarithmic scale), this follows from the results of
[9] and [10, 11]. It would be nice to get the tightness for other groups. Let me formulate
it as a conjecture:
Conjecture. Consider all k-tuples of subspaces X1 , . . . , Xk of F[G] such that X1 . . . Xk =
0. Take a minimal value Mk (G) of the sum of their codimensions. Here the minimum is
taken over the choice of the field also. Then, for large |G|, the maximal cardinality of a
k-colored product-free subset of G is Mk (G) · |G|o(1) .
This may seem to strong: At first, we actually use in the proof only that [1]X1 . . . Xk =
0, not X1 . . . Xk = 0. But I do not know whether it allows to improve the bound. At
second, the minimizing over F looks somehow strange. Possibly, we should start from the
case of p-groups and the group rings over Fp .
5
Ellenberg’s refinement
In [12] the following nice further generalization is proved: for any sets A, B in the additive
group of Fnp there exist subsets A1 ⊂ B, B1 ⊂ B for which A+ B ⊂ (A1 + B) ∪(A+ B1) and
|A1 | + |B1 | 6 M. Here M is the [3] and [9] bound: three times the number of monomials
in n variables of degree at most n(p − 1)/3. For an arbitrary group we get the following:
Theorem 3. Let G be a finite group, and X0 , X1 , . . . , Xk be linear subspaces of the group
algebra F[G] such that X0 X1 · . . . · Xk = 0. Denote ti = codim Xi . Let A1 , . . . , Ak be
arbitrary subsets of G. Then there exist subsets Bi ⊂ Ai , i = 1, . . . , k, and C ⊂ G such
that |C| 6 t0 , |Bi | 6 ti for all i = 1, . . . , k, and
A1 A2 · . . . · Ak ⊂ C ∪ B1 A2 · . . . · Ak ∪ A1 B2 · . . . · Ak ∪ . . . ∪ A1 A2 · . . . · Ak−1 Bk .
Moreover, we may additionally require that Bi = Bj (1 6 i < j 6 k) whenever simultaneously Xi = Xj and Ai = Aj .
Following [12] and [8] (Ellenberg uses the cited result of Roy Meshulam; we do not, but
the concept of lexicographic leaders of multi-linear forms is still important for us.)
Let F be a field, A be a linearly ordered finite set of size |A| = d. Denote by FA
the d-dimensional space of functions f : A → F. For non-zero element z ∈ FA we define
the leader ℓ(z) and the outsider out(z) as the minimal, corr. maximal, a ∈ A such that
z(a) 6= 0.
Lemma 2. Let W ⊂ FA be a linear subspace. Then there are exactly dim X different
leaders of elements of W (and, of course, as many different outsiders).
Proof. Gauss elimination.
Proof of Theorem 3. Consider the linear space W of F-valued functions f on A1 × A2 ×
. . . × Ak such that f (a1 , . . . , ak ) = ϕ(a1 . . . ak ) for a certain function ϕ on G. We introduce
arbitrary linear orderings on all Ai ’s, it produces a lexicographic ordering on A1 × . . . × Ak ,
hence we may define the leaders of non-zero elements of W . What we actually prove is
that all but at most t0 of these leaders may be covered by the sets B1 × A2 × . . . × Ak ,
A1 × B2 × . . . × Bk , . . . , A1 × A2 × . . . × Ak−1 × Bk for certain subsets Bi ⊂ Ai , |Bi | 6 ti .
The claim of Theorem immediately follows.
P
Consider the linear subspace W0 ⊂ W defined by the condition that
ϕ(g)g −1 ∈ X0 .
The codimension of W0 in W does not exceed t0 , thus by Lemma 2 all but at most t0
leaders of the elements of W are the leaders of the elements of W0 . By Lemma 2 for all
i = 1, . . . , k there exist subsets Bi ⊂ Ai , |Bi | 6 ti such that the outsiders of the different
functions from FAi ∩ Xi take all the possible values from Ai \ Bi (here FAi is a subspace
of F[G] spanned by the elements of Ai ⊂ G ⊂ F[G].) We prove that for this choice of the
subsets Bi the sets B1 ×A2 ×. . .×Ak , A1 ×B2 ×. . .×Bk , . . . , A1 ×A2 ×. . .×Ak−1 ×Bk cover
the leaders of all the elements of W0 . Assume
the contrary. Then there exists a non-zero
P
element of W0 defined by a function ϕ,
ϕ(g)g −1 ∈ X0 , such that the leader (c1 , . . . , ck )
of a function ϕ(a1 . . . ak ) on A1 × . . . × Ak satisfies ci ∈
/ Bi for all i = 1, . . . , k. Thus there
exist the elements ηi ∈ FAi ∩ Xi with outsiders out(ηi ) = ci . Look at the constant term of
the product
X
[1]
ϕ(g)g −1 · η1 · η2 · . . . · ηk = 0.
It equals
X
ai ∈Ai
ϕ((a1 . . . ak )−1 )[a1 ]η1 [a2 ]η2 . . . [ak ]ηk ,
and by the lexicographic reasoning the onliest non-zero summand is
ϕ((c1 . . . ck )−1 )[c1 ]η1 [c2 ]η2 . . . [ck ]ηk 6= 0,
a contradiction.
Note that Theorem 1 follows from Theorem 3 in many cases (for example, for 3progressions), but there are cases when the estimate of Theorem 1 is bit better (for example,
if n1 = n2 , n3 = n4 , X1 = X2 , X3 = X4 we get |A| 6 t1 + t2 using Theorem 1 , but 2t1 + t2
using Theorem 3.
6
Abelian p-groups
Q
Let p be a prime. Let G = ni=1 CNi be a finite Abelian p-group with n generators
g1 , . . . , gn , gi generates CNi , each Ni is a power of p.
How to choose zero-product subspaces
codimension in Fp [G]? Group algebra
Q withmlow
i
Fp [G] is generated by the products P
(1 − gi ) , where mi ∈ {0, 1, . . . , Ni − 1}. Let
η1 , . . . , ηk be positive reals such that
ηi = 1 (usually the choice ηi = 1/k is optimal).
Fix also positive parameters λ1 , . . . , λn . Let Xi , i = 1, . . . , k, be a subspace generated by
monomials for which
n
X
mj
− ηi > 0
λj
Nj − 1
j=1
Any product f1 . . . fk for fi ∈ Xi has some 1 − gj in a power strictly greater than Nj − 1,
but (1 − gj )Nj = 0.
In order to estimate the codimension ti of Xi we may use Chernoff bound (which
is essentially tight for large n by Cramer theorem). Take η1 = · · · = ηk = 1/k, then
X1 = · · · = Xk , t1 = · · · = tk := t. If ξj are random variablesP
uniformly distributed in
the set Nj1−1 {0, 1, . . . , Nj − 1}, then t/|G| is the probability that j λj (ξj − k1 ) 6 0. If this
inequality holds, for any x ∈ (0, 1) we have
P
x
λj ξj
> x(λ1 +···+λn )/k .
For N > 2 denote
SN (x) :=
N −1
1 X j/(N −1)
1
x
= Exξ , ξ uniformly distributed on
{0, 1, . . . , N − 1}
N j=0
N −1
Then Chebyshev inequality implies that
t/|G| 6
n
Y
j=1
x−λj /k SNj (xλj ).
Now we fix the value of x and parameters λ1 , . . . , λn so that this rewrites as
t/|G| 6
n
Y
j=1
min x−1/k SNj (x).
x∈(0,1]
In particular, for k = 3, where we denoted above N · minx∈(0,1] x−1/k SN (x) = κN , we get
the following
Q
Theorem 4. If all Ni are powers of the Q
same odd prime p, then the size of a set A ⊂ CNi
without 3-term progressions is at most i κNi .
We have S2 (x) > S3 (x) > S4 (x) > . . . for any positive x 6= 1 (this may be proved
by Karamata majorization inequality or, most likely, somehow differently), limn Sn (x) =
(x − 1)/ log x. Thus the sequence aN (k) = minx∈(0,1] x−1/k SN (x) decreases with N. For
example, if k = 3, it varies from a2 (3) = 0.9449 . . . to a∞ (3) = 0.8414 . . . . It means
that increasing Nj makes our bound for t/|G| better. Say, for C9n we get exponentially
better bounds on |A| than for 3n copies of C3n (onto which we could partition C9n in order
to get some exponential bound): if k = 3, improvement is t/|G| 6 0.872n compared to
t/|G| 6 0.919n .
We may also rephrase the original result by Croot, Lev and Pach for G = C4 n and
solutions of the equation xyz −2 = 1 with mutually different x, y, z in the same spirit (and
get the same bound as they get). The difference with the situation considered above is that
exponent 2 is not coprime with |G|, so A2 is in general situation much less than A, and
we do not count x2 = y 2 , x 6= y, as a non-trivial solution. This is handled by a partition
onto classes modulo G2 and writing such products for a, b, c in the same class, as Croot,
Lev and Pach do. This does not work as easy for, say, C8 , since the kernel and image of
the homomorphism g → g 2 no longer coincide. I suppose that Theorem 4 should hold also
for general 2-groups.
Another proofs for Abelian p-groups are proposed by Will Sawin and Eric Naslund
(they use divisibility of binomial coefficients) and by David Speyer (Witt vectors), see the
explanation in [13].
7
Matrix groups
Heisenberg group G = Hn−1 (Fp ) of order p2n−1 consists of (n + 1) × (n + 1) matrices
(aij )06i,j6n over Fp satisfying conditions aii = 1; aij = 0 unless i = j or i = 0 or j = n:
1 a01 . . . . . . a0n a0n
0 1 0 0 . . . a1n
0 0 1 0 . . . a2n
.
.. .. .. . . . .
..
. . .
.
.
.
0 0 0 ....... 1
It is generated by the elements g1 , . . . , gn−1 , h1 , . . . , hn−1 , s where gi = id+e0,i , hi = id+ei,n ,
s = id + e0,n , where id denotes identity matrix, ei,j denote matrix units. They satisfy
relations gi hj = hi gj unless i = j, gi hi = shi gi , s commutes with all gi and hi .
βn−1 α1
αn−1
We may uniquely write each element in the form sγ hβ1 1 . . . hn−1
g1 . . . gn−1
, 0 6 γ, βi , αi 6
p
p − 1. Denote Fp [G] elements: z = s − 1, yi = hi − 1, xi = gi − 1. Then z = yip = xpi = 0.
βn−1 α1
αn−1
Elements of the form z γ y1β1 . . . yn−1
x1 . . . xn−1
are called reduced. reduced elements for
which 0 6 γ, βi , αi 6 p − 1 form a basis of the group ring. Our relations rewrite as
xi yi = yi xi + zyi xi + z + zxi + zyi . Define a degree of an arbitrary word (sequence) in
the alphabet {z, x1 , . . . , xn−1 , y1 , . . . , yn−1} as twice number of z’s plus number of all other
letters. Key lemma is that any word of degree D > 2n(p − 1) equals 0 being evaluated as
a group ring element. Indeed, replacing xi yi = yi xi + zyi xi + z + zxi + zyi and using other
commutativity relations we may reduce each word to a sum of reduced words of the same
or greater degree. Any such a word equals 0 provided that its degree exceeds 2n(p − 1).
Now define a subspace X of Fp [G] formed by reduced monomials of degree strictly greater
than 2n(p − 1)/k. We have X k = 0 and dimension of X is exponentially small for large n
(if k > 3) by essentially the same reasons as in previous section.
The group G = UT (n, Fp ), |G| = pn(n−1)/2 , of upper unitriangular n × n-matrices over
Fp may be treated similarly. Namely, we may choose generators gij = id + eij , i < j in
G; each element of G has unique representation as a product of these generators taken in
inverse lexicographic order and in powers at most p − 1
α
α
α
n−1,n
n−2,n
gn−1,n
gn−2,n
. . . g1,21,2 , 0 6 αi,j 6 p − 1.
gij and gkl commute unless j = k or i = l. In this case we have relations gij gjl = gjl gij gil .
Denote xij = gij − 1 in a group ring Fp [G], we have xpij = 0 and there is a basis in Fp [G]
formed by the elements
α
α
α
n−1,n
n−2,n
xn−1,n
xn−2,n
. . . x1,21,2 , 0 6 αi,j 6 p − 1.
Define a degree of any word in alphabet {xij }’s as a sum of (j − i) over all used letters
(multiplicity counted of course). If i < j < l we have (1 + xij )(1 + xjl ) = (1 + xjl )(1 +
xij )(1 + xil ), thus xij xjl = xjl xij + xil + xij xil + xjl xil + xjl xij xil . Using these relations we
may reduce each word to a sum of reduced words of the same or greater degree. It remains
to define define a subspace
X of Fp [G] formed by reduced monomials of degree at strictly
P
greater than (p − 1)( i<j (j − i))/k = (p − 1)(n3 − n)/(6k). Using Chernoff bound we get
an estimate for the maximum of A of order |G|1−c with some c > 0 depending on p, but
bounded from both sides.
I suppose that the huge subspaces with zero cube must exist in the group rings of all
finite groups with small exponent. Specific conjecture: for any N there exists λ < 1 such
that any group G in which g N = 1 for all g ∈ G, the group algebra K[G] has a subspace
with zero cube of codimension at most |G|λ. Here K may depend on G, I guess that
K = Fp for some prime divisor p of N should work.
Even if true, this does not cover the upper unitriangular matrices case, so possibly
something even better (what?) holds.
8
Acknowledgments
I am grateful to Vsevolod Lev, Jordan Ellenberg, Ilya Bogdanov, Ilya Shkredov, Roman
Karasev, Roman Mikhailov, Cosmin Pohoata, Alexander Efimovich Zalesskii and Anatoly
Moiseevich Vershik for fruitful discussions.
References
[1] Noga Alon. Combinatorial Nullstellensatz, Combin. Probab. Comput. 8 (1999), 7–29.
[2] Ernie Croot, Vsevolod Lev, and Peter Pach. Progression-free sets in Zn4 are exponentially small. Ann. of Math. 185 (2017), no. 1, 331–337.
[3] Jordan Ellenberg and Dion Gijswijt. On large subsets of Fq n with no three-term arithmetic progression. Ann. of Math. 185 (2017), no. 1, 339–343
[4] Eric Naslund. Upper Bounds for Sunflower-Free Sets in Znd .
URL:https://goo.gl/Tc5V6I
[5] Ilya Bogdanov, the answer to Mathoverflow question.
URL: http://mathoverflow.net/questions/239222/vector-with-many-non-zero-coordinates
[6] John E. Olson. A Combinatorial Problem on Finite Abelian Groups I, Journal of
Number Theory, 1(1) (1969), 8–10.
[7] Kevin O’Bryant. Sets of integers that do not contain long arithmetic progressions. Electronic Journal of Combinatorics 18 (1) (2011).
URL:http://www.combinatorics.org/ojs/index.php/eljc/article/download/v18i1p59/pdf
[8] Roy Meshulam..On the maximal rank in a subspace of matrices. The Quarterly Journal
of Mathematics 36 (1985), no. 2, 225–229.
[9] Robert Kleinberg, Will Sawin, David Speyer. The Growth Rate of Tri-Colored SumFree Sets, arXiv preprint arXiv:1607.00047 (2016).
[10] Sergey Norin. A distribution on triples with maximum entropy marginal, arXiv
preprint arXiv:1608.00243 (2016).
[11] Luke Pebody. Proof of a Conjecture of Kleinberg-Sawin-Speyer, arXiv preprint
arXiv:1608.05740 (2016).
[12] Jordan Ellenberg. Sumsets as unions of sumsets of subsets, arXiv preprint
arXiv:1612.01929 (2016).
[13] David Speyer, blog post. URL:https://sbseminar.wordpress.com/2016/07/08/boundsfor-sum-free-sets-in-prime-power-cyclic-groups-three-ways/
| 4math.GR
|
arXiv:1708.03372v1 [math.ST] 10 Aug 2017
Optimal Prediction for Additive Function-on-Function Regression
Matthew Reimherr∗ and Bharath Sriperumbudur
Bahaeddine Taoufik
Pennsylvania State University
Lynchburg College
Abstract
As with classic statistics, functional regression models are invaluable in the analysis of functional data. While there are now extensive tools with accompanying theory available for linear
models, there is still a great deal of work to be done concerning nonlinear models for functional
data. In this work we consider the Additive Function-on-Function Regression model, a type of
nonlinear model that uses an additive relationship between the functional outcome and functional covariate. We present an estimation methodology built upon Reproducing Kernel Hilbert
Spaces, and establish optimal rates of convergence for our estimates in terms of prediction error. We also discuss computational challenges that arise with such complex models, developing
a representer theorem for our estimate as well as a more practical and computationally efficient
approximation. Simulations and an application to Cumulative Intraday Returns around the
2008 financial crisis are also provided.
1
Introduction
Functional data analysis (FDA) concerns the statistical analysis of data where one of the variables
of interest is a function. FDA has seen rapidly increasing interest over the last few decades and
has successfully been applied to a variety of fields, including economics, finance, the geosciences,
and the health sciences. One of the most fundamental tools in statistics is linear regression, as
such, it has been a major area of research in FDA. While the literature is too vast to cover here,
we refer readers to Ramsay and Silverman (2006); Ramsay et al. (2009); Horváth and Kokoszka
(2012); Kokoszka and Reimherr (2017), which provide introductions to FDA, as well as Morris
(2015), which provides a broad overview of methods for functional linear regression.
∗
Corresponding
author:
Matthew
Reimherr,
411
[email protected]
1
Thomas
Building,
University
Park,
PA
16802,
A major challenge of functional regression is handling functional predictors. At least conceptually, a functional predictor means having a large number (theoretically infinite) of predictors
that are all highly correlated. To handle such a setting, certain regularity conditions are imposed
to make the problem tractable. Most of these conditions are directly or indirectly related to the
smoothness of the parameter being estimated. However, the convergence rates of the resulting
estimators then depend heavily on these assumptions, and the rates are not parametric when the
predictor is infinite dimensional.
One of the most well studied models in FDA is the functional linear model. Commonly, one
distinguishes between function-on-scalar, scalar-on-function, and function-on-function regression
when discussing such models, with first term denoting the type of response and the second term
denoting the type of covariate. The convergence rates for function-on-scalar regression are usually
much faster than for the scalar-on-function or function-on-function. Methodological, theoretical,
and computational issues related to functional linear models are now well understood. More recently, there has been a growing interest in developing nonlinear regression models. While it is
natural to begin examining nonlinear models after establishing the framework for linear ones, there
is also a practical need for such models. Functional data may exhibit complicated temporal dynamics, which may exhibit nonlinear patterns and are not well modeled assuming linearity; Fan
et al. (2015) examine this issue deeply.
Nonlinear regression methods for FDA have received a fair amount of attention for the scalaron-function setting, while function-on-function regression models, where the relationship between
the response and covariates is believed to be nonlinear, have received considerably less attention. Concerning nonlinear scalar-on-function regression, James and Silverman (2005) introduced
a functional single index model, where the outcome is related to a linear functional of the predictor
through a nonlinear transformation. This work would later be extended in Fan et al. (2015), allowing for a potentially high-dimensional number of a functional predictors. Preda (2007) explored
fitting a fully nonlinear model using reproducing kernel Hilbert spaces (RKHS). In contrast, Müller
et al. (2013) simplified the form of the nonlinear relationship by introducing the functional additive
model, which combines ideas from functional linear models and scalar additive models Hastie and
Tibshirani (1990). Optimal convergence rates for the functional additive model were then established by Wang and Ruppert (2015), which generalized the work of Cai and Yuan (2012) in the
linear case. An alternative to the functional additive model was given in Zhu et al. (2014) who
first expressed the functional predictor using functional principal components analysis, FPCA, and
2
then built an additive model between the outcome and scores.
Moving to function-on-function regression, Lian (2007) extended the work of Preda (2007) to
functional outcomes, which was then also considered in Kadri et al. (2010). Most relevant to the
present paper is the work of Scheipl et al. (2015) who extended the work of Müller et al. (2013) by
introducing an additive model for function-on-function regression. They used a general trivariate
tensor product basis approach for estimation, which allowed them to rely on GAM from the MGCV
package in R to carry out the computation, as is implemented in the Refund package. Ma and
Zhu (2016), examining the same model, considered a binning estimation technique combined with
FPCA. In addition, they were able to prove convergence of their estimators, but made no mention
of optimality while also needing a great deal of assumptions which are challenging to interpret.
Another estimation technique was examined in Kim et al. (2017), which was similar to the trivariate
tensor product approach of Scheipl et al. (2015), but two of the bases are explicitly assumed to be
orthogonal B-splines, while the third comes from an FPCA expansion. However, as with Scheipl
et al. (2015), no theoretical justification is provided. Lastly, in very recent work, Sun et al. (2017)
considered the case of using an RKHS framework to estimate a function-on-function linear model.
Extending the work the Cai and Yuan (2012), they were able to establish the optimality of their
procedure. Our work can viewed as extending this work to nonlinear relationships via a functionon-function additive model.
The goal of this work is to develop a penalized regression framework based on Reproducing
Kernel Hilbert Spaces, RKHS, for fitting the additive function-on-function regression model, AFFR
Scheipl et al. (2015). A major contribution of this work is to provide optimal convergence rates of
our estimators in terms of prediction error, and that this rate is the same as for the scalar outcome
setting Wang and Ruppert (2015). We also discuss computational aspects of our approach, as
the RKHS structure allows for a fairly efficient computation as compared to the trivariate tensor
product bases that have been used previously. Background and the model are introduced in Section
2. Computation is discussed in Section 3, while theory is presented in Section 4. We conclude with
a numeric study consisting of simulations and an application to financial data.
2
Model and Background
We assume that we observe i.i.d pairs {(Xi (t), Yi (t)) : i = 1, . . . , n, t ∈ [0, 1]}. The functions
could be observed on other intervals, but as long as they are closed and bounded, then they can
3
always be rescaled to be [0, 1], thus it is common in FDA to work on the unit interval. Both the
outcome, Yi (t), and Xi (t) are assumed to be completely observed functions, a practice sometimes
referred to as dense functional data analysis Kokoszka and Reimherr (2017); practically this means
that the curve reconstruction contributes a comparatively small amount of uncertainty to the final
parameter estimates. The additive function-on-function regression model is defined as
Z
1
g(t, s, Xi (s)) ds + εi (t).
Yi (t) =
0
We assume that the functions Xi , εi , and Yi are elements of L2 [0, 1], which is a real separable
Hilbert space. The trivariate function, g(t, s, x) is assumed to be an element of an RKHS, denoted
K.
Recall that an RKHS is a Hilbert space that possesses the reproducing property, namely, we
assume that K is a Hilbert space of functions from [0, 1] × [0, 1] × R → R, and that there exists a
kernel function k(t, s, x, t0 , s0 , x0 ) = kt,s,x (t0 , s0 , x0 ) that satisfies
f (t, s, x) = hkt,s,x , f iK ,
for any f ∈ K. There is a one-to-one correspondence between K and k, thus choosing the kernel
function completely determines the resulting RKHS. The functions in K inherit properties from k,
in particular, one can choose k so that the functions in K possess some number of derivatives, or
satisfy some boundary conditions. In addition, many Sobolev spaces, which are commonly used
to enforce smoothness conditions, are also RKHS’s. We refer an interested reader to Berlinet and
Thomas-Agnan (2011) for further details.
We propose to estimate g by minimizing the following penalized objective:
(1)
RSSλ (g) =
n Z
X
i=1
0
1
2
1
Z
Yi (t) −
g(t, s, Xi (s)) ds
0
dt + λkgk2K ,
i.e.,
ĝ = arg inf RSSλ (g),
g∈K
where λ > 0. As we will see in the next section, an explicit solution to this minimization problem
exists due to the reproducing property. However, we will also discuss using FPCA to help reduce
the computational burden.
4
3
Computation
One of the benefits of using RKHS methods is that one can often get an exact solution to the
corresponding minimization problem such as the one in (1), due to the representer theorem Kimeldorf and Wahba (1971). This also turns out to be the case here, however, later on we will discuss
using a slightly modified version that still works well and is easier to compute. The expression we
derive is quite a bit simpler than the analogs derived in Cai and Yuan (2012); Wang and Ruppert
(2015); Sun et al. (2017); this is partly due to our use of functional principal components, which
simplify the expression and also provide an avenue for reducing the computational complexity of
the problem, and also due to our use of the RKHS norm penalty when fitting the model (where as
others used a more general penalty term).
Using the reproducing property we have
hkt,s,Xi (s) , giK = g(t, s, Xi (s))
for
i = 1, 2, ..., n.
We then have that
Z
Z 1
Z 1
(2)
g(t, s, Xi (s))ds =
hg, kt,s,Xi (s) iK ds = g,
0
0
0
1
,
kt,s,Xi (s) ds
K
which is justified by the integrability constraints inherent in Assumption 1(iii ), discussed in the
next section. Let v̂1 , v̂2 , ..., v̂n denote the empirical functional principal components, EFPC’s,
of Y1 , Y2 , ..., Yn .
Then, assuming the Yi have been centered, it is a basic fact of PCA that
span{v̂1 , . . . , v̂n } = span{Y1 , . . . , Yn }. Recall that it is also a basic fact from linear algebra that the
v̂1 , v̂2 , ..., v̂n can be completed to form a full orthonormal basis (all of the additional functions will
have an empirical eigenvalue of 0). We then apply Parseval’s identity to obtain
Z 1Z 1
2
2
Z 1
n X
∞
n Z 1
X
X
hYi , v̂j i − g,
kt,s,X(s) v̂j (t)dtds
.
Yi (t) −
g(t, s, Xi (s))ds dt =
i=1
0
0
0
i=1 j=1
Define the subspace (of K)
Z 1 Z
H1 = span
0
0
K
1
k ((w, x, y); (t, s, Xi (s))) v̂j (t)dtds, i = 1, 2, ..., n, j ≥ 1 ,
0
as well as its orthogonal compliment H1⊥ . The space K can be decomposed into the direct sum:
K = H1 ⊕ H1⊥ , which means that we can write any function g ∈ K as g = g1 + g1⊥ , with g1 ∈ H1
and g1⊥ ∈ H1⊥ . Using this decomposition we have that, for j ≤ n,
Z 1Z 1
Z 1Z 1
(3)
g,
kt,s,Xi (s) v̂j (t)dtds
= g1 ,
kt,s,Xi (s) v̂j (t)dtds
.
0
0
0
K
5
0
K
Since kgk2K = kg1 k2K + kg1⊥ k2K , it follows from (1) and (3) that ĝ ∈ H1 and has the form
ĝ(t, s, x) =
n X
n
X
Z
1Z 1
αij
0
i=1 j=1
k (t, s, x); (t0 , s0 , Xi (s0 )) v̂j (t0 )dt0 ds0 ,
0
i.e., the sum over j terminates at n instead of tending to infinity. Note that this same expression
would hold if we replaced the {vj (t)} with {Yj (t)} (since they span the same space), however, it
would not hold for an arbitrary basis. We use the FPCs for computational reasons as we discuss at
the end of the section. To compute the estimate, ĝ, we need only compute the coefficients {αij }.
As usual, the coefficients αij can be computed via a type of ridge regression. Note that
Z 1Z 1
kt,s,Xi (s) v̂j (t)dtds
ĝ,
0
=
=
0
n X
n
X
i0 =1 j 0 =1
n X
n
X
K
1Z 1Z 1Z 1
Z
α i0 j 0
0
0
0
0
1Z 1Z 1Z 1
Z
α i0 j 0
0
i0 =1 j 0 =1
0
0
hkt,s,Xi0 (s) , kt0 ,s0 ,Xi (s0 ) iK v̂j 0 (t)v̂j (t0 )dtdsdt0 ds0
k(t, s, Xi0 (s); t0 , s0 , Xi (s0 ))v̂j 0 (t)v̂j (t0 )dtdsdt0 ds0 .
0
Define
Z
1Z 1Z 1Z 1
Aiji0 j 0 =
0
0
0
k(t, s, Xi0 (s); t0 , s0 , Xi (s0 ))v̂j 0 (t)v̂j (t0 )dtdsdt0 ds0 .
0
Turning to the norm in the penalty we can use the same arguments to show that
kĝk2K = hĝ, ĝiK =
X
αij Aiji0 j 0 αi0 j 0 .
iji0 j 0
Thus the minimization problem can be phrased as
2
n X
n
X
X
X
Yij −
Aiji0 j 0 αi0 j 0 + λ
αij Aiji0 j 0 αi0 j 0 .
i=1 j=1
i0 j 0
iji0 j 0
We now vectorize the problem by stacking the columns of Yij and αij , denoted as YV and αV . We
also turn the array Aiji0 j 0 into a matrix AV , by collapsing the corresponding dimensions. We can
then phrase the minimization problem as
(YV − AV αV )> (YV − AV αV ) + λα>
V AV αV .
Thus, the final estimate can be expressed as
−1
α̂V = (A>
V AV + λAV ) AV YV .
Note that we are estimating n2 parameters and inverting an n2 × n2 matrix. Thus for computational convenience, it is often useful to truncate the EFPCs at some value J < n. However,
6
even without truncating this approach still has the potential to lead to significantly less parameters
than the basis methods of Scheipl et al. (2015), where the number of parameters to estimate is m3
with m being the number of basis functions used in their tensor product basis. In contrast, our
approach yields only n2 parameters, and combined with an FPCA, this can be reduced to nJ with
relatively little loss in practical predictive performance. There is also the possibility of using an
eigen-expansion on k to reduce the computational complexity even further Parodi and Reimherr
(2017), though we don’t purse that here.
4
Asymptotic Theory
In this section, we demonstrate that the excess risk, <n (defined below), of our estimator converges
to zero at the optimal rate. Optimal convergence of <n , for scalar-on-function linear regression
was established by Cai and Yuan (2012), while optimal convergence for the continuously additive
scalar-on-function regression model was established in Wang and Ruppert (2015). In both cases an
RKHS estimation framework was used. Because our model involves a functional response, the form
of the excess risk <n is different and requires some serious mathematical extensions over previous
works. However, we will show that the convergence rate for our model is the same as the one found
in Wang and Ruppert (2015).
We begin by defining the excess risk, <n . Let Xn+1 (t) be new predictor which is distributed
as, but independent of (Xi (t))ni=1 . We let E∗ denote the expected value, conditioned on the data
{(Yi , Xi ) : 1 ≤ i ≤ n}. Then the excess risk is defined as
Z 1 Z 1
<n = E∗
(ĝ(t, s, Xn+1 (s)) − g(t, s, Xn+1 (s)))2 dtds .
0
0
Note that <n is still a random variable as it is a function of the data. Intuitively, this quantity can be
thought of as prediction error, namely, for a future observation, how far away is our prediction from
the optimal one where the true g is known. For ease of exposition, we present all of assumptions
below, even the ones discussed previously.
Assumption 1. We make the following assumptions.
(i ) The observations {Yi (t), Xi (t)} are assumed to satisfy
Z
Yi (t) =
g(t, s, Xi (s)) ds + εi (t)
7
where {Xi } and {εi } are independent of each other and iid across i = 1, . . . , n.
(ii ) Denote by Lk the integral operator with k as its kernel:
Z
(Lk f )(t, s, x) :=
k(t, s, x; t0 , s0 , x0 )f (t0 , s0 , x0 ) dt0 ds0 dx0 .
The kernel, k, which also defines the RKHS, K, is assumed to be symmetric, positive definite, and
square integrable.
(iii ) Assume that there exists a constant c > 0 such that for any f ∈ K and t ∈ [0, 1] we have
Z
E
4
1
f (t, s, X(s)) ds
" Z
≤c E
0
1/2
(iv ) Let Lk
1/2
−1/2
kt,s,x := Lk
1
2 #2
f (t, s, X(s)) ds
< ∞.
0
denote a square–root of L (which exists due to Assumption 1(ii )) and define
kt,s,x . Define the operator, C, as
Z Z Z
1/2
1/2
0
C(f ) = E
kt,s,Xi (s) hkt,s0 ,Xi (s0 ) , f iL2 dsds dt .
Assume that the eigenvalues {ρk : k ≥ 1} of C satisfy ρk k −2r for some constant r > 1/2.
(v ) There exists a constant M > 0 such that, for all t ∈ [0, 1] and i = 1, . . . , M
E(2i (t)) ≤ M < ∞.
(vi ) The function g lies in Ω, which we assume is a closed bounded ball in K.
We are now in a position to state our main result.
2r
Theorem 1. If Assumption 1 holds and the penalty parameter, λ, is chosen such that λ n− 2r+1
then we have that
−2r
lim lim sup P <n ≥ An− 2r+1 = 0.
A→∞ n→∞ g∈Ω
Before interpreting this result, let us discuss each of the assumptions individually. Assumption
1(i ) explicitly defines the model we are considering. Assumption 1(ii ) ensures that the kernel has
a spectral decomposition via Mercer’s theorem, which will be used extensively. Assumption 1(iii )
is fairly typical in these sorts of asymptotics, assuming that the fourth moment is bounded by a
constant times the square of the second. Assumption 1(iv ) introduces a central quantity that is
used extensively in the proofs. While not immediately obvious, this assumption basically states
how “smooth” or “regular” the function g is, as g must lie in K, whose kernel contributes to C. In
such results it is common for X to contribute to the asymptotic behavior as the prediction error
8
1/2
depends on the complexity of the X. Note that kt,s,x is a well defined quantity and it is easy to
show via the reproducing property that it is an element of L2 ([0, 1]2 × R). The operator C does
depend on the choice of the square-root K 1/2 (which is not a unique choice), however its eigenvalues
do not. Assumption 1(v ) simply assumes that the point-wise variance of the errors is bounded,
while the last assumption requires that the true function lie in a ball in K, which is used to control
the bias of the estimate.
The rate given in Theorem 1 is the same as was found in the scalar outcome case in Wang
and Ruppert (2015), thus we know that this is the minimax rate of convergence. In our case, as
well as in Wang and Ruppert (2015) and Cai and Yuan (2012), it is the interaction between the
covariance of X and the kernel k which determines the optimal rate. The proof is quite extensive
and given in the appendix. The idea of the proof is to rephrase the estimate using operator notation
instead of the representation theorem. The difference between the estimate and truth is then split
into a bias/variance decomposition. Bounding the bias turns out to be relatively straight forward.
Bounding the variance is done by decomposing it into five more manageable pieces, and then
bounding each of them separately. Our task is complicated by the fact that the errors and response
are now functions, where as in both Wang and Ruppert (2015) and Cai and Yuan (2012) they were
scalars. This requires extending many of the lemmas to this new setting, as well as using some
completely new arguments to get the necessary bounds in place.
5
Simulation Study
Here we investigate the prediction performance of AFFR. We compare it with a linear model
estimated in one of two ways. The first way will be denoted as LM R (Reduced linear model) and
LM F (Non-reduced linear model), where both use FPCA to reduce the dimension of the predictors,
but LM R also reduces the dimension of the outcome, while LM F does not. To implement our
approach we relied heavily on the TensorA package van den Boogaart (2007) in R, which allowed
us to carryout various tensor products very quickly.
We consider three different settings for g(t, s, x) one linear and two nonlinear forms:
(a) Model (a): g(t, s, x) = tsx,
(b) Model (b): g(t, s, x) = t + s + x2 ,
(c) Model (c): g(t, s, x) = tsx2 + x4 .
9
In all settings, the predictors Xi (t) and errors i (t) are taken to be iid Gaussian processes with
mean 0 and the following covariance function from the Matérn family:
!
!
√
√
5|t − s| 5|t − s|2
5|t − s|
C(t, s) = 1 +
,
+
exp −
ρ
3ρ2
ρ
where ρ = 1/4. For the RKHS we utilized the Gaussian kernel
h
i
0
0
0
−δ (x−x )2 +(y−y )2 +(z−z )2
k (x, y, z), (x , y , z ) = e
,
0
0
0
where δ is the tuning parameter of the Gaussian kernel. We will examine the sensitivity of our
approach to this parameter in Table 2. All of the curves (Xi (t), Yi (t), and εi (t)) were simulated on a
M = 50 equispaced grid between 0 and 1. The data is approximated by a B-spline of K = 100 basis
functions, using the fda package. We denote by JX and JY the number of principal components of
Xi and Yi .
To evaluate the different approaches, we used 1000 repetitions of every scenario. In each case we
generate 150 curves to fit the different models and then generated another 150 curves to evaluate
out-of-sample prediction error. The metric for determining prediction performance we denote as
RPE, for relative prediction error. This metric denotes the improvement of the predictions over
just using the mean. An RPE of 0 means that the model shows no improvement over just using
the mean, while an RPE of 1 means the predictions are perfect. More precisely, we first compute
the Mean Squared Prediction error as:
M SP E =
n
X
kYi − Ybi k2L2 ,
i=1
where Ybi is a predicted value using one of the three discussed models or simply the mean. The
RPE is then defined as
RP E =
M SP Emean − M SP E
,
M SP Emean
where M SP Emean denotes the MSPE using a mean only model. Note that even in the mean only
model, all parameters are estimated on the initial 150 curves and prediction is then evaluated on
the second 150. Therefore, it is actually possible to have a numerically negative RPE if an approach
isn’t predicting any better than just using the mean.
The RPEs of LM R and LM F for the three models (a), (b), and (c) are summarized in Table
1. For both models, we took JX = 3, which explained over 85% of the variability of the predictors
and for LM R we took JY = 3 PCs for the outcome as well. The RPEs for our approach with
10
δ = {2−3 , 2−2 , 2−1 , 1, 2} and JY = 3, 6, 9 are summarized in Table 2. An initial look at the tables
confirms much of what one would expect. When the true model is linear, the two linear approaches
work best, resulting in about half the RPE of AFFR. However, when moving to the two nonlinear
models, the AFFR approach does substantially better. This increased performance is seen for any
choice of JY and δ. Furthermore, the prediction performance seems relatively robust to the choice
of JY and δ. In the case of JY this is not so surprising as over 90% of the variability of the Yi is
explained by the first three FPCs. In contrast, there is some sensitivity to the choice of δ, but it is
relatively weak given how much we are changing δ in each row. In our application section we set
δ using a type of median, but one could also refit the model with a few different δ and choose the
one with the best prediction performance. Given how consistent the AFFR predictions are, trying
a few δ appears to be satisfactory, and large grid searches can be avoided.
Model (a)
Model (b)
Model (c)
LM R
0.045
0.030
0.060
LM F
0.045
0.029
0.060
Table 1: Relative prediction errors, RPE, for the two linear models. For both, the number of FPCs
for the predictor is JX = 3. LMR also reduces the dimension of the outcome with JY = 3 FPCs.
Model (a)
Model (b)
Model (c)
JY = 3
JY = 6
JY = 9
JY = 3
JY = 6
JY = 9
JY = 3
JY = 6
JY = 9
δ = 2−3
0.025
0.026
0.026
0.379
0.379
0.379
0.840
0.840
0.845
2−2
0.024
0.025
0.025
0.370
0.370
0.370
0.816
0.804
0.815
δ = 2−1
0.023
0.024
0.023
0.360
0.361
0.361
0.847
0.831
0.830
20
0.022
0.023
0.021
0.346
0.347
0.347
0.83
0.83
0.83
δ = 21
0.020
0.021
0.019
0.328
0.328
0.400
0.808
0.808
0.790
P EV
90.45%
99.12%
99.88%
90.82%
99.10%
99.84%
91.25%
99.22%
99.87%
δ=
δ=
Table 2: Relative prediction error, RPE, for AFFR with different kernel parameter values, δ. In
every case the penalty parameter, λ, is chosen using cross-validation. PEV indicates the proportion
of explained variance of Y for the corresponding number of FPCs, JY .
6
Application to Cumulative Intraday Data
We conclude with an illustration of our approach applied to real data. Cumulative Intra-Day
Returns (CIDR’s) consist of daily stock prices that are normalized to start at zero at the beginning
11
of each trading day. FDA methods have been useful in analyzing such data Gabrys et al. (2010);
Kokoszka and Reimherr (2013); Horváth et al. (2014), given the density at which stock prices can
be observed. Let Pi (tj ) denote the price of a stock on day i and time of day tj . The CIDRs are
then defined as
Ri (tj ) = 100 [ln Pi (tj ) − ln Pi (t1 )] , i = 1, ..., n, j = 1, ..., M.
The CIDRs are observed each minute throughout the trading day. This corresponds to M = 390
minutes (9:30 am-4:00 pm EST) of trading time for each trading day of the New York Stock
Exchange, NYSE. In this application study, we deal with the CIDR’s of two of the most important
US market indexes: Standard & Poor’s 500 Index (S&P 500) and the Dow Jones Industrial Average
(DJ). Also, we consider two individual stocks: General Electric Company (GE) and International
Business Machines Corporation (IBM). The study period of the data consists of three periods
in relation to the 2007–2008 financial crisis, denoted as Before (06/13/2006-04/10/2007), During
(11/01/2007-07/28/2008), and After (01/04/2010-10/1/2010). These periods each contain 270
calendar days.
We investigate the performance of the market indexes, S&P 500 and DJ, in predicting GE
and IBM for the three periods. Understanding such relationships is imperative for developing
financial portfolios as many strategies consist of balancing buying/shorting certain stocks with
buying/shorting market indices Nicholas (2000). We fit three different models; the first is our
discussed model, AFFR, while the other two are linear models. The first linear model uses an
FPCA on both the outcome and predictor (5 PCs for both) and then fits a multivariate linear
model, while the second linear model only uses FPCA on the predictor (5 PCs) Kokoszka and
Reimherr (2017). To evaluate the prediction performance for each period we split each period into
3 equal folds and use a type K-fold cross-validation. The model is fit on two folds, while prediction
is then evaluated on the third. We use the Gaussian kernel from Section 5 and the smoothing
parameter selected via Generalized Cross-Validation. Prediction performance is then averaged over
the 3 folds. To provide a more readily interpretable metric for prediction performance, we use
the same RPE metric given in Section 5, which denotes the relative performance of a model with
respect to a mean only model. A value of 1 means perfect prediction, while a value of 0 indicates
that the model is doing no better than just using the mean. The results are summarized in Table
3.
As we can see, all models perform better, during and after the crisis. This suggests that the
behavior of the market had not returned to its pre-crisis characteristics. Looking at Figure 1, we
12
can clearly see that the volatility increases during and after the financial crises. This suggests
that the overall “market” effect on the stocks is stronger during periods of high-volatility. When
comparing the three different models, the linear models do nearly the same, which is to be expected
since 5 PCs explains over 90% of the variability of the stocks. The AFFR model is not too far
behind, but does noticeably worse in every setting. This suggests that the relationship between the
discussed stocks and the indices is approximately linear; if there are any nonlinear relationships
then they are either very minor deviations from linearity or are not well captured by an additive
structure.
Period
Before
During
After
Model
AFFR
LMR
LMF
AFFR
LMR
LMF
AFFR
LMR
LMF
GE on DJ
0.081
0.249
0.247
0.429
0.537
0.536
0.354
0.474
0.474
GE on SP
0.078
0.199
0.198
0.322
0.479
0.479
0.398
0.465
0.465
IBM on DJ
0.024
0.249
0.248
0.416
0.551
0.550
0.293
0.454
0.454
IBM on SP
0.034
0.194
0.193
0.249
0.435
0.434
0.248
0.399
0.399
Table 3: Prediction performance of three models: AFFR (our model), LMR (linear model with PCA
in both the outcome and predictor), and LMF (linear model with PCA on the predictor only). The
top row corresponds to predicting GE based on DJ, the second corresponds to prediction GE from
SP, and so on. Each number denotes the relative increase in out-of-sample prediction performance
over a mean only model, with 100% denoting perfect prediction and 0% denoting no increase over
just using the mean.
A
A.1
Proof of Theorem 1
Excess Risk
We begin by expressing the excess risk in an alternative form. Recall that k(t, s, x; t0 , s0 , x0 ) is the
kernel function used to define the RKHS, K. This kernel can be viewed as the kernel of an integral
operator, Lk , which maps L2 ([0, 1]2 × R) → K ⊂ L2 ([0, 1]2 × R). In particular
Z Z Z
(Lk f )(t, s, x) =
k(t, s, x; t0 , s0 , x0 )f (t0 , s0 , x0 ) dt0 ds0 dx0 .
From here on, for simplicity, we will denote L2 ([0, 1]2 × R) as simply L2 . By Assumption 1, Lk is a
positive definite, compact operator, which is also self-adjoint in the sense that hf, Lk gi = hLk f, gi,
13
Figure 1: Plots of the intraday cumulative returns for the Dow Jones Index (top) and General
Electric (bottom) before (left), during (middle), and after (right) the 2008 financial crisis.
1/2
for any f and g in L2 . We can therefore define a square-root of Lk , denoted as Lk
−1
that satisfies
1
f1 ∈ K ⇐⇒ Lk 2 f1 ∈ L2 and Lk2 f2 ∈ K ⇐⇒ f2 ∈ L2 .
−1/2
Note that if Lk has a nontrivial null space, then Lk
can still be well defined since assuming
f ∈ K means that f is orthogonal to the null space of L. Recall that one can also move between
the K and L2 inner product as follows
−1/2
hf, giK = hLk
−1/2
f, Lk
giL2 = hf, L−1
k giL2 .
We refer the interested reader to Kennedy and Sadeghi (2013) for more details.
Let ĝ denote our estimate of the true function, g. We then define the following
1
−1
2
kt,s,X(s)
= Lk 2 kt,s,X(s) ,
−1
−1
h = Lk 2 g and ĥ = Lk 2 ĝ.
Using the reproducing property, we have that
1
1
2
g(t, s, X(s)) = hkt,s,X(s)
, hiL2
2
and ĝ(t, s, X(s)) = hkt,s,X(s)
, ĥiL2 .
Now define the random operator, T : L2 → L2 as
1/2
1/2
Tt,s,s0 = kt,s,Xn+1 (s) ⊗ kt,s0 ,Xn+1 (s0 ) ,
where ⊗ denotes the tensor product, and the resulting object is interpreted as an operator:
1/2
1/2
Tt,s,s0 (f ) = kt,s,Xn+1 (s) hkt,s0 ,Xn+1 (s0 ) , f iL2 .
14
We also define a second operator, which integrates out t, s, and s0 , and takes an expectation over
Xn+1 :
Z
(4)
1Z 1Z 1
C=E
0
Tt,s,s0 dsds dt .
0
0
0
Note that C is a symmetric, positive definite, compact operator, and thus has a spectral decomposition
C=
(5)
∞
X
ρk (φk ⊗ φk ),
k=1
where ρk ≥ 0 and φk ∈ L2 are, respectively, the eigenvalues and eigenfunctions of C. This decomposition will be used later on.
As we said before, denote by E∗ the expected value conditioned on the data (X1 , Y1 ), ..., (Xn , Yn ).
The excess risk can be written as
2
Z 1 Z 1
[ĝ(t, s, Xn+1 (s)) − g(t, s, Xn+1 (s)))] ds dt
<n = E∗
0
∗
Z
0
1 Z 1
hkt,s,Xn+1 (s) , ĥλ iL2 − hkt,s,Xn+1 (s) , hiL2
=E
0
∗
Z
0
1 Z 1
0
∗
Z
1Z
0
1Z 1
=E
0
= E∗
Z
0
0
0
1Z 1Z 1
0
2
ds dt
2
1
2
hkt,s,Xn+1 (s) , ĥλ − hiL2 ds
=E
1
2
1
2
dt
1
1
0
2
2
hkt,s,X
, ĥ − hiL2 hkt,s
0 ,X
0 , ĥ − hiL2 dsds dt
n+1 (s)
n+1 (s )
hTt,s,s0 (ĥ − h), ĥ − hiL2 dsds0 dt
0
= hC(ĥ − h), ĥ − hiL2 = kĥ − hk2C .
Thus, the excess risk can be expressed as sort of a weighted L2 norm, where the operator C defines
the weights, which is composed of the kernel and the distribution of Xn+1 .
A.2
Re-expressing the Estimator
In this section we define an alternative form for the estimator ĝ, which was given in Section 3.
In particular, instead of using the reproducing property, we will write down the estimator using
operators. To do this, we will take derivatives of RSSλ (g) with respect to g. Since these are
functions, we mean the Fréchet derivative or strong derivative. Note that RSSλ (g) is a convex
differentiable functional over K. However, so that we are working with L2 instead of K, we use
15
] λ (h) := RSSλ (L1/2 h), where h = L−1/2 g:
RSS
k
k
] λ (h) =
RSS
n Z
X
i=1
1
1
Z
Yi (t) −
0
0
1/2
hh, kt,s,Xi (s) iL2 ds
2
dt + λkhk2L2 .
] λ (h) is a convex differentiable functional over L2 . Thus, when taking the derivative, we
Now RSS
are using the topology of L2 not K.
] λ (h) we first focus on the penalty, which is easier. We have that
To take the derivative of RSS
∂
khk2L2 = 2h.
∂h
] λ (h) we first define the empirical quantities
Turning to the first term in RSS
1/2
1/2
Ti;t,s,s0 = kt,s,Xi (s) ⊗ kt,s0 ,Xi (s0 )
and
n
(6)
1X
Cn =
n
1Z 1Z 1
Z
i=1
0
0
Ti;t,s,s0 dsds0 dt.
0
Now we can apply a chain rule to obtain
" n Z
2 #
Z 1
∂ 1X 1
1/2
Yi (t) −
hh, kt,s,Xi (s) iL2 ds dt
∂h n
0
i=1 0
n Z Z
1
2X 1 1
2
=−
Yi (t)kt,s,X
dsdt + 2Cn h.
i (s)
n
0
0
i=1
For notational simplicity, define
n
Γk1/2 ,Y
1X
=
n
i=1
Z
0
1Z 1
0
1
2
dsdt.
Yi (t)kt,s,X
i (s)
So, we finally have that
∂ ]
RSS λ (h) = −2Γk1/2 ,Y + 2Cn h + 2λh,
∂h
which yields the estimate
(7)
ĥ = (Cn + λI)−1 Γk1/2 ,Y ,
where I is the identity operator.
16
A.3
Proof of Theorem 1 - Controlling Bias
Using Assumption 1 we can express
Z
1/2
hkt,s,Xi (s) , hiL2 + εi (t).
Yi (t) =
and we therefore have that
Γk1/2 ,Y = Cn (h) + fn
where
n
1X
fn =
n
Z
1Z 1
0
i=1
0
1
2
dsdt.
i (t)kt,s,X
i (s)
This implies that ĥ from (7) can be expressed as
ĥ = (Cn + λI)−1 Cn (h) + (Cn + λI)−1 fn .
We introduce an intermediate quantity, hλ , which is given by
hλ = (C + λI)−1 C(h),
where C is defined in (4). The difference between hλ and h represents the bias of the estimator ĥ.
Balancing this quantity with the variance, discussed in the next section, is called the bias-variance
trade off a common term in nonparametric smoothing. Inherently, the idea is that to achieve an
optimal ĥ we have to balance both the bias and variance so that neither one is overly large.
Using the eigenfunctions of C as a basis, we can write
h=
∞
X
ak φk .
k=1
Since C and C + I have the same eigenfunctions, it follows that we can express
C + λI =
∞
X
−1
(λ + ρk )(φk ⊗ φk ) =⇒ (C + λI)
=
k=1
∞
X
(λ + ρk )−1 (φk ⊗ φk ).
k=1
So we have that hλ can be expressed as
hλ = (C + λI)−1 C(h) =
∞
X
ak ρk
φk .
λ + ρk
k=1
So the difference, hλ − h can be written as
(8)
hλ − h = −
∞
X
λak
φk .
λ + ρk
k=1
17
The bias is therefore given by
khλ − hk2C =
∞
∞
X
X
λ2 a2k ρk
ρk
ρk
2
≤
λ
max
.
a2k = λ2 khk2L2 max
2
2
k≥1 (λ + ρk )
k≥1 (λ + ρk )2
(λ + ρk )
k=1
k=1
It is easy to verify that the maximum of F (x) = x/(λ + x)2 is achieved at x = λ with the maximum
value being
1
4λ .
We can therefore bound the bias as
khλ − hk2C ≤
λkhk2L2
.
4
In the statement of Theorem 1 we assume that λ n2r/(2r+1) , which implies that the bias is of the
2r
order n− 2r+1 O(1). We will show in the next section that the variance of our estimate achieves the
same order.
A.4
Proof of Theorem 1 - Controlling Variability
Controlling the variability of the estimates, kĥ − hλ kC follows similar arguments as controlling the
bias. However, there are many more terms which must be analyzed separately. In particular, we
decompose ĥ − hλ into five separate components:
ĥ − hλ = T1 + T2 + T3 + T4 + T5 ,
(9)
where the Ti terms are given by
T1 = (C + λI)−1 C(hλ − h),
T2 = λ(C + λI)−2 C(h),
T3 = −(C + λI)−1 fn ,
T4 = (C + λI)−1 (Cn − C)(hλ − h),
T5 = (C + λI)−1 (C − Cn )(hλ − ĥ).
While a bit tedious, it only requires linear algebra and repeated calls to the definitions of ĥ and hλ to
verify (9), we thus omit the details here. We now develop bounds for each term, kTi kC , separately.
For the first four, it turns out to be convenient to bound kC ν Ti kL2 for 0 < ν ≤ 1/2, as these bounds
will be needed for the final term T5 . Notice that when ν = 1/2 we have kC ν Ti kL2 = kTi kC .
1. Using the eigenfunctions of C to express hλ − h as in (8), we get that
∞
X
λak ρk
φk .
T1 = −
(λ + ρk )2
k=1
18
We then have that
kC ν T1 k2L2 =
∞
2(1+ν)
X
λ2 a2 ρ
k k
(λ + ρk )4
k=1
2(1+ν)
≤ λ2 max
k≥1
ρk
khk2L2 .
(λ + ρk )4
Again, it is a basic calculus exercise to show that
2(1+ν)
1+ν
2(1+ν)
λ
ρ
1−ν
(1 − ν)2(1−ν) (1 + ν)2(1+ν) 1
max k
≤
=
.
4
k≥1 (λ + ρk )4
16
λ2−2ν
λ + λ 1+ν
1−ν
We thus have the bound
kC ν T1 k2L2 ≤ cλ2ν khk2L2 ,
(10)
where c is a constant that depends only on ν.
2. Using the same arguments as in the previous step, we have that
(11)
kC ν T2 k2L2 =
∞
2(1+ν)
X
λ2 a2 ρ
k k
k=1
(λ + ρk )4
≤ λ2 max
k≥1
∞
2(1+ν) X
ρk
a2k ≤ cλ2ν khk2L2 .
(λ + ρk )4
k=1
3. Turning to T3 , we apply Lemma 1 with 0 < ν ≤ 1/2 to obtain
kC ν T3 k2L2 = kC ν (C + λI)−1 fn kL2 =
1
nλ1−2ν+1/2r
Op (1),
1
where r is defined as in Assumption 1. By the statement of Theorem 1 it follows that nλ1+ 2r
tends to a nonzero constant, meaning that
1
(12)
nλ
1
1−2ν+ 2r
λ2ν → 0,
since λ → 0. Thus we have that kC ν T3 kL2 = Op (λν ).
4. To bound T4 we first fix a second value ν > ν2 > 0 that satisfies 2r(1−2ν2 ) > 1, or equivalently
ν2 < (2r −1)/4r, as well as 4r(2ν2 +2ν) > 1, which is possible as long as r > 1/2 (Assumption
1). We now apply a basic operator inequality
kC ν T4 kL2 = kC ν (C + λI)−1 (Cn − C)(hλ − h)kL2
≤ kC ν (C + λI)−1 (Cn − C)C −ν2 kop kC ν2 (hλ − h)kL2 .
and then apply Lemmas 3 and 4 to obtain
−1
1
ν
2
1−2ν+ 2r
(13)
kC T4 kL2 ≤ Op
nλ
Op λ2ν2 =
1
nλ
1
1−2ν+ 2r
op (1) ,
since ν2 > 0 and λ → 0. Using (12) we conclude that kC ν T4 kL2 = Op (λν ).
19
5. The last term is the most involved to bound and the reason why the previous four bounds
involved C ν Ti . We begin by expressing
1
kT5 kC = kC 2 (C + λI)−1 (C − Cn )(hλ − ĥ)kL2
1
≤ kC 2 (C + λI)−1 (C − Cn )C −ν kop kC ν (hλ − ĥ)kL2 .
Here ν > 0 is chosen to satisfy 2r(1 − 2ν) > 1. Applying Lemma 3, we have that
kT5 k2C ≤
1
Op (1)kC ν (h − ĥ)k2L2 .
nλ1/2r
We have now, in some sense, looped back and are dealing with the term h − ĥ. Using (9) we
have
(14)
kC ν (h − ĥ)kL2 ≤ kC ν T1 kL2 + kC ν T2 kL2 + kC ν T3 kL2 + kC ν T4 kL2 + kC ν T5 kL2 .
The first four terms we already have bounds for, so we need only focus on the last, which
again, has looped back to our original term. We now apply Lemma 2 to obtain
kC ν T5 k2L2 = kC ν (C + λI)−1 (C − Cn )(h − ĥ)k2L2 ≤
1
nλ
1
1−2ν+ 2r
kC ν (h − ĥ)k2L2 .
Combining the above with (14) we have that
1
ν
≤ kC ν T1 kL2 + kC ν T2 kL2 + kC ν T3 kL2 + kC ν T4 kL2 .
kC (h − ĥ)kL2 1 −
1
1−2ν+ 2r
nλ
Using (12) it thus follows that
kC ν (h − ĥ)kL2 = Op (1)(kC ν T1 kL2 + kC ν T2 kL2 + kC ν T3 kL2 + kC ν T4 kL2 ),
and applying steps 1-4 we get that
kC ν (h − ĥ)kL2 = Op (λν ) = op (1),
and we finally have that
kT5 kC =
1
1
nλ 2r
op (1).
We can now combine Steps 1-4, taking ν = 1/2, with step 5 to finally conclude that
2r
<2n = kĥ − hk2C = λ2 Op (1) = n− 2r+1 Op (1).
Combined with the results of Section A.3, this concludes the proof.
20
A.5
Auxiliary Lemmas
Here we state four lemmas which are generalizations of ones used in Cai and Yuan (2012) and Wang
and Ruppert (2015).
Lemma 1. If Assumption 1 holds then for any 0 ≤ ν ≤
ν
kC (C + λI)
−1
fn kL2 = Op
1
2
nλ
1
1−2ν+ 2r
− 1
2
.
Lemma 2. Let Assumption 1 hold. Then for any ν > 0 such that 2r(1 − 2ν) > 1, we have that
−1
ν
kC (C + λI)
(Cn − C)C
−ν
kop = Op
nλ
1
1−2ν+ 2r
− 1
2
,
where k.kop represents the usual operator norm i.e., kAkop = suph:khkL2 =1 kAhk.
Lemma 3. Let Assumption 1 hold and fix 0 < ν < ν2 to be any two values that satisfy 2r(1−2ν) > 1
and 4r(ν2 + ν) > 1, then we have that
−1
ν2
kC (C + λI)
(Cn − C)C
−ν
kop = Op
nλ
1
1−2ν2 + 2r
− 1
2
.
Lemma 4. (Lemma 1, Cai and Yuan (2012)) For any 0 < ν < 1,
kC ν (hλ − h)kL2 ≤ (1 − ν)1−ν ν ν λν khkL2 .
Lemma 5. Fix ν > 0 and ν2 > 0 such that 4r(ν2 +ν) > 1. If there exist constants 0 < c1 < c2 < ∞
such that c1 k −2r < sk < c2 k −2r , then there exist constants c3 , c4 > 0 depending only on c1 , c2 such
that
−1
c4 λ 2r −1+2ν2 ≤
∞
X
j=1
2 +2ν
s2ν
j
(λ + sj
)1+2ν
−1
≤ c3 (1 + λ 2r −1+2ν2 ).
Proof of Lemma 1
Recall that
n
fn =
1X
n
i=1
Z
0
1Z 1
0
1
2
i (t)kt,s,X
dsdt.
i (s)
Using Parseval’s identity we have that
kC ν (C + λI)−1 fn k2L2 =
∞
X
k=1
21
ρ2ν
k
hfn , φk i2 .
(λ + ρk )2
Taking expected values yields
∞
ν
E kC (C + λI)
−1
fn k2L2
ρ2ν
1X
k
=
E
n
(λ + ρk )2
1Z 1
Z
0
k=1
0
2
1
2
(t)hkt,s,X(s) , φk iL2 dsdt
.
By Jensen’s inequality we have
1Z 1
Z
0
0
2
1
2
(t)hkt,s,X(s) , φk iL2 dsdt
Z
1 Z 1
≤
(t)hkt,s,X(s) , φk iL2 ds
0
Z
2
1
2
1Z
0
1Z 1
=
0
0
0
1
dt
1
∗
2
2
2 (t)hkt,s,X(s)
, φk iL2 hkt,s
∗ ,X(s∗ ) , φk iL2 dsds dt.
Using the assumed independence between and X, as well as the assumption that E(2 (t)) ≤ M ,
where M is a constant, we obtain
Z
1Z 1
E
0
0
2
1
2
(t)hkt,s,X(s) , φk iL2 dsdt
1Z 1Z 1
Z
≤ME
0
0
0
hkt,s,X(s) , φk iL2 hkt,s∗ ,X(s∗ ) , φk iL2 dsds dt
1
2
1
2
∗
= M hC(φk ), φk i = M ρk .
Since 0 ≤ ν ≤
1
2
and both ρk and λ are positive, we can obtain the bound
−1
ν
E kC (C + λI)
fn k2L2
∞
∞
ρ2ν+1
M X ρ2ν+1
M X
k
k
≤
≤
.
n
(λ + ρk )2
nλ1−2ν
(λ + ρk )1+2ν
k=1
k=1
Now we apply Lemma 5 with ν2 = 1/2 to obtain
E kC ν (C + λI)−1 fn k2L2 ≤
c∗
1
nλ1−2ν+ 2r
,
where c∗ is a constant. An application of Markov’s inequality completes the proof.
Proof of Lemma 2
By definition
kC ν (C + λI)−1 (Cn − C)C −ν k2op =
sup
f :kf kL2 =1
kC ν (C + λI)−1 (Cn − C)C −ν f k2L2 .
Fix f ∈ L2 such that kf kL2 = 1. We can expand f as
f=
∞
X
fk φk ,
k=1
By Parseval’s identity we have
ν
−1
kC (C + λI)
(Cn − C)C
−ν
2
fk =
∞
X
j=1
"
∞
ρνj X
fk ρ−ν
k h(Cn − C)φk , φj iL2
ρj + λ
22
k=1
#2
.
Applying the Cauchy-Schwartz inequality and using the fact that kf kL2 = 1 we have that
!1/2
∞
∞
X
X
−ν
−2ν
2
fk ρk h(Cn − C)φk , φj iL2 ≤
.
ρk h(Cn − C)φk , φj iL2
k=1
k=1
So we can bound the operator norm as
kC ν (C + λI)−1 (Cn − C)C −ν k2op ≤
∞ X
∞
2ν
X
ρ−2ν
k ρj
hφj , (Cn − C)φk i2L2 .
(λ + ρj )2
k=1 j=1
Applying Jensen’s equality we get that
1
1
2
2
∞ X
∞
∞ X
∞
−2ν 2ν
−2ν 2ν
X
X
ρk ρj
ρk ρj
2
2
hφj , (Cn − C)φk iL2
Ehφj , (Cn − C)φk iL2
E
≤
.
(λ + ρj )2
(λ + ρj )2
k=1 j=1
k=1 j=1
Using the definition of Cn from (6) we have that
Z 1 Z 1 Z 1 1
2
1
1
2
2
∗
2
2
Ehφj , (Cn − C)φk iL2 ≤ Ehφj , Cn φk iL2 = E
hkt,s,X(s) , φj ihkt,s∗ ,X(s∗ ) , φk idsds dt .
n
0
0
0
Note the first inequality follows from the fact that C is the mean of Cn and thus replacing C above
with any other quantity cannot decrease it (since it is minimized when using C). One can show
this using basic calculus arguments over Hilbert spaces, thus we omit the details here. By applying
Cauchy-Schwartz inequality and Fubini’s theorem we have
Ehφj , (Cn − C)φk i2L2
"Z Z
#!
2 # "Z 1 Z 1 1
2
1
1
1
1
2
hkt,s,X(s)
hkt2∗ ,s∗ ,X(s∗ ) , φk ids∗ dt∗
, φj ids dt
≤ E
n
0
0
0
0
"Z
2 Z 1 1
2 #
Z Z
1
1
1 1 1
2
E
hkt,s,X(s)
, φj ids
hkt2∗ ,s∗ ,X(s∗ ) , φk ids∗
dtdt∗ .
=
n 0 0
0
0
Using Cauchy-Schwartz inequality again
Ehφj , (Cn − C)φk i2L2
Z 1 1
Z 1 1
4
4
Z Z
1
1 1 1 1
∗
2
2
≤
E2
hkt,s,X(s) , φj ids E 2
hkt∗ ,s∗ ,X(s∗ ) , φk ids
dtdt∗ .
n 0 0
0
0
Note that we can move to the K inner product to obtain:
1
1/2
1/2
2
, φk iL2 = hkt,s,X(s) , Lk φk iK = (Lk φk )(t, s, X(s))
hkt,s,X(s)
and L1/2 φk is a function in K, thus we can apply Assumption 1.4 to obtain
"Z Z
#
2 # "Z 1 Z 1 1
2
1
1
1
c
2
Ehφj , (Cn − C)φk i2L2 ≤ E
hkt,s,X(s)
, φj ids dt E
hkt2∗ ,s∗ ,X(s∗ ) , φk ids∗ dt∗ .
n
0
0
0
0
23
It is easy to see that
"Z
1 Z 1
E
0
0
2
1
2
hkt,s,X(s) , φj ids
#
dt = ρj .
Now we obtain
Ehφj , (Cn − C)φk i2L2 ≤ cn−1 ρj ρk .
Therefore,
1
1
2
2
1−2ν 1+2ν
∞ X
∞
∞ X
∞
2ν
X
X
ρ
ρ
ρ−2ν
ρ
c
j
j
k
k
.
hφj , (Cn − C)φk iL2 ≤
E
(λ + ρj )2
n
(λ + ρj )2
k=1 j=1
k=1 j=1
Note that
∞ X
∞
X
ρ1+2ν
ρ1−2ν
j
k
k=1 j=1
(λ + ρj )2
=
∞
X
ρ1−2ν
k
k=1
∞
X
ρ1+2ν
j
j=1
(λ + ρj )2
.
Since 2r(1 − 2ν) > 1 and ρk < c2 k −2r we have
∞
X
≤ c2
ρ1−2ν
k
∞
X
k −2r(1−2ν) = c∗∗ < ∞.
k=1
k=1
Finally, by applying Lemma 5 with ν2 = 1/2 we obtain
1
1
E kC ν (C + λI)−1 (Cn − C)C −ν kop ≤ γ(nλ1−2ν+ 2r )− 2 ,
An application of Markov’s inequality completes the proof.
Proof of Lemma 3
Recall that
kC ν2 (C + λI)−1 (Cn − C)C −ν k2op =
sup
h:khkL2 =1
kC ν2 (C + λI)−1 (Cn − C)C −ν hk2L2 .
Note that
C ν2 (C + λI)−1 =
∞
X
ρνj 2
j=1
ρj + λ
(φj ⊗ φj ),
and recall that from the proof of Lemma 2,
(Cn − C)C
−ν
h=
∞
X
k=1
24
ak ρ−ν
k (Cn − C)φk .
Therefore
−1
ν2
C (C + λI)
(Cn − C)C
−ν
h=
∞
X
ρνj 2
ρj + λ
j=1
h
∞
X
ak ρ−ν
k (Cn − C)φk , φj iL2 φj .
k=1
By Parseval’s identity we obtain
kC ν2 (C + λI)−1 (Cn − C)C −ν k2op =
∞
X
ρνj 2
"
j=1
ρj + λ
h
∞
X
#2
ak ρ−ν
k (Cn − C)φk , φj iL2
k=1
By Cauchy-Schwartz inequality and using the same steps in the proof of Lemma 2, we have
ν2
kC (C + λI)
−1
(Cn −
C)C −ν k2op
≤
2ν2
∞ X
∞
X
ρ−2ν
k ρj
k=1 j=1
(λ + ρj )2
hφj , (Cn − C)φk i2L2 .
Using the same arguments in the proof of Lemma 2, we obtain
1
1
2
2
2ν2
1−2ν 1+2ν2
∞ X
∞
∞ X
∞
X
X
ρ
ρ−2ν
ρ
ρ
c
j
j
k
k
.
E
hφj , (Cn − C)φk iL2 ≤
(λ + ρj )2
n
(λ + ρj )2
k=1 j=1
k=1 j=1
Note that
∞ X
∞
2
X
ρ1−2ν
ρ1+2ν
j
k
(λ + ρj )2
k=1 j=1
=
∞
X
ρ1−2ν
k
∞
2
X
ρ1+2ν
j
j=1
k=1
(λ + ρj )2
.
1
Note that the condition 2r(1 − 2ν) > 1 implies ν < 12 − 2r
< 12 . We therefore have
λ + ρj 2ν−1
≤ 1.
ρj
It follows that
∞ X
∞
2
X
ρ1+2ν
ρ1−2ν
j
k
(λ + ρj )2
k=1 j=1
≤
∞
X
ρk1−2ν
∞
X
j=1
k=1
2
ρ2ν+2ν
j
(λ + ρj )2ν+1
.
Recall that
∞
X
ρ1−2ν
k
k=1
≤ c2
∞
X
k −2r(1−2ν) = c∗∗ < ∞.
k=1
Therefore
−1
ν2
E kC (C + λI)
(Cn − C)C
−ν
kop ≤
∞
cc∗∗ X
n
j=1
1
2
2
ρ2ν+2ν
j
(λ + ρj )2ν+1
.
By applying Lemma 5 we have that
1
1
1
E kC 2 (C + λI)−1 (Cn − C)C −ν kop ≤ β(nλ 2r +1−2ν2 )− 2 ,
where β is a constant. An application of the Markov inequality completes the proof.
25
A.6
Proof of Lemma 5
Using the same arguments as in Cai and Yuan (2012), we get that
∞
X
j=1
2 +2ν
s2ν
j
(λ + sj )1+2ν
=
≤
∞
X
j=1
∞
X
j=1
s1+2ν
j
(λ + sj )1+2ν
sj2ν2 −1
c1+2ν
k −2r(1+2ν) −2r(2ν2 −1)
1
k
(λ + c2 k −2r )1+2ν
= c1+2ν
1
∞
X
j=1
≤ c1+2ν
1
= c1+2ν
1
k −2r(2ν2 −1)
(λk 2r + c2 )1+2ν
!
x−2r(2ν2 −1)
dx
(λx2r + c2 )1+2ν
1
!
Z ∞
−2r(2ν2 −1)
y
1
+ λ2ν2 −1−1/2r
dy .
2r
1+2ν
c2
λ1/2r (y + c2 )
1
+
c2
Z
∞
For the integral to be finite, it is enough if 2r(2ν2 + 2ν) ≥ 1 + δ, for some δ > 0, as the integrand
will go to zero faster than y −(1+δ) . The argument for the lower bound follows the same arguments.
26
References
A. Berlinet and C. Thomas-Agnan. Reproducing Kernel Hilbert spaces in Probability and Statistics.
Springer Science & Business Media, 2011.
T. T. Cai and M. Yuan. Minimax and adaptive prediction for functional linear regression. Journal
of the American Statistical Association, 107(499):1201–1216, 2012.
Y. Fan, G. M. James, and P. Radchenko. Functional additive regression. The Annals of Statistics,
43(5):2296–2325, 2015.
R. Gabrys, L. Horváth, and P. Kokoszka. Tests for error correlation in the functional linear model.
Journal of the American Statistical Association, 105(491):1113–1125, 2010.
T. Hastie and R. Tibshirani. Generalized additive models. Wiley Online Library, 1990.
L. Horváth and P. Kokoszka. Inference for functional data with applications, volume 200. Springer,
2012.
L. Horváth, P. Kokoszka, and G. Rice. Testing stationarity of functional time series. Journal of
Econometrics, 179(1):66–82, 2014.
G. M. James and B. W. Silverman. Functional adaptive model estimation. Journal of the American
Statistical Association, 100(470):565–576, 2005.
H. Kadri, E. Duflos, P. Preux, S. Canu, and M. Davy. In Y. W. Teh and M. Titterington, editors,
Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics,
volume 9 of Proceedings of Machine Learning Research, pages 374–380, 2010.
R. A. Kennedy and P. Sadeghi. Hilbert Space Methods in Signal Processing. Cambridge University
Press, 2013.
J. S. Kim, A.-M. Staicu, A. Maity, R. J. Carroll, and D. Ruppert. Additive function-on-function
regression. Journal of Computational and Graphical Statistics, (Accepted), 2017.
G. S. Kimeldorf and G. Wahba. Some results on Tchebycheffian spline functions. Journal of
Mathematical Analysis and Applications, 33:82–95, 1971.
P. Kokoszka and M. Reimherr. Predictability of shapes of intraday price curves. The Econometrics
Journal, 16(3):285–308, 2013.
27
P. Kokoszka and M. Reimherr. Introduction to Functional Data Analysis. CRC Press, 2017.
H. Lian. Nonlinear functional models for functional responses in reproducing kernel Hilbert spaces.
Canadian Journal of Statistics, 35(4):597–606, 2007.
H. Ma and Z. Zhu. Continuously dynamic additive models for functional data. Journal of Multivariate Analysis, 150:1–13, 2016.
J. S. Morris. Functional regression. Annual Review of Statistics and Its Application, 2:321–359,
2015.
H.-G. Müller, Y. Wu, and F. Yao. Continuously additive models for nonlinear functional regression.
Biometrika, pages 607–622, 2013.
J. G. Nicholas. Market neutral investing. Bloomberg Press Princeton, NJ, 2000.
A. Parodi and M. Reimherr. FLAME: Simultaneous variable selection and smoothing for functionon-scalar regression. Technical report, Pennsylvania State University, 2017.
C. Preda. Regression models for functional data by reproducing kernel Hilbert spaces methods.
Journal of Statistical Planning and Inference, 137(3):829–840, 2007.
J. O. Ramsay and B. Silverman. Functional data analysis. Wiley Online Library, 2006.
J. O. Ramsay, G. Hooker, and S. Graves. Functional Data Analysis with R and MATLAB. Springer
Science & Business Media, 2009.
F. Scheipl, A.-M. Staicu, and S. Greven. Functional additive mixed models. Journal of Computational and Graphical Statistics, 24(2):477–501, 2015.
X. Sun, P. Du, X. Wang, and P. Ma. Optimal penalized function-on-function regression under a
reproducing kernel Hilbert space framework. Journal of the American Statistical Association,
page Accepted, 2017.
K. G. van den Boogaart. tensorA: Advanced tensors arithmetic with named indices. R package
version 0.31, URL http://CRAN. R-project. org/package= tensorA, 2007.
X. Wang and D. Ruppert. Optimal prediction in an additive functional model. Statistica Sinica,
25:567–589, 2015.
28
H. Zhu, F. Yao, and H. H. Zhang. Structured functional additive regression in reproducing kernel
Hilbert spaces. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76
(3):581–603, 2014.
29
| 10math.ST
|
Positioning via Direct Localization in C-RAN Systems
Seongah Jeong1, Osvaldo Simeone2 , Alexander Haimovich2, Joonhyuk Kang3,*
arXiv:1511.07637v3 [cs.IT] 18 May 2017
1
School of Engineering and Applied Sciences (SEAS), Harvard university, 29 Oxford street, Cambridge, MA 02138, USA
2
Department of Electrical & Computer Engineering, New Jersey Institute of Technology (NJIT),
Newark, NJ 07102, USA
3
Department of Electrical Engineering, Korea Advanced Institute of Science and Technology
(KAIST), 291, Daehak-ro, Yuseong-gu, Daejeon 34141, Korea
*
[email protected]
Abstract: Cloud Radio Access Network (C-RAN) is a prominent architecture for 5G wireless
cellular system that is based on the centralization of baseband processing for multiple distributed
radio units (RUs) at a control unit (CU). In this work, it is proposed to leverage the C-RAN architecture to enable the implementation of direct localization of the position of mobile devices
from the received signals at distributed RUs. With ideal connections between the CU and the RUs,
direct localization is known to outperform traditional indirect localization, whereby the location of
a source is estimated from intermediary parameters estimated at the RUs. However, in a C-RAN
system with capacity limited fronthaul links, the advantage of direct localization may be offset
by the distortion caused by the quantization of the received signal at the RUs. In this paper, the
performance of direct localization is studied by accounting for the effect of fronthaul quantization
with or without dithering. An approximate Maximum Likelihood (ML) localization is developed.
Then, the Cramér-Rao Bound (CRB) on the squared position error (SPE) of direct localization
with quantized observations is derived. Finally, the performance of indirect localization and direct
localization with or without dithering is compared via numerical results.
1.
Introduction
Positioning of radio frequency (RF) sources by nodes of a wireless network is currently an active
research area due to its relevance in Global Positioning System (GPS)-denied environments, such
as military applications, indoors localization, disaster response, emergency relief, surveillance and
tactical systems [1]-[12]. The importance of the topic is attested by the recent Federal Communications Commission (FCC) specification for indoor positioning accuracy [13] and by the inclusion
of various localization techniques in the Long Term Evolution (LTE) standard [14]. The traditional
1
approach to source localization has been indirect, or two-step, localization. Accordingly, a set of
distributed base stations (BSs) estimates position-related parameters, such as time of arrival (TOA)
[1]-[5], [7], time difference of arrival (TDOA) [1], [3], angle of arrival (AOA) [1], [4], [6]-[11] or
received signal strength (RSS) [4], [12]. The estimated parameters are then transmitted to a control unit (CU) that determines the source’s position. For instance, in LTE, positioning is based on
TDOA in the uplink, which is referred to as uplink TDOA (UTDOA) [14].
Cloud radio access network (C-RAN), a cloud-based network architecture, has emerged as a
leading technology for 5G and beyond [15]. In C-RANs, BSs are connected to a CU by means of
so called fronthaul links, as shown in Fig. 1. The fronthaul links carry sampled and quantized baseband signals to and from the CU. Thanks to the fronthaul links, the CU can carry out all baseband
processing on behalf of the connected BSs. Since the BSs implement only radio functionalities,
they are typically referred to as radio units (RUs). C-RANs hence enable the joint processing at
the CU of the signals received by the RUs, making it possible to implement direct, or one-step, localization [1], [3], [7]-[11], whereby the source’s position is determined directly from the received
signals at the RUs. The concept of direct localization was first introduced by Wax and Kailath
[10] in the 1980s. However, it was only recently that efficient algorithms were proposed for its
implementation [1].
With ideal fronthaul links, direct localization can generally outperform indirect localization [1],
[7]-[9]. However, in a C-RAN, due to the limited capacity of the fronthaul links, the baseband
signals received by the RUs are quantized prior to transmission to the CU. In this work, we study
the performance of direct localization by accounting for the distortion caused by fronthaul quantization. While in [1], [7]-[11], direct localization in C-RAN systems was studied by assuming
ideal compression within an information-theoretic framework, here we consider a more practical
2
scenario in which the RUs perform scalar quantization as in state-of-the-art C-RAN systems.
In particular, in considered C-RAN systems, fronthaul transmission is based on scalar quantization that allows for an arbitrary additive dither. Dithering, that is the addition of a random
offset, known at the receiver, prior to quantization, is known to provide potential gains over conventional quantization [16], [17]. In the context of distributed detection, as for the C-RAN systems
at hand, the performance gain of dithering can be ascribed to the capability to effectively enhance
the resolution of the observation at the CU. To see this, consider the case in which each RU has
a one-bit threshold quantizer per quadrature component. The use of independent random dithering at the RUs allows each RU to effectively use a different, and randomly selected, threshold for
quantization, hence improving the resolution of the effective quantizer between the RUs and the
CU.
The main contributions of this paper are summarized as follows.
• To the best of our knowledge, this is the first work in the literature, except our prior work [3],
to analyze direct localization in C-RAN with limited capacity links. Unlike the informationtheoretic analysis of [3], here we assume practical fronthaul processing based on scalar quantization, enabling also dithering.
• We derive an approximate Maximum Likelihood (ML) solution for direct localization in CRAN system.
• We derive the Cramér-Rao Bound (CRB) on the accuracy of direct localization by taking into
account the impact of quantization. The analysis allows us to analytically quantify the performance loss caused by the fronthaul quantization. We also show that for asymptotically large
fronthaul capacity, that CRB tends to that of direct localization from unquantized received
signals.
3
• The performances of the proposed direct localization schemes with and without dithering are
compared with direct and indirect localization via ideal fronthaul links as well as with the
relevant CRBs in C-RAN systems, via numerical results.
In the following, after introducing the system and signal models, we shortly overview indirect
localization in Section 2. Novel direct localization schemes based on approximate ML are proposed for the C-RAN system in Section 3, and the analytical performance of direct localization
via CRB is derived in Section 4. Numerical examples and concluding remarks are presented in
Section 5 and Section 6, respectively.
2.
System, Signal Models and Preliminaries
In this section, we describe system and signal models, and review indirect localization.
2.1.
System Model
We consider the C-RAN system illustrated in Fig. 1, which consists of a single active source, Nr
distributed RUs and a CU. The RUs may consist of different types of infrastructure nodes such
as macro/femto/pico base stations, relay stations or distributed antennas, and the source may be
a mobile device. Each RU is equipped with an M-element antenna array, while the source is a
single-antenna node. The set of RUs, denoted as Nr = {1, . . . , Nr }, is located within a D × D
square area, and all the RUs are assumed to be synchronous (e.g., via GPS). The source is located
at a position p = [x y]T , which is known a priori to lie in a given region A contained within overall
square region. Each RU j ∈ Nr is located at a position p j = [xj yj ]T , and the positions of all
RUs are assumed to be known to the CU. The distance and angle between the source and RU j are
defined as
dj (pp) = kpp − p j k,
y − yj
−1
φj (pp) = tan
,
x − xj
4
(1a)
(1b)
D
đ
đ
A
p [ x y ]T
D
đ
M
đ
pj
[ x j y j ]T
Bj
đ
Control unit
(CU)
: Radio unit (RU)
: Source
: Fronthaul
Fig. 1. Illustration of the considered C-RAN system used for localization.
respectively. We assume far-field conditions, namely that the array aperture is much smaller than
the distance dj (pp).
The RUs are connected to the CU via finite-capacity fronthaul links, and the CU aims at locating
the source based on quantized data transmitted from the RUs to the CU over the fronthaul links.
Specifically, each RU j sends a message at a rate of Bj bps/Hz to the CU, where the normalization
is with respect to the bandwidth of the signal transmitted by the source, as we will further discuss
below. Note that the fronthaul links can be either wireless, e.g., a microwave link, or wired, e.g., a
coaxial cable or a fiber optics link.
2.2.
Signal Model
The source transmits a scaled version of a baseband waveform s(t) with unit energy and single
sided bandwidth W Hz at an unknown time t0 . The waveform s(t) is known to the CU and it may
represent a training or synchronization signal. During the observation time T , the M × 1 received
5
signal vector at the RU j ∈ Nr is hence expressed as
r j (t) = bj α j (pp)s(t − τj (pp) − t0 ) + z j (t), 0 ≤ t ≤ T,
(2)
where bj is an unknown complex scalar coefficient whose average power E[|bj |2 ] is only available
at the CU and that describes the channel attenuation from the source to RU j; α j (pp) is the antenna
array response of RU j, which, assuming a uniform linear antenna array (ULA) at each RU, is given
√
by α j (pp) = 1/ M[1 e−i2π∆A cos φj (pp)/λ · · · e−i2π(M −1)∆A cos φj (pp)/λ ]T , with the wavelength of the
signal λ and the antenna separation ∆A ; τj (pp) is the propagation delay along the path between the
source and RU j, which depends on the source’s position as τj (pp) = dj (pp)/c, with the propagation
speed c; and z j (t) represents the contribution of noise and interference whose components are
assumed to a zero-mean, white, complex Gaussian stochastic process with power σj2 . We assume
that the source moves slowly enough and that the observation time T is sufficiently long so that the
inequality T ≫ maxj∈Nr τj (pp) + t0 is satisfied.
After sampling at Nyquist rate, the discrete-time received signal (2) is given as
r j [n] , r j (nTs ) = bj α j (pp)sj [n] + z j [n],
0 ≤ n ≤ Ns − 1,
(3)
where Ts = 1/2W is the sampling period; Ns = ⌈T /Ts ⌉ is the number of time samples; and
the sampled versions of the signal and the noise are denoted sj [n] , s(nTs − τj (pp) − t0 ) and
z j [n] , z j (nTs ), respectively. Taking the discrete Fourier transform (DFT) of the discrete signal
(3), the received signal is expressed in the frequency domain as
R j (k) = bj α j (pp)S(k)e−iwk (τj (pp)+t0 ) + Z j (k),
0 ≤ k ≤ Ns − 1,
(4)
where wk = 2πk/(Ns Ts ); and R j (k), S(k) and Z j (k) indicate the DFT coefficients of the respective time-domain signals r j [n], s(nTs ), and z j [n]. Given the assumptions above, we have the
6
equality
PNs −1
k=0
|S(k)|2 = 1, and the DFT coefficients of the noise Z j (k) are uncorrelated and
have the power σj2 , for 0 ≤ k ≤ Ns − 1. We observe that, in (4), the information about the source’s
location is embedded in two ways in the received signal, namely in time delay τj (pp) and in the
array response α j (pp).
The RUs quantize and forward the received signals in (4) to the CU for positioning under the
fronthaul constraint. As a practical solution, prescribed, e.g., by the CPRI standard [19], we assume
that each RU applies the same uniform quantizer for all frequencies and antennas. To evaluate the
localization accuracy, we adopt the SPE [2], [3], [18]
ρ = E kp̂p − p k2 ,
(5)
as the performance criterion of interest, where p̂p is the estimated source’s position.
2.3.
Indirect Localization in C-RAN
In the conventional, indirect localization method, each RU j estimates the parameters τj (pp) + t0
and φj (pp), which are referred to as TOA and AOA measurements, respectively. Subsequently,
these estimates are sent to the CU, which locates the source, for example, by maximum likelihood
based on TOA and AOA estimates, where the dependence on the unknown parameter t0 can be
eliminated. Note that, under fronthaul capacity constraints, the TOA and AOA measurements at
RU j should be quantized prior to transmission to the CU. Nevertheless, assuming that the number
of samples Ns is large enough, the number of bits Ns Bj available for quantization of the estimates
is large enough to make the corresponding distortion due to quantization negligible.
3.
Direct Localization in C-RAN
Here, we derive an approximate ML solution for direct localization in C-RAN system with limited
fronthaul links, for which a practical fronthaul processing is considered based on scalar quantiza7
tion with dithering. The approximation is taken so as to account for the effect of scalar quantization
and for the optimal solution of the likelihood function, whose details are discussed as follows.
With direct localization, the position of the source is estimated directly from the signals at the
RUs [1], [3], [7]-[9]. Due to the fronthaul capacity limitation of the C-RAN system, we assume
that each RU j sends the quantized version of the received signal R j (k) in (4) to the CU over
limited-capacity fronthaul links with Bj bits per sample. The CU performs localization based
on the quantized signals. Specifically, as discussed below, we adopt a standard uniform dithered
quantization [16], [17], where the dither is known to the CU. The addition of a dither signal is
known to potentially improve the accuracy of a quantized signal as discussed in Section 1.
Scalar quantization at RU j operates on the dithered signal
R j (k) + D j (k),
(6)
where D j (k) is a dither signal, independent of the received signal R j (k). Quantization uses
Bj /2M bits, or equivalently Lj = 2Bj /2M quantization levels, for either the real or imaginary parts
of the dithered signal (6) at any of the M received antennas. The dynamic range [−Rjmax , Rjmax ]
of the quantizer applied at each RU j is fixed, and may be based on preliminary Monte Carlo
experiments aimed at guaranteeing that the peak-to-peak interval includes the received signals
Rj (k)} with high probability, when considering a given random distribution of the source, and
{R
of the channel coefficients, as well as the distribution of the noise (see also Section 5). Each real
R j (k) in the set of
and imaginary entry of the dithered signal (6) is quantized to the closest value R̂
representation points {−Rjmax + (l − 1)∆j } for l = 1, . . . , Lj , where ∆j = 2Rjmax /(Lj − 1) is the
step size of the quantizer. We take the dithers D j (k) to be independent and uniformly distributed
in the interval [−∆j /Ij , ∆j /Ij ], where Ij is a parameter that can be chosen to minimize the SPE
(5).
8
Rj (k)} of the dithered received signal, i.e., of {R
Rj (k)+
The CU recovers the quantized versions {R̂
D j (k)}, from all RU j ∈ Nr . Then, it subtracts the dither to obtain
Rj (k) = R̂
Rj (k) − D j (k).
R̃
(7)
Rj (k)}
Differently from [1], [7]-[9], the estimator under consideration is applied to the signals {R̃
Rj (k)} due to the limited fronthaul capacity. We are
in (7), rather than to the received signals {R
Rj (k)}
interested in approximating the ML estimate of the position p based on the received signal {R̃
in (7).
To this end, we account for the effect of quantization as an additive Gaussian noise term, as in
Rj (k) = R j (k) + Q j (k),
R̃
(8)
where Q j (k) represents the quantization noise whose components have the zero-mean Gaussian
2
distribution with the variance σq,j
(k). In this way, from (4), the effective Gaussian noise on the
Rj (k)} of RU j is given by Z j (k) + Q j (k), whose entries are distributed i.i.d. as
observation {R̃
2
zero-mean complex Gaussian variables with variance γj2 (k) = σj2 (k) + σq,j
(k).
With reference to (7), an approximate ML estimate of the unknown parameters ({bj }, t0 , p ) is
hence obtained by solving the problem
minimize
Nr
X
Cj (bj , t0 , p ),
(9)
j=1
with
Cj (bj , t0 , p ) =
N
s −1
X
k=0
1
Rj (k) − bj α j (pp)S(k)e−iwk (τj (pp)+t0 )
R̃
2
γj (k)
2
(10)
over the position p , the transmit time t0 and the channel attenuation parameters {bj }.
2
2
To evaluate the power of the quantization noise σq,j
(k), and hence γj2 (k) = σj2 (k) + σq,j
(k),
as a function of the fronthaul capacity Bj , we leverage arguments from rate-distortion theory [20].
9
Note that this provides an approximation of the effect of quantization since scalar quantization is
adopted here, as opposed to the vector quantization with arbitrarily large blocks that is assumed in
2
rate-distortion theory. This approximation allows us to compute σq,j
(k) as the value that guarantees
the equality [21, eq. (10.11)]
Rj (k)]m ; [R̃
Rj (k)]m ) =
I([R
Bj
,
M
(11)
Rj (k)]m ; [R̃
Rj (k)]m ) is the mutual information between the input of the quantizer and its
where I([R
Rj (k)]m to be distributed
output, after subtraction of the dither. In order to evaluate (11), we take [R
as CN (0, E[|bj |2 ]|S(k)|2 + σj2 ), hence treating the parameter bj as complex Gaussian with zero
mean and power E[|bj |2 ], i.e., Rayleigh fading. The net effect of these choices is that the weights
1/γj2 (k) in (10) are given by
γj2 (k) = σj2 +
Rj (k)]m |2 ]
E[|[R
2
Bj
M
−1
= σj2 +
E[|bj |2 ]|S(k)|2 + σj2
2
Bj
M
.
(12)
−1
As per (12), the weight given to each RU j in (10) increases with the capacity Bj of its fronthaul
link.
The minimization (9) is first solved over the amplitude parameter bj for fixed (t0 , p ), yielding
b∗j (t0 , p )
= argmin
bj
=
=
Nr
X
Cj (bj , t0 , p )
j=1
1
γ
H γ
S j (t0 , p ) ⊗ α j (pp) R̃
Rj ,
S̄
2
γ
S j (t0 , p )
S̄
1
γ
S j (t0 , p )
S̄
α j (pp)k2
kα
γ
H γ
p
Rj ,
S
R̃
2 S̄ j (t0 , p ) ⊗ α j (p )
(13)
γ
γ
Rj =
S j (t0 , p ) = [S(0)e−iw0 (τj (pp)+t0 ) /γj (0) · · · S(Ns − 1)e−iwNs −1 (τj (pp)+t0 ) /γj (Ns − 1)]T ; R̃
where S̄
T
T
Rj (0)/γj (0) · · · R̃
Rj (Ns − 1)/γj (Ns − 1)]T ; and the last equality can be obtained by recalling that
[R̃
α j (pp)k2 = 1 for j ∈ Nr . By substituting (13) in (10) and simplifying the notation as
we have kα
γ
γ
S j k2 = kS̄
S j (t0 , p )k2 =
kS̄
PNs −1
k=0
|S(k)|2 /γj2 (k), we obtain that the minimization (9) is equivalent
10
to the problem
maximize C̃(t0 , p ),
(14)
with
Nr
s −1 iw (τj (p
X
p) NX
αH
e k p)+t0 ) S ∗ (k)
j (p
p
R(k)
C̃(t0 , ) =
R̃
γ
2
γ
(k)
S
S̄
j
j
j=1
k=0
=
=
S H U (pp)
z H (t0 )S
Nr
X
2
2
(15a)
(15b)
2
V j (pp) ,
z H (t0 )V
(15c)
j=1
γ
u1 (pp)/kS̄
S1k
where z (t0 ) = [1 e−iw1 t0 · · · e−iwNs −1 t0 ]T ; S = diag{S(0), . . . , S(Ns −1)}; and U (pp) = [u
γ
p)R̃
Rj (0)/γj2 (0)
S Nr k], with u j (pp) = [eiw0 τj (pp) α H
· · · u Nr (pp)/kS̄
j (p
p)R̃
R j (Ns −1)
· · · eiwNs −1 τj (pp)α H
j (p
/γj2 (Ns − 1)]T ; and V j (pp) is the jth column of V (pp) = S H U (pp).
Since the vector z (t0 ) in (15c) has the structure of a column of a Ns -point Discrete Fourier
Transform (DFT), the evaluation of (15c) for different potential value of t0 can be computed by
means of Fast Fourier Transform (FFT), allowing us to efficiently estimate t0 in the following way
[1], [7]. Define a grid of possible values for t0 as {0, 1, . . . , qt0 Ns − 1}Ts /(qt0 Ns ), where qt0 is an
integer that determines the resolution of the estimate of t0 . Apply the qt0 Ns -point FFT matrix F to
F ]k,n = e−i2π(k−1)(n−1)/(qt0 Ns ) for 1 ≤ k, n ≤ qt0 Ns , with
each column V j (pp) for j ∈ Nr , where [F
A]k,n defining the (k, n)th element of matrix A . Note that, when qt0 ≥ 2, each vector V j (pp) is
[A
padded with trailing zeros to length qt0 Ns . Then, the maximization (14) can be carried out jointly
over t0 and p , yielding the position estimate p ∗ , by solving the following problem
"N
#
r
X
F V j (pp)|2
p ∗ = argmax
max
|F
,
p 1≤k≤qt0 Ns
j=1
(16)
k
where [aa]k is the kth element of vector a ; and | · | represents the entry-wise absolute value of its
argument.
11
In summary, the proposed solution is referred to an approximate ML because (i) the effect of
scalar quantization in approximated by a Gaussian noise term in (8), whose variance is obtained by
means of rate-distortion arguments; (ii) the optimization over t0 is obtained by means of the grid
search described above.
4.
CRB for Direct Localization
In this section, we derive the CRB [18] on the SPE (5) for the direct localization. The CRB in the
presence of dithering generally depends on the distribution of the dither and its calculation appears
to be intractable. Nevertheless, in the high SNR regime, where the CRB becomes tight, the impact
of dithering is less pronounced. For these reasons, we concentrate here on the derivation of the
CRB for standard quantization without dithering. The CRB provides a lower bound on the SPE (5)
as
ρ ≥ tr J −1 (pp) ,
(17)
where J (pp) is the Equivalent Fisher Information Matrix (EFIM) [2], [3] for the estimation of the
source’s position p .
To elaborate, we denote the uniform quantization function at RU j as Qj (x) = {l|qj,l−1(∆j ) <
x ≤ qj,l (∆j ); l = {1, . . . , Lj }}, where qj,l (∆j ) = −Rjmax + (l − 0.5)∆j is the lth quantization
threshold for the RU j with qj,0 = −∞ and qj,Lj = ∞. The signal S = {S(k)} is known to the
CU, and we also assume that the transmit time t0 is known in order to obtain a lower bound on the
SPE. The probability of the quantized signals available at the CU for a given unknown parameter
ℑ T
vector θ = [ppT bT1 · · · bTNr ]T with bj = [bℜ
j bj ] is then given as
R; θ ) =
P (R̂
M
Nr NY
s −1 Y
Y
Lj
Y Y
ζ
Rj (k)]m )−l)
R ζj (k)]m ); θ j )δ(Qj ([R
Pj,l (Qj ([R
,
j=1 k=0 m=1 ζ∈{ℜ,ℑ} l=1
12
(18)
T
T
T
T
Rj (Ns − 1)]T ; δ(·) is the Kronecker-delta funcRj = [R̂
Rj (0) · · · R̂
RNr ]T with R̂
R = [R̂
R1 · · · R̂
where R̂
Rζj (k)]m ); θ j ) is the probability that
tion, i.e., δ(0) = 1 and δ(x) = 0 for x 6= 0; and Pj,l (Qj ([R
Rℜ
Rℑ
either the real or imaginary part of the received signal, [R
j (k)]m or [R
j (k)]m , for the mth antenna
Rζj (k)]m ) = l,
at the RU j takes on a specific value l among the Lj quantization levels, i.e., Qj ([R
given as
Rζj (k)]m ); θ j )
Pj,l(Qj ([R
=Φ
ζ
qj,l (∆j ) − fj,k,m
(θθ j )
√
σj / 2
!
−Φ
ζ
qj,l−1(∆j ) − fj,k,m
(θθ j )
√
σj / 2
!
,
(19)
with ζ being the parameter to indicate the real and imaginary part of the signal as ζ ∈ {ℜ, ℑ} and
θ j = [ppT b Tj ]T . In (19), Φ(·) is the complementary cumulative distribution function of the standard
Gaussian distribution Φ(x) =
Rx
2 /2
e−t
−∞
√
/ 2πdt, and we have defined the noiseless received
α j (pp)]m S(k)e−iwk (τj (pp)+t0 ) .
signal (see (4)) as fj,k,m(θθ j ) = bj [α
R; θ ), and the FIM J (θθ ) for the unknown
The log-likelihood function is given by L(θθ ) = ln P (R̂
T
θ )] [18]. The CRB in (17) can be then
parameter vector θ can be written as J (θθ ) = −ER̂
R [∇θ ∇θ L(θ
obtained from the EFIM for the location p , which is given as
J (pp) = X − Y Z −1Y T ,
(20)
J (θθ )](1:2,1:2) ; Y = [J
J (θθ )](1:2,3:2Nr +2) ; and Z = [J
J (θθ )](3:2Nr +2,3:2Nr +2) , with [A
A](a:b,c:d)
where X = [J
being the sub-matrix of A corresponding to from the ath to the bth rows and from the cth to
the dth columns. In Appendix 9.1, we calculate these matrices as X =
P Nr
T
j=1 U j Ψ j U j ;
Y =
V Ψ 1V T , . . . , V ΨNr V T }, where V = [002 02 I 2 ]; U j =
U 1Ψ1V T · · · U Nr ΨNr V T ]; and Z = diag{V
[U
[cos φj (pp)/c − sin φj (pp)/dj (pp) 0 T2 ; sin φj (pp)/c cos φj (pp)/dj (pp) 0 T2 ]; and the matrix Ψ j repre-
13
sents the FIM of RU j with respect to θ̃θ j = [τj (pp) φj (pp) b Tj ]T and is defined as
Ψj ]p,q =
[Ψ
2
ζ
ζ
ζ
ζ
X Γj,k,m,l − Γj,k,m,l−1 ∇[θ̃θ j ]p fj,k,m(θθ j )∇T[θ̃θ j ]q fj,k,m(θθ j )
Rζj (k)]m ); θ j )
πσj2 Pj,l (Qj ([R
k,m,ζ,l
,
(21)
for 1 ≤ p, q ≤ 4, with being zero if the corresponding parameters are orthogonal,
−
ζ
θ j ))
(θ
(qj,l (∆j )−fj,k,m
σ2
j
Γζj,k,m,l = e
2
,
(22)
and the required derivatives
ℜ
∇τj (pp) fj,k,m
(θθ j ) = Re{−iwk fj,k,m(θθ j )}
ℑ
∇τj (pp) fj,k,m
(θθ j ) = Im{−iwk fj,k,m(θθ j )}
i2π∆A sin φj (pp)(m − 1)fj,k,m(θθ j )
ℜ
∇φj (pp) fj,k,m(θθ j ) = Re
λ
p
i2π∆A sin φj (p )(m − 1)fj,k,m(θθ j )
ℑ
∇φj (pp) fj,k,m(θθ j ) = Im
λ
ℜ
∇bℜj fj,k,m
(θθ j ) = Re{[αj (pp)]m S(k)e−iwk (τj (pp+t0 )) }
ℑ
∇bℜj fj,k,m
(θθ j ) = Im{[αj (pp)]m S(k)e−iwk (τj (pp+t0 )) }
ℜ
∇bℑj fj,k,m
(θθ j ) = −Im{[αj (pp)]m (k)e−iwk (τj (pp+t0 )) }
ℑ
∇bℑj fj,k,m
(θθ j ) = Re{[αj (pp)]m S(k)e−iwk (τj (pp+t0 )) }.
(23)
It is noted that the CRB in (17) with the EFIM in (20) is increased by the presence of quantization as compared to the CRB of direct localization with unquantized signals derived in [8], [9],
which is defined with the same EFIM J (pp) in (20) except for the different FIM Ψ j of each RU j
calculated with the nonzero elements for the non-orthogonal parameters as
Ψj ]p,q =
[Ψ
X 2
ζ
∇ f ζ (θθ )∇T[θ̃θ j ]q fj,k,m
(θθ j ).
2 [θ̃θ j ]p j,k,m j
σ
j
k,m,ζ
(24)
We can compare the FIM Ψ j of RU j for quantized signals in (21) and for unquantized signals
14
in (24) that appear in the EFIM J (pp), in the low SNR regime, where the power of channel noise
is much larger than that of the desired signal, i.e., σj ≫ |fj,k,m(θθ j )|. For this purpose, we assume
that all the RUs have the same fronthaul capacity constraints Bj = B and that the channel noise
power satisfies σj2 = σ 2 for j ∈ Nr . Thus, we use the same uniform quantization with L = 2B/2M
quantization levels and quantization threshold {ql (∆)} for all RUs, for 1 ≤ l ≤ L. Then, the FIM
Ψ j of RU j in (21) for quantized case can be approximated as
Ψj ]p,q = LQ
[Ψ
X 2
ζ
ζ
(θθ )∇T[θ̃θ j ]q fj,k,m
∇[θ̃θ j ]p fj,k,m
(θθ ),
2
σ
k,m,ζ
(25)
where, by comparison to (24) in the unquantized case, the factor LQ can be interpreted as the
quantization loss of the CRB for quantized signals and is defined as
2
2
q 2 (∆)
q (∆)
− l 2
− l−12
e σ −e σ
L
1 X
≤ 1.
LQ =
ql−1 (∆)
2π l=1 Φ ql (∆)
√
√
− Φ σ/ 2
σ/ 2
(26)
The factor LQ results in a decrease on the Fisher Information at each RU j due to the operation
over quantized signals, and we have that, from (20), the EFIM J (pp) of quantized case is equivalent
to LQ times that of unquantized case. By the expression (17) of the CRB, it can be seen that LQ
directly evaluates the ratio between the CRBs of unquantized and quantized cases at low SNR, so
we can approximately write
LQ ≈
CRBUQ
,
CRBQ
(27)
where CRBUQ and CRBQ are the actual CRBs of direct localization in unquantized and quantized
cases with the EFIM J (pp) being given by (24) and (21), respectively. The approximation (27) will
be validated via numerical results in the next section.
As a final remark, it is noted that, when the fronthaul capacity Bj for j ∈ Nr becomes arbitrarily
large, namely the step size of quantizer ∆j goes to zero and the quantization level Lj approaches
15
to infinity, the CRB for quantized and unquantized signal models, CRBUQ and CRBQ , can be seen
to coincide. This can be proved by the fact that the FIM Ψ j of each RU j ∈ Nr in (21) approaches
to (24), as Bj turns to infinity (see Appendix 9.2).
5.
Numerical Results
In this section, we evaluate the performance of direct localization with the aim of assessing the
impact of quantization. For the indirect localization, as discussed in Section 2.3, the source is
located by ML based on TOA and AOA estimates both of which are obtained by means of the
multiple signal classification (MUSIC) algorithm [5], [6], as in [8], but other super-resolution
techniques are also possible. We consider a network with Nr = 4 RUs placed at the vertices of
a square area with side of length D = 4 km, while the source is uniformly and randomly located
within a 3 × 3 km2 area of A centered in the entire region. Each RU is assumed to be equipped
with an M = 8 element ULA with antenna separation ∆A = λ/2. The sampled waveform is
√
s(n) = sinc(t/Ts )/ Ns |t=nTs with bandwidth W = 1/(2Ts ) = 200 kHz as in GSM, so that the
sampling period is Ts = 2.5 µs, and each location estimation is based on Ns = 8 samples. We
assume that the channel coefficients bj are independent and Rician, with Rician factor K = 20 dB.
Also, we assume sensors with equal noise variance. The dynamic range of the uniform quantizer
applied at each RU j to all antennas is set based on Monte Carlo experiments guaranteeing that
the received signals are included in the peak-to-peak interval with the probability of 95% given
conditions.
We impose the fronthaul capacity constraints as B1 = B2 = B3 = B and B4 = 2B, and
we set qt0 = 1 in the direct localization scheme for estimating t0 (see Section 3). We evaluate
qP
N
the root mean squared (RMS) error as performance metric defined as
n=1 ρn /N, where ρn =
kp̂pn − p k2 with N being the number of experiments and p̂pn being the estimated source location
16
65
Indirect (Ideal)
Direct (Ideal)
Direct (w/o Dithering)
Direct (w/ Dithering)
Direct, CRB (Ideal)
Direct, CRB (w/o Dithering)
60
RMS error (m)
55
50
45
40
35
30
25
2
3
4
5
6
7
B/M (bps/Hz/antenna)
8
9
10
Fig. 2. RMS error versus fronthaul capacity B/M bps/Hz/antenna for SNR (per antenna) = 0 dB
(Nr = 4, M = 8, Ns = 8 and Ts = 2.5 µs).
of the nth experiments. For reference, we also present the performance of direct localization with
ideal fronthaul links and the CRBs for quantized and unquantized signals derived in Section 4.
Fig. 2 and Fig. 3 show the RMS error versus the fronthaul capacity constraint B/M (bps/Hz
/antenna) when SNR per antenna is 0 dB and 5 dB, respectively. First, we observe that, when
the SNR is sufficiently large, as in Fig. 3 (SNR= 5 dB), indirect localization may outperform
direct localization if the fronthaul capacity is small enough. This is due to the distortion caused
by fronthaul quantization when the quantization is coarse. Note that this is not the case at lower
SNR, where at SNR= 0 dB, the signal degradation due to the additive noise overwhelms the loss
due to quantization. Furthermore, as long as the fronthaul capacity is large enough, as expected,
direct localization has the potential to significantly outperform indirect localization, with additional
marginal gains achievable via dithering. Also, the gain of dithering is more relevant at lower
fronthaul capacity. For instance, for SNR= 0 dB, the gain of dithering (reduction in localization
17
28
26
Indirect (Ideal)
Direct (Ideal)
Direct (w/o Dithering)
Direct (w/ Dithering)
Direct, CRB (Ideal)
Direct, CRB (w/o Dithering)
RMS error (m)
24
22
20
18
16
2
3
4
5
6
7
B/M (bps/Hz/antenna)
8
9
10
Fig. 3. RMS error versus fronthaul capacity B/M bps/Hz/antenna for SNR (per antenna) = 5 dB
(Nr = 4, M = 8, Ns = 8 and Ts = 2.5 µs).
RMS) increases from 8.5% at B/M = 4 to 13.3% at B/M = 2. Finally, the CRB, as well as the
performance with ideal fronthaul, is approached by the direct scheme as the fronthaul capacity and
the SNR increase.
Next, we turn to the evaluation of the performance loss of direct localization due to quantization
by means of the analysis in Section 4. Specifically, we compare the quantization loss LQ in (26)
to the actual ratio between the CRBs of the unquantized and quantized cases, i.e., CRBUQ /CRBQ ,
obtained by numerical simulations using the same parameters as discussed above, except that we
impose an equal fronthaul capacity constraint Bj = B, and use the same uniform quantizer for all
RU j ∈ Nr .
Fig. 4 shows the discussed ratio between CRBs versus SNR per antenna for B/M = 4
bps/Hz/antenna. It is seen that, the factor LQ provides a close approximation of the actual performance loss, as expected in (27), not only in the low-SNR regime in which it was derived. In
18
0.86
0.84
Actual Ratio
LQ
0.82
CRBUQ/CRBQ
0.8
0.78
0.76
0.74
0.72
0.7
0.68
0.66
−5
−3
0
SNR (dB)
3
5
Fig. 4. Ratio between CRBs for unquantized and quantized observations versus SNR (per antenna)
(dB) for B/M = 4 bps/Hz/antenna (Nr = 4, M = 8, Ns = 8 and Ts = 2.5 µs).
particular, the analytical loss (26) properly quantifies the quantization loss due to quantization,
which becomes significant at the higher SNR.
6.
Concluding Remarks
The C-RAN architecture, with its centralized baseband processing at a control unit (CU), enables
the implementation of direct, or one-step, localization. While direct localization outperforms the
traditional indirect, or two-step, localization in the presence of ideal connections between the CU
and the radio units (RUs), this may not be the case in the C-RAN due to the distortion caused by
quantization on the fronthaul links. In this paper, we have studied the performance of direct and
indirect localization in the C-RAN by means of the calculation of the CRB as well as quantization
effect on the localization accuracy via analytical and numerical results. Interesting open problems
concern the study of the optimal scalar quantization design for the RUs to maximize the localization
accuracy.
19
7.
Acknowledgments
This work has been supported by the National GNSS Research Center program of Defense Acquisition Program Administration and Agency for Defense Development.
8.
References
[1] Weiss, A. J.: ’Direct position determination of narrowband radio frequency transmitters’, IEEE
Signal Processing Letters, 2004, 11, (5), pp. 513–516
[2] Jeong, S., Simeone, O., Haimovich, A., et al.: ’Beamforming design for joint localization
and data transmission in distributed antenna system’, IEEE Transactions on Vehicular Technology,
2015, 64, (1), pp. 62–76
[3] Jeong, S., Simeone, O., Haimovich, A., et al.: ’Optimal fronthaul quantization for cloud radio
positioning’, IEEE Transactions on Vehicular Technology, 2015, PP, (99), pp. 1–7
[4] Gezici, S., Tian, Z., Giannakis, G. B., et al.: ’Localization via ultra-wideband radios: a look at
positioning aspects for future sensor networks’, IEEE Signal Processing Magazine, 2005, 22, (4),
pp. 70–84
[5] Li, X., Pahlavan, K.: ’Super-resolution TOA estimation with diversity for indoor geolocation’,
IEEE Transactions on Wireless Communications, 2004, 3, (1), pp. 224–234
[6] Schmidt, R. O.: ’Multiple emitter location and signal parameter estimation’, IEEE Transactions
on Antennas and Propagation, 1986, 34, (3), pp. 276–280
[7] Bar-Shalom, O., Weiss. A. J.: ’Transponder-aided single platform geolocation’, IEEE Transactions on Signal Processing, 2013, 61, (5), pp. 1239–1248
20
[8] Weiss, A. J., Amar, A.: ’Direct position determination of multiple radio signals’, Eurasip
Journal Applied Signal Processing, 2005, 2005, (1), pp. 37–49
[9] Amar, A., Weiss, A. J.: ’Direct position determination in the presence of model errors –known
waveforms’, Digital Signal Processing, 2006, 16, (1), pp. 52–83
[10] Wax, M., Kailath, T.: ’Optimum localization of multiple sources by passive arrays’, IEEE
Transactions on Acoustics, Speech and Signal Processing, 1983, 31, (5), pp. 1210–1217
[11] Wax, M., Kailath, T.: ’Decentralized processing in sensor arrays’, IEEE Transactions on
Acoustics, Speech, and Signal Processing, 1985, 33, (5), pp. 1123–1129
[12] Niu, R., Varshney, P. K.: ’Target location estimation in sensor networks with quantized data’,
IEEE Transactions on Signal Processing, 2006, 54, (12), pp. 4519–4528
[13] Hateld, D. N., ’A report on technical and operational issues impacting the provision of wireless
enhanced 911 services’ (Federal Communications Commission, Washington, D.C., 2002), pp. 1–
54. Available at http://www.emergentcomm.com/documents/Supporting Documents/hatfield report1.pdf,
[14] 3GPP TS 36.455, ’E-UTRA: LTE positioning protocol A (LPPa) v12.2.0’, 2015
[15] Chen, K., Duan, R., ’C-RAN: The road towards green RAN’ (China Mobile Research Institute, White Paper v2.5, 2011), http://labs.chinamobile.com/cran/wp-content/uploads/CRAN white
paper v2 5 EN.pdf, pp. 1–48
[16] Schuchman, L.: ’Dither signals and their effect on quantization noise’, IEEE Transactions on
Communication Technology, 1964, 12, (4), pp. 162–165
[17] Gray, R. M., Stockham, T. G.: ’Dithered quantizers’, IEEE Transactions on Information
Theory, 1993, 39, (3), pp. 805–812
21
[18] Kay, S. M.: ’Fundamentals of signal processing-estimation theory’ (Prentice Hall, NJ, 1993)
[19] CPRI: ’CPRI specification v6.1’ (2014), http://www.cpri.info/jp/spec.html
[20] Cover, T. M., Thomas, J. A.: ’Element of information theory’ (John Wiley & Sons, NJ, 2006)
[21] Zhang, F.: ’The Schur complement and its applications’ (Springer, NY, 2005), vol. 4.
[22] Zhang, H., Shen, Y., Chai, L., et al.: ’State dimension reduction and analysis of quantized
estimation systems’, Signal Processing, 2014, 105, pp. 363–375
[23] Olver, F. W. J.: ’NIST handbook of mathematical functions’ (Cambridge University Press,
NY, 2010)
9.
Appendix
9.1.
Calculation of EFIM for Direct Localization based on Quantized Signals
In this appendix, we calculate the EFIM J (pp) of direct localization in (20). Similar to [2], [3],
T
T
since the source is localizable, the mapping of θ = [ppT b T1 · · · b TNr ]T to θ̃θ = [θ̃θ 1 · · · θ̃θ Nr ]T with
T T , where
θ̃θ j = [τj (pp) φj (pp) b Tj ]T is a bijection. Then, we can have the relationship J (θθ ) = T J (θ̃θ )T
T
θ )],
T = ∂θ̃θ /∂θθ = [U
U 1 · · · U Nr ; V 1 · · · V Nr ] is the Jacobian matrix; and J (θ̃θ ) = −ER̂
R [∇θ̃θ ∇θ̃θ L(θ
V j ]2(j−1)+1:2j,3:4 = I 2 .
with U j defined in (20) and V j ∈ ℜ2Nr ×4 having all zeros except for [V
Ψ1 , . . . , Ψ Nr }, and using the relationship
Referring [12], J (θ̃θ ) can be calculated as J (θ̃θ ) = diag{Ψ
J (θθ ) = T J (θ̃θ )T
T T , J (θθ ) can be written as J (θθ ) = [X
X Y ; Y T Z ], where X , Y and Z are given in
(20). Finally, by applying the Schur complement (see, e.g., [21]), we can have the EFIM J (pp) in
(20).
22
9.2.
Proof of Asymptotic Convergence of CRB for Large Fronthaul Capacity
Here, we show that the matrix Ψ j (21) that appears in the CRB for quantized observations approaches the counterpart quantity (24) for the case with unquantized observations as the fronthaul
capacity Bj for j ∈ Nr goes to infinity. This allows us to conclude, as desired, that the CRB with
quantized observations tends to the CRB with unlimited fronthaul as Bj tends to infinity. To this
end, we first prove the limit
2
ζ
ζ
X Γj,k,m,l − Γj,k,m,l−1
Lj
lim
∆j →0,Lj →∞
l=1
Rζj (k)]m ); θ j )
Pj,l (Qj ([R
→ 2π,
(28)
for ∀j, k, m, ζ. For simplicity of the notation, we will omit the subscript k, m and superscript ζ
in the following. The proof is similar to the proof of Theorem 2 in Section 4.2 of [22]. Denote
ζ
˜ j,l = q̃j,l − q̃j,l−1. Also, we define the probability density function
q̃j,l = qj,l (∆j ) − fj,k,m
(θθ j ), and ∆
of the standard normal distribution as φ(x) = e−x
2 /2
√
˜ j,l → 0 when ∆j → 0,
/ 2π. Noting that ∆
we can calculate the limit (28) as
2
q̃j,l−1
q̃j,l
Lj 2π φ
√
√
−
φ
X
σ / 2
σ / 2
j
j
lim
q̃
q̃
˜
j,l
∆j,l →0,Lj →∞
√
Φ σ /√2 − Φ σ j,l−1
l=1
j
j/ 2
q̃
˜ j,l 2
˜ j,l −1
q̃j,l −∆
q̃j,l −∆
q̃j,l
j,l
∞
√
√
√
√
−
φ
−
Φ
φ
Φ
X
˜ j,l
∆
σj / 2
σj / 2
σj / 2
σj / 2
√
=
lim 2π
˜ j,l
˜ j,l
∆
∆
˜ j,l →∞
∆
σ
/
2
√
√
j
l=1
σj / 2
σj / 2
2
−1
∞
X
˜ j,l
∆
d
d Φ(t)
√
lim 2π φ(t)
=
q̃j,l
q̃j,l
˜ j,l →∞
dt
dt
∆
σ
/
2
√
√
j
t=
t=
l=1
σj /
∞
X
2
2
σj / 2
q̃ 2
− j,l
σ2
j
˜ j,l
∆
q̃j,l
√
√
e
˜ j,l →∞
∆
σ
/
2
σ
/
2
j
j
l=1
2
2 q̃j,l
Z √
− 2 dq̃j,l
q̃j,l
√
√
e σj
=
2π
σj / 2
σj / 2
R
Z √
r2
=
2πr 2 e− 2 dr → 2π.
=
lim
√
2π
(29)
R
23
√
The last equation can be obtained by denoting r = q̃j,l /(σj / 2) and by an Euler integral of Gamma
function
R
R
lim
r 2 e−r
∆j →0,Lj →∞
2 /2
√
/ 2πdr = 1 [23]. Given the limit (28), we now have that the limit
ζ
ζ
X X Γj,k,m,l − Γj,k,m,l−1
Lj
k,m,ζ l=1
2
ζ
ζ
(θθ j )∇T[θ̃θ ] fj,k,m
∇[θ̃θ j ]p fj,k,m
(θθ j )
j q
Rζj (k)]m ); θ j )
πσj2 Pj,l (Qj ([R
=
X 2
ζ
∇ f ζ (θθ )∇T[θ̃θ j ]q fj,k,m
(θθ j )
2 [θ̃θ j ]p j,k,m j
σ
k,m,ζ j
(30)
holds, which concludes the proof since (30) shows that the EFIM for the case with quantized
observations tends to that of the case with unquantized observations, as the fronthaul capacity Bj
for j ∈ Nr goes to infinity.
24
| 7cs.IT
|
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
SANG-IL OUM
arXiv:1601.03800v2 [math.CO] 27 Jul 2016
Abstract. Rank-width is a width parameter of graphs describing whether it
is possible to decompose a graph into a tree-like structure by ‘simple’ cuts. This
survey aims to summarize known algorithmic and structural results on rank-width
of graphs.
1. Introduction
Rank-width was introduced by Oum and Seymour [71]. Roughly speaking, the
rank-width of a graph is the minimum integer k such that the graph can be decomposed into a tree-like structure by recursively splitting its vertex set so that each
cut induces a matrix of rank at most k. The precise definition will be presented in
the next section.
Tree-width is a better known and well-studied width parameter of graphs introduced by Robertson and Seymour [74]. Tree-width has been very successful for
deep theoretical study of graph minor structures as well as algorithmic applications.
In particular, many NP-hard problems can be solved efficiently if the input graph
belongs to a class of graphs of bounded tree-width. However, graphs of bounded
tree-width have bounded average degree and therefore the application of tree-width
is mostly limited to ‘sparse’ graph classes.
Clique-width [23, 25] aims to extend tree-width by allowing more ‘dense’ graphs to
have small clique-width. There are many efficient algorithms based on dynamic programming assuming that a clique-width decomposition is given, see [42]. However,
there is no known polynomial-time algorithm to determine whether the clique-width
of an input graph is at most k for fixed k ≥ 4. (There is a polynomial-time algorithm
to decide clique-width at most 3 by Corneil et al. [19].)
Rank-width solves this dilemma in some way; there is a polynomial-time algorithm
to decide whether the rank-width of an input graph is at most k for each fixed k.
In addition, for each graph of rank-width k and clique-width c, the inequality
k ≤ c ≤ 2k+1 − 1
holds [71] and therefore a class of graphs has bounded clique-width if and only
if it has bounded rank-width. Furthermore a rank-decomposition of width k can
be translated into a clique-width decomposition of width at most 2k+1 − 1 [71] so
that we can use known algorithms for graphs of bounded clique-width. Thus, if
we could extend theorems on tree-width to rank-width, then it becomes useful for
some dense graphs. Moreover there is a correspondence between branch-width of
binary matroids and rank-width of bipartite graphs in terms of vertex-minors and
Date: July 28, 2016.
Supported by Basic Science Research Program through the National Research Foundation of
Korea (NRF) funded by the Ministry of Science, ICT & Future Planning (2011-0011653).
1
2
SANG-IL OUM
pivot-minors, discussed in [63]. This correspondence allows us to apply the matroid
theory to prove statements on rank-width, for instance in [26].
In 2008, the author with Hliněný, Seese, and Gottlob wrote a survey [42] on
various width parameters including rank-width. Our (incomplete) survey aims to
supplement that by focusing on algorithmic problems of finding rank-width and
structural aspects related to rank-width.
2. Definitions and equivalent width parameters
All graphs in this survey are assumed to be simple, meaning that they have no
parallel edges and no loops. We focus on finite undirected graphs in this article,
though Kanté and Rao [52] worked on rank-width of directed graphs and edgecolored graphs and Courcelle [21] extended rank-width and linear rank-width to
countable graphs.
In later sections, we discuss and use relations to matroids and branch-width. For
the definition of matroids, please refer to the book of Oxley [73]. Robertson and
Seymour [77] introduced the notion of branch-width for both graphs and matroids.
Key concepts of binary matroids related to rank-width can be found in [63].
2.1. Rank-width. In order to define rank-width, we first need a measure to describe the complexity of a cut, which we call the cut-rank function ρG . For a graph
G and a subset X of the vertex set V (G) of G, let us define ρG (X) to be the rank of
a |X| × |V (G) − X| 0-1 matrix AX over the binary field where the entry of AX on
the i-th row and j-th column is 1 if and only if the i-th vertex in X is adjacent to
the j-th vertex in V (G) − X. (If X = ∅ or X = V (G), then ρG (X)
= 0.)
Note that
1 1 0
this matrix is over the binary field; thus, for instance, the rank of 0 1 1 is equal
1 0 1
to 2, as 1 + 1 = 0. The cut-rank function satisfies the following inequalities.
(1)
(2)
ρG (X) = ρG (V (G) − X) for all X ⊆ V (G),
ρG (X) + ρG (Y ) ≥ ρG (X ∪ Y ) + ρG (X ∩ Y ) for all X, Y ⊆ V (G).
The inequality (2) is called the submodular inequality.
A tree is subcubic if every node has degree 1 or 3. A rank-decomposition of a graph
G is a pair (T, L) of a subcubic tree T with at least two nodes and a bijection L from
V (G) to the set of all leaves of T . For each edge e of T , T − e induces a partition
(Ae , Be ) of the leaves of T and we say that the width of e is ρG (L−1 (Ae )). By (1), the
choice of Ae , Be does not change the width of e. The width of a rank-decomposition
(T, L) is the maximum width of edges in T .
The rank-width of a graph G, denoted by rw(G), is the minimum width over all
rank-decompositions of G. (If G has less than two vertices, then G has no rankdecompositions; in this case we say that G has rank-width 0.)
2.2. Linear rank-width. Let us define a linearized variant of rank-width as follows.
A linear layout of an n-vertex graph G is a permutation v1 , v2 , . . . , vn of the vertices
n−1
of G. The width of a linear layout is mini=1
ρG ({v1 , v2 , . . . , vi }). The linear rankwidth of a graph G, denoted by lrw(G), is the minimum width over all linear layouts
of G. (If G has less than two vertices, then we say the linear rank-width of G is 0.)
An alternative way to define the linear rank-width is to define a linear rankdecomposition (T, L) to be a rank-decomposition such that T is a caterpillar and
then define linear rank-width as the minimum width over all linear rank-decompositions.
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
3
(A caterpillar is a tree having a path such that every leaf of the tree is either in
the path or adjacent to a vertex in the path.) This immediately implies that the
rank-width is always less than or equal to the linear rank-width.
2.3. Equivalent width parameters. Let us write cw(G) to denote the cliquewidth, introduced in [23, 25]. As we recall in Section 1, a class of graphs has
bounded rank-width if and only if it has bounded clique-width by the following
inequality.
Theorem 2.1 (Oum and Seymour [71]). rw(G) ≤ cw(G) ≤ 2rw(G)+1 − 1.
One can also prove, by the same idea, that linear clique-width of a graph G is at
least lrw(G) and at most 2lrw(G) + 1.
The inequalities in Theorem 2.1 are essentially tight:
• The n × n grid has rank-width n − 1, shown by Jelı́nek [45] and clique-width
n + 1, shown by Golumbic and Rotics [36].
• For every k, there exists a graph G such that its rank-width is at most k + 1
and the clique-width is at least 2⌊k/2⌋−1 . This is due to a theorem by Corneil
and Rotics [20], that constructs a graph of tree-width k whose clique-width
is at least 2⌊k/2⌋−1 . Oum [66] showed that a graph of tree-width k has rankwidth at most k + 1.
We may say that clique-width and rank-width are equivalent in the sense that
one is bounded if and only if the other is bounded. There are a few other width
parameters of graphs that are equivalent to rank-width. We list them below.
• NLC-width was introduced by Wanke [81] and is convenient to design an
algorithm based on dynamic programming on graphs of bounded cliquewidth or rank-width. Let us write nlc(G) to denote the NLC-width of G.
Johansson [48, 49] proved that
cw(G) ≤ nlc(G) ≤ 2 cw(G).
It can be easily shown that
rw(G) ≤ nlc(G) ≤ 2rw(G) .
• Boolean-width, denoted by boolw(G), was introduced by Bui-Xuan, Telle,
and Vatshelle [16]. They showed that
log2 cw(G) ≤ boolw(G) ≤ cw(G) and
rw(G)2
+ O(rw(G)).
4
In addition, they showed that for each k, there exist graphs G and H such
that the rw(G) ≥ k and boolw(G) ≤ 2 log2 k + 4 and rw(H) = k and
boolw(H) ≥ (k + 1)/6. It is not yet known [16] whether rw(G)2 can be
reduced to O(rw(G)) in the above inequality in general.
The definition of boolean-width is a slight variation of rank-width, only
changing the cut-rank function into an alternative function called the booleandimension. For a graph G and a subset A of vertices of G, let U(A) be the
set of all subsets of V (G) − A that can be represented as N(X) − A for
some X ⊆ A, where N(X) is the set of all vertices in V (G) − X that are
adjacent to a vertex in X. Then the boolean-dimension of A is defined to
be log2 |U(A)|. It turns out that |U(A)| = |U(V (G) − A)| (see [54]) and
log2 rw(G) ≤ boolw(G) ≤
4
SANG-IL OUM
therefore the boolean-dimension is symmetric. However, it is not necessarily
integer-valued.
• One may consider defining rank-width over a different field other than the
binary field GF (2) by changing the definition of cut-rank function. For a
field F, let us say that F-rank-width of a graph G, denoted by rwF (G), is
the obtained width parameter from rank-width by changing the cut-rank
function to evaluate the rank of a 0-1 matrix over F instead of the binary
field. The proof of Oum and Seymour [71] can be extended easily to F-rankwidth to prove
rwF (G) ≤ cw(G) ≤ 2rwF (G)+1 − 1.
Thus rank-width and F-rank-width are equivalent.
Furthermore Oum, Sæther, and Vatshelle [70] observed that if F is the
field of all rational numbers (or more generally, F is a field of characteristic
0 or 2), then one can enhance the above inequality by showing
rw(G) ≤ rwF (G) ≤ cw(G) ≤ 2rw(G)+1 − 1.
Most of other width parameters mentioned above do not have a direct algorithm
to find a corresponding decomposition and therefore we simply employ algorithms
on finding rank-decompositions and convert them. But F-rank-width is similar to
rank-width in the sense that one can use the theorem of Oum and Seymour [71] to
design its own fixed-parameter approximation algorithm. This was used to design
a faster algorithm for certain vertex partitioning problems [70]. When Kanté and
Rao [52] extended rank-width to directed graphs and edge-colored graphs, they use
the idea of F-rank-width.
3. Discovering Rank-width and linear rank-width
We skip the discussion on algorithms on graphs of small rank-width, except the
important meta theorem of Courcelle, Makowsky, and Rotics [24]. Their meta theorem states the following.
Theorem 3.1 (Courcelle, Makowsky, and Rotics [24]). Let k be a fixed constant.
For every closed monadic second-order formula of the first kind on graphs, there is
a O(n3 )-time algorithm to determine whether the input graph of rank-width at most
k satisfies the formula.
Here, the monadic second-order formula of the first kind on graphs are logic
formulas which allow ∃, ∀, ¬, ∧, ∨, ∈, (, ), true, and adj with first-order variables
each representing a vertex and set variables each representing a set of vertices, where
adj(x, y) is true if two vertices x and y are adjacent. If such a formula has no free
variable, then it is called closed. We refer the readers to the previous survey [42]. We
would like to mention that Grohe and Schweitzer [37] announced a polynomial-time
algorithm to check whether two graphs are isomorphic when the input graphs have
bounded rank-width.
We often talk about fixed-parameter tractable (FPT) algorithms. This means that
for a parameter k, the running time of the algorithm is at most f (k)nc for some c
where n is the length of the input. For more about parametrized complexity, please
see the following books [30, 27]. For us, the parameter is usually the rank-width
of the input graph. A fixed-parameter tractable algorithm is thought to be better
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
5
than XP algorithms, that is an algorithm whose running time is bounded above by
f (k)ng(k) .
3.1. Hardness. Computing rank-width is NP-hard, as mentioned in [64]. This can
be deduced easily from the following three known facts.
(1) Oum [63] showed that the branch-width of a binary matroid is equal to the
rank-width of its fundamental graph plus one. Note that every cycle matroid
is binary.
(2) Hicks and McMurray Jr. [38] and independently Mazoit and Thomassé [60]
showed that if a graph G is not a forest, then the branch-width of the cycle
matroid M(G) is equal to the branch-width of G.
(3) Seymour and Thomas [80] proved that computing branch-width of graphs is
NP-hard.
It is also NP-hard to compute linear rank-width. This is implied by the following
facts.
(1) Kashyap [53] proved that it is NP-hard to compute the path-width of a
binary matroid given by its representation.
(2) By Oum [63], it is straightforward to deduce that the path-width of a binary
matroid is equal to the linear rank-width of its fundamental graph plus one.
3.2. Computing rank-width. Oum and Seymour [72] presented a generic XP
algorithm for symmetric submodular integer-valued functions. Their algorithm implies that in time O(n8k+12 log n) we can find a rank-decomposition of width at most
k, if it exists, for an input n-vertex graph.
But we definitely prefer to remove k in the exponent of n. Courcelle and Oum [26]
constructed a fixed-parameter tractable algorithm to decide, for fixed k, whether
the rank-width is at most k, in time O(g(k)n3). However their algorithm does not
provide a rank-decomposition of width at most k directly, because their algorithm
uses forbidden vertex-minors, which we discuss in the next section.
This problem was solved later. Hliněný and Oum [41] devised an algorithm to
find a rank-decomposition of width at most k if it exists, for an input n-vertex graph
in time O(g(k)n3). Here g(k) is a huge function. The proof uses the correspondence
between branch-width of binary matroids and rank-width of graphs and still uses a
huge list of forbidden minors in matroids.
For bipartite circle graphs, one can compute rank-width in polynomial time. This
is due to the following results.
• Seymour and Thomas [80] proved that branch-width of planar graphs can
be computed in polynomial time.
• Hicks and McMurray Jr. [38] and independently Mazoit and Thomassé [60]
showed that if a graph G is not a forest, then the branch-width of the cycle
matroid M(G) is equal to the branch-width of G.
• Oum [63] showed that the branch-width of a binary matroid is equal to the
rank-width of its fundamental graph plus one. Note that every cycle matroid
is binary.
• De Fraysseix [28] proved that a graph is a bipartite circle graph if and only
if it is a fundamental graph of the cycle matroid of a planar graph.
It would be interesting to extend the class of graphs on which rank-width can be
computed exactly.
6
SANG-IL OUM
Question 1. Can we compute rank-width of circle graphs in polynomial time?
3.3. Fixed-parameter approximations. For many algorithmic applications, we
restrict our input graph to have bounded rank-width. Most of such algorithms
assume that the input graph is given with a rank-decomposition of small width and
for such algorithms, it is usually enough to have an efficient algorithm that outputs
a rank-decomposition of width at most f (k) for some function f or confirms that
the rank-width of the input graph is larger than k.
The first such algorithm was devised by Oum and Seymour [71] with f (k) = 3k +1
and the running time O(8k n9 log n), where n is the number of the vertices in the
input graph. It was later improved by Oum [64] with f (k) = 3k + 1 and the running
time O(8k n4 ). This allows us to compute rank-width within a constant factor in
polynomial time when the input graph is known to have rank-width at most c log n
for some constant c.
In the same paper, Oum [64] presented another algorithm with f (k) = 3k − 1
whose running time is O(g(k) n3) for some huge function g(k).
Question 2. Does there exist an algorithm with a function f (k) and a constant c
that finds a rank-decomposition of width at most f (k) or confirms that the rank-width
of the n-vertex input graph is larger than k, in time O(ck n3 )?
One may hope to find a faster approximation algorithm for rank-width. For
tree-width, Bodlaender, Drange, Dregi, Fomin, Lokshtanov, and Pilipczuk [7] presented an algorithm that, for an n-vertex input graph and an integer k, finds a
tree-decomposition of width at most 5k + 4 or confirms that the tree-width is larger
than k in time 2O(k) n. However, for rank-width, it seems quite non-trivial to reduce
n3 into nc for some c < 3. This may be justified by the observation that one needs
to compute the rank of matrices and computing a rank of a low-rank matrix may
be one of the bottlenecks. There are some algorithms computing rank faster [17]
but it is not clear whether those algorithms can be applied to answer the following
question.
Question 3. Does there exist an algorithm with functions f (k), g(k) and a constant
c < 3 that finds a rank-decomposition of width at most f (k) or confirms that the
rank-width of the n-vertex input graph is larger than k, in time O(g(k)nc)?
3.4. Computing linear rank-width. Let us turn our attention to linear rankwidth. Nagamochi [61] presented a simple generic XP algorithm analogous to Oum
and Seymour [72] for path-width. His algorithm implies that we can find a linear
rank-decomposition of width at most k by calling the algorithm to minimize submodular functions [62, 43, 44, 79] at most O(n2k+4 ) times, where n is the number of
vertices of the input graph. In fact, for the linear rank-width, one can use a direct
combinatorial algorithm in [64] to avoid using the generic algorithm to minimize
submodular functions and improve its running time.
Until recently there was no known algorithm that, in time O(g(k)nc ), finds a
linear rank-decomposition of width at most k or confirms that a graph has linear
rank-width larger than k for an input n-vertex graph. By the well-quasi-ordering
property to be discussed in Subsection 4.6, there exists an algorithm that decides
in time O(g(k)nc) whether the input n-vertex graph has linear rank-width at most
k, but such an algorithm would not easily produce a linear rank-decomposition of
small width. Hliněný [39] claimed that one can use the self-reduction technique [32]
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
7
to convert a decision algorithm for linear rank-width at most k into a construction
algorithm.
Recently, Jeong, Kim, and Oum [46] devised a direct algorithm for graphs of rankwidth at most k that runs in time O(g(k)ℓ n3 ) to output a linear rank-decomposition
of width at most ℓ or confirm that the linear rank-width is larger than ℓ. Now
observe the following inequality shown by Kwon [55] (also in Adler, Kanté, and
Kwon [4]).
Theorem 3.2 (Kwon [55]; Adler, Kanté, and Kwon [4]). Let G be an n-vertex graph.
Then lrw(G) ≤ rw(G)⌊log2 n⌋.
Thus ℓ ≤ k⌊log2 n⌋ and so we can conclude the following.
Theorem 3.3 (Jeong, Kim, and Oum [46]). For each fixed integer k, there is a
polynomial-time (XP) algorithm to compute the linear rank-width of an input graph
of rank-width at most k.
However, it is not known whether there is a fixed-parameter tractable algorithm
to compute linear rank-width of a graph of rank-width k.
Theorem 3.3 generalizes the polynomial-time algorithm of Adler, Kanté, and
Kwon [4] that computes the linear rank-width of graphs of rank-width at most 1.
3.5. Exact algorithms. Though the number of rank-decompositions of an n-vertex
graph is a superexponential function, we can still determine the rank-width of an
n-vertex graph in time O(2n poly(n)) where poly(n) is some polynomial in n, shown
by Oum [67]. This algorithm can find an optimal rank-decomposition as well.
Question 4. Is it possible to compute rank-width of an n-vertex graph in time O(cn )
for some c < 2?
It is easy to compute the linear rank-width of an n-vertex graph in time O(2n poly(n))
by dynamic programming [8]. All algorithms mentioned so far in this subsection use
exponential space.
3.6. Software. Krause implemented a simple dynamic programming algorithm to
compute the rank-width of a graph.1 This software is now included in the open
source mathematics software package called SAGE; see the manual.2
Friedmanský wrote the Master thesis [35] about implementing the exact exponentialtime algorithm of Oum [67].
Bui-Xuan, Raymond, and Trébuchet [15] implemented, in SAGE, the algorithm
of Oum [64] that either outputs a rank-decomposition of width at most 3k + 1 or
confirms that rank-width is larger than k in time O(n4 ) for an input n-vertex graph.
3.7. Other results. Lee, Lee, and Oum [57] showed that asymptotically almost
surely the Erdős-Rényi random graph G(n, p) has rank-width ⌈n/3⌉ − O(1) if p is
a constant between 0 and 1. Furthermore, if n1 ≪ p ≤ 12 , then the rank-width is
⌈ n3 ⌉ − o(n), and if p = c/n and c > 1, then rank-width is at least rn for some
r = r(c).
1http://pholia.tdi.informatik.uni-frankfurt.de/
~philipp/software/rw.shtml
http://www.sagemath.org/doc/reference/graphs/sage/graphs/graph_decompositions/rankwidth.html
2
8
SANG-IL OUM
4. Structural aspects
As tree-width is closely related to graph minors, rank-width is related to vertexminors and pivot-minors, observed in [63]. We first review the definition of pivotminors and vertex-minors and then discuss various theorems, mostly motivated by
the theory of tree-width and graph minors.
4.1. Vertex-minors and pivot-minors. For a graph G and its vertex v, G ∗ v is
the graph such that V (G ∗ v) = V (G) and two vertices x, y are adjacent in G ∗ v if
and only if
(i) both x and y are neighbors of v in G and x is non-adjacent to y, or
(ii) x is adjacent to y and v is non-adjacent to at least one of x and y.
Such an operation is called the local complementation at v. Clearly G ∗ v ∗ v = G.
Two graphs are locally equivalent if one is obtained from another by a sequence of
local complementations. We say that a graph H is a vertex-minor of a graph G if
H is an induced subgraph of a graph locally equivalent to G.
Local complementation is a useful tool to study rank-width of graphs, because
local complementation preserves the cut-rank function, thus preserving rank-width
and linear rank-width [63]. This implies that if H is a vertex-minor of G, then
the rank-width of H is less than or equal to the rank-width of G and the linear
rank-width of H is less than or equal to the linear rank-width of G.
Pivot-minors are restricted version of vertex-minors. For an edge uv, we define
G ∧ uv = G ∗ u ∗ v ∗ u. This operation is called a pivot. Two graphs are pivot
equivalent if one is obtained from another by a sequence of pivots. A graph H is a
pivot-minor of a graph G if H is an induced subgraph of a graph pivot equivalent
to G. There is a close relation between the pivot operation and the operation of
switching a base in a binary matroid, which motivates the study of pivot-minors.
Every pivot-minor of a graph G is also a vertex-minor of G but not vice versa.
Vertex-minors of graphs can be seen as minors of isotropic systems, a concept
introduced by Bouchet [11, 14, 10]. Similarly pivot-minors of graphs can be seen as
minors of binary delta-matroids, also introduced by Bouchet [9, 12].
4.2. Relation to tree-width and branch-width. Let G be a graph whose treewidth is k. Kanté [50] showed that rank-width of G is at most 4k + 2 by showing
how to convert a tree-decomposition into a rank-decomposition constructively. This
inequality was improved by Oum [66], who showed that rank-width is at most k + 1.
In fact, the following are shown. The incidence graph I(G) of a graph G is the
graph obtained from G by subdividing each edge exactly once.
Theorem 4.1 (Oum [66]). Let G be a graph of branch-width k. Then the incidence
graph I(G) of G has rank-width k or k − 1, unless k = 0.
The converse does not hold; the complete graph Kn has tree-width n − 1 and
rank-width 1 if n ≥ 2. We remark that Courcelle [22] proved that if G is a graph of
tree-width k, then its incidence graph I(G) has clique-width at most 2k + 4.
Since G is a vertex-minor of I(G), we deduce that the rank-width of G is less than
or equal to the branch-width of G, unless G has branch-width 0. This inequality
can be generalized to an arbitrary field.
Theorem 4.2 (Oum, Sæther, Vatshelle [70]). For every field F and a graph G,
rwF (G) is less than or equal to branch-width of G, unless G has branch-width 0.
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
9
The line graph L(G) of a simple graph G is a vertex-minor of the incidence graph
I(G), obtained by applying local complementations at vertices of G in I(G). This
implies that the rank-width of L(G) is less than or equal to the rank-width of I(G).
The following theorem proves that they are very close to each other.
Theorem 4.3 (Oum [68]). Let G be a simple graph of branch-width k. Then the
line graph of G has rank-width k, k − 1, or k − 2.
Surprisingly, Adler and Kanté [2] showed that for trees, linear rank-width is equal
to path-width.
Another relation to tree-width was discovered by Kwon and Oum [56]. They have
shown that graphs of small rank-width are precisely pivot-minors of graphs of small
tree-width.
Theorem 4.4 (Kwon and Oum [56]).
(1) Every graph of rank-width k is a pivotminor of a graph of tree-width at most 2k.
(2) Every graph of linear rank-width k is a pivot-minor of a graph of path-width
at most k + 1.
Hliněný, Kwon, Obdržálek, and Ordyniak [40] found an analogue of Theorem 4.4
to relate tree-depth and shrub-depth of graphs by vertex-minors.
Fomin, Oum, and Thilikos [33] showed that when graphs are planar, or H-minorfree, then having bounded tree-width is equivalent to having bounded rank-width.
This is already proven in Courcelle and Olariu [25] without explicit bounds because
they use logical tools. For instance, in [33], it is shown that the tree-width of a
planar graph G is at most 72 rw(G) − 2, and if a graph G has no Kr minor with
r > 2, then its tree-width is at most 2O(r log log r) rw(G). The last bound on tree-width
can be improved to 2O(r) rw(G) by using a recent result of Lee and Oum [58].
4.3. Chromatic number. A class I of graphs is χ-bounded if there exists a function f such that χ(G) ≤ f (ω(G)) for all graphs G ∈ I. Here, ω(G) is the maximum
size of a clique in G and χ(G) is the chromatic number of G. Dvořák and Král’ [31]
proved that the class of graphs of rank-width at most k is χ-bounded as follows.
Theorem 4.5 (Dvořák and Král’ [31]). There exists a function f : Z × Z → Z such
that for every graph G of rank-width at most k, χ(G) ≤ f (ω(G), k).
More generally, Geelen (see [31, 18]) conjectured that for each fixed graph H, the
class of graphs with no H vertex-minor is χ-bounded. This conjecture has been
verified when H is the wheel graph on 6 vertices [31] and a fan graph [18].
4.4. Duality. As rank-width is an instance of branch-width of some symmetric
submodular function, we can use a theory developed for symmetric submodular
functions by [77, 72, 5, 59, 29]. In particular, one can use the concept called a
tangle to verify that a graph has large rank-width.
For a graph G, a ρG -tangle of order k is a set T of subsets of V (G) satisfying the
following three axioms.
(T1) For A ⊆ V (G), if ρG (A) < k, then either A ∈ T or V (G) − A ∈ T .
(T2) If A, B, C ∈ T , then A ∪ B ∪ C 6= V (G).
(T3) For all v ∈ V (G), V (G) − {v} ∈
/T.
A theorem of Robertson and Seymour [77] implies the following.
10
SANG-IL OUM
Theorem 4.6 (Robertson and Seymour [77]). For an integer k, a graph G has a
ρG -tangle of order k if and only if its rank-width is at least k.
A similar concept can be defined for linear rank-width. A ρG -obstacle of order k
is a set O of subsets of V (G) satisfying the following three axioms.
(B1) For all A ∈ O, ρG (A) < k.
(B2) If X ⊆ Y ∈ O and ρG (X) < k, then X ∈ O.
(B3) If A ∪ B ∪ C = V (G), A ∩ B = ∅, ρG (A) < k, ρG (B) < k and |C| ≤ 1, then
exactly one of A and B is in O.
The following theorem is implied by Fomin and Thilikos [34, Theorem 3].
Theorem 4.7 (Fomin and Thilikos [34]). For an integer k, a graph G has a ρG obstacle of order k if and only if its linear rank-width is at least k.
4.5. Large rank-width. Robertson and Seymour [75] proved that every graph with
sufficiently large tree-width contains a large grid graph as a minor. An analogous
conjecture was made in [68] as follows.
Question 5. Is it true that for each fixed bipartite circle graph H, every graph G
with sufficiently large rank-width contains a pivot-minor isomorphic to H?
This has been confirmed when G is a line graph, a bipartite graph, or a circle
graph, see [68].
Adler, Kanté and Kwon [3] proved that for each tree T , every distance-hereditary
graph with sufficiently large linear rank-width contains a vertex-minor isomorphic
to T . It would be interesting to determine whether every graph of sufficiently large
linear rank-width contains a vertex-minor isomorphic to a fixed tree T .
4.6. Well-quasi-ordering. A pair (Q, ) of a set Q and a relation on Q is a
quasi-order if
(1) x x for all x ∈ Q and
(2) x z if x y and y z.
A quasi-order (Q, ) is a well-quasi-order if every infinite sequence q1 , q2 , . . . of Q
has a pair qi , qj (i < j) such that qi qj . We say that Q is well-quasi-ordered by
the relation .
Motivated by the celebrated graph minor theorem by Robertson and Seymour [78]
and its special case on tree-width [76], Oum [65] proved that graphs of bounded
rank-width are well-quasi-ordered under taking pivot-minors as follows.
Theorem 4.8 (Oum [65]). For all positive integers k, every infinite sequence G1 , G2 , . . .
of graphs of rank-width at most k admits a pair Gi , Gj (i < j) such that Gi is isomorphic to a pivot-minor of Gj .
Theorem 4.8 has been generalized to skew-symmetric or symmetric matrices over
a fixed finite field (Oum [69]) and σ-symmetric matrices over a fixed finite field
(Kanté [51]).
Courcelle and Oum [26] constructed a modulo-2 counting monadic second-order
logic formula to decide whether a fixed graph H is isomorphic to a vertex-minor of an
input graph. This can be combined with the meta-theorem of Courcelle, Makowsky,
and Rotics [24] to deduce the following.
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
11
Theorem 4.9 (Courcelle and Oum [26]). Let I be a set of graphs closed under
taking vertex-minors. If I has bounded rank-width, then there exists an algorithm
that decides whether an input n-vertex graph is in I in time O(n3 ).
We remark that a positive answer to the following question would imply the graph
minor theorem of Robertson and Seymour.
Question 6. Are graphs well-quasi-ordered under taking pivot-minors? In other
words, does every infinite sequence G1 , G2 , . . . of graphs admit a pair Gi , Gj (i < j)
such that Gi is isomorphic to a pivot-minor of Gj ?
We would also like to have an algorithm to detect pivot-minors and vertex-minors
of a graph.
Question 7. For a fixed graph H, can we decide in polynomial time whether an
input graph has a pivot-minor (or a vertex-minor) isomorphic to H?
4.7. Forbidden vertex-minors. By Theorem 4.8, we can easily deduce that there
exists a finite set Fk of graphs such that a graph G has rank-width at most k if and
only if Fk contains no vertex-minor of G, because otherwise there will be an infinite
anti-chain of vertex-minor-minimal graphs having rank-width k + 1. However, the
well-quasi-ordering theorem does not provide an upper bound on |Fk |.
Oum [63] showed that pivot-minor-minimal graphs of rank-width k + 1 have at
most (6k+1 − 1)/5 vertices. This gives an explicit upper bound on |Fk | as well as a
constructive enumeration algorithm for graphs in Fk .
For k = 1, F1 is known. Oum [63] showed that graphs of rank-width at most 1
are precisely distance-hereditary graphs; a graph G is distance-hereditary if in every
connected induced subgraph, a shortest path between two vertices is also a shortest
path in G [6]. (This also proves that distance-hereditary graphs have clique-width
at most 3 as an easy consequence of Theorem 2.1, which was initially proved by
Golumbic and Rotics [36].) A theorem of Bouchet [10, 13] implies that a graph
G has rank-width at most 1 if and only if C5 (the cycle graph of length 5) is not
isomorphic to a vertex-minor of G. A similar argument can be used to show that a
graph G has rank-width at most 1 if and only if neither C5 nor C6 is isomorphic to
a pivot-minor of G.
For the linear rank-width, we can still use Theorem 4.8 to deduce that there
exists a finite set Lk of graphs such that a graph G has linear rank-width at most k
if and only if Lk contains no vertex-minor of G. Unlike the situation in rank-width,
so far no upper bound on the size of graphs in Lk is known. Adler, Farley, and
Proskurowski [1] determined L1 as a set of three particular graphs. Jeong, Kwon,
k
and Oum [47] proved that |Lk | ≥ 3Ω(3 ) . Adler, Kanté and Kwon [3] presented an
algorithm to construct, for each k, a finite set L′k of distance-hereditary graphs of
k
linear rank-width k + 1 such that |L′k | ≤ 3O(3 ) , every proper vertex-minor of a graph
in L′k has linear rank-width at most k, and every distance-hereditary graph of linear
rank-width k + 1 contains at least one graph in L′k .
Acknowledgments. The author would like to thank Mamadou M. Kanté, Eun
Jung Kim, O-joung Kwon, and Jisu Jeong for their helpful comments on an early
draft of this paper. The author would also like to thank anonymous referees for
their helpful suggestions.
12
SANG-IL OUM
References
[1] I. Adler, A. M. Farley, and A. Proskurowski. Obstructions for linear rankwidth at most 1.
Discrete Appl. Math., 168:3–13, 2014.
[2] I. Adler and M. M. Kanté. Linear rank-width and linear clique-width of trees. Theoret. Comput. Sci., 589:87–98, 2015.
[3] I. Adler, M. M. Kanté, and O. Kwon. Linear rank-width of distance-hereditary graphs II.
Vertex-minor obstructions. arXiv:1508.04718, 2015.
[4] I. Adler, M. M. Kanté, and O. Kwon. Linear rank-width of distance-hereditary graphs I. A
polynomial-time algorithm. Algorithmica, 2016. doi:10.1007/s00453-016-0164-5.
[5] O. Amini, F. Mazoit, N. Nisse, and S. Thomassé. Submodular partition functions. Discrete
Appl. Math., 309(20):6000–6008, 2009.
[6] H.-J. Bandelt and H. M. Mulder. Distance-hereditary graphs. J. Combin. Theory Ser. B,
41(2):182–208, 1986.
[7] H. L. Bodlaender, P. G. Drange, M. S. Dregi, F. V. Fomin, D. Lokshtanov, and M. Pilipczuk.
A ck n 5-approximation algorithm for treewidth. SIAM J. Comput., 45(2):317–378, 2016.
[8] H. L. Bodlaender, F. V. Fomin, A. M. C. A. Koster, D. Kratsch, and D. M. Thilikos. A note
on exact algorithms for vertex ordering problems on graphs. Theory Comput. Syst., 50(3):420–
432, 2012.
[9] A. Bouchet. Greedy algorithm and symmetric matroids. Math. Programming, 38(2):147–159,
1987.
[10] A. Bouchet. Isotropic systems. European J. Combin., 8(3):231–244, 1987.
[11] A. Bouchet. Graphic presentations of isotropic systems. J. Combin. Theory Ser. B, 45(1):58–
76, 1988.
[12] A. Bouchet. Representability of ∆-matroids. In Combinatorics (Eger, 1987), volume 52 of
Colloq. Math. Soc. János Bolyai, pages 167–182. North-Holland, Amsterdam, 1988.
[13] A. Bouchet. Transforming trees by successive local complementations. J. Graph Theory,
12(2):195–207, 1988.
[14] A. Bouchet. Connectivity of isotropic systems. In Combinatorial Mathematics: Proceedings of
the Third International Conference (New York, 1985), volume 555 of Ann. New York Acad.
Sci., pages 81–93, New York, 1989. New York Acad. Sci.
[15] B.-M.
Bui-Xuan,
J.-F.
Raymont,
and
P.
Trébuchet.
Implantation
des
algorithmes
oum-seymour
et
oum.
A
slide
of
a
talk
at
JGA2011.
http://www.mimuw.edu.pl/~jr348692/files/2011-11-17_diapos_jga.pdf, 2011.
[16] B.-M. Bui-Xuan, J. A. Telle, and M. Vatshelle. Boolean-width of graphs. Theoret. Comput.
Sci., 412(39):5187–5204, 2011.
[17] H. Y. Cheung, T. C. Kwok, and L. C. Lau. Fast matrix rank algorithms and applications. J.
ACM, 60(5):Art. 31, 25, 2013.
[18] I. Choi, O.-j. Kwon, and S. Oum. Coloring graphs without fan vertex-minors and graphs
without cycle pivot-minors. arXiv:1512.03481, 2015.
[19] D. G. Corneil, M. Habib, J.-M. Lanlignel, B. Reed, and U. Rotics. Polynomial-time recognition
of clique-width ≤ 3 graphs. Discrete Appl. Math., 160(6):834–865, 2012.
[20] D. G. Corneil and U. Rotics. On the relationship between clique-width and treewidth. SIAM
J. Comput., 34(4):825–847 (electronic), 2005.
[21] B. Courcelle. Clique-width and edge contraction. Inform. Process. Lett., 114(1-2):42–44, 2014.
[22] B. Courcelle. Clique-width and tree-width of sparse graphs. manuscript, 2015.
[23] B. Courcelle, J. Engelfriet, and G. Rozenberg. Handle-rewriting hypergraph grammars. J.
Comput. System Sci., 46(2):218–270, 1993.
[24] B. Courcelle, J. A. Makowsky, and U. Rotics. Linear time solvable optimization problems on
graphs of bounded clique-width. Theory Comput. Syst., 33(2):125–150, 2000.
[25] B. Courcelle and S. Olariu. Upper bounds to the clique width of graphs. Discrete Appl. Math.,
101(1-3):77–114, 2000.
[26] B. Courcelle and S. Oum. Vertex-minors, monadic second-order logic, and a conjecture by
Seese. J. Combin. Theory Ser. B, 97(1):91–126, 2007.
[27] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk,
and S. Saurabh. Parameterized algorithms. Springer, Cham, 2015.
RANK-WIDTH: ALGORITHMIC AND STRUCTURAL RESULTS
13
[28] H. de Fraysseix. Local complementation and interlacement graphs. Discrete Math., 33(1):29–
35, 1981.
[29] R. Diestel and S. Oum. Unifying duality theorem for width parameters in graphs and matroids.
I. Weak and strong duality. arXiv:1406.3797, 2014.
[30] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Springer, 2013.
[31] Z. Dvořák and D. Král’. Classes of graphs with small rank decompositions are χ-bounded.
European J. Combin., 33(4):679–683, 2012.
[32] M. R. Fellows and M. A. Langston. Fast self-reduction algorithms for combinatorial problems
of VLSI design. In VLSI algorithms and architectures (Corfu, 1988), volume 319 of Lecture
Notes in Comput. Sci., pages 278–287. Springer, New York, 1988.
[33] F. V. Fomin, S. Oum, and D. M. Thilikos. Rank-width and tree-width of H-minor-free graphs.
European J. Combin., 31(7):1617–1628, 2010.
[34] F. V. Fomin and D. M. Thilikos. On the monotonicity of games generated by symmetric
submodular functions. Discrete Appl. Math., 131(2):323–335, 2003.
[35] M. Friedmanský. Implementation and optimization of a graph algorithm for computing rankwidth. Master’s thesis, Masarykova Univerzita, 2011.
[36] M. C. Golumbic and U. Rotics. On the clique-width of some perfect graph classes. Internat. J.
Found. Comput. Sci., 11(3):423–443, 2000. Selected papers from the Workshop on Theoretical
Aspects of Computer Science (WG 99), Part 1 (Ascona).
[37] M. Grohe and P. Schweitzer. Isomorphism testing for graphs of bounded rank width.
arXiv:1505.03737. Presented at FOCS 2015, 2015.
[38] I. V. Hicks and N. B. McMurray Jr. The branchwidth of graphs and their cycle matroids. J.
Combin. Theory Ser. B, 97(5):681–692, 2007.
[39] P. Hliněný. Simpler self-reduction algorithm for matroid path-width. arXiv:1605.09520, 2016.
[40] P. Hliněný, O.-j. Kwon, J. Obdržálek, and S. Ordyniak. Tree-depth and vertex-minors. European J. Combin., 56:46–56, 2016.
[41] P. Hliněný and S. Oum. Finding branch-decompositions and rank-decompositions. SIAM J.
Comput., 38(3):1012–1032, 2008.
[42] P. Hliněný, S. Oum, D. Seese, and G. Gottlob. Width parameters beyond tree-width and their
applications. The Computer Journal, 51(3):326–362, 2008.
[43] S. Iwata. A faster scaling algorithm for minimizing submodular functions. SIAM J. Comput.,
32(4):833–840 (electronic), 2003.
[44] S. Iwata, L. Fleischer, and S. Fujishige. A combinatorial strongly polynomial algorithm for
minimizing submodular functions. Journal of the ACM (JACM), 48(4):761–777, 2001.
[45] V. Jelı́nek. The rank-width of the square grid. Discrete Appl. Math., 158(7):841–850, 2010.
[46] J. Jeong, E. J. Kim, and S. Oum. The “art of trellis decoding” is fixed-parameter tractable.
arXiv:1507.02184. Presented at SODA2016, 2016.
[47] J. Jeong, O.-j. Kwon, and S. Oum. Excluded vertex-minors for graphs of linear rank-width at
most k. European J. Combin., 41:242–257, 2014.
[48] Ö. Johansson. Clique-decomposition, NLC-decomposition, and modular decomposition—
relationships and results for random graphs. In Proceedings of the Twenty-ninth Southeastern
International Conference on Combinatorics, Graph Theory and Computing (Boca Raton, FL,
1998), volume 132, pages 39–60, 1998.
[49] Ö. Johansson. Graph decomposition using node labels. PhD thesis, Royal Institute of Technology, 2001.
[50] M. M. Kanté. Vertex-minor reductions can simulate edge contractions. Discrete Appl. Math.,
155(17):2328–2340, 2007.
[51] M. M. Kanté. Well-quasi-ordering of matrices under Schur complement and applications to
directed graphs. European J. Combin., 33(8):1820–1841, 2012.
[52] M. M. Kanté and M. Rao. The rank-width of edge-coloured graphs. Theory Comput. Syst.,
52(4):599–644, 2013.
[53] N. Kashyap. Matroid pathwidth and code trellis complexity. SIAM Journal on Discrete Mathematics, 22(1):256–272, 2008.
[54] K. H. Kim. Boolean matrix theory and applications, volume 70 of Monographs and Textbooks
in Pure and Applied Mathematics. Marcel Dekker, Inc., New York, 1982. With a foreword by
Gian-Carlo Rota.
14
SANG-IL OUM
[55] O.-j. Kwon. On the structural and algorithmic properties of linear rank-width. PhD thesis,
KAIST, 2015.
[56] O.-j. Kwon and S. Oum. Graphs of small rank-width are pivot-minors of graphs of small
tree-width. Discrete Appl. Math., 168:108–118, 2014.
[57] C. Lee, J. Lee, and S. Oum. Rank-width of random graphs. J. Graph Theory, 70(3):339–347,
July/August 2012.
[58] C. Lee and S. Oum. Number of cliques in graphs with a forbidden subdivision. SIAM J.
Discrete Math., 29(4):1999–2005, 2015.
[59] L. Lyaudet, F. Mazoit, and S. Thomassé. Partitions versus sets: a case of duality. European
J. Combin., 31(3):681–687, 2010.
[60] F. Mazoit and S. Thomassé. Branchwidth of graphic matroids. In A. Hilton and J. Talbot,
editors, Surveys in Combinatorics 2007, volume 346 of London Math. Soc. Lecture Note Ser.,
pages 275–286. Cambridge Univ. Press, Cambridge, 2007.
[61] H. Nagamochi. Linear layouts in submodular systems. In K.-M. Chao, T.-s. Hsu, and D.T. Lee, editors, ISAAC ’12, volume 7676 of Lecture Notes in Comput. Sci., pages 475–484.
Springer Berlin Heidelberg, 2012.
[62] J. B. Orlin. A faster strongly polynomial time algorithm for submodular function minimization. Math. Program., 118(2, Ser. A):237–251, 2009.
[63] S. Oum. Rank-width and vertex-minors. J. Combin. Theory Ser. B, 95(1):79–100, 2005.
[64] S. Oum. Approximating rank-width and clique-width quickly. ACM Trans. Algorithms,
5(1):Art. 10, 20, 2008.
[65] S. Oum. Rank-width and well-quasi-ordering. SIAM J. Discrete Math., 22(2):666–682, 2008.
[66] S. Oum. Rank-width is less than or equal to branch-width. J. Graph Theory, 57(3):239–244,
2008.
[67] S. Oum. Computing rank-width exactly. Inform. Process. Lett., 109(13):745–748, 2009.
[68] S. Oum. Excluding a bipartite circle graph from line graphs. J. Graph Theory, 60(3):183–203,
2009.
[69] S. Oum. Rank-width and well-quasi-ordering of skew-symmetric or symmetric matrices. Linear
Algebra Appl., 436(7):2008–2036, 2012.
[70] S. Oum, S. H. Sæther, and M. Vatshelle. Faster algorithms for vertex partitioning problems
parameterized by clique-width. Theoret. Comput. Sci., 535:16–24, 2014.
[71] S. Oum and P. Seymour. Approximating clique-width and branch-width. J. Combin. Theory
Ser. B, 96(4):514–528, 2006.
[72] S. Oum and P. Seymour. Testing branch-width. J. Combin. Theory Ser. B, 97(3):385–393,
2007.
[73] J. G. Oxley. Matroid theory. Oxford University Press, New York, 1992.
[74] N. Robertson and P. Seymour. Graph minors. II. Algorithmic aspects of tree-width. J. Algorithms, 7(3):309–322, 1986.
[75] N. Robertson and P. Seymour. Graph minors. V. Excluding a planar graph. J. Combin. Theory
Ser. B, 41(1):92–114, 1986.
[76] N. Robertson and P. Seymour. Graph minors. IV. Tree-width and well-quasi-ordering. J.
Combin. Theory Ser. B, 48(2):227–254, 1990.
[77] N. Robertson and P. Seymour. Graph minors. X. Obstructions to tree-decomposition. J. Combin. Theory Ser. B, 52(2):153–190, 1991.
[78] N. Robertson and P. Seymour. Graph minors. XX. Wagner’s conjecture. J. Combin. Theory
Ser. B, 92(2):325–357, 2004.
[79] A. Schrijver. A combinatorial algorithm minimizing submodular functions in strongly polynomial time. J. Combin. Theory Ser. B, 80(2):346–355, 2000.
[80] P. Seymour and R. Thomas. Call routing and the ratcatcher. Combinatorica, 14(2):217–241,
1994.
[81] E. Wanke. k-NLC graphs and polynomial algorithms. Discrete Appl. Math., 54(2-3):251–266,
1994.
Department of Mathematical Sciences, KAIST, Daejeon, South Korea
E-mail address: [email protected]
| 8cs.DS
|
12th International Conference on CFD in Oil & Gas, Metallurgical and Process Industries
SINTEF, Trondheim, NORWAY
May 30th - June 1st, 2017
CFD 2017
IMPLEMENTATION, DEMONSTRATION AND VALIDATION OF A USER-DEFINED WALL
FUNCTION FOR DIRECT PRECIPITATION FOULING IN ANSYS FLUENT
Sverre G. JOHNSEN1∗ , Tiina M. PÄÄKKÖNEN2 , Stein T. JOHANSEN13 , Riitta L. KEISKI2 , Bernd
WITTGENS1
1 SINTEF
2 University
3 NTNU,
Materials and Chemistry, NO-7465 Trondheim, NORWAY
of Oulu, Environmental and Chemical Engineering, FI-90014 Oulu, FINLAND
Dept. of Energy and Process Engineering, NO-7491 Trondheim, NORWAY
∗
E-mail: [email protected]
arXiv:1706.01453v1 [cs.CE] 2 Jun 2017
ABSTRACT
In a previous paper (Johnsen et al., 2015) and presentation (Johnsen
et al., 2016), we developed and demonstrated a generic modelling
framework for the modelling of direct precipitation fouling from
multi-component fluid mixtures that become super-saturated at the
wall. The modelling concept involves the 1-dimensional transport
of the fluid species through the turbulent boundary layer close to
the wall. The governing equations include the Reynolds-averaged
(RANS) advection-diffusion equations for each fluid species, and
the axial momentum and energy equations for the fluid mixture.
The driving force for the diffusive transport is the local gradient
in the species’ chemical potential. Adsorption mechanisms are not
modelled per se, but the time-scale of adsorption is reflected in the
choice of Dirichlet boundary conditions for the depositing species,
at the fluid-solid interface.
In this paper, the modelling framework is implemented as a userdefined function (UDF) for the CFD software ANSYS Fluent, to
act as a wall boundary condition for mass-transfer to the wall. The
subgrid, 1-dimensional formulation of the model reduces the computational cost associated with resolving the fine length-scales at
which the boundary-layer mass transfer is determined, and allows
for efficient modelling of industry-scale heat exchangers suffering
from fouling.
The current paper describes the modelling framework, and demonstrates and validates its applicability in a simplified 2D heat exchanger geometry (experimental and detailed CFD modelling data
by Pääkkönen et al. (2012, 2016)). By tuning the diffusivity, only,
good agreement with the experimental data and the detailed CFD
model was obtained, in terms of area-averaged deposition rates.
Keywords: CFD, Heat Exchangers, Mass transfer, Multiscale,
UDF, Wall function, Fouling .
NOMENCLATURE
Greek Symbols
∂y+ ≡ ∂/∂y+ Dimensionless derivative in the wall-normal direction, [−].
κ = 0.42 von Kármán constant, [−].
µ
Dynamic viscosity, [Pas].
ρf
Fluid mixture mass density, [kg/m3 ].
τw
Wall shear stress, [Pa].
Latin Symbols
cP
Specific heat capacity, [J/kgK].
C
Concentration, [kg/m3 ].
Ea
Activation energy, [J/mol].
g
Gravity vector, [m/s2 ].
Specific sensible enthalpy, [J/kg].
Deposition rate, [kg/m2 s].
Mass flux vector, [kg/m2s].
Pre-exponential factor, m4/kgs2 .
Surface integration rate constant, m4/kgs2 .
Thermal conductivity, [W/mK].
Number of species, [−].
Pressure, [Pa].
Prandtl number, [−].
Wall heat flux, [W/m2 ].
R
Universal gas constant, [8.3144598J/Kmol].
Sc
Schmidt number, [−].
T
p Absolute temperature, [K[.
uτ ≡ τw/ρ f ,w Shear velocity, [m/s].
u f ,x
Fluid velocity parallel to the wall, [m/s].
uf
Mass-averaged advective fluid velocity vector, [m/s].
x
Cartesian coordinate, parallel to the wall, [m].
X
Mass fraction, [kg/kg].
y
Cartesian coordinate, normal to the wall, [m].
z
Mole fraction, [mol/mol].
hsens
jdep
j
k0
kr0
k
N
P
Pr
qw
Sub/superscripts
+
Dimensionless variable.
a, b, c Curve-fit parameters.
bulk
Value in the bulk.
d
Diffusive.
f
Property of the fluid mixture.
i
Species index.
I
Solid-fluid interface.
in
Value at inlet.
reg
Regression value.
Sat
Saturation value.
t
Turbulent.
w
Value at the wall.
INTRODUCTION
Fouling of solid surfaces and heat exchanger surfaces in particular, is a common and much studied problem in most process industries, as reflected in the review paper by MüllerSteinhagen (2011). Fouling is defined as the unwanted accumulation of solid (or semi-solid) material on solid surfaces.
A similar phenomenon is the desired accumulation of solids
e.g. in chemical vapor deposition (Krishnan et al., 1994;
Kleijn et al., 1989). A common and costly problem in many
industrial applications is the direct precipitation of super saturated fluids on heat exchanger surfaces. Typical examples
S. G. Johnsen, T. M. Pääkkönen, S. T. Johansen, R. L. Keiski, B. Wittgens
are found in e.g. the high-temperature off-gas from waste
incineration, metal production, or in power plants, where efficient heat recovery is key to sustainable production, and
where a combination of direct precipitation and deposition
of e.g. solid metal oxides is a major showstopper. Similar
issues can be found in almost all process industries, and in
the current work we study the deposition of a low-solubility
salt (calcium carbonate, CaCO3 ) from liquid water. By precipitation, we understand all types of phase transitions from
a fluid to a relatively denser phase, e.g. gas → liquid (condensation), gas → solid (sublimation), liquid → solid (solidification). For some materials, the precipitate may have a
crystalline structure (crystallization)(e.g. CaCO3 ).
In our modelling work, fouling due to mass deposition from
a fluid phase is grouped into two different classes; 1) particulate fouling, where particles carried by the fluid phase
penetrate through the laminar boundary layer and stick to
the wall (e.g. precipitates, dust, or soot particles) (Johansen,
1991; Johnsen and Johansen, 2009); and 2) direct precipitation where the fluid is super-saturated close to the wall and a
phase-transition occurs at the wall (current paper). The direct
precipitation on solid surfaces is due to the molecular diffusion through the stagnant boundary layer close to the wall.
This is a complex physical process where the diffusion flux
of each species is coupled to the diffusion fluxes and thermodynamic/chemical properties of all the species present. Commonly, a combination of 1 and 2 takes place. Fouling can
only occur if the adhesive forces between the foulant and the
wall are strong enough to overcome the flow-induced shear
forces at the wall.
In previous papers, we developed frameworks for the mathematical modelling of particle deposition and re-entrainment
(Johansen, 1991; Johnsen and Johansen, 2009) and direct precipitation (Johnsen et al., 2015). In presentations
(Johnsen et al., 2010, 2016), it was demonstrated how these
models could be employed as wall boundary conditions
(mass sinks) for CFD models. Pääkkönen et al. (2016) compared CFD simulations with experimental results with respect to CaCO3 deposition in a lab-scale heat exchanger setup. In the current paper we apply the wall function approach
published in (Johnsen et al., 2015), in a coarse grid CFD
model, and test it against the detailed CFD modelling results
and experimental data obtained by Pääkkönen et al. (2012,
2016).
EXPERIMENTAL SETUP
The modelling results are validated against experimental data
from crystallization fouling on a heated surface. The experimental setup includes a flow-loop with a test-section (a
rectangular flow channel), with ohmically heated test surfaces. In the present work, we investigate the case where
the wall heat flux was a constant qw = 52.5kW/m2 . A waterbased test liquid, supersaturated with respect to CaCO3 , is
circulated from a mixing tank and is filtered before entering the test section (average inlet velocities ranging from
u f ,x,in = 0.2 − 0.4m/s and temperature of Tin = 303K), where
CaCO3 precipitates and deposits on the heated test surface.
The growth of the fouling layer is monitored by measuring
the temperature at the test surface. The decreased overall
heat transfer coefficient due to the fouling layer (fouling resistance) will cause the test-section surface temperature to
increase. Details of the experimental setup, procedure and
results were described by Pääkkönen et al. (2012).
Center
Cross‐
section
(a)
(b)
Figure 1: Computational geometry and fine-mesh (a), for two-step
fouling model (Pääkkönen et al., 2016), and coarse mesh
(b), for fouling wall function model.
MODEL DESCRIPTION
In the present paper, CFD is used to model experiments performed in the aforementioned experimental setup. Two different modelling approaches are employed; 1) Two-step fouling model (Pääkkönen et al., 2016); and 2) Fouling wall
function for direct precipitation fouling (Johnsen et al., 2015,
2016). These two differ fundamentally in the way they approach the problem. Model 1 relies on a detailed CFD mesh
close to the wall in order to be able to model the boundary
layer phenomena correctly, and employs the traditional twostep approach (see e.g. (Mullin, 2001)) to model the deposition rate. Model 2, on the other hand, relies on a relatively
coarse mesh, where the cell centers of the cells residing at the
wall are in the log-layer. This approach employs a subgrid
model to calculate the deposition rates from a set of simplified governing equations. For more details, see descriptions
below as well as mentioned references. The main objective
of the current paper is to shed light on the applicability of
the wall function approach, since the successful application
of such a method would be an essential step towards the costefficient modelling of many industry scale applications.
USER-DEFINED WALL FUNCTION FOR DIRECT PRECIPITATION FOULING IN ANSYS FLUENT/ CFD 2017
Geometry and Computational Mesh
Figure 1 presents the 2D geometry used in the CFD simulations. The figure shows a 2D-representation of the liquidfilled gap between two parallel, vertical heat transfer surfaces. The liquid enters from the top and exits through the
bottom. For more details about the experimental set-up, refer to Pääkkönen et al. (2012).
Two different meshes were applied for the two different modelling approaches described below; namely a fine mesh, as
shown in Figure 1a (Pääkkönen et al., 2016), and a coarse
mesh as shown in Figure 1b. The coarse mesh was used
with the wall function model (Johnsen et al., 2015) whereas
the fine mesh was used with the two-step model (Pääkkönen
et al., 2016). With the fine mesh, the y+ value at the surface
is about 0.08, and the total number of cells is 76000. In the
coarse mesh, the y+ value at the wall is between 20 (for the
u f ,x,in = 0.2m/s case) and 36 (for the u f ,x,in = 0.4m/s case),
and the total number of cells is 276. In addition, the wall
function utilizes a 1-dimensional, logarithmic subgrid consisting of 300 computational nodes, the first node at a walldistance equal to 1/10000th of the distance to the cell center in
the coarse CFD mesh (∼ 2.34 · 10−7 m).
Model Fluid
The test liquid in the experiments was a mixture of various
salts dissolved in water. Refer to (Pääkkönen et al., 2015)
for details. In the current modelling work it was assumed
that the test fluid was a pure calcium carbonate, CaCO3 , solution in water. Thus, the mixture was considered as a dilute, electrically quasi-neutral ideal mixture with no chemical reactions. In the present paper, the CaCO3 mass-fraction
of 4.197 · 10−4 kg/kg was used for the test fluid entering
the model geometry. Temperature-dependent fluid properties (mass density, viscosity, diffusivity) were modelled in
accordance with Table 2 in (Pääkkönen et al., 2015).
Fouling Models
Traditionally mass deposition at the wall surface, in crystallization fouling, is modelled based on a two-step approach.
In the two-step modelling approach, the fouling process consists of 1) transport from the bulk to the vicinity of the wall,
and 2) surface integration (i.e. adsorption onto the fouling
layer). The species transport to the vicinity of the crystalfluid interface, is based on the difference between the bulk
and interface concentrations. The mass transfer coefficient is
typically estimated from empirical correlations. At the surface, the integration of the species into the crystal body is
modelled as a pseudo chemical reaction driven by the difference between the interface and saturation concentrations.
When the two steps are combined, the interfacial concentration, which is often unknown, cancels out of the model.
The two-step approach has been used as a stand-alone model
(Bansal et al., 2008; Helalizadeha et al., 2005; Augustin and
Bohnet, 1995) as well as part of a CFD model (Mwaba et al.,
2006; Brahim et al., 2003).
Two-step fouling model
Pääkkönen et al. (2016) implemented the two-step model
into CFD by utilizing the ability of CFD to model the transport of species to the vicinity of the surface, and thus provide the interfacial concentration difference between the surface and the fluid. To account for the wall shear-stress dependency of the adhesion probability seen in experiments
(Pääkkönen et al., 2015), a time scaling factor was included
in the model to scale the fluid residence time at the wall.
The mass deposition rate to the surface, based on the twostep approach, including the effect of the residence time
(Pääkkönen et al., 2015) can be expressed as
jdep = β
1 βρ f u2τ
+ (Cb −CSat ) −
2 kr0 µ f
s
2
βρ f u2τ
1 βρ f u2τ
+ 0
(Cb −CSat ) . (1)
−
4 kr0 µ f
kr µ f
From the experiments, it was determined that the fouling process was controlled by surface integration (Pääkkönen et al.,
2012). Thus, Eq. (1) reduces to
jdep = kr0 (Cb −CSat )2
µf
,
ρ f u2τ
(2)
where the rate constant for the surface integration can be determined from
kr0 = k0 exp (−Ea/R T ) .
(3)
The pre-exponential factor k0 = 1.62 · 1022 m4/kgs2 , and the activation energy Ea = 148kJ/mol were determined from the experiments, for the surface integration controlled fouling process (Pääkkönen et al., 2015). The two-step fouling model
was implemented into CFD as mass and momentum sink
terms.
Fouling wall function
The core idea of the fouling wall function approach is to formulate the species transport equations on one-dimensional
form by applying appropriate approximations and simplifications in the turbulent boundary layer. Next, the simplified
governing equations are solved on a local subgrid for each
grid cell residing at the wall, to obtain the cell-specific deposition mass flux. Thus, the calculated species mass fluxes, at
the wall, can be used as mass sinks in the CFD grid cells next
to the wall.
The set of steady-state governing equations consists of the
Advection-Diffusion equation (ADE) for each species,
∇ · (ρ f Xi u f ) + ∇ · j d,i = 0 ,
(4)
the fluid mixture momentum and energy equations,
∇ · (ρ f u f u f ) = −∇
∇P + ∇ τ + ρ f g ,
(5)
!
∇ · (ρ f hsens, f u f ) = ∇ (k f ∇ T ) − ∇
∑ j i,d hsens,i
,
(6)
i
and the restriction that the mass- and mole-fractions must
sum to unity,
(7)
∑ Xi = ∑ zi = 1 .
i
i
Introducing turbulence, dimensionless variables and appropriate simplifications, the simplified governing equations are
obtained:
+
ν
+
=0
(8)
∂y+ t ρ+f ∂y+ Xi + ∂y+ jd,i,y
Sct
gives the mass-fraction profiles;
∂y+ u+f ,x = 1// µ+ + µt+
(9)
S. G. Johnsen, T. M. Pääkkönen, S. T. Johansen, R. L. Keiski, B. Wittgens
No. of cells across channel
Inlet velocity
0 . 2 m/s
0 . 4 m/s
4
30
49
6
20
36
8
15
25
10
11
20
gives the dimensionless axial fluid mixture velocity profile;
and
i
h
+
+ +
∂y+ T + = 0
∂y+ K(0)
T + K(1)
(10)
gives the dimensionless temperature profile.
N
+
+
+
≡ k+f ,c + k+f ,t ∂y+ ln c+
K(0)
P − Prw ∑ i,d,y cP,i ,
(11)
i=1
and
+
K(1)
≡ k+f + k+f ,t + k+f ,c
(12)
express the dimensionless groups in Eq. (10). For more details, refer to (Johnsen et al., 2015).
Due to the assumed weak effect of thermophoresis (due to
small temperature gradients) and the lack of good estimates
of the thermophoretic diffusivity, only diffusiophoresis (concentration gradient diffusion) was considered in the current
work. Furthermore, it was assumed that the model fluid
could be treated as a dilute, ideal mixture. This reduces
the Maxwell-Stefan diffusion model to the Fickian diffusion
model. The mixture mass density and viscosity was modelled in accordance with (Pääkkönen et al., 2015), while
constant mixture thermal conductivity and specific heat capacity 0.6637W/m2 of and 4182J/kgK , respectively, were used.
The turbulent Schmidt number was set to 1. The MaxwellStefan binary diffusivity was tuned so that the area averaged deposition rate matched that of the experiments, for
the u f ,x,in = 0.2m/s data-point, and was kept constant for the
other inlet velocities. This resulted in a Fickian diffusivity of
3.64 · 10−5 m2/s.
CFD Models
CFD modelling was performed using the ANSYS FLUENT
16.2 CFD software. Turbulence is modelled with the standard k − ε turbulence model. In the fine-mesh CFD model,
the Enhanced Wall Treatment is employed to resolve the near
wall region in the fine mesh model.
Temperature dependent fluid properties were implemented
via user-defined functions (UDFs) in accordance with
(Pääkkönen et al., 2015). The fouling models were also implemented via UDFs and hooked into ANSYS Fluent via the
adjust function hook. Due to the low deposition rates observed, it was expected that the mass transfer to the wall
would have a very small effect on the bulk conditions in the
coarse mesh. Thus, the fouling wall function was not utilized
as a mass source, but was run on a frozen flow field.
Coarse-Mesh Velocity Wall Function
The fouling wall function was designed to work on grids
where the grid cells residing on the wall are in the log-layer.
The main reason for this is that its bulk boundary conditions
were chosen to be valid for fully developed turbulent flow.
In the current experimental set-up, however, due to the low
Reynolds numbers, such a stringent requirement of the wall
20
Dimensionless Velocity Parallel to the Wall, [‐]
Table 1: Wall y+ values at the centre cross-section, for selected
coarse grids with uniform node spacing, for inlet velocities 0.2 and 0.4m/s .
Ashrafian, 2007
Fine‐Mesh
Coarse‐Mesh
Fouling Wall Function ‐ Subgrid Model
18
16
14
12
0.4m/s
10
0.2m/s
8
6
4
2
0
0.1
1
10
100
1000
Dimensionless Distance to the Wall, y+, [‐]
Figure 2: Comparison of dimensionless velocities as functions of
dimensionless wall distance at the center cross-section
(isothermal conditions), for the coarse- (circles) and finemesh (solid lines) CFD models, the fouling wall function
subgrid model (dashed, black line), and theoretical velocity profile (Ashrafian and Johansen, 2007) (dotted, black
line).
y+ left us with very coarse meshes. In Table 1, the approximate wall y+ value at the center cross-section (see Figure 1)
is shown for various coarse meshes where the node spacing
is constant across the channel.
In order to predict the wall shear stress and general velocity profile accurately, on the coarse mesh, the wall function
proposed by Ashrafian and Johansen (2007), was employed;
11.4 arctan y+ ,
y+ ≤ y+
∗
11.4
u+f ,x (y+ ) =
, (13)
+ > y+
1 ln 1+κy++ + u+ (y+
),
y
∗
∗
f
,x
κ
1+κy
∗
with the dimensionless turbulent kinematic viscosity
(
νt+
=
y+
11.4
κy+
2
y+ ≤ y+
∗ ,
y+ > y+
∗
(14)
where dimensionless velocity is defined as u+f ,x = u f ,x /uτ ,
dimensionless wall distance is defined as y+ = uτ y/ν , y+
∗ =
51.98, and κ = 0.42 is the von Kármán constant.
A sensitivity study was done to investigate how the coarse
meshes performed against the fine-mesh CFD model and the
Ashrafian-Johansen wall function. It was determined that the
mesh with 6 cells across the channel reproduced the finemesh velocity and temperature profiles quite well and at the
same time gave an acceptable wall y+ value. In Figure 2,
it is shown how the fine-mesh and coarse-mesh CFD models perform against the profile published by Ashrafian and
Johansen (2007) under isothermal conditions (no heating),
in terms of dimensionless variables. The deviations at high
y+ values are due to the effect of the opposing channel wall
and the relatively low Reynolds numbers investigated. For
the coarse-mesh CFD model, the fouling wall function subgrid model is included for validation of the subgrid velocity
profile. In Figure 3a, the coarse and fine-mesh axial velocity profiles at the center cross-section are compared, and in
Figure 3b, the temperature profiles are compared. It can be
seen that generally, the axial velocity was underpredicted, in
the coarse-mesh CFD model, whereas the temperature was
overpredicted.
USER-DEFINED WALL FUNCTION FOR DIRECT PRECIPITATION FOULING IN ANSYS FLUENT/ CFD 2017
355
0.45
Coarse‐Mesh CFD Model
Fine‐Mesh CFD Model
Coarse‐Mesh CFD Model
Fine‐Mesh CFD Model
350
345
0.35
0.4m/s
340
0.30
Temperature, [K]
Velocity Parallel to the Wall, [m/s]
0.40
0.25
0.20
0.2m/s
0.15
335
330
325
320
315
0.10
0.2m/s
310
0.05
305
0.4m/s
300
0.00
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19
0
1
2
3
4
5
6
7
8
9
Distance to Heated Wall, [mm]
Distance to Heated Wall, [mm]
(a)
(b)
Figure 3: Comparison of parallel-to-wall flow velocity profiles (a) and temperature profiles (b) at the center duct cross-section (x = 100mm),
for coarse and fine-mesh CFD models, for wall heat flux of 52.5kW /m2 and inlet velocities 0.2m/s and 0.4m/s .
Boundary Conditions for the Fouling Wall Function
The fouling wall function requires boundary conditions for
temperature and species mass-fractions at the wall as well as
axial velocity, temperature and species mass-fractions in the
bulk. The bulk values as well as the wall temperature are
taken directly from the CFD model via the inbuilt macro library in ANSYS Fluent, and utilized as Dirichlet boundary
conditions in the subgrid model. The species-specific massfraction boundary conditions at the wall, however, require
special attention. First, the type of boundary condition depends on whether the species is depositing or not; second,
they depend on which diffusive transport mechanisms are
dominating close to the wall (Johnsen et al., 2017).
E.g., consider the case where diffusion due to mass-fraction
gradients (diffusiophoresis) is the sole transport mechanism
close to the wall. For the non-depositing species, the massfraction gradient at the wall must be zero to ensure zero deposition flux, and we employ the Neuman BC for the ADE,
at the wall. For the depositing species, however, we do not
have a priori knowledge of the deposition flux, so we cannot
use the mass-fraction gradient as a BC. We have to use the
Dirichlet BC. That is, we need to specify the mass-fractions
of the depositing species, at the wall.
The mass-fractions at the wall (interface mass-fractions) are
consequences of the balance between transport through the
turbulent boundary layer and the species integration into the
crystal lattice. Therefore, it is a function of e.g. temperature, temperature gradient, composition, composition gradients, wall shear stress, crystal properties, etc. Thus, the interface mass-fraction is not just a fixed boundary condition,
but is in fact part of the solution itself. If the kinetics of the
surface reaction are known, it is possible to estimate the interface mass-fractions. Then, an iterative procedure can be
employed to find the interface mass-fraction that ensures that
the transport rate through the boundary layer and the integration rate into the crystal are identical (Johnsen et al., 2017).
Lacking accurate predictions of the surface reaction rates, the
current wall function model employed interface concentrations obtained from the fine-mesh CFD model (see Figure
4). These concentrations are dependent on both wall temperature and inlet velocity (wall shear stress). By curve fitting
the Logistic function,
a
XI,reg =
,
(15)
1 + (Tw /b )c
Table 2: Curve fit polynomial coefficients for velocity dependence
of interface mass-fractions (Eqs. 16-18).
a
0
1
2
3
b
c
0.251 654 342.436
409.600
1.284 76
26.6133 −2179.69
−3.577 31 −112.872
4968.05
3.414 71
to the fine-mesh CFD data, we obtained good representations
of the interface mass-fractions for each inlet velocity case.
The inlet velocity-dependent fitting parameters, a, b, and c,
are shown in Figure 5, and could be accurately described in
terms of 3rd and 2nd order polynomials;
a
= a0 + a1 u f ,x,in + a2 u2f ,x,in + a3 u3f ,x,in ,
b =
c
=
b0 + b1 u f ,x,in + b2 u2f ,x,in ,
c0 + c1 u f ,x,in + c2 u2f ,x,in .
(16)
(17)
(18)
The coefficients are given in Table 2. Employing Eq. (15)
with coefficients given by Eqs. 16-18, we got a good, general
representation of the CFD-data (see black circles in Figure
5).
Figure 6 shows the temperature dependence of the calculated
interface mass-fraction (Eq. (15)), for selected inlet velocity
cases. It is seen that the interface mass-fraction drops from
close to the bulk value to zero, at a certain threshold temperature, which appears to be dependent on the inlet velocity. In reality, this is a consequence of the complex interplay
between mass deposition rate, interface mass-fraction, wall
temperature, and wall shear stress. We will be content, however, to consider this as an inlet velocity dependent feature.
At temperatures below the threshold, the deposition regime
is interface controlled, whereas at higher temperatures it is
diffusion controlled. Pääkkönen et al. (2012) concluded that
the fouling regime was interface controlled, in these experiments, since the over-all deposition rate is not increasing for
increasing flow-velocities, as would be expected for a mass
transfer controlled fouling regime. However, various segments of the heated wall may be in different fouling regimes
depending on the local flow conditions and wall temperature,
as indicated in Figure 7. In general, the higher the difference
between the bulk and interface mass-fractions, the higher
CaCO3 Mass‐Fractions @ Heated Wall, [10‐4 kg/kg]
S. G. Johnsen, T. M. Pääkkönen, S. T. Johansen, R. L. Keiski, B. Wittgens
4.5
Ux,in=0.4 m/s
4.0
3.5
3.0
2.5
2.0
1.5
1.0
Fine‐Mesh CFD Data
Curve‐Fitted Data
0.5
0.0
0
20
40
60
80
100
120
140
160
180
200
Position Along Streamwise Direction of Heated Wall, [mm]
Figure 4: Comparison of fine-mesh CFD (black circles) and bestfit (red lines) CaCO3 mass-fractions at the wall plotted
against the axial position along the heated wall, for selected inlet velocities.
400
0.42
300
250
0.41
200
0.40
150
a
b
c
Polynomial trend‐lines
0.40
0.39
0.20
0.25
0.30
0.35
100
50
Curve‐Fit Parameter Values (b,c)
Curve‐Fit Parameter Values (a)
350
0.41
0
0.40
Inlet Velocity, [m/s]
CaCO3 Mass‐Fractions @ Heated Wall, [10‐4 kg/kg]
Figure 5: Curve-fit parameters a, b, and c (see Eq. (15)) plotted as
functions of inlet velocity, along with best-fit polynomial
trend-lines (see Eqs. 16-18 and Table 2).
4.5
Inlet Bulk Mass‐frac.
4.0
3.5
3.0
2.5
2.0
1.5
1.0
Figure 7: Wall temperature vs. position along the heated wall, for
the u f ,x,in = 0.333m/s case, for the fine-mesh model
(red), coarse-mesh model (solid black), and adjusted
coarse-mesh wall-temperature (−6.5K) (dashed black).
The relationship between the wall temperature and the
interface mass-fraction is shown by the white curve. The
contour plot in the background corresponds to the interface mass-fraction values at given wall temperatures
(dark gray corresponds to low XI , and light gray corresponds to high XI ) and links the modelled, local wall temperatures with an expected local interface mass-fraction.
prone to overpredict the wall temperature, as was discussed
above (see Figure 3b), there is a risk that the interface massfraction is severely underpredicted if the true wall temperature is lower than, but close to the threshold temperature. To
reduce the risk of overprediction of deposition rates, a fixed
6.5K was subtracted from the wall temperature when calculating the interface mass-fraction from Eq. (15). Figure 7
shows that a greater part of the overpredicted wall temperature curve (solid black) is in the low interface mass-fraction
region (dark gray area) than the fine-mesh model wall temperature curve (red). Hence, a greater part of the wall will
have low interface mass-fraction in the coarse-mesh model
than in the fine-mesh model. The corrected wall temperature
curve (dashed black), however, is more similar to the finemesh model temperature curve. Furthermore, interface massfractions below the saturation mass-fraction indicate that the
fluid is undersaturated at the crystal surface. Physically this
means that deposition is unfavorable with respect to minimizing the Gibbs free energy, thus no deposition will take
place (Johnsen et al., 2017). Therefore, the Dirichlet boundary condition for the CaCO3 was set to
0.5
XI,CaCO3 = max (XI,reg , XSat ) .
0.0
320
325
330
335
340
345
350
355
360
Wall Temperature, [K]
Figure 6: Comparison of temperature dependency of CaCO3 bulk
(red line), saturation (dashed red line), and interface
(calculated for various inlet velocities, from Eq. (15):
black= 0.4m/s, light gray= 0.2m/s) mass-fractions, at the
wall.
the deposition rate (mind that at interface mass-fractions below the metastable equilibrium mass-fraction, fouling might
not take place at all). Thus, the deposition rate at locations with wall temperatures above the threshold can be expected to dominate. Since the coarse-mesh CFD model is
(19)
RESULTS AND DISCUSSION
The ambition in the current work was to demonstrate the applicability of a previously developed fouling wall function
framework (Johnsen et al., 2015), in practice. To approach
this objective, its implementation, as a user-defined function
in ANSYS Fluent 16.2, was employed to demonstrate how
it performs against a more traditional two-step fouling modelling approach (Pääkkönen et al., 2016), in the context of
a well-controlled laboratory experiment (Pääkkönen et al.,
2012).
The main motivation for developing the fouling wall function
was to eliminate the need to resolve the turbulent boundary
layer and enable efficient fouling modelling in industry scale
USER-DEFINED WALL FUNCTION FOR DIRECT PRECIPITATION FOULING IN ANSYS FLUENT/ CFD 2017
7
Position Along Streamwise Direction of Heated Wall, [mm]
6
0
1E‐03
Fine-Mesh Std. K-epsilon
Coarse-Mesh Wall Function Model
5
Local Mass eposition Rate, [kg/m2s]
Mass deposition rate [10-6 kg/m2s]
Experiments
4
3
2
1
0
0.20
0.25
0.30
0.35
50
100
150
200
Fine‐Mesh Model
Coarse‐Mesh Model
1E‐04
1E‐05
0.2m/s
1E‐06
1E‐07
0.267m/s
0.333m/s
0.4m/s
1E‐08
0.40
Inlet Velocity, [m/s]
Figure 8: Comparison of the area averaged mass deposition rates
from the fine-mesh two-step model and the coarse-mesh
fouling wall function model with the experimental data.
CFD simulations. Hence, the modelling framework relies on
relatively high wall y+ values in the CFD cells residing at the
wall. This proved to be challenging in the employment of
the above cited experimental and simulation data, for comparison. Due to the low Reynolds numbers encountered in
the data from Pääkkönen et al., it was necessary to confide in
a very coarse CFD mesh as basis for the fouling wall function
modelling (see Figure 1).
The mass deposition rates predicted by the fouling wall function are depending directly on the wall function boundary
conditions;
• wall mass-fractions for the depositing species,
• bulk and wall temperatures,
• bulk velocity parallel to the wall,
where bulk refers to the center of the CFD grid cells residing
at the wall. Thus, accurate prediction of the deposition rates
rely heavily on the accurate CFD modelling of these quantities.
By utilizing the wall function published by Ashrafian and Johansen (2007), we managed to reproduce the fine-mesh CFD
model velocity and temperature profiles fairly well, qualitatively. However, the quantitative discrepancy turned out to be
the major source of error in the fouling wall function modelling results. In Figure 2, it can be seen how the dimensionless velocity profiles are comparable in the absence of heating, and in Figure 3 it can be seen how dimensional velocity and temperature profiles are comparable under constant
heating of 52.5kW /m2 . The effect on the velocity profiles,
by turning on/off heating was minimal.
Since thermophoresis was neglected in the current work, the
role of the temperature was to provide the temperature dependent fluid properties (mass density, viscosity, and saturation
mass-fraction), and interface mass-fraction for the depositing species. Although the inaccurate prediction of any of
these will affect the predicted mass deposition rate to some
extent, it seemed that the effect of the inaccurate prediction
of the interface mass-fraction was the most severe. As was
indicated in figures 6 and 7, even modest errors in the local
wall temperature could result in a severely miss-represented
interface mass-fraction. Since the mass deposition rate is expected to scale approximately linearly with the difference between the bulk and interface mass-fractions, the mass deposition rate can be off by an order of magnitude by just a slight
overprediction of the wall temperature, as seen in Figure 6.
To avoid underpredicting the interface mass-fraction due to
overprediction of the wall temperature, the temperature was
1E‐09
Figure 9: Comparison of the local mass deposition rates calculated
by the fine-mesh two-step model (lines) and the coarsemesh fouling wall function model (circles), at selected
inlet velocities.
subtracted a fixed 6.5K when calculating the interface massfractions (see Figure 7).
Figure 8 presents a comparison between the experimental
data, the fine-mesh two-step fouling model data, and the data
obtained from the coarse-mesh fouling wall function model.
In the absence of reliable measurements/calculations of the
diffusivity, it was treated as a calibration parameter, for the
fouling wall function. The fouling wall function data were
thus obtained with a tuned diffusivity of 3.64 · 10−5 m2/s, reproducing the u f ,x,in = 0.2m/s experimental data point. The
same, constant diffusivity was used in all grid cells along the
wall, for all the inlet velocity cases. Despite the issues with
predicting the required boundary conditions for the fouling
wall function model accurately, the modelling results compared very well with the results from the fine-mesh two-step
fouling modelling and the experimental data, in terms of the
area averaged mass deposition rate.
In Figure 9, the local deposition rates are compared for the
fine-mesh two-step model and the coarse-mesh fouling wall
function model. It can be seen that even if the area-averaged
values compared well, the local values differs significantly.
The mismatch seems primarily to be due to
• inaccurate prediction of interface mass-fraction;
• inaccurate prediction of wall temperature in the coarse
mesh;
• inaccurate prediction of bulk velocity in the coarse
mesh.
The most crucial improvement to the fouling wall function
model would be to get accurate interface mass-fractions. An
in-depth study of these effects are left to future investigations. In the meantime, we are content to summarize that
the fouling wall function approach performed very well in
a scenario, slightly outside the design specifications of the
modelling framework, with respect to the Reynolds number.
The model fluid used in the current paper is a coarse simplification of the actual fluid employed in the cited experiments. The real fluid was a salt-water solution involving a
multitude of chemically reacting ions and molecules. This
is reflected by the fact that the content of dissolved CaCO3
in the model fluid, is much higher than the saturation concentration. Thus, the modelled CaCO3 may be seen as a
pseudo-component representing e.g. the true CaCO3 fraction in addition to Ca2+ , CO2−
3 and possibly other species.
In the present case, at relatively low concentrations, this simplification seems to be justified in both modelling approaches
S. G. Johnsen, T. M. Pääkkönen, S. T. Johansen, R. L. Keiski, B. Wittgens
employed. However, this may be part of the explanation of
the local difference between deposition rates resulting from
the two modelling methods.
The present demonstration case indicates that in industryscale applications, where very fine meshes are infeasible, the
wall function approach may provide a means to do physically
detailed simulations of complex fluids, in complex geometries, at reasonable computational cost. In particular, if it can
be assumed that the deposition rates are so small that they do
not affect the flow field significantly, the savings in computational cost will be great. Then, the fouling wall function can
be run on a frozen flow-field, and sensitivity studies or optimization studies on e.g. diffusivities, wall surface properties,
etc., that does not affect the macro scale flow-fields can be
performed without the need to update the frozen flow-field.
Establishing the frozen flow-field on the coarse mesh, without the fouling wall function activated is very efficient due
to the low number of computational cells needed. Then, running multiple fouling scenarios can be done on that flow-field
just by changing input parameters to the fouling wall function and running one single CFD iteration, for each fouling
scenario, with the fouling wall function activated.
CONCLUSION
Two different CFD modelling approaches were compared
with experimental data on mass deposition rates in an experimental heat exchanger set-up. The two CFD strategies
resolved the fine length-scales determining the mass transfer
through the turbulent boundary layer, in two different ways:
1) the refinement was done in the 2D CFD mesh, resulting in a relatively high number of grid cells and a wall y+
of ca. 0.08; and 2) the refinement was taken into account
in a wall function utilizing a 1-dimensional subgrid, allowing for a coarse CFD mesh with wall y+ of about 30. The
fine-mesh CFD model utilized a traditional two-step modelling approach for the mass deposition modelling, complemented with the fluid residence-time at the wall, whereas the
coarse-mesh CFD model wall function solved the coupled
Advection-Diffusion, momentum and energy equations on a
local subgrid to estimate the mass deposition rates.
The coarse-mesh model performed very well compared to
the fine-mesh model and experimental data, with respect to
area average deposition rates. Significant mismatch was observed, however, in the local deposition rates. The lacking
accuracy in the coarse-mesh model was mainly due to the
challenges in predicting interface mass-fractions, wall temperatures and bulk velocities, on the very coarse mesh.
The over-all good performance of the coarse-mesh model
gives strong support to the idea that the wall function approach may provide a means to do physically detailed simulations of complex fluids, in complex, industry-scale geometries, at a reasonable computational cost.
ACKNOWLEDGEMENTS
This work was funded by the Research Council of Norway
and The Norwegian Ferroalloy Producers Research Association, through the SCORE project (Wittgens, 2013). Sverre
expresses his gratitude towards the University of Oulu, Environmental and Chemical Engineering, FINLAND, for hosting him and his family during August 2016.
REFERENCES
ASHRAFIAN, A. and JOHANSEN, S.T. (2007). “Wall
boundary conditions for rough walls”. Progress in Computational Fluid Dynamics, 7(2-4), 230–236.
AUGUSTIN, W. and BOHNET, M. (1995). “Influence of
the ratio of free hydrogen ions on crystallization fouling”.
Chemical Engineering and Processing: Process Intensification, 34(2), 79–85.
BANSAL, B.,
CHEN, X.D. and MÜLLERSTEINHAGEN, H. (2008).
“Analysis of ’classical’
deposition rate law for crystallisation fouling”. Chem. Eng.
Process.: Process Intensification, 47, 1201–1210.
BRAHIM, F., AUGUSTIN, W. and BOHNET, M. (2003).
“Numerical simulation of the fouling process”. International
Journal of Thermal Sciences, 42(3), 323 – 334.
HELALIZADEHA, A., MÜLLER-STEINHAGEN, H.
and JAMIALAHMADIA, M. (2005). “Mathematical modelling ofmix ed salt precipitation during convective heat
transfer and sub-cooled flowboiling”. Chemical Engineering
Science.
JOHANSEN, S.T. (1991). “The deposition of particles on
vertical walls”. International Journal of Multiphase Flow,
17(3), 355–376.
JOHNSEN, S.G. and JOHANSEN, S.T. (2009). “Deposition modelling from multi-phase dispersed flow - a boundary
layer wall function approach”. Heat Exchanger Fouling and
Cleaning VIII - 2009.
JOHNSEN, S.G., ÅBERG, M. and JOHANSEN, S.T.
(2010). “Implementation and demonstration of a boundary
condition wall function for industrial scale particulate fouling cfd modeling”. ICMF 2010 7th International Conference
on Multiphase Flow, Tampa, Florida, USA.
JOHNSEN, S.G., JOHANSEN, S.T. and WITTGENS,
B. (2015). “A wall-function approach for direct precipitation/crystallization fouling in cfd modelling”. Heat Exchanger Fouling and Cleaning XI - 2015.
JOHNSEN, S.G., JOHANSEN, S.T. and WITTGENS, B.
(2016). “Implementation and demonstration of a boundary
condition wall function for direct precipitation fouling cfd
modelling”. ICMF 2016 9th International Conference on
Multiphase Flow, Firenze, Italy.
JOHNSEN, S.G., PÄÄKKÖNEN, T.M., ANDERSSON,
S., JOHANSEN, S.T. and WITTGENS, B. (2017). “On the
wall boundary conditions for species-specific mass conservation equations in mathematical modelling of direct precipitation fouling from supersaturated, multi-component fluid mixtures”. arXiv:1703.01448 [physics.flu-dyn].
KLEIJN, C.R., VAN DER MEER, T.H. and HOOGENDOORN, C.J. (1989). “A mathematical model for lpcvd in a
single wafer reactor”. Journal of The Electrochemical Society, 136(11), 3423–3433.
KRISHNAN, A., NING, Z. and PRZEKWAS, A. (1994).
“A computational model for chemical vapor deposition processes in industrial reactors”. 4th InterSociety Conference
on Thermal Phenomena in Electronic Systems (I-THERM),
222–236.
MÜLLER-STEINHAGEN, H. (2011). “Heat transfer
fouling: 50 years after the kern and seaton model”. Heat
Transfer Engineering, 32(1), 1–13.
MULLIN, J.W. (2001). Crystallization. ButterworthHeinemann.
MWABA, M.G., GOLRIZ, M.R. and GU, J. (2006). “A
semi-empirical correlation for crystallization fouling on heat
exchange surfaces”. Applied Thermal Engineering, 26(4),
440–447.
PÄÄKKÖNEN, T.M., RIIHIMÄKI, M., SIMONSON,
C.J., MUURINEN, E. and KEISKI, R.L. (2012). “Crystallization fouling of CaCO3 - analysis of experimental thermal
USER-DEFINED WALL FUNCTION FOR DIRECT PRECIPITATION FOULING IN ANSYS FLUENT/ CFD 2017
resistance and its uncertainty”. International Journal of Heat
and Mass Transfer, 55(23-24), 6927 – 6937.
PÄÄKKÖNEN, T.M., RIIHIMÄKI, M., SIMONSON,
C.J., MUURINEN, E. and KEISKI, R.L. (2015). “Modeling
CaCO3 crystallization fouling on a heat exchanger surface definition of fouling layer properties and model parameters”.
International Journal of Heat and Mass Transfer, 83, 84–98.
PÄÄKKÖNEN, T.M., OJANIEMI, U., PÄTTIKANGAS,
T., MANNINEN, M., MUURINEN, E., KEISKI, R.L. and
SIMONSON, C.J. (2016). “CFD modelling of CaCO3 crystallization fouling on heat transfer surfaces”. International
Journal of Heat and Mass Transfer, 97, 618 – 630.
WITTGENS, B. (2013).
“Score - staged combustion for energy recovery in ferroalloy industry”.
http://www.sintef.no/en/projects/score-staged-combustionfor-energy-recovery-in-fer/.
| 5cs.CE
|
arXiv:1502.06286v1 [cs.PL] 22 Feb 2015
StarL: Towards a Unified Framework for Programming,
Simulating and Verifying Distributed Robotic Systems
ABSTRACT
We developed StarL as a framework for programming, simulating, and verifying distributed systems that interacts with
physical processes. StarL framework has (a) a collection of
distributed primitives for coordination, such as mutual exclusion, registration and geocast that can be used to build
sophisticated applications, (b) theory libraries for verifying
StarL applications in the PVS theorem prover, and (c) an execution environment that can be used to deploy the applications on hardware or to execute them in a discrete event simulator. The primitives have (i) abstract, nondeterministic
specifications in terms of invariants, and assume-guarantee
style progress properties, (ii) implementations in Java/Android that always satisfy the invariants and attempt progress
using best effort strategies. The PVS theories specify the invariant and progress properties of the primitives, and have
to be appropriately instantiated and composed with the application’s state machine to prove properties about the application. We have built two execution environments: one
for deploying applications on Android/iRobot Create platform and a second one for simulating large instantiations of
the applications in a discrete even simulator. The capabilities are illustrated with a StarL application for vehicle to
vehicle coordination in a automatic intersection that uses
primitives for point-to-point motion, mutual exclusion, and
registration.
1.
INTRODUCTION
Programs that monitor and control physical processes over
a network are becoming common in robotics [22, 36, 39],
smart homes [11], and flexible manufacturing [24]. An execution of such a program (consider, for example, a robotic
swarm [36]) is not determined by the underlying computing stack alone, but it also depends on the physical and the
network environment. Therefore, to support useful formal
reasoning about these types of programs the semantic framework should account for the nondeterminism arising from
concurrency, message delays, failures, and the uncertainties
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, or republish, to post on servers or to
redistribute to lists, requires prior specific permission and/or a fee. Request
permissions from [email protected].
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
in the physical world. In swarm robotics, for instance, there
is a big gap between the semantics of the models used for
proving theorems and the real environment in which the
systems run. The former is typically a synchronous network
without message delays, collision-free physics, etc., while actual implementations use ad hoc strategies for dealing with
message losses, noise, and obstacle avoidance, etc.
We are developing the StarL framework [1] to bridge this
gap by providing a nondeterministic programming abstraction that is sufficiently detailed for proving theorems about
reliability and performance of the system, and yet does not
overwhelm the application developers. The core of StarL is
a collection of primitives—mutual exclusion, point-to-point
motion, leader election, geocast, set-agreement, and many
more—that are useful for building distributed robotic applications (StarL applications). Each primitive has a hardwareindependent, abstract, and nondeterministic specification
and an open source implementation in Java. The motionrelated primitives also have platform specific implementations. StarL applications running on robots are written in
Java and use these primitives to accomplish sophisticated
coordination tasks. Example StarL applications we have
built include a distributed search in which a collection of
robots coordinate to find tagets in a building, a light painting application in which a given diagram’s outline is painted
collaboratively by a collection of robots, and a traffic intersection coordination protocol (see Section 3). The primitives
not only allow us to develop verifiable code for distributed
robotics, but offer easy code reuse and maintenance.
The second component of StarL is the StarL PVS library
of theories modeling the abstract specifications of the primitives in the language of the PVS theorem prover [30]. These
specifications are nondeterministic and have two parts. The
first part asserts an invariant property of the primitive and
the second part asserts an assume-guarantee style progress
property [16]. For example, an abstract specification of the
mutual exclusion primitive is parameterized by a set of identifiers for participating processes, and the identity of the
critical section(s). The specification states (a) no two participating processes occupy the critical section simultaneously
(invariant), and (b) that if there exists a time bound within
which acquired critical sections are released (assumption)
then there exists a time bound within which any requesting
process gains access to the critical section. In this paper, we
show how the PVS theorem prover can be used to develop
theories and to verify key invariant properties of StarL applications using the above mentioned primitive theories and
OVERVIEW OF STARL
This work builds up on the work of Zimmerman’s master’s thesis [40, 13]. We concretize the concept of primitives, build the connection to PVS, and develop several new
applications.
Hardware Layers
2.
StarL Portable Layers
their nondeterministic specifications1 .
The third component of StarL is a collection of StarL execution environments that can be used to deploy the applications on actual hardware or in a simulator. We have
built two execution environments for StarL: one for deploying Applications on our Android/iRobot platform and a second one for simulating large instances of the applications in
a discrete even simulator. To our knowledge, StarL is the
first framework that enables the creation of verified software
for distributed robotic systems.
We provide an overview of StarL in Section 2. Then we
illustrate application development in StarL with one detailed
example in Section 3. The PVS translation and verification
of this application in Section 3.3. Section 4 describes some of
the other StarL primitives. Section 5 describes the execution
environments and the simulator. Finally, we discuss related
work in Section 6 and conclude in Section 7.
StarL Application
StarL Primitives
Messaging
Interface
Motion
Interface
Communication
Protocol
Motion
Automation
Communication
Drivers
Motor
Drivers
Communication
Channel
Hardware
Chassis
Application
Layer
Interface Layer
Sensors
Interface
Logic Layer
Position & Sensor
Data Processing
Platform Layer
Sensor/GPS
Drivers
Hardware Layer
Localization
& Sensors
Figure 1: StarL Architecture.
1
2
3
4
5
6
7
8
mc = M otionControl(...);
mc.do_move(t, A);
while (¬mc.motionf lag && ¬mc.f ailed)
// wait
if (mc.motionf lag)
{
// reached
}
Primitives.
Deterministic abstractions are easier to program than nondeterministic ones. Programs for a distributed robotic system, however, have to deal with nondeterminism from communication, dynamics, and failures. We make the choice of
exposing these nondeterminisms to the programmer through
the StarL primitives. We ameliorate the loss of determinism
by making the primitives uniform in the following way: The
StarL architecture defines a special set of write-one, readmany objects that are stored in a part of the heap called the
global variable holder (gvh) for each participating process.
A StarL program interacts with a primitive by invoking a
set of StarL functions that access these StarL objects in gvh.
For example, the StarL M utex primitive implements a distributed mutual exclusion algorithm that allows fixed set of
processes PList to access an object in a mutually exclusive
fashion. A StarL application uses this primitive as follows:
1
2
3
4
5
6
7
8
9
mux = M utex(id, PList); //exclusive with PList
mux.do_mutex(myreq);
while (¬mux.crit && ¬mux.f ailed)
// wait
if (mux.crit)
{
// use
mux.release(myreq);
}
The variables mux.crit and mux.f ailed in process i’s gvh
are written by the mutual exclusion algorithm to indicate
that i does or does not have access to the requested set
myreq, or whether the mutex algorithm has failed. Similarly,
the StarL M otionControl primitive implements a path planning and motion control primitive that interfaces with the
low-level motors and actuators and enables the robot running the process to move towards a target t while avoiding
a region A. An application uses this primitive as follows:
1
Progress properties are verified by composing assumptions
and guarantees but this will be the topic of a future paper.
The variables mc.motionf lag and mc.f ailed in process i’s
gvh are updated by the motion control algorithm to indicate
to the application if i has arrived at the target or whether
the motion control has failed.
Verification.
Each primitive not only has a Java implementation but
also has formal specifications that state their key invariants and assume-guarantee style progress properties. These
properties are written in the language of the PVS theorem
prover [31]. In Section 3.5, we describe how these primitive
theories are composed with the specification of the application to create complete PVS theories that can then be
verified using the PVS prover. In this paper, we focus on
the invariant properties which are proved inductively using the Timed Automaton Library for PVS [3, 5]. The
progress properties involve compositional assume-guarantee
proofs that are commonly used in the proof of self-stabilizing
algorithms [12] and will be the subject of a future paper.
Architecture.
A robot interacts with the physical environment through
sensors and motors. It also interacts with other robots
through the communication channels. All of these constitute the execution environment of a StarL application and
are organized into five layers as shown in Figure 1. The
bottom two layers are hardware platform specific and the
top three layers are portable (see Section for more details).
For deploying StarL Applications on our Android/iRobot
platform, the platform layer implements the functions for
controlling motion of the iRobot Create robots, wireless
communication, and for reading data from the OptiTrack
indoor positioning system. For simulating the applications,
the platform layer is simulated using models of robot motion
and communication channels. The logic layer wraps the low
level methods into high level methods that will be provided
to construct the interface layer. The interface layer consti-
D ZONE (A0)
A ZONE (B1)
A ZONE (A1)
A
B
D ZONE (B0)
D ZONE (C0)
C
D
A ZONE (D1)
A ZONE (C1)
D ZONE (D0)
Figure 2: The four-way automatic intersection.
tutes the global variable holder (gvh) and the various StarL
functions to pass data in and out of rest of the stack. It is
an organized collection of all underlying StarL functionality.
Through the interface layer, applications may access each
part of the framework. The top layer is the application layer.
This includes StarL primitives (Section 4) as well as the
StarL applications. The StarL primitives are constructed
using methods from the interface layer. The StarL application uses both interface layer methods and StarL primitives
to accomplish more complicated tasks.
3.
AUTOMATIC INTERSECTION
We discuss the key facets of StarL with an automatic intersection application. Automatic intersection protocols that
exploit vehicle to vehicle (V2V) communication have been
proposed at various levels of detail in the context of smart
cities and autonomous cars [14, 18]. We use a toy version of
this application to illustrate improvements in programmability and verifiability with StarL.
Automatic intersection layout.
Consider a four-way, double-lane, intersection that will be
navigated by autonomous robotic vehicles through communication (see Figure 2). Each vehicle arrives at one of the
arrival zones A0 , B0 , C0 , D0 with a designated departure
zone A1 , B1 , C1 , D1 . It coordinates with the other vehicle according to a intersection coordination protocol (ICP)
and proceeds to move through a sequence of critical zones
A, B , C , D following certain right-hand traffic rules (e.g., no
backing or U-turns). For example, a vehicle with source destination pair (A0 , D1 ) will have the path A0 , A, C , D, D1 .
The requirements from the system are:
(a) (traffic safety) No two vehicles occupy the same critical
zone at the same time.
(b) (traffic progress) There exists a time-bound within which
every approaching vehicle departs.
We would also like the protocol to permit concurrent safe
traversals. For examples, vehicles with paths A0 , A, A1 and
D0 , D, B , B1 should not block each other.
3.1
Intersection Coordination using StarL
A protocol for intersection works as follows: the participating vehicles agree on the set of participants, then they
request access to the sequence of zones needed for traversal in the intersection from the set of agreed-upon participants; once they have access to the entire sequence, they
start traversing; when a zone is crossed it is released. For
the sake of simplicity, in this presentation we assume that
processes do not fail and robots do not get stuck.
Figure 3 shows the code for implementing this ICP using
StarL primitives and Figure 4 shows the actual Java implementation. Each vehicle participating in the coordination
runs an instance of this protocol with the same identifier
xid that uniquely identifies the intersection. For the process
at vehicle i, the local variable plist i is a list of identifiers
of participating process initialized to the emptylist. The local variable myseq i is a list of zones; it is initialized to the
sequence of zones that i must traverse to go from its current position (mypos) to its destination. Here, mypos is a
StarL variable in the gvh storing the position of the vehicle
and is updated by the location sensors. This protocol uses
two StarL primitives called Registration and M utex. More
details about these primitives, their interfaces, and the conditional guarantees they provide are described in Section 4.
In brief, Registration allows a set of processes in a neighborhood to agree on a subset that contains participating
processes; M utex allows mutually exclusive access to one or
a set of shared resources. reg and mux are instances of these
primitives with the identifier xid . Finally, the loc variable,
of the enumerated type, is initialized to the value S0.
The protocol waits in a loop until loc becomes S0. If loc
is S0 then the do register() function is invoked to start the
registration process and loc is set to reg_wait. If and when
the registration process returns successfully, the StarL variable reg.rList is set to a non-null value. From reg_wait, the
process moves to mutex_wait only if registration completes
(reg.rList nonempty) and in that case the list is copied to
local variable plist and do mutex is invoked to obtain exclusive access to the sequence of zones mid (myseqi ) (except
the first and the last) from the processes in plist. If and
when the mutex process returns successfully, the StarL variable mux.crit is set to true. From mutex_wait, the process moves to move_wait only if mutex returns successfully
(mux.crit true) and in that case do move is invoked which
sends from plist a sequence of points to M otioncontrol. In
move_wait, when the vehicle traverses the zone plist[1 ] and
reaches plist[2 ], the zone plist[1 ] is removed from the list
and release(plist[1]) is called to release that zone to the
mutual exclusion. When the vehicle i reaches its destination
zone, the loc is changed to S1.
3.2
Java Implementation of ICP
Figure 4 shows a fragments of Java implementation of ICP
that highlights the usage of the M utex and M otioncontrol
primitives. Line 1 in Figure 4 (corresponds to line 5 in Figure 3) enumerates the program locations. Line 4 (corresponds to Line 1 in Figure 3) creates the variable plist to
store the list of vehicles that will be returned from in the
Registration primitive. Then it creates a variable (myseq)
to hold the list of wanted zones. It is initialized by computing the sequence of critical zones plus the departure zone the
vehicle needs to go through this intersection. Line 6 (line 3)
creates a instance of the registration primitive in the gvh using the intersection ID. Similarly, a instance of the mutual
1
2
3
4
5
plist: List[PIDS] := {};
myseq: List[Zones] := path(mypos,dest)
reg = Registration(xid);
mux = M utex(xid);
loc enum {S0,reg_wait,mutex_wait,move_wait,S1} := S0
private enum Location {
S_0, REGISTER_WAIT, MUTEX_S, MUTEX_WAIT,
MOVE_WAIT, S_1, DONE, ...
};
LinkedList<ItemPosition> plist;
LinkedList<ItemPosition> myseq = getMyseq();
RegPrim Reg = new Registration(gvh,I_ID);
MutexPrim Mutex = new M_Mutex(gvh, I_ID);
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
while (state != done)
switch case state
S0: loc = reg_wait; reg.do_register();
reg_wait: if reg.rlist != null then
state = mutex_wait;
plist = reg.rList;
mux.do_mutex(mid(myseq), plist);
mutex_wait: if mux.crit = myseq then
state = move_wait;
do_move(plist)
move_wait: if pos ∈ seq[2] then
mux.release(seq[1]); myseq = tail(myseq);
if myseq = [dest] then state := S1
S1: //done
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
Figure 3: General Algorithm.
23
24
exclusion primitive is created in the gvh in the next line.
When access is granted by Mutex (line 23), it sends the
motion command do move and changes loc to move wait.
The the Java code, gvh.plat.moat refers to the motion automation that controls the movements of the robot. The
vehicle stops when it has either reached the neighborhood
of the destination or failed. In this application, since we
want to ensure safety, the program is interrupted if collision is detected. Therefore, the vehicle stops if and only if
it has reached the neighborhood of the destination (line 28,
(line 16)). Then, line 29, (line 17) releases(release) the previous critical zone. Line 34,(line 18) moves (do move) to
the next critical zone in myseq. Additionally, location is
changed to to S 1 if there are no more zone in myseq.
Line 42 waits until the vehicle has reached the neighborhood of the departure zone, then the last critical zone is released and the unRegister method is called. Line 46 freezes
the robot, preventing any more motion command to be executed. Line 53 makes the execution of the while loop wait
so that the states are updated once roughly every 100 milliseconds.
25
26
27
28
29
30
31
32
33
34
@Override
public List<Object> callStarL() {
while(location != DONE) {
switch(location) {
// implementation of some locations not shown
case REGISTER_WAIT:
if(Reg.getList() != null){
plist=Reg.getList();
Mutex.do_mutex(myseq,plist);
location = Location.MUTEX_WAIT;
}
break;
case MUTEX_WAIT:
if(Mutex.od_mutex()){
gvh.plat.moat.doMove(currentDestination);
location = Location.MOVE_WAIT;
}
break;
case MOVE_WAIT:{
if(!gvh.plat.moat.inMotion)
Mutex.release(CSname(preDestination));
preDestination = new
ItemPosition(currentDestination);
myseq.remove()
if(!myseq.isEmpty()){
currentDestination =
(ItemPosition)myseq.peek();
gvh.plat.moat.doMove(currentDestination);
};
else{
location = Location.S_1;
}
35
36
37
38
}
break;
case S_1:
if(!gvh.plat.moat.inMotion) {
Mutex.release(CSname(preDestination));
Reg.unRegister();
preDestination = null;
gvh.plat.moat.motion_stop();
location = Location.DONE;
}
break;
case DONE:
break;
}
sleep(100);
}
39
40
41
42
43
44
45
46
47
48
49
50
3.3
StarL PVS Library
The StarL application code and the primitives can be
translated to the PVS theorem prover’s language of high order logic, for rigorously proving safety and progress properties with appropriate environmental assumptions. Figure 5
shows the key part of the PVS theory specifying a system
running the ICP application. It defines the semantics of
the system in terms of a timed automaton [19]. Although
the pseudo code of Figure 3 and its Java implementation
Figure 4 are for an individual processes, this PVS theory
(together with its supporting and importing theories) specify the behavior of the entire system with arbitrarily number
of asynchronously evolving processes.
3.4
Overview of the PVS Theories
The theory uses the TAME library [4, 6] for modeling
timed automata in PVS. The body of the theory defines the
states, the start states, the actions, and the transitions of
the automaton—a special action dt models the passage of
time. By importing the time machine theory with these
parameters (Line 2), the generic timed automaton theory
51
52
53
54
55
}
Figure 4: ICP Java Implementation.
is instantiated and that gives the instances of the relevant
definitions and theorems (e.g., the notion of reachable states,
invariants, and inductive proof rules) for this model.
The interface and the implementation of each StarL primitive is defined in separate parameterized PVS theories such
as Mutex decls and Registration decls. The Traffic decls theory
imports appropriate instances of these theories. In order to
define a timed automaton that is the composition of several primitives (Mutex and Registration), in this paper we
develop an approach for compositional modeling of timed
automata in PVS. For the sake of brevity, the theory presented here excludes the registration process and we drop
the parts related to timing behavior.
Line 6 defines the state components of this automaton.
time loc and myseq variables correspond to the variables
with the same name in Figure 3. However, notice that
here they are arrays indexed by the process (PID), i.e., they
model the location and the sequence of zones for all the processes in the system. The variable timer is a global clock
used to prove time-bound properties2 The timer move variable is a stopwatch that tracks, for each vehicle, the duration
of physically traversing zones. The state component mux is
the state of the imported Mutual exclusion primitive.
The action datatype defines the names and types of all the
state transition. The enabled (a, s) predicate defines whether
the action a can occur in state s and the trans function
function defines the post-state of a occurring at s. dt models the progress of real time; do mutex (i, Z) models the ith
process requesting the set of zones Z; od mutex (i, Z) models successful completion, i.e., the mutual exclusion primitive granting i access to Z; and release(i, Z) models process
i crossing a zone and releasing it to the mutual exclusion
primitive. Note the enabling condition for completing Mutex (od mutex , Line 22): it is a conjunction of a condition
from the ICP and a condition from Mutex; this captures
composition of the two automata. Similarly in Lines 31-32,
the transition function for the two actions are combining the
transitions of ICP and Mutex.
2
Traffic decls : THEORY BEGIN
IMPORTING Mutex decls[PIDS,Zones ]
4
sd: array[PIDS → (Valid sd?) ]
6
states: TYPE = [# loc: array[PIDS → Locations ],
myseq: array[PIDS → ZoneList ],
timer move: array[PIDS → nonnegreal ],
timer: nonnegreal,
mux: Mutex decls.states # ]
8
10
actions: DATATYPE BEGIN
dt(delta t:nonnegreal): dt?
do mutex(i:PIDS,RS:Zoneset): do mutex?
od mutex(i:PIDS,RS:Zoneset): od mutex?
release(i:PIDS,RS:Zoneset):release?
END actions
12
14
16
18
enabled(a:actions, s:states):bool = CASES a OF
dt(delta t): ...
do mutex(i,RS): loc(i,s) = S0 AND RS =
list2set(Path(sd(i))),
od mutex(i,RS): loc(i,s) = mutex wait AND
Mutex decls.enabled(od mutex(i,RS),mux(s)),
release(i,RS): loc(i,s) = move wait AND RS =
car(myseq(i,s)) AND (NOT myseq(i,s) = null)
ENDCASES
20
22
24
26
trans(a:actions, s:states):states = CASES a OF
dt(delta t): s WITH ...
do mutex(i,RL): s WITH
[loc:= loc(s) WITH [(i):= mutex wait ],
mux:= mutex decls.trans(do mutex(i,RL),mux(s)) ],
od mutex(i,RL): s WITH
[loc:= loc(s) WITH [(i):= move wait ],
mux:= mutex decls.trans(od mutex(i,RL),mux(s)) ],
release(i,RL): IF myseq(i,s) = null THEN s WITH
[loc:= loc(s) WITH [(i):= S1 ] ]ELSE s WITH
[myseq:= myseq(s) WITH [(i):= cdr(myseq(i,s)) ],
timer move := timer move(s) WITH [(i):= 0 ],
mux:= Mutex decls.trans(release(i,RL),mux(s)) ]
ENDIF
ENDCASES
28
30
3.5
Proving Theorems about ICP
The above PVS theory defines a timed automaton model
and its semantics for a system with an arbitrary number
of processes executing the ICP which in turn involves those
processes participating in the M utex primitive. We give a
sketch of the key invariants that are used to prove safety of
ICP using the PVS theorem prover. The Mutex primitive is
not presented in detail in this paper. It involves a set of processes and has a key component critset : [PIDS → Zoneset]
that records the (possibly empty) set of zones that each process has exclusive access to. Its key invariant property is
stated in Line 2 of Figure 6. It asserts that at any reachable
state s of the Mutex primitive, for any pair of processes i
and j, crit set(i, s) ∩ crit set(j , s) = ∅.
The next inductive invariant (Line 8) states what we found
to be the key property needed for proving safety of ICP: it asserts that for any process j, (a) if j is in S0, then myseq(j, s)
is the list of zones from its current position to the destination, (b) if j is in mutex wait, then myseq(j, s) is same as the
list in (a) and it has requested to M utex exclusive access to
all the elements in this list, and (c) if j is in move wait, (i.e.,
M utex has completed and j is moving), then myseq(j, s) is a
subset of crit set(j, s). Using this invariant, the main safety
invariant (Line 10) is proved: it states that for any two processes that are moving, the occupy different zones.
4.
STARL PRIMITIVES
StarL currently includes the implementation of the following primitives: path planning, distributed path planning,
geocast, leader election, registration, mutual exclusion, barrier synchronization. In this section, we enumerate the interfaces and specifications for some of them.
Some of the specifications involve timing properties that
are stated with respect to certain intervals defined over realtime. These intervals are defined using constants d, d1 , d2 ,
2
The details of timing analysis will be presented in a future
paper.
32
34
36
38
40
42
44
IMPORTING
time machine[states,actions,enabled,trans,start ]
END Traffic decls
Figure 5: PVS theory for (part of) ICP.
etc. The role of these constants play in verification are different from the role in implementation. For verifying StarL
applications, these constants appear as existentially quantified parameters in the lemma statements (see Line 14). We
assume that certain progress making events happen within
some time bound, e.g., delivery of messages, to prove existence of time bounds of other events such as traversal
through intersection. In the actual implementation of the
primitives, the progress time bounds may be violated, but
they still provide a guideline tuning best-effort strategies.
4.1
Motion Control
The M otioncontrol primitive allows the application to direct the robot towards a specific target point while avoiding
a bad region. Both the target and the region are specified in
the current coordinate system. The motion completes successfully if the robot reaches a neighborhood of the target
while avoiding the bad region and this is indicated to the
program. The interface includes:
(a) htarget, avoidi variable pair in the gvh that stores the
target and the bad region,
(b) gotopoint function is invoked to set target and avoid,
2
4
6
8
10
Inv mux safety(s):bool = FORALL (i,j): NOT (i = j)
IMPLIES disjoint?(crit set(i,s),crit set(j,s))
lemma mux safety: LEMMA FORALL (s): reachable(s) ⇒
Inv mux safety(s);
Inv list crit(s):bool = FORALL (j):
(loc(j,s) = move wait IMPLIES
subset?(list2set(myseq(j,s)),crit set(j,s))) AND
(loc(j,s) = mutex wait IMPLIES list2set(myseq(j,s))=
req set(j,s) AND myseq(j,s) = Path(sd(j))) AND
(loc(j,s) = S0 IMPLIES myseq(j,s) = Path(sd(j)))
lemma list crit: LEMMA FORALL (s): reachable(s) ⇒
Inv list crit(s);
Inv safety(s):bool = FORALL (i,j): NOT (i = j) AND
loc(i,s) = loc move wait AND loc(j,s) = move wait
IMPLIES (NOT (car(myseq(j,s)) = car(myseq(i,s)))
OR myseq(j,s) = null OR myseq(i,s) = null)
lemma safety: LEMMA FORALL (s): reachable(s) ⇒
Inv safety(s);
12
14
16
% Progress lemmas
lemma move entry: LEMMA FORALL (s): FORALL (i):
EXISTS (d1 :nonnegreal): reachable(s) AND timer(s)
> = d1 IMPLIES loc(i,s) = loc move wait;
lemmap rogress: LEMMA FORALL (s): FORALL (i):
EXISTS (d2 :nonnegreal): reachable(s) AND timer(s)
> = d2 IMPLIES loc(i,s) = loc s1;
Figure 6: PVS theory with key ICP invariants.
(c) motionf lag is a boolean variable that is set to done
when the motion completes successfully, and it is set
to fail to indicate that the lower-level motion controller
cannot move the robot to the target.
The following properties summarize the specification of Motion control.
(a) (safety) The position of the robot is always outside the
region in avoid.
(b) (progress) If motionf lag is set to true then the position
of the robot is locates near target.
The motion control primitive is implemented using lowerlevel control and path planning algorithms. More details
about implementations are provided in Section 5.2.
4.2
Geocast and Broadcast
The Geocast primitive allows a process to send a message m to all other processes/robots in its’ neighborhood A;
here A is defined by distance, and d is a timing parameter.
The following properties specify the behavior of the geocast
primitive. The interface includes
(a) do geocast(m, A, d) function to start geocast of message
m over area A with timing parameter d (explained below),
(b) hGcastf lagi is a variable in the gvh that indicates that
the geocast has completed.
The following properties summarize the properties of the
primitive. If a message m is send through geocast at time
t0 then the following hold:
(a) (exclusion) Any process continuously located outside A
during the time interval [t0 , t0 + d] will not deliver m.
(b) (inclusion) Any process located within A during the time
[t0 , t0 + d] will receive m within d time of the geocast.
For a robot moving in or out of A during the geocast period, the message may or may not be delivered; but a robot
outside A is guaranteed not to receive the message. The
implementation of geocast over a wireless network involves
details like tagging the message with the location of the originating process before sending, resending messages in the absence of acknowledgments, and dropping the messages based
on the receiver’s location. Of course, (b) can only be guaranteed under additional assumptions about messages being
delivered in a timely fashion.
BCast(m, d) or broadcast is a special geocast in which
the A defines the entire network. The second condition then
requires that all process that are non-faulty over the interval
[t0 , t0 + d] receive the message.
4.3
Registration
The Register primitive solves a set-valued distributed consensus problem for a set of processes to agree on the identity
of the participants. If registration completes successfully,
then the agreed upon set contains a process’s identifier if
and only if it is a participating process. The interface includes:
(a) Register function for creating a register object,
(b) do register function for starting registration,
(c) hrList, tsi pair stores in the gvh; rList is the agreed
set and ts is the time-stamp for when the computation
finishes; otherwise rList stores a null value.
The following properties summarize the nondeterministic
specification of the Register primitive.
(a) (agreement) For any two processes i and j with agreement timestamps (ts) within d of each other, the corresponding rList’s are identical.
(b) (soundness) For any process i, i is contained in rList
with time stamp ts only if i invoked do register at most
d1 time before t.
(c) (progress) For any process i, if i invokes do register then
within at most d2 time registration completes with i,
that is, rList contains i.
The Register is implemented using the Geocast primitive.
To support multiple registered lists inside an application,
each registration object is invoked with an identifier. A registered process may unregister from the list and this essentially restarts a registration process among the remaining
processes. The rList value can be updated with a new time
stamp and in the interim it may be null .
4.4
Leader Election
The Election primitive elects a leader and conveys the
leader’s identity to set of participating processes. If the election fails then the participating processes learn about this
as well. The interface includes:
(a) Election function for creating an election object; it takes
the list of participants as a parameter,
(b) do election function starts the election,
(c) hLeaderi stores the identity of the leader in the gvh,null
if the election is in progress, and fail if the election fails.
The following properties summarize the nondeterministic
specification of the Election primitive.
(a) (agreement) For any two processes i and j that start
election within d time of each other, if Leader is not
null or fail for either of the two processes, then Leader
has identical value for both.
(b) (soundness) For any process i, Leader = i only if i invoked do election at most d1 time before t.
(c) (progress) For any process i, if i invokes do election then
within at most d2 time election completes successfully,
that is, Leader equals a valid identifier.
Currently, one of the implementations of leader election
is based on randomized ballot creation and a second implementation is based on a version of the Bully algorithm [9].
4.5
Mutual Exclusion
The M utex primitive allows a fixed set of processes to
access an object (or a set of objects)in a mutually exclusive
fashion. If a process requests multiple objects, then it gains
access to all of them at the same time, but it may release
them one at a time. The interface includes:
(a) M utex function for creating an mutual exclusion object
for a list of participating processes and a list of critical
sections.
(b) do election is invoked to request a set of critical sections,
(c) crit stores in gvh a boolean value indicating whether
access to all the requested critical sections have been
granted to this process.
The following properties summarize the specification of the
Mutex primitive.
(a) (safety): For any two processes, the set of critical sections they have access to are disjoint.
(b) (progress): if there exists a time bound d1 within which
critical sections are released then there exists a time
bound d2 within which any requesting process gains access to its critical section(s).
(c) (non-interference): If no process holds the critical sections being requested by i, then i gains access with time
d3 (d3 d2 ).
The intersection coordination protocol described in Section 3
uses the Mutex primitive. Currently, mutual exclusion is
implemented using a modification of Ricart & Agrawala’s
algorithm [9].
In summary, all the primitives provide a same type of abstraction to the programmer: an set of invariant properties
that restrict what the nondeterministic environment can do,
and a set of assume-guarantee style progress property. The
primitives are invoked by calling the interface functions, and
progress can be detected by reading the appropriate variables in the gvh.
5.
EXECUTION ENVIRONMENTS
In order to run a StarL application on a hardware platform or inside a simulation environment, it has to be connected with an execution environment. We have developed
two execution environments: (1) for running applications
on a collection of Android smart phones that control iRobot
Create robots and (2) for simulating the applications in a
discrete event simulation environment. Recall, the execution environment define the lowest two layers of Figure 1
(platform and physical layers), and the rest of the software
stack is portable.
5.1
Deploying Applications on HW Platforms
For deploying StarL Applications on our Android/iRobot
platform, the platform layer implements the functions for
controlling motion of the iRobot Create robots, wireless
communication, and for reading data from the OptiTrack
indoor positioning system.
The sensor data and the location data from the positioning system are processed, filtered, through the different layers and are recorded in currentlocation variables in
the gvh. When the application calls the do move() in the
M otioncontrol, a motion controller is started that decides
when and what command to send to robot while making use
of positioning data and sensor data and updating motionf lag
in the gvh. When the controller decides that the robot has
to go straight or arc, the platform layer issues the appropriate wheel speed command to the iRobot Create chassis.
The motion interface also provides underlying motion automation. For example, one can specify a robot’s type so
that the robot can behave differently when it collides with
an object. There are implementations of stop on collision,
back away from collision point, discover objects around the
initial collision point.
The message interface provides basic send and receive
functions over a Wi-Fi network using our built in protocols. These low-level functions are used to build the Geocast
primitive.
5.2
Simulating Applications
The same StarL code can also be simulated in a discrete
event simulator that we have built. This is useful for testing
applications on many virtual robots and without a hardware
platform. resume here. The simulator features a custom
implementation of the platform layer which directs motion,
message, and trace commands into a coordinating thread
referred to as the simulation engine. The simulator can execute an arbitrary number of copies of a StarL application
code to run and interact simultaneously through simulated
messages and robotic platforms.
The StarL simulator allows a developer to run an application under a broad range of conditions and with any number of participating robots. A visualizer displays the current
position of each agent and can be extended to display additional application specific information. Even we could now
produce a simulation environment same to the real robotic
platform, the challenges robots face are realistic. A large
set of simulating parameters can be tuned. Message delays, message loss rate, obstacles in the physical environment, robots crash failures and even adversary robots are
among the tunable simulation parameters.
Creating the simulation in StarL is simple. Using our
simulation template, one need to specify the application (fig-
Figure 9: The light painting simulation.
Figure 7: A snapshot of the ICP simulation.
Figure 8: A simulation of the distributed search application.
ure 4) to simulate along with some simulation parameters.
For example, one can simulate ICP with 4 robots, 100 milliseconds average message delay, the obstacles in the physical environment, shown in figure reffig:sim1. One can also
customize the visualizer to display some extra application
specific information, such as the state of the robots.
5.3
The ICP Application
Screen shot for simulating ICP general solution using four
robots is shown in figure 7. Robot 2 starts in B0, and intends to turn left. Robot 1 starts in D0 and intends to go
straight. Robot 0, starting at C0, and robot 3, starting in
A0, both intend to turn right. The dotted lines are intended
zone sequence. Robot 1 and robot 2 are in the intersection
concurrently since their set of critical zones are disjoint.
5.4
Other Applications
There are four other demo applications in StarL, including Race App, Maze App, Distributed Search App, Light
Painting App. Each of them demonstrates some aspects of
the StarL primitives.
In Race App, there is a sequence of destination points.
Every robot picks the same destination point and tries to
reach it before any other robot does. When robots collide
with each other, they stop and turn until they are facing
away from each other. When a robot reaches one destination
point, it announces that through Broadcast so that every
robots starts to race to the next destination point. This
application demonstrates how to make use of the motion
and communication interface.
In Maze App, robots are put into a Maze like environment,
which contains obstacles shown to robots and obstacles hidden to robots. The robot’s goal is to navigate through the
maze thus reaching the destination point. The robot uses the
path planning primitive to find a possible path to the destination. When the robot’s bump sensor detects an unseen
obstacle, the robot updates its’ obstacle map and recalculates path to the destination. This application demonstrates
the path planning primitive as well as different built-in motion automation.
In Distributed Search App, a group of robots search a
house to find an item. They first start the leader election
primitive to elect a leader. Then the leader assigns rooms
to each robot. Each robot goes to its assigned rooms and
searches for the item. If the item is found, the robot announces it’s finding to the group. A simulation screen shot
is shown in figurer̃effig:sim2. The item to be found is at the
top right corner shown in a blue circle. Thin gray lines are
robots’ movement traces. The first three robots have entered
their assigned room and started searching, the pink robot is
moving towards it’s assigned room at the bottom right corner. The yellow robot is still waiting for its assignment from
the leader.
In Light Painting App, a simple diagram is given to a
group of robots. The robots will try to plan their path
to paint the lines in the diagram, with one or more colors.
This application makes use of the distributed path planning
primitive. A simulation for drawing a cube is shown in figurer̃effig:sim3. The red tube is the distributed path planning
reach tube for each robot. The painted lines are shown in
green. On the left, the robots started to paint; on the right,
the robots have finished the painting.
6.
RELATED WORK
Robotic systems and theory.
There is a large body of theoretical work spanning control
theory, computer science, and robotics that deals with development of distributed algorithms for flocking, coverage,
and formation control for robotic swarms [10, 38, 33, 28, 25,
8, 35]. The safety and convergence properties of algorithms
are typically analyzed by hand (as opposed to verified with
a computer), under various simplifying assumptions. Control theorists and roboticists typically capture the details of
the dynamics of the robots and abstract away the communication delays and issues arising from asynchrony, while the
computer scientists make the complementary assumptions.
In the last ten years, these algorithms have been used to
create spectacular robotic systems [36, 26] and demonstrations [29, 23] for SLAM, flocking, collaborative search, and
even construction. In building these systems, each group
uses its own specific, and often proprietary hardware and
software architecture to implement the algorithms, with limited scope for reuse and no support for formal reasoning. In
fact, currently there are no frameworks or tools supporting
modular design, implementation, and formal verification of
distributed robotic systems.
Programming languages.
Currently robotic systems are programmed using standard programming languages like C, C++, and Python. It
is also common to design low-level controllers using MATLAB/Simulink and then automatically generate C-code. The
Robot Operating System (ROS) [34] provides a popular set
of libraries for building applications. It provides device drivers,
message-passing and other low-level libraries for interfacing
with sensors and actuators, and therefore, it could be used
to build the lower layers in a StarL application. Several
synchronous programming languages like Lustre [15], Esterel [7], Signal [21], and the Time-triggered framework [37]
have been developed over the past two decades. These languages are not only used in practice for signal processing,
automotive, aerospace, and manufacturing applications, but
they also provide strong formal semantics and support for
verification. However, they all provide a deterministic programming abstraction and in one way or another we found
them to be too restrictive for distributed robotic systems
that work in highly dynamic environments.
Formal verification for distributed systems.
There is a large body of work on formal models for distributed systems or communicating state machines. A very
general framework with limited verification support through
PVS is the hybrid I/O automaton framework [19, 27]. Differential dynamic logic [32] with the related Keymera theorem prover is another well-developed framework. There are
several less expressive models that have been developed for
completely automatic verification under the umbrella of parameterized verification (see, for example, [2, 20, 17] and the
references therein).
7.
CONCLUSIONS
We presented what is to our knowledge a design of the
first programming framework for distributed robotic systems that also supports simulations and rigorous verification. Since a robotic system is essentially an open system
with many sources of nondeterminism, our primitives sacrifice determinism in the programming abstraction and instead provide a uniform way of interacting with physical
environment, communication channels and other programs.
The proposed StarL framework also provides theory libraries
for verifying StarL applications in the PVS theorem prover,
and two execution environments: one that is used to deploy
the applications on smart phones that control robots, and
the other for running discrete event simulations with many
participating robots. The capabilities are illustrated with
a StarL application for vehicle to vehicle coordination in a
automatic intersection that uses StarL primitives for pointto-point motion, mutual exclusion, and registration.
The future directions of research include expansion of the
StarL-PVS library further to include failure models and to
support the verification of progress properties. Another direction is to develop a compiler for generating both the PVS
theories (Figure 5) and the Java implementation (Figure 4)
from the StarL programs (Figure 3).
Acknowledgments
We thank Adam Zimmerman for developing and documenting an earlier version of StarL development for his masters
thesis research and Nitin Vaidya for several valuable discussions. This work is sponsored in part by the National
Science Foundation.
8.
REFERENCES
[1] The StarL framework, February 2015. Available for
download from
https://github.com/lin187/StarL_LCTES.
[2] P. A. Abdulla and B. Jonsson. Verifying networks of
timed processes. In Tools and Algorithms for the
Construction and Analysis of Systems, pages 298–312.
Springer, 1998.
[3] M. Archer. TAME: PVS Strategies for special purpose
theorem proving. Annals of Mathematics and Artificial
Intelligence, 29(1/4), February 2001.
[4] M. Archer, C. Heitmeyer, and S. Sims. TAME: A PVS
interface to simplify proofs for automata models. In
Proceedings of UITP ’98, July 1998.
[5] M. Archer, H. Lim, N. Lynch, S. Mitra, and
S. Umeno. Specifying and proving properties of timed
I/O automata in the TIOA toolkit. In In Fourth
ACM-IEEE International Conference on Formal
Methods and Models for Codesign (MEMOCODE’06).
IEEE, 2006.
[6] M. Archer, H. Lim, N. Lynch, S. Mitra, and
S. Umeno. Specifying and proving properties of timed
I/O automata using Tempo. Design Automation for
Embedded Systems, 2008.
[7] G. Berry and L. Cosserat. The esterel synchronous
programming language and its mathematical
semantics. In Seminar on Concurrency,
Carnegie-Mellon University, pages 389–448, London,
UK, UK, 1985. Springer-Verlag.
[8] J. Cortes, , S. Martinez, T. Karatas, and F. Bullo.
Coverage control for mobile sensing networks. IEEE
Transactions on Robotics and Automation,
20(2):243–255, 2004.
[9] G. Coulouris, J. Dollimore, T. Kindberg, and G. Blair.
Distributed Systems: Concepts and Design.
Addison-Wesley Publishing Company, USA, 5th
edition, 2011.
[10] X. Défago and A. Konagaya. Circle formation for
oblivious anonymous mobile robots with no common
sense of orientation. In Proc. 2nd Int’l Workshop on
Principles of Mobile Computing (POMC’02), pages
97–104, Toulouse, France, October 2002. ACM.
[11] C. Dixon, R. Mahajan, S. Agarwal, A. Brush, B. Lee,
S. Saroiu, and P. Bahl. An operating system for the
home. In NSDI. USENIX, April 2012.
[12] S. Dolev. Self-stabilization. MIT Press, Cambridge,
MA, USA, 2000.
[13] P. S. Duggirala, T. T. Johnson, A. Zimmerman, and
S. Mitra. Static and dynamic analysis of timed
distributed traces. In RTSS, pages 173–182, 2012.
[14] M. R. Hafner, D. Cunningham, L. Caminiti, and D. D.
Vecchio. Cooperative collision avoidance at
intersections: Algorithms and experiments. IEEE
Transactions on Intelligent Transportation Systems,
14(3):1162–1175, 2013.
[15] N. Halbwachs, P. Caspi, P. Raymond, and D. Pilaud.
The synchronous dataflow programming language
lustre. In Proceedings of the IEEE, pages 1305–1320,
1991.
[16] T. A. Henzinger, S. Qadeer, and S. K. Rajamani.
Decomposing refinement proofs using
assume-guarantee reasoning. In Proceedings of the
2000 IEEE/ACM International Conference on
Computer-aided Design, ICCAD ’00, pages 245–253,
Piscataway, NJ, USA, 2000. IEEE Press.
[17] T. Johnson and S. Mitra. Parameterized verification of
distributed cyber-physical systems:an aircraft landing
protocol case study. In ACM/IEEE Third
International Conference on Cyber-Physical Systems,
April 2012, Beijing, China, 2012.
[18] T. Johnson, S. Mitra, and K. Manamcheri. Safe and
stabilizing distributed cellular flows. In Proceedings of
IEEE Internaitonal Conference on Distributed
Computing Systems (ICDCS 2010), 2010.
[19] D. K. Kaynar, N. Lynch, R. Segala, and
F. Vaandrager. The Theory of Timed I/O Automata.
Synthesis Lectures on Computer Science. Morgan
Claypool, November 2005. Also available as Technical
Report MIT-LCS-TR-917.
[20] P. Krcál and W. Yi. Communicating timed automata:
The more synchronous, the more difficult to verify. In
Computer Aided Verification, 18th International
Conference, CAV, volume 4144 of Lecture Notes in
Computer Science, pages 249–262. Springer, 2006.
[21] P. Le Guernic, A. Benveniste, P. Bournai, and
T. Gautier. Signal–a data flow-oriented language for
signal processing. Acoustics, Speech and Signal
Processing, IEEE Transactions on, 34(2):362–374,
1986.
[22] Q. Lindsey, D. Mellinger, and V. Kumar. Construction
with quadrotor teams. Auton. Robots, 33(3):323–336,
2012.
[23] Q. Lindsey, D. Mellinger, and V. Kumar. Construction
with quadrotor teams. Autonomous Robots,
33(3):323–336, 2012.
[24] M. R. Lucas and D. M. Tilbury. A study of current
logic design practices in the automotive manufacturing
industry. Int. J. Hum.-Comput. Stud., 59(5):725–753,
2003.
[25] M. Mesbahi and M. Egerstedt. Graph-theoretic
Methods in Multiagent Networks. Princeton University
Press.
[26] N. Michael, D. Mellinger, Q. Lindsey, and V. Kumar.
The grasp multiple micro-uav testbed. Robotics &
Automation Magazine, IEEE, 17(3):56–65, 2010.
[27] S. Mitra. A Verification Framework for Hybrid
Systems. PhD thesis, Massachusetts Institute of
Technology, Cambridge, MA 02139, September 2007.
[28] R. Olfati-Saber, J. Fax, and R. Murray. Consensus
and cooperation in networked multi-agent systems.
Proceedings of the IEEE, 95(1):215–233, January 2007.
[29] R. Oung and R. DâĂŹAndrea. The distributed flight
array. Mechatronics, 21(6):908–917, 2011.
[30] S. Owre, S. Rajan, J. Rushby, N. Shankar, and
M. Srivas. PVS: Combining specification, proof
checking, and model checking. In R. Alur and T. A.
Henzinger, editors, Computer-Aided Verification, CAV
’96, number 1102 in LNCS, pages 411–414, New
Brunswick, NJ, July/August 1996. Springer-Verlag.
[31] S. Owre, N. Shankar, J. M. Rushby, and D. W. J.
Stringer-Calvert. PVS Language Reference. Computer
Science Laboratory, SRI International, Menlo Park,
CA, Sept. 1999.
[32] A. Platzer. Differential logic for reasoning about
hybrid systems. In A. Bemporad, A. Bicchi, and G. C.
Buttazzo, editors, HSCC, volume 4416 of Lecture
Notes in Computer Science, pages 746–749. Springer,
2007.
[33] G. Prencipe. Corda: Distributed coordination of a set
of autonomous mobile robots. In ERSADS, pages
185–190, May 2001 2001.
[34] M. Quigley, K. Conley, B. P. Gerkey, J. Faust,
T. Foote, J. Leibs, R. Wheeler, and A. Y. Ng. Ros: an
open-source robot operating system. In ICRA
Workshop on Open Source Software, 2009.
[35] M. Schwager, J. McLurkin, and D. Rus. Distributed
coverage control with sensory feedback for networked
robots. In Robotics: Science and Systems,
Philadelphia, Pennsylvania, August 2006. The MIT
Press.
[36] C. Steiner. Bot in the delivery:kiva systems. Forbes
Magazine, March 2009. http://www.forbes.com/
forbes/2009/0316/040_bot_time_saves_nine.html.
[37] W. Steiner and B. Dutertre. Automated formal
verification of the TTEthernet synchronization quality.
In NASA Formal Methods - Third International
Symposium, NFM, volume 6617 of Lecture Notes in
Computer Science, pages 375–390. Springer, 2011.
[38] I. Suzuki and M. Yamashita. Distributed autonomous
mobile robots: Formation of geometric patterns.
SIAM Journal of computing, 28(4):1347–1363, 1999.
[39] M. Turpin, K. Mohta, N. Michael, and V. Kumar.
Goal assignment and trajectory planning for large
teams of interchangeable robots. Auton. Robots,
37(4):401–415, 2014.
[40] A. Zimmerman. Starl for programming reliable robotic
networks. Master’s thesis, University of Illinois at
Urbana-Champaign, 2013.
| 6cs.PL
|
Are Clouds Ready to Accelerate Ad hoc Financial Simulations?
arXiv:1412.4556v1 [cs.DC] 15 Dec 2014
Blesson Varghese and Adam Barker
School of Computer Science, University of St Andrews, UK
Email: {varghese, adam.barker}@st-andrews.ac.uk
Abstract—Applications employed in the financial services
industry to capture and estimate a variety of risk metrics are
underpinned by stochastic simulations which are data, memory
and computationally intensive. Many of these simulations are
routinely performed on production-based computing systems.
Ad hoc simulations in addition to routine simulations are
required to obtain up-to-date views of risk metrics. Such
simulations are currently not performed as they cannot be
accommodated on production clusters, which are typically over
committed resources. Scalable, on-demand and pay-as-you go
Virtual Machines (VMs) offered by the cloud are a potential
platform to satisfy the data, memory and computational
constraints of the simulation. However, “Are clouds ready to
accelerate ad hoc financial simulations?”
The research reported in this paper aims to experimentally
verify this question by developing and deploying an important
financial simulation, referred to as ‘Aggregate Risk Analysis’
on the cloud. Parallel techniques to improve efficiency and performance of the simulations are explored. Challenges such as
accommodating large input data on limited memory VMs and
rapidly processing data for real-time use are surmounted. The
key result of this investigation is that Aggregate Risk Analysis
can be accommodated on cloud VMs. Acceleration of up to 24x
using multiple hardware accelerators over the implementation
on a single accelerator, 6x over a multiple core implementation
and approximately 60x over a baseline implementation was
achieved on the cloud. However, computational time is wasted
for every dollar spent on the cloud due to poor acceleration
over multiple virtual cores. Interestingly, private VMs can offer
better performance than public VMs on comparable underlying
hardware.
Keywords-cloud computing; heterogeneous computing; Aggregate Risk Analysis; financial risk; risk simulation
I. I NTRODUCTION
Financial applications in the risk industry are underpinned
by large-scale stochastic simulations which are data, memory and computationally intensive [1]. These simulations are
run on a weekly, monthly or quarterly basis on production
systems such as commodity clusters to generate risk metrics
including Probable Maximum Loss (PML) [2] and Tail
Value-at-Risk (TVaR) [3] due to catastrophic events such
as earthquakes, hurricanes and floods. The results obtained
from these simulations are then interpreted by actuaries for
key decision making and planning a financial year.
The simulations that run on a routine basis are sufficient
if the risk metrics do not have to be updated before their
next run. Consider a simulation that takes into account the
fluctuation of one parameter, for example, currency, on a
weekly basis. The simulation is performed on a weekly
basis to update the risk metrics. However, this is not the
case in real-time scenarios where risk metrics will need
to be obtained on an ad hoc basis before the next routine
run. For example, consider real-time online pricing when
an underwriter needs to close a deal with a client over the
telephone.
Production systems are not the right type of infrastructure
for simulating on an ad hoc basis since they are firstly optimised to run routine simulations and accommodate data of
known sizes, and secondly, over committed and at best fully
utilised resources with no scope to satisfy the data, memory and computational requirements of ad hoc simulations.
Consequentially, even if ad hoc simulations are performed
on production clusters they tend to be slow. One solution
to this problem would be to use dedicated systems for ad
hoc simulations. However, this is not always possible since
there is an additional investment on top of the maintenance
costs of production clusters. An alternative solution to reduce
the cost of investment is by using hardware accelerators
as coprocessors in heterogeneous clusters [4], [5]. Though
computation can be accelerated to suit ad hoc simulations the
memory and data requirements cannot be always satisfied.
Hardware accelerators have limited memory and thereby
cannot handle large data in memory.
Cloud computing infrastructure has the potential to address the above challenges. Maintenance costs can be eliminated and resources can be scaled on-demand, which can
satisfy the requirements of ad hoc risk simulations [6],
[7]. In this paper, the research question, “Are clouds ready
to accelerate ad hoc financial simulations?” is explored.
One application, namely Aggregate Risk Analysis, widely
employed in the risk industry is developed and deployed
on the cloud. Parallel techniques to accelerate the analysis
and techniques to efficiently accommodate data and handle
memory on cloud VMs are investigated. The experimental
studies on the cloud indicate that the application achieves
up to a 60x acceleration on VMs with hardware accelerators
but with poor acceleration due to wasted computational time
per dollar spent. Nevertheless the cloud can accommodate
financial simulations.
The remainder of this paper is organised as follows.
Section II considers research related to risk applications.
Section III presents Aggregate Risk Analysis. Section IV
presents an experimental study of sequential and parallel
implementations on cloud VMs. This paper concludes in
Section V by considering future work.
II. R ELATED W ORK
The domain of computational finance and risk addresses
problems related to achieving fast computations, surmounting challenges of data management and efficiently handling
memory of computing systems. Therefore, this domain is
dependent on the advances in high-performance computing.
Research of financial applications on production-based computing systems have progressed from small scale clusters
[8], [9] to large supercomputers [10], [11], and the typical
problem addressed is achieving fast computations. These
applications are hosted either on in-house clusters or on
supercomputing infrastructures to which the owners of the
application have access.
A number of financial applications are being migrated
from small clusters to be hosted on multiple core processors
and many core coprocessor which are available at a low
budget [12]. For example, research related to financial applications exploiting the Cell BE processor [13] [14], FPGAs
[15] [16] and GPUs [17], [18], [19], [20]. In all the above
research, the need for speeding up financial applications are
presented and is achieved. However, ad hoc analytics in
financial risk is important which is now possible with the
availability of scalable on-demand VMs provided by cloud
computing and the development of big data techniques.
Given that the cloud is a potential high-performance computing platform to address big data problems it is now ripe to
explore risk applications in the cloud context [21]. There is
limited research exploring the feasibility of accelerating and
accommodating financial simulations for ad hoc analysis on
the clouds. The research reported in this paper is motivated
towards this end.
III. AGGREGATE R ISK A NALYSIS ON THE C LOUD
Financial applications are underpinned by large-scale simulations which are both data, memory and computationally
intensive. One such simulation is a Monte-Carlo like simulation performed on a portfolio of risk held by a reinsurer,
referred to as Aggregate Risk Analysis [22], [23], [24].
This simulation provides actuaries and decision makers with
millions of alternate views of catastrophic events, such as
earthquakes, that can occur and the order in which they
can occur in a year for portfolio risk management and realtime pricing. Millions of trials are simulated, with each trial
comprising a set of possible future earthquake events and
the probable loss for each trial is estimated.
Although Aggregate Risk Analysis is an embarrassingly
parallel problem, there are significant challenges in achieving efficient parallelism. One major challenge is sharing
large input data between the processing cores constrained
by limited memory bandwidth. Further, the challenge of
accommodating input data in limited memory hardware is
constrained by the complex memory architecture of accelerating hardware such as GPUs.
Large and small sized data along with metadata are
required for performing Aggregate Risk Analysis. The large
data required is the Year Event Table, denoted as Y ET ,
which contains the occurrence of earthquake events for a
year. The YET is obtained from a catalogue of possible
future earthquakes that is generated using earthquake models. The frequency and the physical characteristic of the
potential earthquake, and the damage the earthquake will
cause are estimated by the hazard and vulnerability modules
respectively of the earthquake model. The YET provides a
million distinct views of potential earthquakes that can occur
in a year.
Each record in a YET is called a Trial, denoted as Ti ,
which represents a possible sequence of event occurrences
for any given year. The sequence of events is defined by a set
of tuples containing the ID of an event and the time-stamp of
its occurrence in a trial, Ti = {(Ei,1 , ti,1 ), . . . , (Ei,k , ti,k )}.
The set is ordered by ascending time-stamp values. A typical
YET may comprise one million trials, and each trial may
have one thousand event time-stamp pairs. The YET is
represented as
Y ET
=
{Ti = {(Ei,1 , ti,1 ), . . . , (Ei,k , ti,k )}},
where i = 1, 2, . . . and k = 1, 2, . . . , 800, . . . , 1500.
The small data required for Aggregate Risk Analysis is
the Event Loss Tables, denoted as ELT , which represents
the collection of specific events and their corresponding
losses. Each record in an ELT is denoted using EventLoss pairs ELi = {Ei , li } and a set of financial terms
associated with the ELT FT 1 = (FT 11 , FT 12 , . . . ). A
typical Aggregate Risk Analysis may comprise ten thousand
ELTs, each containing tens of thousands of event losses with
exceptions even up to a few million event losses. The ELTs
is represented as
ELT =
ELi
FT 1
= {Ei , li },
= (FT 11 , FT 12 , . . . )
where i = 1, 2, . . . , 10, 000, . . . , 30, 000.
The metadata is defined as a Portfolio, denoted as P F ,
which contains a group of Programs, denoted as P represented as P F = {P1 , P2 , · · · , Pn }, with n = 1, 2, . . . , 10.
Each Program in turn covers a set of Layers, denoted as
L, that cover a collection of ELTs under a set of financial
terms of the Layer. A single layer Li is composed of the
set of ELTs E = {ELT1 , ELT2 , . . . , ELTj }, and two set
of Layer Terms, denoted as FT 2 = (FT 21 , FT 22 ) and
FT 3 = (FT 31 , FT 32 ).
A typical Layer covers approximately three to thirty
individual ELTs. The Layer can be represented as
= {ELT1 , ELT2 , . . . , ELTj },
E
FT 2 = (FT 21 , FT 22 ),
L=
FT 3 = (FT 31 , FT 32 )
where j = 1, 2, 3, . . . , 30.
The algorithm (line no. 1-17 shown in Algorithm 1) for
aggregate analysis has two stages. In the first stage, data
is loaded into local memory what is referred to as the
preprocessing stage in this paper. In this stage Y ET , ELT
and P F , are loaded into memory.
Algorithm 1: Aggregate Risk Analysis
Input : Y ET , ELT , P F
Output: Y LT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
for each Program, P , in P F do
for each Layer, L, in P do
for each Trial, T , in Y ET do
for each Event, E, in T do
for each ELT covered by L do
Lookup E in the ELT and find
corresponding loss, lE
Apply Financial Terms FT 1 to lE
lT ← lT + lE
end
Apply Financial Terms FT 2 to lT
Apply Financial Terms FT 3 to lT
end
end
end
end
Populate Y LT using lT
In the second stage, the four step simulation executed for
each Layer and for each trial in the YET is performed as
shown below and the resulting Year Loss Table (Y LT ) is
produced.
In the first step shown in line no. 6 in which each event
of a trial its corresponding event loss in the set of ELTs
associated with the Layer is determined. In the second step
shown in line nos. 7-9, secondary uncertainty is applied to
each loss value of the Event-Loss pair extracted from an
ELT. A set of contractual financial terms are then applied
to the benefit of the layer. For this the losses for a specific
event’s net of financial terms I are accumulated across all
ELTs into a single event loss shown in line no. 9. In the third
step in line no. 11 the event loss for each event occurrence
in the trial, combined across all ELTs associated with the
layer, is subject to occurrence terms. In the fourth step in
line no. 12 aggregate terms are applied.
The financial terms FT 2 and FT 3 applied on the loss
values combined across all ELTs associated with the layer
are Occurrence and Aggregate terms. Two occurrence terms,
namely (i) Occurrence Retention, denoted as FT 21 , which
is the retention or deductible of the insured for an individual
occurrence loss, and (ii) Occurrence Limit, denoted as
FT 21 , which is the limit or coverage the insurer will pay
for occurrence losses in excess of the retention are applied.
Occurrence terms are applicable to individual event occurrences independent of any other occurrences in the trial. The
occurrence terms capture specific contractual properties of
’eXcess of Loss’ [25] treaties as they apply to individual
event occurrences only. The event losses net of occurrence
terms are then accumulated into a single aggregate loss
for the given trial. The occurrence terms are applied as
lT = min(max(lT − F T 21 ), FT 22 ).
Two aggregate terms, namely (i) Aggregate Retention,
denoted as FT 31 , which is the retention or deductible
of the insured for an annual cumulative loss, and (ii)
Aggregate Limit, denoted as FT 32 , which is the limit or
coverage the insurer will pay for annual cumulative losses
in excess of the aggregate retention are applied. Aggregate
terms are applied to the trial’s aggregate loss for a layer.
Unlike occurrence terms, aggregate terms are applied to
the cumulative sum of occurrence losses within a trial and
thus the result depends on the sequence of prior events
in the trial. This behaviour captures contractual properties
as they apply to multiple event occurrences. The aggregate
loss net of the aggregate terms is referred to as the trial
loss or the year loss. The aggregate terms are applied as
lT = min(max(lT − F T 31 ), FT 32 ).
The analysis generates loss values associated with each
trial of the YET which populates the Year Loss Table (YLT).
Important risk metrics such as the Probable Maximum Loss
(PML) and the Tail Value-at-Risk (TVaR) which are used
for both internal risk management and reporting to financial
regulators and rating agencies can be derived from the output
of the analysis. Furthermore, these metrics flow into a final
stage of the risk analytics pipeline, namely Enterprise Risk
Management, where liability, asset, and other forms of risks
are combined and correlated to generate an enterprise wide
view of risk. Additional functions can be used to generate
reports that will aid actuaries and decision makers, for
example, reports presenting Return Period Losses (RPL).
The simulations can also be extended for providing vital
information required for disaster recovery and management.
IV. E XPERIMENTAL S TUDIES
In this section, the cloud computing platform investigated
for implementing sequential and parallel Aggregate Risk
Analysis is presented. The data structures chosen for representing the input and intermediate data to address memory
bottlenecks and optimisations to improve the performance
of the analysis are considered. An empirical study of the
analysis both sequentially and in parallel on (single and
multiple core) CPU and (single and multiple) GPU VMs
on the cloud are presented.
A. Platform
Aggregate risk analysis was performed on public VMs
available from the Amazon Elastic Compute Cloud (EC2)1
and on private VMs. Table I shows the underlying hardware
and specifications of the VMs used for this research.
The VMs offered by Amazon are referred to as instances.
Five types of instances are offered by Amazon, namely
the (i) general purpose, (ii) memory optimised, (iii) cluster
compute, (iv) storage optimised, and (v) GPU instances are
used for the analysis. Only instances with at least 15 GB of
RAM are used (this is a requirement for the analysis).
The general purpose instances employed are the m1
and m3 instances, namely m1.xlarge, m3.xlarge and
m3.2xlarge. The m1 instance is a VM of Intel(R)
Xeon(R) CPU E5-2650 and m3 instances are VMs abstracted
over Intel(R) Xeon(R) CPU E5-2670. Each virtual CPU
(vCPU) of the m3 instances is a hardware hyperthread on
the underlying processor.
The memory optimised instances employed are the m2
and cr1 instances, namely m2.xlarge, m2.2xlarge,
m2.4xlarge and cr1.8xlarge. The m2 instances are
VMs of Intel(R) Xeon(R) CPU E5-2665 and the cr1
instance abstracts Intel(R) Xeon(R) CPU E5-2670. Each
virtual CPU (vCPU) of the cr1 instance is a hardware
hyperthread on the underlying processor.
The compute optimised instance employed are the
cc1 and cc2 instances, namely cc1.4xlarge and
cc2.8xlarge. Both cc1 and cc2 instances are abstraction of Intel(R) Xeon(R) CPU X5570. Each virtual CPU
(vCPU) of the cc2 instance is a hardware hyperthread on
the underlying processor.
The storage optimised instance employed are the
hi1 and hs1 instances, namely hi1.4xlarge and
hs1.8xlarge. The hi1 instance abstracts Intel(R)
Xeon(R) CPU E5620 and hs4 is a VM over the Intel(R)
Xeon(R) CPU E5-2650.
The GPU instance employed is cg1.4xlarge backed
by two Intel(R) Xeon(R) CPU X5570 and two NVIDIA
Tesla M2050 GPUs. Each GPU consists 448 processor cores
and 3 GB of total global memory yielding 2.625 GB of
user available memory and a memory bandwidth of 148.4
GB/sec.
Two private VMs vm1 and vm2 are employed. vm1 is
backed by Intel(R) Xeon(R) E5-2620 which is relatively old
hardware when compared to the underlying processor for
Amazon instances. vm2 is backed by Intel(R) Xeon(R) E52665 similar to the underlying CPU on the m2 instances.
vm2 is also supported by the same GPU on the Amazon
1 http://aws.amazon.com/ec2/
Instance
Type
No. of
Memory Processor Type
Vir(GiB)
tual
CPUs
(VCPU)
Amazon cloud VMs
m1.xlarge 4
15.0
Intel Xeon E5-2650
m2.xlarge 2
17.1
Intel Xeon E5-2665
m2.2xlarge 4
34.2
Intel Xeon E5-2665
m2.4xlarge 8
68.4
Intel Xeon E5-2665
m3.xlarge 4
15.0
Intel Xeon E5-2670
m3.2xlarge 8
30.0
Intel Xeon E5-2670
cr1.4xlarge 32
244.0
Intel Xeon E5-2670
cc1.4xlarge 16
23.0
Intel Xeon X5570
cc2.8xlarge 32
60.5
Intel Xeon X5570
hi1.4xlarge 16
60.5
Intel Xeon E5620
hs1.4xlarge 16
117.0
Intel Xeon E5-2650
16
22.5
Intel Xeon X5570
cg1.4xlarge
448
3.0
NVIDIA
Tesla
M2050
Private VMs
vm1
12
128.0
Intel Xeon E5-2620
16
64.0
Intel Xeon E5-2665
vm2
448
3.0
NVIDIA
Tesla
M2050
Clock
Speed
(GHz)
2.00
2.40
2.40
2.40
2.60
2.60
2.60
2.93
2.93
2.40
2.00
2.93
0.575
2.00
2.40
0.575
Table I: VMs employed for Aggregate Risk Analysis
GPU instance. The vCPUs of both VMs are hardware hyperthreads on the underlying processor and the Xen hypervisor
is used.
The Ubuntu 13.10 cloud image is used on all VMs.
Sequential and parallel versions of Aggregate Risk Analysis
were implemented. C++ was used for the sequential implementation, OpenMP was used in the parallel implementation
on multiple core instances, and CUDA was used for the
sequential and parallel implementations on GPU instances.
On the CPU instances, both versions were compiled using
the GNU Compiler Collection g++ 4.6 and optimised using
‘-O3’; the parallel implementation required -fopenmp
during compilation for including the OpenMP directive. The
NVIDIA CUDA Compiler (nvcc 5.0) was used for compiling
on the GPU instances. Message Passing Interface (MPI) was
employed and added nearly 10% to the execution times and
was also found to lower the efficiency of the VMs; hence
those results are not presented. However, for a multiple GPU
instance a combination of MPI to communicate between
the instances, OpenMP for exploiting parallelism of the
virtual cores on each instance and CUDA programming for
exploiting the GPU were employed.
B. Implementation on the Cloud
In all the implementations each trial in the YET is
executed using a single thread. All data required for the
analysis is available as an Amazon Elastic Block Storage
(EBS)2 volume which is attached onto an instance. Nearly
fifteen hours of continuous data transfer was required to the
EBS volume.
2 http://aws.amazon.com/ebs/
(a) Time taken on VMs
(b) Time taken when number of Trials are varied on the fastest public and
private VMs
Figure 1: Sequential performance of Aggregate Risk Analysis on VMs
One important factor for obtaining good performance in
the Aggregate Risk Analysis algorithm is the selection of
a data structure for representing ELTs. The ELTs function
as dictionaries with key-value pairs requiring fast random
key lookup. A sparse representation of ELTs covered by a
Layer using direct access tables was implemented. Although
fast lookups are obtained a sparse representation is at the
expense of high memory utilisation. If a YET consists of
1,000,000 events and an ELT consists of 10,000 event-loss
pairs, then the direct access table would contain 1,000,000
event entries of which 990,000 events would have zero
loss values. Considering that one Layer would cover fifteen
ELTs in a typical analysis, 15 million event-loss pairs need
to be generated in the memory of the instance of which
14,850,000 events have zero loss values.
Nevertheless a direct access table was employed in all
implementations. Alternate compact representations were
not chosen for the following reasons: (i) A search operation
is required to find an event-loss pair in a compact representation. Sequential and binary search require O(n) and
O(log(n)) memory accesses respectively to locate an eventloss pair. Even if a constant-time space-efficient hashing
scheme requiring a constant number of memory accesses
is adopted there is considerable implementation and run-
time performance complexity. This overhead will be high on
GPUs with complex memory hierarchies consisting of global
and shared memories. (ii) To perform aggregate analysis
on a YET of one million trials, each trial comprising one
thousand events, and for a layer covering fifteen ELTs,
there are fifteen billion event-loss pairs. Direct access tables,
although require large memory space, allow for the least
number of memory accesses as each lookup in an ELT
requires only one memory access per search operation.
Two data structure implementations of the ELTs were
considered. In the first implementation, each ELT is an
independent table, and therefore, in a read cycle, each thread
independently looks up its events from the ELTs. All threads
within a block access the same ELT. By contrast, in the
second implementation, the ELTs are combined as a single
table. Consequently, the threads then use the shared memory
to load entire rows of the combined ELTs at a time. The latter
implementation performs poorly compared to the former
since for the threads to collectively load from the combined
ELT each thread must first write which event it requires.
This results in additional memory overheads.
On the CPU instance offering multiple virtual cores the
entire data required for the analysis is processed in memory.
The GPU implementation uses the GPU’s global memory to
store all of the required data structures. The parallel implementation on the GPU requires high memory transactions
and leads to inefficient performance on the GPU platform.
To surmount this challenge shared memory can be utilised
over global memory.
The algorithm is optimised in the following four ways.
Firstly, by chunking, which refers to processing a block of
events of fixed size (or chunk size) for the efficient use of
shared memory.
Chunking is more beneficial in the GPU implementation
than in the CPU implementations. In the case of the GPU
implementation looking up events in a trial and applying
financial terms to losses at the Event and Layer level are
chunked. Further, the financial and layer terms are stored in
the streaming multi-processor’s constant memory.
If the intermediate losses are represented in global memory, then while applying the financial terms at the Event and
Layer level would require the global memory to be accessed
and updated adding considerable memory overheads. The
memory overhead is minimised by chunking when (i) the
financial terms are applied, and (ii) reading events in a
trial from the YET. Chunking reduces the number of global
memory update and global read operations. Moreover, the
benefits of data striding can also be used to improve speedup.
Secondly, the implementation are optimised by loop unrolling, which refers to the replication of blocks of code
included within for loops by the compiler to reduce the
number of iterations performed by the for loop. This is done
using the pragma directive.
Thirdly, the implementations on the CPU and GPU are
optimised by using single precision operations when possible. Read operations are faster using float variables as they
are only half the size of a double variable. The performance
of single precision operations tend to be approximately twice
as fast as double precision operations.
Fourthly, in the case of the GPU a further optimisation
can be achieved by migrating data from both shared and
global memory to the kernel registry. The kernel registry has
the lowest latency compared to all other forms of memory
available in the GPU architecture.
C. Empirical Analysis
The results obtained from the experimental studies are
presented in this section. All data required for the analysis
is stored as an EBS volume and attached onto the instances
considered in Table I. Figure 1 to Figure 4 are results
obtained on CPU instances; the multi-core architecture of
the instances are exploited in the parallel implementation.
Figure 5 and Figure 6 are results obtained on the GPU
instance; both single and multiple GPUs are exploited in
the parallel implementation. In all experiments, the analysis
uses as input a YET comprising one million trials, with each
trial consisting of one thousand catastrophic events, and one
Portfolio with one Program comprising one Layer covering
sixteen ELTs. The input parameters are realistic and were
chosen based on industry-wide practices.
1) Results from CPU instances: Figure 1a shows the
graph plotted for the time taken for sequentially performing
aggregate risk analysis on all instances shown in Table I.
Under the general purpose instances, the m1 is the slowest
for performing the analysis requiring 565 seconds; the m3
instance is over 37% faster than the m1 instance. The
memory optimised instance cr1 is the fastest for performing
the analysis requiring 295 seconds which is 37% faster
than the memory optimised m2 instances. The difference
in the performance obtained on storage optimised instances
is just over 20%. Cluster instances cc1 and cc2 perform
comparably to the cg1 instance. The fastest sequential CPU
performance on the cloud requires less than five minutes
which is nearly 50% fastest than the slowest sequential
performance on the cloud. Private VMs vm1 and vm2 have
surprisingly good performance. vm1 takes only 340 seconds
which is nearly 40% faster than m1 instances. The vm2
VM completes the analysis in 288 seconds which is over
2% faster than the best performance on Amazon. Figure 1b
shows the both the increase in the total time taken for the
fastest sequential analysis on the cloud when the number
of trials are varied between two hundred thousand and one
million trials.
The parallel implementation of the analysis on the CPU
requires multiple threads to be executed on the instance
which can be done in two ways. Firstly, by executing a
single thread per virtual core, and secondly, by executing
multiple threads per core.
Figure 2 shows the graphs obtained from the parallel
implementation of the analysis when one thread is executed
per virtual core on the instance. The graphs are organised
based on the number of virtual cores on the instance. The instance with two virtual cores obtains nearly a 96% efficiency
when two threads are employed (Figure 2a). Instances with
four virtual cores obtain upto 87.5% efficiency (Figure 2b).
The two instances with eight virtual cores have an average
efficiency of over 70% (Figure 2c).
The storage optimised, cc1 and cg1 instances with
sixteen cores each exhibit very little speedup and efficiency
for more than eight cores (Figure 2d). Surprisingly, there
is no hardware acceleration obtained which is expected.
Beyond eight cores it would seem that the cost of hardware
and the use of virtualised hardware on the sixteen core
VMs do not benefit the analysis. Another reason is that as
the number of cores are increased the bandwidth to access
memory is not equally increased which is a limiting factor.
Similarly, in the case of thirty two core instances no
acceleration is obtained beyond sixteen cores (Figure 2e).
The fastest parallel execution on the CPU is obtained on the
cluster compute instance cc2.8xlarge taking 27 seconds
with a speedup of nearly 11x over the fastest sequential
implementation. The performance of cr1.8xlarge is second to the cc2 instance requring 40 seconds when multiple
threads are employed though it performs the sequential
analysis the fastest.
The private VMs again outperform the public instances
(Figure 2f). vm1 takes 44 seconds achieving a speedup
of 7.5x over its sequential performance and vm2 takes 22
seconds achieving a speedup of 13x over its sequential
performance.
Figure 3 shows the graphs obtained from the parallel
implementation of the analysis when multiple threads are
executed on the instances. In all cases, multiple threads
per Amazon core do not provide any acceleration for the
analysis. Increasing the number of threads per core results
in an increase in the communication cost between threads.
The private VMs vm1 and vm2 achieve a speedup of 9%
and 5% respectively when multiple threads are employed
per virtual core.
Figure 4a shows the graph plotted for the best time
taken for performing parallel aggregate risk analysis on
all instances shown in Table I. Under the general purpose
instances, though the m1 instance is the slowest for performing the sequential analysis and for parallel analysis the
m2.xlarge is the slowest requiring 240 seconds. Virtual
core acceleration is achieved on the m1 instance which is
over 1.5x faster than m2.xlarge. The m3.2xlarge is
nearly 2 times faster than m3.xlarge and m1.xlarge.
The cluster instance cc2 followed by cr1 are the fastest
requiring 27 seconds and 40 seconds respectively. Hence
(a) Two virtual core public instance
(b) Four virtual core public instances
(c) Eight virtual core public instances
(d) Sixteen virtual core public instances
(e) Thirty two virtual core public instances
(f) Private VMs
Figure 2: Parallel execution (single thread per virtual core) of Aggregate Risk Analysis on public (a) to (e) and private (f)
VMs
(a) Two virtual core public instance
(b) Four virtual core public instances
(c) Eight virtual core public instances
(d) Sixteen virtual core public instances
(e) Thirty two virtual core public instances
(f) Private VMs
Figure 3: Parallel execution (multiple threads per virtual core) of Aggregate Risk Analysis on public (a) to (e) and private
(f) VMs
(a) Time taken on VMs
(b) Time taken when number of Trials are varied on the fastest public and
private VMs
Figure 4: Parallel execution of Aggregate Risk Analysis on
public and private VMs
upto a 21x speedup is obtained for parallel analysis by
exploiting the multi-core architecture over the sequential
analysis and upto a 9x speedup over the slowest parallel
analysis. Again, private VMs outperform public instances.
The best performance of of vm2 is 21 seconds using multiple
threads which is 22% faster than the best performance
achieved by the cc2 instance. Similarly, vm1 takes 38
seconds for the analysis on multiple threads which is 5%
faster than the second best performance by cr1 instances
on public instances.
Figure 4b shows the increase in the total time taken for
the fastest parallel analysis on the cloud when the number
of trials are varied between two hundred thousand and one
million trials.
2) Results from GPU instance: Single and multiple GPU
instances (cg1.4xlarge) are considered for risk analysis on the cloud. CUDA provides an abstraction over the
streaming multi-processors of the GPU, referred to as a
CUDA block. Unlike the parallel implementations on the
CPU instance an additional parameter that needs to be
considered in the GPU implementations is the number of
threads executed per CUDA block. To represent 1,000,000
trials of the analysis on the GPU instance consider each trial
Figure 5: Execution of Aggregate Risk Analysis on a single
GPU on public and private VMs
is executed on one thread. If 128 threads are executed on one
streaming multi-processor there will be approximately 7813
blocks which need to be executed on 14 streaming multiprocessors; each streaming multi-processor will therefore
need to execute 558 blocks. All threads executed on a
streaming multi-processor share fixed allocations of shared
and constant memory. Therefore, there is a trade-off for
optimal performance; each thread can access larger amounts
of shared and constant memory if there are fewer threads
per block, but then the global memory will required to
be accessed more resulting in increased global memory
overheads.
Figure 5 shows the time taken for parallel risk analysis
on a single GPU instance when the number of threads
per CUDA block are varied between 1 and 512. At least
16 threads per block are required to achieve performance
comparable to the best parallel implementation which is
noted on cc2.8xlarge instance. To exploit the full potential of hardware acceleration offered by GPUs at least 64
threads are required. An improvement in the performance is
observed with 256 threads per block beyond which performance starts to diminish. The best time for performing the
analysis on a single GPU is 19.93 seconds which is around
15x faster than the best sequential performance on the CPU,
nearly 1.4x faster than the best multiple core performance on
the CPU, and over 6x faster than the sequential performance
on a GPU. On the private VM vm2 takes only 16.86 seconds.
This is nearly 16% faster than the GPU on the public
instance.
The performance of the analysis on multiple GPU instances is shown in Figure 6a. In the multiple GPU implementation the workload for the analysis is decomposed and
made available to the multiple instances that employ GPUs.
Each CPU thread schedules the workload on the GPUs.
Time taken for the analysis on four Amazon GPU instances
is 5.024 seconds which is approximately 3.97 times faster
than employing a single GPU with close to 97% efficiency.
Compared to the sequential implementation on a single GPU
(a) Performance of multiple GPUs
(b) Performance on four GPUs for varying threads per block
Figure 6: Aggregate Risk Analysis on public and private
VMs with multiple GPUs
a speedup of over 24x is achieved when multiple GPU
instances are used. On the other hand vm2 takes only 4.238
seconds which is 16% faster than the multiple GPUs on the
public instance.
Figure 6b shows the performance of the analysis on four
GPUs when the number of threads per block is varied from
16 to 64. Experiments could not be pursued beyond 64
threads per block due to the limitation on the block size
the shared memory can use. The best performance of 5.024
seconds on the public VM and 4.2375 seconds on the private
VM is achieved when the number of threads per block is
32; the block size is the same as the WARP size of the
GPU, therefore, an entire block of threads can be swapped
when high latency operations occur. Increasing the number
of threads per block does not improve the performance since
there is shared memory overflow.
D. Summary
The experimental studies indicate that the public clouds
are a suitable platform for accommodating risk simulations.
The data, memory and computational requirements can be
met on the cloud VMs. Risk simulation can be accelerated
on the public cloud although the simulations do not scale
well over the virtual cores of the instances; for example,
for thirty two core instances no acceleration is achieved
beyond sixteen cores. This results in wasted computational
time per dollar spent on the simulation. Hence, maximum
performance cannot be achieved thereby not fully exploiting the potential of the public cloud. Nevertheless, a 60x
speedup is achieved on public instances over a baseline
implementation. Interestingly, the private VMs are faster
than the public instances. For example, the sequential CPU
implementation, parallel CPU implementation and the parallel GPU implementation on private VMs are up to 40%,
22% and 16% faster than the best performance achieved on
public instances.
V. C ONCLUSIONS
The cloud is a potential platform for performing ad hoc
risk simulations important to financial applications. Scalable
and on-demand resources of the cloud are attractive for running ad hoc simulations and for meeting their data, memory
and computational requirements. The research reported in
this paper was motivated towards experimentally verifying
whether clouds are ready to accelerate financial simulations.
A typical application employed in the financial industry,
namely Aggregate Risk Analysis, was developed and deployed on a variety of cloud VMs. The implementation
exploited parallelism to accelerate the application and efficient management of data and memory to accommodate the
application on the cloud. The experimental results indicate
that the application can be accommodated to run on the
cloud and an acceleration of up to 60x over a baseline
implementation can be obtained with hardware accelerators
on the cloud. Nevertheless, there is poor efficiency in the
acceleration achieved highlighting the inability to harness
the full potential of all available compute cores resulting
in wasted computational performance. It is noted that the
private VMs perform better than the public VMs.
Migrating financial applications onto the cloud is viable
since the cloud provides a suitable platform to accommodate
the computational, data and memory demands of ad hoc
simulations. This is of significant benefit to the financial
industry as well as its associated industries since the scalability and availability of resources on an on-demand basis
reduce maintenance costs. However, while acceleration was
achieved for the simulation, in our experience it could not
be run most efficiently on the public cloud since there was
wasted computational time for every dollar spent.
R EFERENCES
[1] N. H. Chan and H. Y. Wong, “Simulation Techniques in
Financial Risk Management,” John Wiley & Sons, 2006.
[2] G. Woo, “Natural Catastrophe Probable Maximum Loss,”
British Actuarial Journal, Vol. 8, 2002.
[3] P. Glasserman, P. Heidelberger and P. Shahabuddin, “Portfolio
Value-at-Risk with Heavy-tailed Risk Factors,” Mathematical
Finance, Vol. 12, No. 3, 2002, pp. 239-269.
[4] K. H. Tsoi and W. Luk, “Axel: A Heterogeneous Cluster with
FPGAs and GPUs,” Proceedings of the 18th ACM/SIGDA International Symposium on Field Programmable Gate Arrays,
2010, pp. 115-124.
[15] N. Singla, M. Hall, B. Shands and R. D. Chamberlain, “Financial Monte Carlo Simulation on Architecturally Diverse
Systems,” Proceedings of Workshop on High Performance
Computational Finance, 2008.
[5] M. Showerman, J. Enos, A. Pant, V. Kindratenko, C. Steffen, R. Pennington and W. -M. Hwu, “QP: A Heterogeneous Multi-Accelerator Cluster,” Proceedings of the 10th
LCI International Conference on High-Performance Clustered
Computing, 2009.
[16] S. Weston, J. -T. Marin, J. Spooner, O. Pell and O. Mencer,
“Accelerating the Computation of Portfolios of Tranched
Credit Derivatives,” Proceedings of Workshop on High Performance Computational Finance, 2010.
[6] D. Kondo, B. Javadi, P. Malecot, F. Cappello and D. P.
Anderson, “Cost-benefit Analysis of Cloud Computing Versus Desktop Grids,” Proceedings of the IEEE International
Symposium on Parallel and Distributed Processing, 2009.
[7] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. Katz, A.
Konwinski, G. Lee, D. Patterson, A. Rabkin, I. Stoica and M.
Zaharia, “A View of Cloud Computing,” Communications of
the ACM, Vol. 53, Issue 4, April 2010, pp. 50-58.
[8] A. Srinivasan, “Parallel and Distributed Computing Issues in
Pricing Financial Derivatives through Quasi Monte Carlo,”
Proceedings of the International Parallel and Distributed Processing Symposium, 2001.
[9] K. Huang and R. K. Thulasiram, “Parallel Algorithm for
Pricing American Asian Options with Multi-Dimensional
Assets,” Proceedings of the 19th International Symposium on
High Performance Computing Symposium, 2005, pp. 177185.
[10] C. Bekas, A. Curioni and I. Fedulova, “Low Cost High Performance Uncertainty Quantification,” Proceedings of Workshop
on High Performance Computational Finance, 2009.
[11] D. Daly, K. D. Ryu and J. E. Moreira, “Multi-variate Finance Kernels in the Blue Gene Supercomputer,” Proceedings
of Workshop on High Performance Computational Finance,
2008.
[12] M. Majmudar, C. Docan, M. Parashar, C. Marty, “Cost vs.
Performance of VaR on Accelerator Platforms,” Proceedings
of Workshop on High Performance Computational Finance,
2009.
[13] V. Agarwal, L. -K. Liu and D. A. Bader, “Financial Modelling on the Cell Broadband Engine,” Proceedings of the
IEEE International Symposium on Parallel and Distributed
Processing, 2008.
[14] C. Docan, M. Parashar, and C. Marty, “Advanced Risk
Analytics on the Cell Broadband Engine,” Proceedings of the
23rd IEEE International Parallel and Distributed Processing
Symposium, 2009.
[17] M. F. Dixon, J. Chong and K. Keutzer, “Accelerating Valueat-Risk Estimation on Highly Parallel Architectures,” Concurrency and Computation: Practice and Experience, Vol. 24,
2012, pp. 895-907.
[18] B. Zhang and C. W. Oosterlee, “Option Pricing with COS
Method on Graphics Processing Units,” Proceedings of the
23rd IEEE International Parallel and Distributed Processing
Symposium, 2009.
[19] L. A. Abbas-Turki, S. Vialle, B. Lapeyre and P. Mercier,
“Pricing Derivatives on Graphics Processing Units Using
Monte Carlo Simulation,” Concurrency and Computation:
Practice and Experience, 2012.
[20] D. M. Dang, C. C. Christara and K. R. Jackson, “An Efficient GPU-Based Parallel Algorithm for Pricing Multi-Asset
American Options,” Concurrency and Computation: Practice
and Experience, Vol. 24, No. 8, 2012, pp. 849-866.
[21] H. Kim, S. Chaudhari, M. Parashar and C. Marty, “Online Risk Analytics on the Cloud,” Proceedings of the 9th
IEEE/ACM International Symposium on Cluster Computing
and the Grid, 2009, pp. 484-489.
[22] R. R. Anderson and W. Dong, “Pricing Catastrophe Reinsurance with Reinstatement Provisions using a Catastrophe
Model,” Casualty Actuarial Society Forum, Summer 1998,
pp. 303-322.
[23] G. G. Meyers, F. L. Klinker and D. A. Lalonde, “The Aggregation and Correlation of Reinsurance Exposure,” Casualty
Actuarial Society Forum, Spring 2003, pp. 69-152.
[24] W. Dong and F. S. Wong, “Portfolio Theory for Earthquake
Insurance Risk Assessment,” Proceedings of the 12th World
Conference on Earthquake Engineering, Paper No. 0163,
2000.
[25] D. Cummins, C. Lewis and Richard Phillips, “Pricing Excessof-Loss Reinsurance Contracts Against Catastrophic Loss,”
The Financing of Catastrophe Risk, Editor: K. A. Froot,
University of Chicago Press, 1999, pp. 93-148.
| 5cs.CE
|
arXiv:1801.10006v1 [math.AC] 30 Jan 2018
PERFECTOID SPACES AND THE HOMOLOGICAL CONJECTURES
Y. ANDRÉ
A BSTRACT. This is a survey of recent advances in commutative algebra, especially in
mixed characteristic, obtained by using the theory of perfectoid spaces. An explanation of
these techniques and a short account of the author’s proof of the direct summand conjecture
are included. One then portrays the progresses made with these (and related) techniques
on the so-called homological conjectures.
1. T HE
DIRECT SUMMAND CONJECTURE
1.1. Let R be a Noetherian (commutative) ring and S a finite ring extension, and let us
consider the exact sequence of finitely generated R-modules
(1.1)
0 → R → S → S/R → 0.
When does this sequence split? Equivalently, when is R → S pure, i.e. remains injective after any base change? This holds for instance when R → S is flat, or when R is a
normal Q-algebra, but not in general (the embedding of Q[x, y]/(xy) in its normal closure
gives a counter-example, since it is no longer an embedding modulo x + y).
The direct summand conjecture, formulated by M. Hochster around 1969, claims that
(1.1) splits whenever R is regular. Hochster proved it when R contains a field [17]. R.
Heitmann proved it in dimension ≤ 3 [14].
1.2. Recently, the author proved it in general [2]:
1.2.1. Theorem. (1.1) splits if R is regular.
This has many (non trivially) equivalent forms. One of them is that every ideal of a
regular ring R is contracted from every finite (or integral) extension of R. Another (more
indirect) equivalent form is the following statement, which settles a question raised by L.
Gruson and M. Raynaud [28, 1.4.3]1:
1.2.2. Theorem. Any integral extension of a Noetherian ring descends flatness of modules.
We will see one more equivalent form below in the framework of the so-called homological conjectures (4.2).
2. T HE ROLE
OF PERFECTOID SPACES
2.1. Some heuristics. After Hochster’s work [18], it is enough to prove the direct summand conjecture in the case when R is a complete unramified local regular ring of mixed
characteristic (0, p) and perfect residue field k. By Cohen’s structure theorem, one may
thus assume R = W (k)[[x1 , . . . , xn ]].
1991 Mathematics Subject Classification. 13D22, 13H05, 14G20.
Key words and phrases. homological conjectures, big Cohen-Macaulay algebra, perfectoid algebra, purity.
1cf. [26] for the equivalence. Gruson and Raynaud settled the case of a finite extension and outlined that the
transition to integral extensions is not a routine exercise.
1
2
Y. ANDRÉ
In characteristic p, all proofs of the direct summand conjecture use the Frobenius endomorphism F in some way. In mixed characteristic, R = W (k)[[x1 , . . . , xn ]] carries a
Frobenius-like endomorphism (acting as the canonical automorphism of W (k) and sending
xi to xpi ), which however does not extend to general finite extensions S of R. To remedy
this, p-adic Hodge theory suggests to “ramify deeply”, by adjoining iterated pth roots of
p, x1 , . . . , xn . Doing this, one leaves the familiar shore of Noetherian commutative algebra
for perfectoid geometry, recently introduced by P. Scholze [30].
To begin with, W (k) is replaced by the non-Noetherian complete valuation ring Ko :=
\1
W (k)[p p∞ ]. The valuation ring Lo of any finite extension L of the field K[ p1 ] satisfies:
x7→xp
1
F : Lo /p → Lo /p is surjective, and Lo is p p∞ -almost finite etale over Ko ,
(2.1)
this being understood in the context of almost ring theory, introduced by G. Faltings and
developped by O. Gabber and L. Ramero [12], which gives precise meaning to “up to
1
1
1
p p∞ -torsion”; for instance, p p∞ -almost etaleness means that p p∞ ΩLo /Ko = 0. Actually,
[12] is much more general: it deals with modules over a commutative ring up to k-torsion,
for some idempotent ideal k. Going beyond the case of a valuation ideal k will be crucial:
1
1
beside “p p∞ -almost” modules, we will have to consider “(pg) p∞ -almost” modules for
some “geometric” discriminant g.
2.2. Perfectoid notions. In perfectoid geometry, one works with certain Banach2 Kalgebras A. One denotes by Ao the Ko -subalgebra of power-bounded elements. One
says that A is uniform if Ao is bounded, and that A is perfectoid if it is uniform and
x7→xp
F : Ao /p → Ao /p is surjective. An example which plays a crucial role in the sequel
1
1
i
1
i
ˆ i W (k)[p pi ][[x1p , . . . , xnp ]][ p1 ] , a deeply ramified avatar of R. Morphisms of perfecis ∪
toid algebras A → B are continuous algebra homomorphisms (one then says that B is a
perfectoid A-algebra).
Perfectoid algebras enjoy three fundamental stability properties [30]:
ˆ A C, and (B ⊗
ˆ A C)o
2.2.1. Tensor product. If B and C are perfectoid A-algebras, so is B ⊗
1
o
o
∞
ˆ Ao C .
is p p -almost isomorphic to B ⊗
2.2.2. Localization. The ring of functions A{ fg } on the subset of the perfectoid space
1
Spa(A, Ao ) where |f | ≤ |g| holds is perfectoid, and A{ fg }o is p p∞ -almost isomorphic to
′
1
Ao h( fg′ ) p∞ i for some approximations f ′ , g ′ of f, g which admit iterated pth -roots in A.
2.2.3. Finite etale extension. Any finite etale extension B of A is perfectoid, and B o is a
1
p p∞ -almost finite etale extension of Ao .
This generalization of 2.1 to perfectoid algebras is Faltings’s “almost purity theorem”
[8] as revisited by Scholze [30] and Kedlaya-Liu [24].
Let us explain how the second assertion of 2.2.3 follows from the first following [1, 3.4.2]. The
idea is to reduce to the case when B is Galois over A with Galois group G, i.e. BG = A and B ⊗A
∼ Q
B → G B. This implies BoG = Ao . On the other hand, since B is a finitely generated projective Aˆ A B, and one deduces from 2.2.1 (assuming B perfectoid) that Bo ⊗
ˆ Ao Bo →
module, B ⊗A B = B⊗
1
Q
o
m
p∞ -almost isomorphism. To get rid of the completion, one passes modulo p : B o /pm
G B is a p
2here and in the sequel, one can work with any perfectoid field K (of mixed char. (0, p)), - i.e. complete,
non-discretely valued, and such that F is surjective on Ko /p. An extensive dictionary between the langage of
commutative algebra and the language of non-archimedean functional analysis is presented in [1, 2.3.1].
PERFECTOID SPACES AND THE HOMOLOGICAL CONJECTURES.
3
is almost Galois over Ao /pm , hence almost finite etale, and a variant of Grothendieck’s “équivalence
remarquable” [12, 5.3.27] allows to conclude that Bo it itself almost finite etale over Ao .
2.3. Direct summand conjecture: the case when S[ p1 ] is etale over R[ p1 ]. Let us go
back to the direct summand conjecture for R = W (k)[[x1 , . . . , xn ]]. The special case
when S[ p1 ] is etale over R[ p1 ] was settled by B. Bhatt [4] and by K. Shimomoto [33], using
2.2.3. Here is a slightly different account, suitable to the sequel.
1
i
1
1
i
ˆ i W (k)[p pi ][[x1p , . . . , xnp ]][ 1p ] and notice
Let us consider the perfectoid algebra A = ∪
1
1
i
1
i
ˆ i W (k)[p pi ][[x1p , . . . , xnp ]] is a faithfully flat extension of R. By assumption
that Ao = ∪
1
B := S ⊗R A is finite etale over A, hence by 2.2.3, B o is p p∞ -almost finite etale, hence
almost pure (in the sense: almost universally injective) over Ao . A fortiori, (1.1) almost
splits after tensoring with Ao . In other words, if e ∈ Ext1R (S/R, R) denotes the class
1
corresponding to (1.1), then p p∞ (e ⊗ 1) = 0 in Ext1R (S/R, R) ⊗R Ao ∼
= Ext1Ao ((S ⊗R
o
o
o
A )/A , A ). One concludes that e = 0 by the following general elementary lemma
1
(applied to M = Re and K = p p∞ Ao ):
2.3.1. Lemma. [2, 1.1.2] Let R be a local Noetherian ring, M a finitely generated Rmodule, A a faithfully flat R-algebra. Let K be an idempotent ideal of A such that K.MA =
0 and R ∩ K 6= 0. Then M = 0.
2.4. The perfectoid Abhyankar lemma. In the general case, S ⊗R A is no longer etale
over A: one must take into account a discriminant g ∈ R of S[ p1 ] over R[ p1 ]. This suggests
to try to generalize 2.2.3 to ramified extensions of perfectoid algebras. It turns out that this
is possible, provided one extracts suitable roots of g in the spirit of Abhyankar’s lemma.
1
1
This leads to replace everywhere “p p∞ -almost” by “(pg) p∞ -almost”, thereby extending the basic setting of almost ring theory beyond the usual situation of a non-discrete
valuation ring. This also leads to introduce the notion of almost perfectoid algebra, where
x7→xp
1
F : Ao /p → Ao /p is only assumed to be (pg) p∞ -almost surjective [1, 3.5.4].
2.4.1. Theorem. [1] Let A be a perfectoid K-algebra, which contains a compatible system
1
of p-power roots g pj of some non-zero-divisor g ∈ Ao . Let B ′ be a finite etale A[ g1 ]1
algebra. Let B be the integral closure of g − p∞ A in B ′ , so that B[ g1 ] = B ′ .
1
Then B is almost perfectoid, and for any n, B o /pm is (pg) p∞ -almost finite etale (hence
1
(pg) p∞ -almost flat) over Ao /pm .
(If g = 1, one may use [12, 5.3.27] again to conclude that B o is itself almost finite etale
over Ao and recover 2.2.3).
j
The basic idea is to look at the pro-system of algebras of functions Aj := A{ pg } on
complements of tubular neighborhoods of the hypersurface g = 0 in the perfectoid space
j
Spa(A, Ao ), resp. at the pro-system B j := B ′ ⊗A[ g1 ] A{ pg }. Each Aj is perfectoid (2.2.2),
1
and each B j is finite etale over Aj , hence perfectoid; moreover B jo is p p∞ -almost finite
etale over Ajo by almost purity (2.2.3). One can show that B o is isomorphic to lim B jo ,
and that the latter has the asserted properties,
However, in the sequel, the identification of (lim B jo )[ p1 ] with the integral closure of
1
g − p∞ A in B ′ plays no role; changing notation, we will set B := (lim B jo )[ p1 ], which is
a uniform Banach algebra, and sktech the proof that B is almost perfectoid and that for
1
every m, B o /pm is (pg) p∞ -almost finite etale over Ao /pm .
4
Y. ANDRÉ
The proof involves six steps.
1
1) For any r ∈ N[ p1 ], lim(Ajo /pr ) is (pg) p∞ -almost isomorphic to Ao /pr . This is essentially
Scholze’s perfectoid version of Riemann’s extension theorem [31, II.3.1] (hint: if R denotes Ao /pr
1
1
j
for short, Ajo /pr is p p∞ -almost isomorphic to Rj := R[( pg ) p∞ ]; the key idea is that for j ′ ≥
P
′
pj s
j + rpk , Rj → Rj factors through Rjk :=
) , so that lim Rj ∼
1 R(
= lim Rjk ; on the
s≤ k
p
g
other hand, the kernel and cokernel of R → Rjk are killed by g raised to a power which tends to 0
− 1
when j, k → ∞). Passing to the limit r → ∞, it follows that lim Ajo = g p∞ A, and this also
holds under the weaker assumption that A is almost perfectoid, cf. [1, 4.2.2].
1
2) lim1 (Bjo /pr ) is (pg) p∞ -almost zero. The technique is similar to the one in 1), cf. [1, 4.4.1].
1
F
3) lim(Bjo /p) → lim(Bjo /p) is (pg) p∞ -almost surjective. Indeed, taking the limit of the exact
sequence 0 → Bjo /p
jo
p−1
p
jo
1
→ Bjo /p → Bjo /p p → 0, one deduces from 2) (for r =
1
p∞
1
p
jo
1
p
p−1
),
p
jo
that
-almost surjective; on the other hand, B /p → B /p is
lim(B /p) → lim(B /p ) is (pg)
an isomorphism because Bj is perfectoid.
4) B is almost perfectoid. From 3), it suffices to show that the natural map Bo /p =
1
(lim Bjo )/p → lim(Bjo /p) is a (pg) p∞ -almost isomorphism. It is easy to see that it is injective [1, 2.8.1], and on the other hand, the composition limF,j (Bjo /p) = B♭o → (lim Bjo )/p →
lim(Bjo /p) is almost surjective by 3).
a
1
j
∼ Bj . The factorization comes
5) B → Bj factors through a g p∞ -almost isomorphism B{ pg } =
j
j
a
from the fact that Bj ∼
= Bj { p }, and one constructs an almost inverse to B{ p } → Bj as follows
g
g
(cf. [1, 4.4.4]): the integral closure of Ao in B′ maps to the integral closure Ajo in Bj , which is
1
p p∞ -almost Bjo (by almost purity). Passing to the limit and inverting pg, one gets a morphism
δ
B′ → B[ 1g ], and the sought for inverse is induced by δ ⊗ 1Aj .
1
6) for every m, Bo /pm is (pg) p∞ -almost finite etale over Ao /pm . This is the decisive step:
how to keep track of the almost finite etaleness of Bjo over Ajo at the limit? As in 2.2.3, the idea
is to reduce to the case when B′ is Galois over A[ g1 ] with Galois group G, i.e. (B′ )G = A[ 1g ] and
∼ Q
ˆ Ajo Bjo →
B′ ⊗A[ 1 ] B′ → G B′ . It follows that Bj is G-Galois over Aj , and (as in 2.2.3) that Bjo ⊗
g
1
Q
jo
is a p p∞ -almost isomorphism. Passing to the limit, (and taking into account the facts
GB
j
ˆ A B is an almost perfectoid algebra and B⊗
ˆ A B{ p } ∼
ˆ Aj Bj , so that
that, by 4) and 5), B⊗
= Bj ⊗
g
1
Q
ˆ Ao Bo → G Bo are (pg) p∞ -almost
one can apply 1)), one concludes that Ao → BoG and Bo ⊗
isomorphisms. To get rid of the completion, one passes modulo pm : Bo /pm is almost Galois over
1
Ao /pm , hence (pg) p∞ -almost finite etale (in constrast to 2.2.3, one cannot conclude that Bo is
1
1
(pg) p∞ -almost finite etale over Ao since p may not belong to (pg) p∞ ).
2.5. Infinite Kummer extensions. In order to extend the strategy of 2.3 to the
general case by means of the perfectoid Abhyankar lemma, one has first to
1
1
1
ˆ i W (k)[p pi ][[x1pi , . . . , xnpi ]][ p1 ] the iterated pth adjoin to the perfectoid algebra ∪
roots of a discriminant g.
At finite level i, it seems very difficult to
1
1
i
1
i
1
control the extension (W (k)[p pi ][[x1p , . . . , xnp ]][g pi , 1p ])o , which is bigger than
1
1
i
1
i
1
W (k)[p pi ][[x1p , . . . , xnp ]][g pi ], but things turn easier at the infinite level, thanks to the
perfectoid theory.
2.5.1. Theorem. [2, 2.5.2] Let A be a perfectoid K-algebra, and let g ∈ Ao be a non-zero
1
1
divisor. Then for any n, Ahg p∞ io /pm is p p∞ -almost faithfully flat over Ao /pm .
PERFECTOID SPACES AND THE HOMOLOGICAL CONJECTURES.
5
1
The basic idea is to add one variable T , consider the perfectoid algebra C := AhT p∞ i
and look at the ind-system of algebras of functions Ci := C{ Tp−g
i } on tubular neighborhoods of the hypersurface T = g in the perfectoid space Spa(C, C o ).
The proof involves three steps.
1
1
\i Cio . This is an easy consequence of the general
7) Ahg p∞ io is p p∞ -almost isomorphic to colim
1
\ i B{ fi }o is p p∞ -almost isomorphic
fact that for any uniform Banach algebra B and f ∈ Bo , colim
p
to (B/f B)o , cf. [1, 2.9.3].
8) C o contains a compatible system of p-power roots of some non-zero-divisor fi such that Cio
1
1
1
\j ChU io /(p pj U − fipj ). This is one instance of Scholze’s
is p p∞ -almost isomorphic to colim
1
approximation lemma [30, 6.7]; one may assume fi ≡ T − g mod p p .
1
j
1
9) ChU io /(p pj U − fip , pm ) is faithfully flat over Ao /pm . One may replace pm by any
1
1
pj
j
p
gij
≡ g mod p. Then fi
1
1
Since C is perfectoid, there is gij ∈ C p∞ io such that
p pj+1 .
positive power of p, e.g.
1
1
1
1
1
1
1
1
j
≡ T pj − gij mod p pj+1 , and ChU io /(p pj U − fip , p pj+1 ) ∼
=
(Ao /p pj+1 )[T p∞ , U ]/(T pj − gij ), a free Ao /p pj+1 -module.
2.6. Conclusion of the proof of the direct summand conjecture. One chooses g ∈ R
1
1
] is etale over R[ pg
]. One then follows the argument of 2.3, replacsuch that S[ pg
1
1
1
ˆ i W (k)[p pi ][[x1pi , . . . , xnpi ]][ p1 ] by the “infinite Kummer extension” A :=
ing A = ∪
1
1
i
1
i
1
ˆ i W (k)[p pi ][[x1p , . . . , xnp ]][ p1 ])hg p∞ i. One introduces the finite etale extension B ′ :=
(∪
S ⊗R A[ 1g ] of A[ g1 ], and one replaces B = S ⊗R A in 2.3 by B := (lim B jo )[ p1 ], noting that
B o is the S-algebra lim B jo . Translating Th. 2.4.1 and Th. 2.5.1 in terms of almost purity
modulo pm , and reasoning as in 2.3 (using the same lemma), one obtains that (1.1) splits
modulo pm for any n [2, §3]. A Mittag-Leffler argument (retractions of R/pm → S/pm
form a torsor under an artinian R/pm -module [17, p. 30]) shows that (1.1) itself splits.
2.7. Derived version. In [5], B. Bhatt revisits this proof and proposes a variant, which
differs in the analysis of the pro-system Ajo /pr occurring in the proof of Th. 2.4.1: he
strengthens step 1) by showing that the pro-system of kernels and cokernels of (A/pr )j →
1
(Ajo /pr )j is pro-isomorphic to a pro-system of (pg) p∞ -torsion modules; this allows to
apply various functors before passing to the limit j → ∞, whence a gain in flexibility.
More importantly, he obtains the following derived version of the direct summand conjecture, which had been conjectured by J. de Jong:
2.7.1. Theorem. [5] Let R be a regular Noetherian ring, and f : X → Spec R be a proper
surjective morphism. Then the map R → RΓ(X, OX ) splits in the derived category D(R).
3. E XISTENCE OF ( BIG ) C OHEN -M ACAULAY
ALGEBRAS
3.1. Cohen-Macaulay rings, and Cohen-Macaulay algebras for non-Cohen-Macaulay
rings. Let S be a local Noetherian ring with maximal ideal m and residue field k. Recall
that a sequence x = (x1 , . . . , xn ) in m is secant3 if dim S/(x) = dim S − n, and regular if
for every i, multiplication by xi is injective in S/(x1 , . . . , xi−1 )S. Any regular sequence
3following Bourkaki’s terminology (for instance); it is also often called “part of a system of parameters”,
although grx S may not be a polynomial ring in the “parameters” xi (it is if x is regular).
6
Y. ANDRÉ
is secant, and if the converse holds, S is said to be Cohen-Macaulay. Regular local rings
have a secant sequence generating m, and are Cohen-Macaulay.
Cohen-Macaulay rings form the right setting for Serre duality and the use of local homological methods in algebraic geometry, and have many applications to algebraic combinatorics [7]. When confronted with a non-Cohen-Macaulay ring S, one may try two
expedients:
1) Macaulayfication: construct a proper birational morphism X → Spec S such that
all local rings of X are Cohen-Macaulay. This weak resolution of singularities, introduced
by Faltings, has been established in general by T. Kawasaki [23]. Hovewer, secant sequences in S may not remain secant (hence not become regular) in the local rings of X;
this motivates the second approach:
2) Construction of a Cohen-Macaulay algebra4: an S-algebra C such that any secant
sequence of S becomes regular in C, and mC 6= C.
The existence of Cohen-Macaulay algebras implies the direct summand conjecture: indeed, if C is a Cohen-Macaulay algebra for a finite extension S of a regular local ring R,
it is also a Cohen-Macaulay R-algebra; this implies that R → C is faithfully flat, hence
pure, and so is R → S.
3.2. Constructions of Cohen-Macaulay algebras. The existence of a (big) CohenMacaulay algebra was established by Hochster and C. Huneke under the assumption that
S contains a field [22]. One may assume that S is a complete local domain. In char. p > 0,
one may then take C to be the total integral closure of S (i.e. the integral closure of S in
an algebraic closure of its field of fractions). This is no longer true in the case of equal
char. 0, which can nevertheless be treated by reduction to char. p >> 0 using ultraproduct
techniques.
The remaining case of mixed characteristic was settled in [2], using the same perfectoid
methods, so that one has:
3.2.1. Theorem. Any local Noetherian ring S admits a (big) Cohen-Macaulay algebra C.
In the case of a complete local domain S of char (0, p) and perfect residue field k
(to which one reduces), one proceeds as follows. Cohen’s theorem allows to present S
as a finite extension of R = W [[x1 , . . . , xn ]]. One first considers the R-algebra A :=
1
1
i
1
i
1
ˆ i W (k)[p pi ][[x1p , . . . , xnp ]][ 1p ])hg p∞ io and the S-algebra Bo = lim Bjo as above 2.6. It fol(∪
1
lows from Th. 2.4.1 and Th. 2.5.1 that Bo is (pg) p∞ -almost isomorphic to a faithfully flat R algebra
1
modulo any power of p. From there, one deduces that the sequence (p, x1 , . . . , xn ) is “(pg) p∞ almost regular” in Bo .
To get rid of “almost”, Lemma 2.3.1 is no longer sufficient: instead, one uses Hochster’s technique of monoidal modifications [19][22]. After m-completion, one gets a S-algebra C in which
(p, x1 , . . . , xn ), as well as any other secant sequence of S, becomes regular.
Subsequently, using the tilting equivalence between perfectoid algebras in char. 0 and
in char.p and applying Hochster’s modifications in char.p rather than in char.0, K. Shimomoto [34] shows that in Th. 3.2.1, in mixed characteristic, C can be taken to be perfectoid5.
In particular, if S is regular, it admits a perfectoid faithfully flat algebra (one may speculate
about the converse6).
4since it would be too restrictive to impose that C is Noetherian, one often speaks of “big” Cohen-Macaulay
algebra.
5actually, Shimomoto’s result is slightly weaker, but can be enhanced.
6this is settled in forthcoming work by Bhatt, Ma, Iyangar.
PERFECTOID SPACES AND THE HOMOLOGICAL CONJECTURES.
7
3.3. Finite and fpqc covers. Since Cohen-Macaulay algebras for regular local rings are
faithfully flat, Th. 3.2.1 implies [2]:
3.3.1. Theorem. Any finite cover of a regular scheme is dominated by a faithfully flat
quasi-compact cover.
If regularity is omitted, Spec(Q[x, y]/(xy)) and its normalization provide a counterexample.
4. H OMOLOGICAL
CONJECTURES
4.1. Origins from intersection theory. Under the influence of M. Auslander, D. Buchsbaum and J.-P. Serre, commutative algebra has shifted in the late 50s from the study of
ideals of commutative rings to the homological study of modules (cf. their characterization
of regular local rings by the existence of finite free resolutions for any finitely generated
module, resp. for the residue field).
Serre proved that for any three prime ideals p, q, r of a regular local ring R such that r is
a minimal prime of p + q, ht r ≤ ht p + ht q [32]. The special case r = m can be amplified:
for any ideals I, J of R such that I + J is m-primary, dim R/I + dim R/J ≤ dim R.
This is no longer true if R is not regular, and attempts to understand the general situation
led to the so-called homological conjectures cf. [7, ch. 9], [20].
4.2. Intersection conjectures. Let (R, m) be local Noetherian ring.
The first “intersection conjecture” was proposed by C. Peskine and L. Szpiro [27],
proved by them when R contains a field by reduction to char. p and Frobenius techniques, and later proved in general by P. Roberts using K-theoretic methods [29]. It states
that if M, N are finitely generated R-modules such that M ⊗ N has finite length, then
dim N ≤ pd M . It implies that R is Cohen-Macaulay if and only if there is an R-module
M of finite length and finite projective dimension (in the spirit of Serre’s characterization
of regular rings, which is the case M = k), resp. if there is an R-module M of finite
injective dimension.
Indeed, one may take N = R and deduce that dim R ≤ pd M ; by the Auslander-Buchsbaum
formula, pd M = depth R − depth M , so that the inequality depth R ≤ dim R is an equality. The
second assertion follows from the fact that id M = depth R [3].
The “new intersection conjecture”, also proved by Peskine, Szpiro [27] and Roberts
[29], states that for any non exact complex F• of free R-modules concentrated in degrees
[0, s] with finite length homology, s > dim R.
The “improved new intersection conjecture” is a variant due to E. Evans and P. Griffith
[11], in which the condition on F• is “slightly” relaxed: the Hi>0 are of finite length and
there exists a primitive cyclic submodule of H0 of finite length. They proved it, assuming
the existence of (big) Cohen-Macaulay algebras7, and showed that it implies their “syzygy
conjecture”. In spite of appearances, the passage from the new intersection conjecture to
its “improved” variant is no small step8: in fact, according to Hochster [20] and S. Dutta
[9], the latter is equivalent to the direct summand conjecture.
7if R is Cohen-Macaulay, the Buchsbaum-Eisenbud criterion gives a condition for the exactness of F in
•
terms of codimension of Fitting ideals of syzygies. In general, the same condition guarantees that F• ⊗ C is
exact for any Cohen-Macaulay R-algebra C [7, 9.1.8].
8
K-theoretic techniques failed to make the leap.
8
Y. ANDRÉ
On the other hand, in the wake of the new intersection conjecture (and motivated by the
McKay correspondence in dimension 3 and the “fact” that threefold flops induce equivalences of derived categories), T. Bridgeland and S. Iyengar obtained a refinement of Serre’s
criterion for regular rings assuming the existence of Cohen-Macaulay algebras [6, 2.4].
By Th. 1.2.1 and Th. 3.2.1, the improved new intersection conjecture and the
Bridgeland-Iyengar criterion thus hold inconditionally:
4.2.1. Theorem. Let R be a Noetherian local ring and F• be a complex of finitely generated free R-modules concentrated in degree [0, s], such that H>0 (F• ) has finite length.
(1) If H0 (F• ) contains a cyclic R-submodule of finite length not contained in
mH0 (F• ), then s ≥ dim R.
(2) If H0 (F• ) has finite length and contains k as a direct summand, and s = dim R,
then R is regular.
And so does the syzygy conjecture:
4.2.2. Theorem. Let R be a Noetherian local ring and M a finitely generated R-module
of finite projective dimension s. Then for i ∈ {1, . . . , s − 1}, the i-th syzygy module of M
has rank ≥ i.
4.3. Further work around the homological conjectures using perfectoid spaces.
4.3.1. In [15], Heitmann and L. Ma show that Cohen-Macaulay algebras can be constructed in a way compatible with quotients S → S/p by primes of height one. Using
arguments similar to Bhatt’s derived techniques, they deduce the vanishing conjecture for
maps of Tor [22]:
4.3.1. Theorem. [15] Let R → S → T be morphisms such that the composed map is
a local morphism of mixed characteristic regular local rings, and S is a finite torsionfree extension of R. Then for every R-module M and every i, the map TorR
i (M, S) →
(M,
T
)
vanishes.
TorR
i
They obtain the following corollary, which generalizes results by Hochster and J.
Roberts, J.-F. Boutot et al.:
4.3.2. Corollary. [15] Let R ֒→ S be a pure, local morphism, with S regular. Then R is
pseudo-rational, hence Cohen-Macaulay.
4.3.2. In [25], Ma and K. Schwede define and study perfectoid multiplier/test ideals in
mixed characteristic, and use them to bound symbolic powers of ideals in regular domains
in terms of ordinary powers:
4.3.3. Theorem. [25] Let R be a regular excellent Noetherian domain and let I ⊂ R
be a radical ideal such that each minimal prime of I has height ≤ h. Then for every n,
I (hn) ⊂ I n .
Here I (hn) denotes the ideal of elements of R which vanish generically to order hn at
I. When R contains a field, the result was proved in [10] and [21].
4.3.3. An efficient and unified way of dealing with questions related to the homological
conjectures in char. p is provided by “tight closure theory”, which has some flavor of
almost ring theory. Using Th. 2.4.1 and Th. 2.5.1 above, Heitmann and Ma give evidence
that the “extended plus closure” introduced in [13] is a good analog of tight closure theory
in mixed characteristic [16].
PERFECTOID SPACES AND THE HOMOLOGICAL CONJECTURES.
9
R EFERENCES
[1] Y. André, Le lemme d’Abhyankar perfectoı̈de, preprint (2016), to appear in Publ. Math. I.H.E.S. doi
10.1007/s10240-017-0096-x
[2] Y. André, La conjecture du facteur direct, preprint (2016), to appear in Publ. Math. I.H.E.S. doi
10.1007/s10240-017-0097-9
[3] H. Bass, On the ubiquity of Gorenstein rings, Math. Z. 82 (1963), 8-28.
[4] B. Bhatt, Almost direct summands, Nagoya math. J. 214 (2014), 195-204.
[5] −, On the direct summand conjecture and its derived variant, preprint arXiv:1608.08882
[6] T. Bridgeland, S. Iyengar, A criterion for regular local rings, C. R. A. S. Paris I 342 (2006), 723-726.
[7] W. Bruns, J. Herzog, Cohen-Macaulay rings, Cambridge Univ. Press 39 (1998).
[8] G. Faltings, Almost étale extensions, Astérisque 279 (2002), 185-270.
[9] S. Dutta, On the canonical element conjecture, Trans. Amer. Math. Soc. 299 (1987), 803-811.
[10] L. Ein, R. Lazarsfeld, K. Smith, Uniform bounds and symbolic powers on smooth varieties, Invent. Math.
144 (2001), 241-252.
[11] E. Evans, P. Griffith, The syzygy problem, Annals of Maths. 114 (1981), 323-333.
[12] O. Gabber, L. Ramero, Almost ring theory, Lecture Notes in Math. 1800, Springer (2003).
[13] R. Heitmann, Extensions of plus closure, J. of Algebra 238 (2001), 801826.
[14] −, The direct summand conjecture in dimension three, Ann. of Maths 156 (2002), 695-712.
[15] −, L. Ma, Big Cohen-Macaulay algebras and the vanishing conjecture for maps of Tor in mixed characteristic, preprint arXiv:1703.08281
[16] −−, Extended plus closure in complete local rings, preprint arXiv:1708.05761
[17] M. Hochster, Contracted ideals from integral extensions of regular rings, Nagoya Math. J., 51 (1973), 25-43.
[18] − , Canonical elements in local cohomology modules and the direct summand conjecture, J. of Algebra 84
(1983), 503-553.
[19] −, Big Cohen-Macaulay algebras in dimension three via Heitmann’s theorem, J. Algebra 254 (2002), 395408.
[20] −, Homological conjectures, old and new, Illinois J. Math. 51, no 1 (2007), 151-169.
[21] −, C. Huneke, Tight closure, invariant theory, and the Briançon-Skoda theorem, J. Amer. Math. Soc. 3
(1990) 31116.
[22] −− Applications of the existence of big Cohen-Macaulay algebras, Adv. in maths 113 (1995), 45-117.
[23] T. Kawasaki, On Macaulayfication of Noetherian Schemes , Trans. A. M. S. 352, No. 6 (2000), 2517-2552.
[24] K. Kedlaya, R. Liu, Relative p-adic Hodge theory, I: Foundations, S. M. F. Astérisque 371 (2015).
[25] L. Ma, K. Schwede, Perfectoid multiplier/test ideals in regular rings and bounds on symbolic powers,
preprint arXiv:1705.02300
[26] T. Ohi, Direct summand conjecture and descent for flatness, Proc. A.M.S. 124, 7 (1996), 1967-1968.
[27] C. Peskine, L. Szpiro, Dimension projective finie et cohomologie locale, Publ. Math. I.H.E.S. 42 (1973),
323-395.
[28] M. Raynaud, L. Gruson, Critères de platitude et de projectivité, Inventiones math. 13 (1971), 1-89.
[29] P. Roberts, Le théorème d’intersection, C. R. Acad. Sc. Paris 304 (1987), 177-180.
[30] P. Scholze, Perfectoid spaces, Publ. Math. I.H.E.S. 116, no 1 (2012), 245-313.
[31] −, On torsion in the cohomology of locally symmetric varieties, Annals of Maths. 182 (2015), 945-1066.
[32] J.-P. Serre, Algèbre locale, multiplicités, Springer LNM 11 (1965).
[33] K. Shimomoto, An application of the almost purity theorem to the homological conjectures, J. pure applied
algebra 220 (2014), 621-632.
[34] −, Integral perfectoid big Cohen-Macaulay algebras via André’s theorem, preprint arXiv:1706.06946
I NSTITUT DE M ATH ÉMATIQUES DE J USSIEU , 4 PLACE J USSIEU , 75005 PARIS F RANCE .
E-mail address: [email protected]
| 0math.AC
|
AEkNN: An AutoEncoder kNN-based classifier with
built-in dimensionality reduction
Francisco J. Pulgar1
· Francisco Charte1
1
Antonio J. Rivera
· Marı́a J. del Jesus1
·
arXiv:1802.08465v2 [cs.LG] 9 Mar 2018
Abstract High dimensionality, i.e. data having a large number of variables, tends
to be a challenge for most machine learning tasks, including classification. A classifier usually builds a model representing how a set of inputs explain the outputs.
The larger is the set of inputs and/or outputs, the more complex would be that
model. There is a family of classification algorithms, known as lazy learning methods, which does not build a model. One of the best known members of this family
is the kNN algorithm. Its strategy relies on searching a set of nearest neighbors, using the input variables as position vectors and computing distances among them.
These distances loss significance in high-dimensional spaces. Therefore kNN, as
many other classifiers, tends to worse its performance as the number of input
variables grows.
In this work AEkNN, a new kNN-based algorithm with built-in dimensionality
reduction, is presented. Aiming to obtain a new representation of the data, having
a lower dimensionality but with more informational features, AEkNN internally
uses autoencoders. From this new feature vectors the computed distances should
be more significant, thus providing a way to choose better neighbors. A experimental evaluation of the new proposal is conducted, analyzing several configurations
and comparing them against the classical kNN algorithm. The obtained conclusions demonstrate that AEkNN offers better results in predictive and runtime
performance.
Keywords kNN · deep learning · autoencoders · dimensionality reduction · high
dimensionality
F. J. Pulgar
Tel.: +34 953 211 956
[email protected]
F. Charte
[email protected]
A. J. Riverea
[email protected]
M. J. del Jesus
[email protected]
1
Department of Computer Science, University of Jaén, 23071 Jaén, Spain
2
F. J. Pulgar et al.
1 Introduction
Classification is a well-known task within the area of machine learning [29]. The
main objective of a classifier is to find a way to predict the label to be assigned
to new data patterns. To do so, usually a model is created from previously labeled data. In traditional classification, each example has a single label. Different
algorithms have been proposed in order to address this work. One of the classic
methodologies is instance-based learning (IBL) [1]. Essentially, this methodology is
based on local information provided by the training instances, instead of constructing a global model from the whole data. The algorithms belonging to this family
are relatively simple, however they have demonstrated to obtain very good results
in facing the classification problem. A traditional example of such an algorithm is
k-nearest neighbors (kNN) [23].
The different IBL approaches, including the kNN algorithm [13], have difficulties when faced with high-dimensional datasets. These datasets are made of samples having a large number of features. In particular, the kNN algorithm presents
problems when calculating distances in high-dimensional spaces. The main reason
is that distances are less significant as the number of dimensions increases, tending
to equate [13]. This effect is one of the causes of the curse of dimensionality, which
occurs when working with high-dimensional data [8, 9]. Another consequence that
emerges in this context is the Hughes phenomenon. This fact implies that the
predictive performance of a classifier decreases as the number of features of the
dataset grows, keeping the number of examples constant [44]. In other words, more
instances would be needed to maintain the same level of performance.
Several approaches have been proposed for facing the dimensionality reduction
task. Recently, a few proposals based on deep learning (DL) [25, 11] have obtained
good results while tackling this problem. The rise of these techniques is produced
by the good performance that DL models have had in many research areas, such
as computer vision, automatic speech processing, or audio and music recognition.
In particular, autoencoders (AEs) are DL networks offering good results due to
their architecture and operation [17, 41, 84, 67, 94, 14].
High dimensionality is usually mitigated by transforming the original input
space into a lower-dimensional one. In this paper an instance-based algorithm,
that internally generates a reduced set of features, is proposed. Its objective is to
obtain a better IBL method, able to deal with high-dimensional data. Specifically,
the present work introduces AEkNN, a kNN-based algorithm with built-in dimensionality reduction. AEkNN projects the training patterns into a lower-dimensional
space, relying in an AE for doing so. The goal is to produce new features of higher
quality from the input data. This approach is experimentally evaluated, and a
comparison between AEkNN and the kNN algorithm is performed considering predictive performance and execution time. The results obtained demonstrate that
AEkNN offers better results in both metrics. In addition, AEkNN is compared
with other traditional dimensionality reduction algorithms. This comparison offers an idea of the behavior of the AEkNN algorithm when facing the task of
dimensionality reduction.
An important aspect of the AEkNN algorithm that must be highlighted is that
it performs a transformation of the features of the input data, against other traditional algorithms that perform a simple selection of the most significant features.
AEkNN performs this transformation taking into account all the characteristics of
AEkNN: An AutoEncoder kNN-based classifier ...
3
the input data, although not all will have the same weight in the new generated
space.
In short, AEkNN combines two reference methods, kNN and AE, in order to
take advantage of kNN in classification and reduce the effects of high dimensionality by means of AE. In this way, the proposed method presents a baseline for
future works. Summarizing, the main contributions of this work are 1) the design
of a new classification algorithm, AEkNN, which combines an efficient dimensionality reduction mechanism with a popular classification method, 2) an analysis of
the AEkNN operating parameters that allows selecting the best algorithm configuration, 3) an experimental demonstration of the improvement that AEkNN
achieves with respect to the kNN algorithm, and 4) an experimental comparison
between the use of an autoencoder for dimensionality reduction with respect to
other classical methods such as PCA and LDA.
This paper is organized as follows. In Section 2 a few fundamental concepts,
such as machine learning, classification, and the kNN algorithm, are briefly introduced, and some details about DL techniques and AEs are provided. Section 3
describes relevant related works, focused on tackling the problem of dimensionality
reduction in kNN. In Section 4, the proposed AEkNN algorithm is introduced. Section 5 defines the experimental framework, as well as the different results obtained
from the experimentation. Finally, Section 6 provides the overall conclusions.
2 Preliminaries
AEkNN, the algorithm proposed in this work, is a kNN-based classification method
designed to deal with high-dimensional data. This section outlines the essential
concepts AEkNN is founded on, such as classification, nearest neighbors classification, DL techniques and AEs. The most basic concepts are introduced in subsection 2.1. The kNN algorithm is discussed in subsection 2.2, while DL and AEs
are briefly described in subsections 2.3 and 2.4.
2.1 Machine Learning and Classification
In general terms, machine learning is a subfield of artificial intelligence whose aim
is to develop algorithms and techniques able to generalize behaviors from information supplied as examples [46, 33]. The different tasks that can be performed on
machine learning can be classified following different criteria. One of these categorizations arises according to the training patterns used to train the machine learning system [3]. In this sense, it can be distinguished between supervised learning,
when patterns are labeled, such as classification and regression, and unsupervised
learning, when patterns are not labeled, such as clustering, among others.
Classification is one of the tasks performed in the data mining phase. It is a
predictive task that usually develops through supervised learning methods [47].
Its purpose is to predict, based on previously labeled data, the class to assign to
future unlabeled patterns. In traditional classification, datasets are structured as
a set of input attributes, known as features, and one output attribute, the class
or label. Depending on the number of values that this output class can take, the
classification problem can be seen as:
4
F. J. Pulgar et al.
– Binary classification, in which each pattern can only belong to one of two
classes.
– Multi-class classification, in which each pattern can only belong to one of a
limited set of classes. Consequently, a binary classification can be seen as a
problem of multi-class classification with only two classes.
Several issues can emerge while designing a classifier, being some of them related to high dimensionality. According to the Hughes phenomenon [44], the predictive performance of a classifier decreases as the number of features increases,
provided that the number of training instances is constant. Another phenomenon
that particularly affects IBL algorithms is the curse of dimensionality. IBL algorithms are based on the similarity of individuals, calculating distances between
them [1]. These distances tend to lose significance as dimensionality grows.
2.2 The kNN Algorithm
kNN is a non-parametric algorithm developed to deal with classification and regression tasks[4, 23]. In classification, kNN predicts the class for new instances
using the information provided by the k nearest neighbors, so that the assigned
class will be the most common among them. Fig. 1 shows a very simple example
on how kNN works with different k values. As can be seen, the prediction obtained
with k = 3 would be B, with k = 5 would be A and with k = 11 would be A.
Fig. 1: k-Nearest neighbors algorithm in a bi-dimensional space.
An important feature of this algorithm is that it does not build a model for
accomplishing the prediction task. Usually, no work is done until an unlabeled data
pattern arrives, thus the denomination of lazy approach [5]. Once the instance to
be classified is given, the information provided by its k nearest neighbors [24] is
used as explained above.
One of kNN’s main issues is its behavior with datasets having a high-dimensional
input space, due to the loss of significance of traditional distances as the dimensionality of the data increases [13]. In such a high-dimensional space distances
between individuals tend to be the same. As a consequence similarity/distancebased algorithms, such as kNN, usually do not offer adequate results.
AEkNN: An AutoEncoder kNN-based classifier ...
5
In this article, kNN has been selected to perform classification tasks. kNN
is very popular since it has a good performance, uses few resources and it is
relatively simple [60, 88, 59, 28]. The objective of this proposal is to present an
algorithm that combines the advantages of kNN to classify with DL models to
reduce dimensionality.
2.3 Deep Learning
DL [87, 25] arises with the objective of extracting higher-level representations
of the analyzed data. In other words, the main goal of DL-based techniques is
learning complex representations of data. The main lines of research in this area
are intended to define different data representations and create models to learn
them [12].
As the name suggests, models based on DL are developed as multi-layered
(deep) architectures, which are used to map the relationships between features in
the data (inputs) and the expected result (outputs) [34, 10]. Most DL algorithms
learn multiple levels of representations, producing an hierarchy of concepts. These
levels correspond to different degrees of abstraction. The following are some of the
main advantages of DL:
– These models can handle a large number of variables and generate new features
as part of the algorithm itself, not as an external step [25, 19, 76].
– Provides performance improvements in terms of time needed to accomplish
feature engineering, one of the most time-consuming tasks [34].
– Achieves much better results than other methods based on traditional tech-
niques [20, 49, 40, 90] while facing problems in certain fields, such as image,
speech recognition or malware detection.
– DL-based models have a high capacity of adaptation for facing new problems
[87, 10].
Recently, several new methods [11, 52, 25] founded on the good results produced by DL have been published. Some of them are focused on certain areas,
such as image processing and voice recognition [52]. Other DL-based proposals
have been satisfactorily applied in disparate areas, gaining advantage over prior
techniques [25]. Due to the great impact of DL-based techniques, as well as the impressive results they usually offer, new challenges are also emerging in new research
lines [11].
There are two main reasons behind the rise of DL techniques, the large amount
of data currently available and the increase in processing power. In this context,
different DL architectures have been developed: AEs (section 2.4), convolutional
neural networks [50], long short-term memory [42], recurrent neural networks [69],
gated recurrent unit [18], deep Boltzmann machines [39], deep stacking networks
[26], deep coding networks [55], deep recurrent encoder [74], deep belief networks
[45, 53], among others [25, 10].
DL models have been widely used to perform classification tasks obtaining good
results [20, 49, 30, 27]. However, the objective of this proposal is not to perform
the classification directly with these models, but use them to the dimensionality
reduction task.
6
F. J. Pulgar et al.
The goal of the present work is to obtain higher-level representations of the
data but with a reduced dimensionality. One of the dimensionality reduction DLbased techniques that has achieved good results are AEs [56]. An AE is an artificial
neural network whose purpose is to reproduce the input into the output, in order
to generate a compressed representation of the original information [10]. Section
2.4 thoroughly describes this technique.
2.4 Autoencoders
An AE is an artificial neural network able to learn new information encodings
through unsupervised learning [17]. AEs are trained to learn an internal representation that allows reproducing the input into the output through a series of
hidden layers. The goal is to produce a more compressed representation of the
original information in the hidden layers, so that it can be used in other tasks.
AEs are typically used for dimensionality reduction tasks by their characteristics
and performance [41, 84, 67, 94, 14, 73, 93]. Therefore, the importance of such
networks in this paper.
The most basic structure of an AE is very similar to that of a multilayer perceptron. An AE is a feedforwark neural network without cycles, so the information
always goes in the same direction. An AE is typically formed by a series of layers:
an input layer, a series of hidden layers and an output layer, being the units in
each layer connected to those in the next one. The main characteristic of AEs is
that the output has the same number of nodes than the input, since the goal is to
reproduce the latter into the former throughout the learning process [10].
Two parts can be distinguished in an AE, the encoder and the decoder. The
first one is made up of the input layer and the first half of hidden layers. The
second is composed of the second half of hidden layers and the output layer. This
is the architecture shown in Fig. 2. As can be seen, the structure of an AE always
is symmetrical.
Fig. 2: Architecture of an autoencoder with three hidden layers.
AEkNN: An AutoEncoder kNN-based classifier ...
7
The encoder and decoder parts in an AE can be defined as functions ω (Eq.
(1)) and β (Eq. (2)), so that:
ω:X→F
(1)
β:F →X
(2)
Where x ∈ Rd = X is the input to the AE, and z ∈ Rp = F is the mapping
contained in the hidden layers of the AE. When there is only one hidden layer (the
most basic case) the AE maps the input X onto Z. For doing this, a weight vector
W and a bias parameter b are used:
z = γ 1 (W x + b )
(3)
Eq. (3) corresponds to the compression function, wherein a encoded input
representation is obtained. Here, γ1 is an activation function such as a rectified
linear unit or a sigmoid function.
The next step is to decode the new representation z to obtain x’, in the same
way as x was projected into z, by means of a weight vector W’ and a bias parameter
b’. Eq. (4) corresponds to the decoder part, where the AE reconstructs the input
from the information contained in the hidden layer.
x’ = γ2 (W 0 z + b0 )
(4)
During the training process, the AE tries to reduce the reconstruction error.
This operation consists of back-propagating the obtained error through the network, and then modifying the weights to minimize such error. Algorithm 1 shows
the pseudo-code of this process.
Algorithm 1 AE training algorithm’s pseudo-code.
1:
2:
3:
4:
5:
6:
7:
8:
9:
Inputs:
T rainData
. For each training instance:
for each instace in T rainData do
. Do a feed-forward through AE to obtain output:
newCod ← feedForwardAE(aeM odel, instance)
. Calculate error:
error ← calculateError(newCod, instance)
. Backpropagate the error through AE and perform weight update:
aeModel ← backpropagateError(eror)
end for
. Train Data
Learning a representation that allows reproducing the input into the output
could seem useless at first sight, but in this case the output is not of interest.
Instead, the concern is in the new representation of the inputs learned in the
hidden layers. Such new codification is really interesting, because it can have very
useful features [34]. The hidden layers learn a higher-level representation of the
original data, which can be extracted and used in independent processes.
Depending on the number of units the hidden layers have, two types of AEs
can be distinguished:
8
F. J. Pulgar et al.
– Those whose hidden layer have less units than the input and output layers.
This type of AE is called undercomplete. Its main objective is to force the
network to learn a compressed representation of the input, extracting new,
higher-level features.
– Those whose hidden layer have more units than the input and output layers.
This type of AE is called overcomplete. The main problem in this case is
that the network can learn to copy the input to the output without learning
anything useful, so when it is necessary to obtain an enlarged representation
of the input it is necessary to use other tools to prevent this problem.
In conclusion, AEs are a very suitable tool for generating a new lower-dimensional
input space made of higher-level features. AEs have obtained good results in the
accomplishment of this task. This is the main reason to choose this technique to
design the AEkNN algorithm described later. However, it is important to note
that there are other methods of dimensionality reduction that produce good results, such as denoising autoencoders [83], restricted Boltzmann machines [70] or
sparse coding [89]. The objective of this proposal is to present an algorithm that
hybridizes kNN with AEs. This establishes a baseline that allows supporting studies with more complex methods.
3 Dimensionality reduction approaches
In this section, an exploration of previous works related to the proposal made
in this paper is carried out. The subsection 3.1 introduces classical proposals to
tackle the dimensionality reduction problem. Some approaches for facing the dimensionality reduction task for kNN are portrayed in subsection 3.2.
In automatic learning, dimensionality reduction is the process aimed to decrease the number of considered variables, by obtaining a subset of main features.
Usually two different dimensionality reduction methods are considered:
– Feature selection [58], where the subset of the initial features that provides
more useful information is chosen. The final features have no transformation
in the process.
– Feature extraction [57], where the process constructs from the initial features
a set of new ones providing more useful and non-redundant information, facilitating the next steps in machine learning and in some cases improving understanding by humans.
3.1 Classical proposals for dimensionality reduction
Most dimensionality reduction techniques can be grouped into two categories,
linear and non-linear approaches [79]. Below some representative proposals found
in the literature, those that can be considered as traditional methods, are depicted.
Commonly, classic proposals for dimensionality reduction were developed using
linear techniques. The following are some of them:
– Principal Components Analysis (PCA) [63, 43] is a well-known solution for di-
mensionality reduction. Its objective is to obtain the lower-dimensional space
AEkNN: An AutoEncoder kNN-based classifier ...
9
where the data are embedded. In particular, the process starts from a series of
correlated variables and converts them into a set of uncorrelated linear variables. The variables obtained are known as principal components, and their
number is less than or equal to the original variables. Often, the internal structure extracted in this process reflects the variance of the data used.
– Factors analysis [75] is based on the idea that the data can be grouped according to their correlations, i.e. variables with a high correlation will be within
the same group and variables with low correlation will be in different groups.
Thus, each group represents a factor in the observed correlations. The potential
factors plus the terms error are linearly combined to model the input variables.
The objective of factor analysis is to look for independent dimensions.
– Classical scaling [78] consists in grouping the data according to their similarity
level. An algorithm of this type aims to place each data in an N-dimensional
space where the distances are maintained as much as possible.
Despite their popularity, classical linear solutions for dimensionality reduction
present the problem that they can not correctly handle complex non-linear data
[79]. For this reason, nonlinear proposals for dimensionality reduction arose. A
compilation of these is presented in [15, 71, 54, 81]. Some of these techniques
are: Isomap [77], Maximum Variance Unfolding [86], diffusion maps [21], manifold
charting [79], among others. These techniques allow to work correctly with complex
non-linear data. This is an advantage when working with real data, which are
usually of this type.
3.2 Proposals for dimensionality reduction in kNN
There are different proposals trying to face the problems of kNN when working
with high-dimensional data. In this section, some of them are collected:
– A method for computing distances for kNN is presented in [91]. The proposed
algorithm performs a partition of the total data set. Subsequently, a reference
value for each partition made is assigned. Grouping the data in different partitions allows to obtain a space of smaller dimensionality where the distances
between the reference points are more significant. The method depends on the
division of the data performed and the selection of the reference. This is a
negative aspect, since a poor choice of parameters can greatly influence the
final results.
– The authors of [48] analyze the curse of dimensionality phenomenon, which
states that in high-dimensional spaces the distances between the data points
become almost equal. The objective is to investigate when the different methods proposed reach their limits. To do this, they perform an experiment in
which different methods are compared with each other. In particular, it is exposed that the kNN algorithm begins to worsen the results when the space
exceeds eight dimensions. A proposed solution is to adapt the calculation of
distances to high-dimensional spaces. However, this approach does not consider
a transformation of the initial data to a lower-dimensional space.
– The proposal in [85] is a kNN-based method called kMkNN, whose objective is
to improve the search of the nearest neighbors in a high-dimensional space. This
problem is approached from another point of view, being the goal to accelerate
10
F. J. Pulgar et al.
the computation of distances between elements without modifying the input
space. To do this, kMkNN relies on k-means clustering and the triangular
inequality. The study shows a comparison with the original kNN algorithm
where it is demonstrated that kMkNN works better considering the execution
time, although it is not as effective when predictive performance is taken into
account.
– A new aspect related to the curse of dimensionality phenomenon, occurring
while working with the kNN algorithm, is explored in [65]. It refers to the
number of times that a particular element appears as the closest neighbor
of the rest of elements. The main objective of the study is to examine the
origins of this phenomenon as well as how it can affect dimensionality reduction
methods. The authors analyze this phenomenon and some of its effects on kNN
classification, providing a foundation which allows making different proposals
aimed to mitigate these effects.
– In [37], the problem of finding the nearest neighbors in a high-dimensional
space is analyzed. This is a difficult problem both from the point of view of
performance and the quality of results. In this study, a new search approach is
proposed, where the most relevant dimensions are selected according to certain
quality criteria. Thus, the different dimensions are not treated in the same way.
This can be seen as an extraction of characteristics according to a particular
criterion. Finally, an algorithm based on the previous approach, that faces the
problem of the nearest neighbor, is proposed. However, this method makes a
selection of the initial features that meet a certain criterion, so it does not
take into account all the input features. Therefore, it could discard important
information in the process.
– A method called DNet-kNN is presented in [62]. It consists in a non-linear
feature mapping based on Deep Neural Network, aiming to improve classification by kNN. DNet-kNN relies on Restricted Boltzmann Machines (RBM)
to perform the mapping of characteristics. RBMs are another type of DL network. The work offers a solution to the problem of high-dimensional data when
using kNN by combining this traditional classifier with DL-based techniques.
The conducted experimentation proves that DNet-kNN improves the results
of the classical kNN. DNet-kNN requires a pre-training of the RBM, which is
an additional phase. In addition, experimentation is performed on datasets of
digits and letters, not on actual images.
The aforementioned proposals represent a list of approaches that are closely
related to the problem dealt with in this article. However, there are many other
proposals that face the problem of dimensionality reduction from other perspectives [68, 7, 16, 36, 92].
In conclusion, different proposals have arisen to analyze and try to address the
problems of IBL algorithms when they have to deal with high-dimensional data.
These methods are affected by the curse of dimensionality, which raises the need to
bring new approaches. Among the previous proposals, there is no one that presents
an hybrid method based on IBL that incorporates the reduction of dimensionality
intrinsically. In addition, none of the proposals obtains improvements both in
predictive performance as well as in execution time. The present work is aimed to
fulfill these aspects.
AEkNN: An AutoEncoder kNN-based classifier ...
11
4 AEkNN: An AutoEncoder kNN-based classifier with built-in
dimensionality reduction
Once the main foundations behind our proposed algorithm have been described,
this section proceeds to present AEkNN. It is an instance-based algorithm with
a built-in dimensionality reduction method, performed with DL techniques, in
particular by means of AEs.
4.1 AEkNN Foundations
As mentioned, high dimensionality is an obstacle for most classification algorithms.
This problem is more noticeable while working with distance-based methods, since
as the number of variables increases these distances are less significant [44, 48].
In such situation an IBL method could loss effectiveness, since the distances between individuals are equated. As a consequence of this problem, different methods
able to reduce its effects have been proposed. Some of them have been previously
enumerated in Section 3.2.
AEkNN is a new approach in the same direction, introducing the use of AEs to
address the problem of high dimensionality. The structure and performance of this
type of neural networks makes it suitable for this task. As explained above, AEs are
trained to reproduce the input into the output, through a series of hidden layers.
Usually, the central layer in AEs has less units than the input and output layers.
Thereby, this bottleneck forces the network to learn a more compact and higherlevel representation of the information. The coding produced by this central layer
can be extracted and used as the new set of features in the classification stage.
In this sense, there are different studies demonstrating that better results are
obtained with AEs than with traditional methods, such PCA or multidimensional
scaling [41, 79]. Also, there are studies analyzing the use of AEs from different
perspectives, either focusing on the training of the network and its parameters
[84] or on the relationship between the data when building the model [67].
AEkNN is an instance-based algorithm designed to address classification with
a high number of variables. It starts working with an N -dimensional space X that
is projected into an M -dimensional space Z, being M < N. This way M new features, presumably of higher level than the initial ones, are obtained. Once the new
representation of the input data is generated, it is possible to get more representative distances. To estimate the output, the algorithm uses the distances between
each test example and the training ones but based on the M higher-level features.
Thus, the drawbacks of high-dimensional data in distances computation can be
significantly reduced. As can be seen, AEkNN is a non-lazy instance-based algorithm. It starts by generating the model in charge of producing the new features,
unlike the lazy methods that do not have a learning stage or model. AEkNN allows to enhance the predictive performance, as well as obtaining improvements in
execution time, when working with data having a large number of features.
12
F. J. Pulgar et al.
4.2 Method description
AEkNN consists of two fundamental phases. Firstly, the learning stage is carried
out using the training data to generate the AE model that allows to produce a
new encoding of the data. Secondly, the classification step is performed. It uses
the model generated in the first phase to obtain the new representation of the test
data and, later, the class for each instance is estimated based on nearest neighbors.
Algorithm 2 shows the pseudo-code of AEkNN, that is thoroughly discussed below,
while Fig. 3 shows the algorithm process in a general way.
Fig. 3: Method description.
The inputs to the algorithm are TrainData and TestData, the train and test
data to be processed, k, the number of neighbors, and PPL, the percentage of
elements per layer (PPL). This latter parameter sets the structure of the AE, i.e.
the number of layers and elements per layer. It is a vector made of as many items
as hidden layers are desired in the AE, indicating each one the percentage of units
in that layer with respect to the number of input features. In section 5, different
configurations are analyzed to find the one that offers the best results.
The algorithm is divided into two parts. The first part of the code (lines 29) corresponds to the training phase of AEkNN. The second part (lines 11-12)
refers to the classification phase. During training AEkNN focuses on learning a
new representation of the data. This is done through an AE, using the training
data to learn the weights linking the AE’s units. This is a process that has to be
repeated for each layer in the AE, stated by the number of elements in the PPL
parameter. This loop performs several tasks:
– In line 5 the function getSizeLayer is used to obtain the number of units in the
layer. This value will depend on the number of characteristics of the training
set (T rainData) and the percentage applied to the corresponding layer, which
is established by the PPL parameter.
– The function getAELayer (called in line 6 and defined in line 14) retrieves a
layer of the AE model. The layer allows to obtain a new representation of the
data given as first parameter (modelData). The number of units in the AE layer
generated in this iteration will be given by the second parameter (sizeLayer).
Firstly, the AE is initialized with the corresponding structure (line 15). The
AEkNN: An AutoEncoder kNN-based classifier ...
13
Algorithm 2 AEkNN algorithm’s pseudo-code.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
Inputs:
T rainData
. Train Data
T estData
. Test Data
PPL
. Percentage per layer
k
. Number of nearest neighbors
. Training phase:
modelData ← TrainData
aeModel ← ()
for each layer in PPL do
sizeLayer ← getSizeLayer(modelData, layer)
aeLayer ← getAELayer(modelData, sizeLayer)
modelData ← applyAELayer(aeLayer, modelData)
aeModel ← addAEModel(aeM odel, aeLayer)
end for
. Classification phase:
result ← classification(T rainData, T estData, k, aeM odel)
return result
function getAELayer(modelData, sizeLayer )
aeLayer ← initializeAEModel(modelData, sizeLayer)
for each instance in modelData do
outPut ← feedForwardAEModel(aeLayer, instance)
error ← calculateDeviation(instance, outP ut)
aeLayer ← updateWeightsAEModel(aeLayer, error)
end for
return aeLayer
end function
function classification(TrainData, TestDAta,k,aeModel)
error ← 0
for each instace in T estData do
newCod ← feedForwardAEModelVector(aeM odel, instance)
outPut ← distanceBasedClassification(newCod, k, T rainData)
if outP ut != realOutP ut(instance) then
error ← error + 1
end if
end for
result ← error / size(T estData)
return result
end function
number of units in the hidden layer is given by the variable computed in the
previous step, and the weights are randomly initialized. Secondly, for each train
instance the following steps take place:
1. The AE is used to obtain the output for the given instance (line 17).
2. The deviation of the given output with respect to the actual one is calculated (line 18).
3. The weights of the network are updated according to the obtained error
(line 19).
Finally, the generated AE layer is returned (line 21).
– The function applyAELayer (line 7) allows to obtain a new representation of
the data given as second parameter (modelData). To do this, the AE layer
previously generated, represented by the first parameter (aeLayer), is used.
14
F. J. Pulgar et al.
– The last step consists in adding the AE layer generated in the current iteration
to the complete AE model (line 8).
During classification (lines 11-12) the function classification is used (lines 2435). The class for the test instances given as the first parameter (T estData) is
predicted. The process performed internally in this function is to transform each
test instance using the AE model generated in the training phase (aeM odel), producing a new instance, more compact and representative. (line 27). This new set
of features is used to predict a class with a classifier based on distances, using for
each new example its k nearest neighbors (line 28). Finally, this function returns
the error rate (result) for the total set of test instances (line 33). As can be seen,
classification is conducted in a lower-dimensional space, mitigating the problems
associated with a high number of variables.
At this point, it should be clarified that the update of weights (lines 16-20)
is done using mini-batch gradient descent [38]. This is a variation of the gradient
descent algorithm that splits the training dataset into small batches that are used
to calculate the model error and update the model coefficients. The reason for
using this technique its better performance when dealing with large dataset.
From the previous description is can be inferred how AEkNN accomplish the
objective of addressing classification with high-dimensional data. On the one hand,
aiming to reduce the effects of working with a large number of variables, AEs have
been used to transform such data into a lower-dimensional space. On the other
hand, the classification phase is founded on the advantages of IBL. In Section 5,
the performance of AEkNN is analyzed.
5 Experimental study
In order to demonstrate the improvements provided by AEkNN, the algorithm
proposed in the present work, an experimental study was conducted. It has been
structured into three steps, all of them using the same set of datasets:
– The objective of the first phase is to determine how the PPL parameter in
AEkNN influences the obtained results. For this purpose, classification results
for all considered configurations are compared in subsection 5.2. At the end,
the value of the PPL parameter that offers the best results is selected.
– The second phase aims to verify whether AEkNN with the selected configuration improves the results provided by the classic kNN algorithm. In subsection
5.3, the results of both algorithms are compared.
– The third phase of the experimentation aims to assess the competitiveness of
AEkNN against traditional dimensionality reduction algorithms, in particular,
principal component analysis (PCA) and linear discriminant analysis (LDA).
In subsection 5.4, the results of the three algorithms are compared.
Subsection 5.1 describes the experimental framework, moreover the following
subsections present the results and their analysis.
5.1 Experimental framework
The conducted experimentation aims to show the benefits of AEkNN over a set
of datasets with different characteristics. Their traits are shown in Table 1. The
AEkNN: An AutoEncoder kNN-based classifier ...
15
datasets’ origin is shown in the column named Ref. For all executions, datasets
are given as input to the algorithms applying a 2×5 folds cross validation scheme.
Table 1: Characteristics of the datasets used in the experimentation.
Dataset
Samples
Number of
Features
Classes
image
drive
coil2000
dota
nomao
batch
musk
semeion
madelon
hapt
isolet
mnist
microv1
microv2
2310
58509
9822
102944
1970
13910
6598
1593
2000
10929
7797
70000
360
571
19
48
85
116
118
128
168
256
500
561
617
784
1300
1300
7
11
2
2
2
6
2
10
2
12
26
10
10
20
Type
Ref
Real
Real
Integer
Real
Real
Real
Integer
Integer
Real
Real
Real
Integer
Real
Real
[6]
[6]
[80]
[6]
[6]
[82]
[6]
[6]
[35]
[66]
[22]
[51]
[6]
[6]
In both phases of experimentation, the value of k for the classifier kNN and
for AEkNN will be 5, since it is the recommended value in the related literature.
In addition, to compare classification results was necessary to compute several
evaluation measures. In this experimentation, Accuracy (5), F-Score (6) and area
under the ROC curve (AUC) (9) were used.
Accuracy (5) is the proportion of true results among the total number of cases
examined.
TP + TN
Accuracy =
(5)
TP + TN + FP + FN
Where TP stands for true positives, instances correctly identified. FP is the
false positives, instances incorrectly identified. TN represents the true negatives,
instances correctly rejected. FN corresponds to false negatives, instances incorrectly rejected.
F-Score is the harmonic mean of Precision (7) and Recall (8), considering
Precision as the proportions of positive results that are true positive results and
Recall as the proportion of positives that are correctly identified as such. These
measures are defined by the Eqs. (6), (7), (8):
F-Score = 2 ∗
P recision ∗ Recall
P recision + Recall
Precision =
Recall =
TP
TP + FP
TP
TP + FN
(6)
(7)
(8)
Finally, AUC is the probability that a classifier will rank a randomly chosen
positive instance higher than a randomly chosen negative one. AUC is given by
the Eq. (9):
16
F. J. Pulgar et al.
Z
−∞
TPR(T )FPR(T )dT
AUC =
(9)
∞
Where TPR stands for the true positive rate and FPR is false positive rate.
The significance of results obtained in this experimentation is verified by appropriate statistical tests. Two different tests are used in the present study:
– In the first part, the Friedman test [31] is used to rank the different AEkNN
configurations and to establish if any statistical differences exist between them.
– In the second part, the Wilcoxon [72] non-parametric sign rank test is used.
The objective is to verify if there are significant differences between the results
obtained by AEkNN and kNN.
These experiments were run in a cluster made up of 8 computers, having 2
CPUs (2.33 GHz) and 7 GB RAM each of them. The AEkNN algorithm and the
experimentation was coded in R language [64], relying on the H2O package [2] for
some DL-related functions.
5.2 PPL Parameter Analysis
AEkNN has a parameter, named PPL, that establishes the configuration of the
model. This parameter allows selecting different architectures, both in number of
layers (depth) and number of neurons per layer.
The datasets used (see Table 1) have disparate number of input features, so the
architectures will be defined according to this trait. Table 2 shows the considered
configurations. For each model the number of hidden layers, as well as the number
of neurons in each layer, is shown. The latter is indicated as a percentage of
the number of initial characteristics. Finally, the notation of the associated PPL
parameter is provided.
Table 2: Configurations used in the experimentation and PPL parameter.
# hidden layers
AEkNN
AEkNN
AEkNN
AEkNN
AEkNN
AEkNN
1
2
3
4
5
6
1
1
1
3
3
3
Number of neurons (%)
Layer 1
Layer 2
Layer 3
25
50
75
150
150
150
25
50
75
150
150
150
PPL parameter
(25)
(50)
(75)
(150, 25, 150)
(150, 50, 150)
(150, 75, 150)
The results produced by the different configurations considered are presented
grouped by metric. Table 3 shows the results for Accuracy, Table 4 for F-Score
and Table 5 for AUC. These results are also graphically represented. Aiming to
optimize the visualization, two plots with different scales have been produced for
each metric. Fig. 4 represents the results for Accuracy, Fig. 5 for F-Score and Fig.
6 for AUC.
AEkNN: An AutoEncoder kNN-based classifier ...
17
Table 3: Accuracy classification results for test data.
Dataset
(0.25)
(0.5)
(0.75)
(1.5, 0.25, 1.5)
(1.5, 0.5, 1.5)
(1.5, 0.75, 1.5)
image
drive
coil2000
dota
nomao
batch
musk
semeion
madelon
hapt
isolet
mnist
microv1
microv2
0.930
0.779
0.929
0.509
0.904
0.995
0.974
0.904
0.532
0.936
0.889
0.963
0.857
0.625
0.945
0.791
0.900
0.516
0.894
0.995
0.979
0.905
0.540
0.946
0.885
0.960
0.863
0.638
0.952
0.862
0.898
0.517
0.890
0.995
0.983
0.909
0.547
0.950
0.882
0.959
0.857
0.629
0.925
0.763
0.928
0.510
0.902
0.996
0.982
0.898
0.520
0.943
0.873
0.950
0.849
0.644
0.938
0.746
0.897
0.515
0.896
0.996
0.980
0.905
0.510
0.947
0.876
0.954
0.856
0.639
0.956
0.677
0.897
0.516
0.894
0.996
0.991
0.896
0.523
0.948
0.876
0.944
0.857
0.629
1.00
0.8
0.7
Accuracy
Accuracy
0.95
0.90
0.6
0.85
0.5
batch
coil2000
HAPT
Image
isolet
MicroV1
MNIST
musk
nomao
semeion
dota
drive
madelon
MicroV2
Dataset
AE configuration
0.25
0.75
1.5_0.5_1.5
0.5
1.5_0.25_1.5
1.5_0.75_1.5
Fig. 4: Accuracy classification results for test data.
The results presented in Table 3 and in Fig. 4 show the Accuracy obtained
for AEkNN with different PPL values. These results indicate that there is no
configuration that works best for all datasets. The configurations with three hidden
layers obtain the best results in 4 out of 14 datasets, whereas the configurations
with one hidden layer win in 10 out of 14 datasets. This trend can also be seen in
the graphs.
Table 4 and Fig. 5 show the F-Score obtained by AEkNN with different PPL
values. The values indicate that the configurations with one single hidden layer get
better results in 11 out of 14 datasets. The configuration with PPL = (0.25) and
with PPL = (0.75) are the ones winning more times (5). The version with PPL =
(0.25) shows disparate results, the best values for some cases and bad results for
other cases, for example with hapt, image or microv2. Although the version with
PPL = (0.75) wins the same number of times, its results are more balanced.
In Table 5 the results for AUC obtained with AEkNN can be seen. Fig. 6 represents those results. For this metric, it can be appreciated that single hidden layer
structures work better, obtaining top results in 12 out of 14 datasets. However, a
configuration that works best for all cases has not been found.
18
F. J. Pulgar et al.
Table 4: F-Score classification results for test data.
Dataset
(0.25)
(0.5)
(0.75)
(1.5, 0.25, 1.5)
(1.5, 0.5, 1.5)
(1.5, 0.75, 1.5)
image
drive
coil2000
dota
nomao
batch
musk
semeion
madelon
hapt
isolet
mnist
microv1
microv2
0.930
0.782
0.963
0.481
0.905
0.995
0.984
0.905
0.549
0.818
0.890
0.963
0.868
0.619
0.945
0.796
0.947
0.487
0.897
0.995
0.988
0.906
0.542
0.829
0.887
0.960
0.872
0.636
0.952
0.863
0.946
0.486
0.893
0.995
0.990
0.910
0.567
0.842
0.883
0.959
0.867
0.625
0.925
0.772
0.962
0.481
0.904
0.995
0.989
0.899
0.531
0.833
0.875
0.950
0.861
0.641
0.938
0.746
0.946
0.488
0.898
0.995
0.988
0.905
0.501
0.839
0.878
0.954
0.867
0.643
0.956
0.683
0.945
0.485
0.897
0.995
0.995
0.896
0.529
0.837
0.878
0.944
0.868
0.628
1.00
0.8
0.7
0.90
F−Score
F−Score
0.95
0.6
0.85
0.5
batch
coil2000
HAPT
Image
isolet
MicroV1
MNIST
musk
nomao
semeion
dota
drive
Dataset
AE configuration
madelon
MicroV2
Dataset
0.25
0.75
1.5_0.5_1.5
0.5
1.5_0.25_1.5
1.5_0.75_1.5
Fig. 5: F-Score classification results for test data.
Table 5: AUC classification results for test data.
Dataset
(0.25)
(0.5)
(0.75)
(1.5, 0.25, 1.5)
(1.5, 0.5, 1.5)
(1.5, 0.75, 1.5)
image
drive
coil2000
dota
nomao
batch
musk
semeion
madelon
hapt
isolet
mnist
microv1
microv2
0.932
0.881
0.526
0.508
0.903
0.997
0.950
0.927
0.532
0.888
0.948
0.974
0.928
0.891
0.943
0.889
0.543
0.514
0.894
0.997
0.958
0.927
0.540
0.898
0.946
0.972
0.934
0.897
0.951
0.922
0.541
0.515
0.890
0.997
0.966
0.928
0.547
0.917
0.942
0.975
0.931
0.887
0.923
0.875
0.526
0.508
0.901
0.997
0.967
0.923
0.520
0.900
0.938
0.965
0.932
0.897
0.936
0.850
0.538
0.514
0.895
0.997
0.962
0.923
0.513
0.916
0.941
0.968
0.927
0.890
0.956
0.826
0.539
0.514
0.893
0.997
0.983
0.922
0.523
0.915
0.940
0.958
0.927
0.885
Summarizing, the results presented before show the metrics obtained for AEkNN
with different PPL values. These results show some variability. Which PPL value is
AEkNN: An AutoEncoder kNN-based classifier ...
19
1.000
0.9
0.975
0.7
AUC
AUC
0.8
0.950
0.925
0.6
0.900
0.5
batch
HAPT
Image
isolet
MicroV1
MicroV2
MNIST
musk
nomao
semeion
coil2000
dota
Dataset
drive
madelon
Dataset
AE configuration
0.25
0.75
1.5_0.5_1.5
0.5
1.5_0.25_1.5
1.5_0.75_1.5
Fig. 6: AUC classification results for test data.
the overall best cannot be determined, since for each dataset there is a setting that
works best. However, some initial trends can be drawn. Single-layer configurations
get more better results than configurations with three hidden layers. Also, the
configurations with PPL = (0.75) and PPL = (0.5) offer close to the best value in
most cases, while the other configurations sometimes are far from the best value.
Once the results have been obtained, it is necessary to determine if there are
statistically significant differences for each one of them in order to select the best
configuration. To do this, the Friedman test [31] will be applied. Average ranks
obtained by applying the Friedman test for Accuracy, F-Score and AUC measures
are shown in Table 6. In addition, Table 7 shows the different p-values obtained
by the Friedman test.
Table 6: Average rankings of the different PPL values by measure
Accuracy
PPL
Ranking
(0.75)
(0.5)
(0.25)
(1.5, 0.75, 1.5)
(1.5, 0.5, 1.5)
(1.5, 0.25, 1.5)
2.679
2.857
3.714
3.821
3.892
4.036
F-Score
PPL
Ranking
(0.5)
(0.75)
(0.25)
(1.5, 0.5, 1.5)
(1.5, 0.25, 1.5)
(1.5, 0.75, 1.5)
2.923
3.000
3.429
3.500
4.071
4.071
AUC
PPL
Ranking
(0.75)
(0.5)
(1.5, 0.25, 1.5)
(0.25)
(1.5, 0.5, 1.5)
(1.5, 0.75, 1.5)
2.357
2.786
3.786
3.857
4.000
4.214
Table 7: Results of Friedman’s test (p-values )
Accuracy
F-Score
AUC
0.236
0.423
0.049
As can be observed in Table 7, for AUC (wich is considerer a stronger performance metric) there are statistically significant differences between the considered
20
F. J. Pulgar et al.
PPL values if we set the p-value threshold to the usual range [0.05, 0.1]. How-
ever, for Accuracy and F-Score there are no statistically significant differences. In
addition, in the rankings obtained, it can be seen that there are two specific configurations that offer better results than the remaining ones. In the three rankings
presented, the results with PPL = (0.75) and PPL = (0.5) appear first, clearly
highlighted with respect to the other values. Therefore, it is considered that these
two configurations are the best ones. Thus, the results of AEkNN with both configurations will be compared against the kNN algorithm.
5.3 AEkNN vs kNN
This second part is focused on determining if the results obtained with the proposed algorithm, AEkNN, improve those obtained with the kNN algorithm. To do
so, a comparison will be made between the results obtained with AEkNN, using
the values of the PPL parameter selected in the previous section, and the results
obtained with kNN algorithm on the same datasets.
First, Table 8 shows the results for each one of the datasets and considered
measures, including running time. The results for both algorithms are presented
jointly, and the best ones are highlighted in bold. Two plots have been generated
for each metric aiming to optimize data visualization, as in the previous phase,
since the range of results was very broad. Fig. 7 represents the results for Accuracy,
Fig. 8 for F-Score, Fig. 9 for AUC, and Fig. 10 for runtime.
Table 8: Classification results of AEkNN (with different PPL) and kNN algorithm
for test data
Dataset
image
drive
coil2000
dota
nomao
batch
musk
semeion
madelon
hapt
isolet
mnist
microv1
microv2
Accuracy
kNN
AEkNN
(0.75) (0.5)
0.937
0.691
0.897
0.507
0.891
0.995
0.956
0.908
0.531
0.951
0.872
0.947
0.800
0.607
0.952
0.862
0.898
0.517
0.890
0.995
0.983
0.909
0.547
0.950
0.882
0.959
0.857
0.629
0.945
0.791
0.900
0.516
0.894
0.995
0.979
0.905
0.540
0.946
0.885
0.960
0.863
0.638
kNN
0.937
0.615
0.945
0.479
0.892
0.995
0.974
0.910
0.549
0.842
0.874
0.946
0.806
0.603
F-Score
AEkNN
(0.75) (0.5)
0.952
0.863
0.946
0.486
0.893
0.995
0.990
0.910
0.567
0.842
0.883
0.959
0.867
0.625
0.945
0.796
0.947
0.487
0.897
0.995
0.988
0.906
0.542
0.829
0.887
0.960
0.872
0.636
kNN
0.934
0.700
0.547
0.416
0.891
0.996
0.934
0.927
0.532
0.903
0.943
0.965
0.890
0.873
AUC
AEkNN
(0.75)
(0.5)
0.951
0.922
0.541
0.515
0.890
0.997
0.966
0.928
0.547
0.917
0.942
0.975
0.931
0.887
Time (seconds)
kNN
AEkNN
(0.75)
(0.5)
0.943
0.074
0.073
0.052
0.889 139.623
36.977 20.479
0.543
3.753
3.262
1.886
0.514 772.219 578.437 370.599
0.894
0.582
0.252
0.200
0.997
7.433
4.829
2.459
0.958
5.317
3.613
2.144
0.927
0.865
0.601
0.381
0.540
3.267
3.150
2.109
0.898
41.673
30.625 16.365
0.946
27.563
24.748 19.538
0.972 1720.547 1213.168 904.223
0.934
1.776
0.977
0.709
0.897
1.542
1.425
0.933
The results shown in Table 8 indicate that AEkNN works better than kNN for
most datasets considering Accuracy. On the one hand, the version of AEkNN with
PPL = (0.75) improves kNN in 11 out of 14 cases, obtaining the best overall results
in 6 of them. On the other hand, the version of AEkNN with PPL = (0.5) obtains
better results than kNN in 11 out of 14 cases, being the best configuration in 6 of
AEkNN: An AutoEncoder kNN-based classifier ...
21
1.00
0.8
0.7
0.90
Accuracy
Accuracy
0.95
0.6
0.85
0.80
0.5
batch
coil2000
HAPT
Image
isolet
MicroV1
MNIST
musk
nomao
semeion
dota
drive
madelon
MicroV2
Dataset
AE configuration
kNN
AEkNN PPL=0.75
AEkNN PPL=0.5
Fig. 7: Accuracy results for test data.
them. In addition, kNN only obtains one best result. Fig. 7 confirms this trend.
It can be observed that the right bars, where AEkNN results are represented, are
higher in most datasets.
1.00
0.8
0.7
0.90
F−Score
F−Score
0.95
0.6
0.85
0.5
0.80
batch
coil2000
HAPT
Image
isolet
MicroV1
MNIST
musk
nomao
semeion
dota
drive
madelon
MicroV2
Dataset
AE configuration
kNN
AEkNN PPL=0.75
AEkNN PPL=0.5
Fig. 8: F-Score results for test data.
Analyzing the data corresponding to the metric F-Score, presented in Table 8,
it can be observed that AEkNN produces an overall improvement over kNN. The
AEkNN version with PPL = (0.75) improves kNN in 11 out of 14 cases, obtaining
the best overall results in 5 of them. The version of AEkNN with PPL = (0.5)
obtains better results than kNN in 10 out of 14 cases, being the best configuration
in 7 of them. kNN does not obtain any best result. How the results of the versions
corresponding to AEkNN produce higher values than those corresponding to kNN
can be seen in Fig. 8.
The data related to AUC, presented in Table 8, also show that AEkNN works
better than kNN. In this case, the two versions of AEkNN improve kNN in 11 out
of 14 cases each one, obtaining the best results in 13 out of 14 cases. Fig. 9 shows
that the trend is increasing towards the versions of the new algorithm. kNN only
gets a best result, specifically with the coil2000 dataset, maybe due to the low
22
F. J. Pulgar et al.
1.00
0.9
0.8
0.95
AUC
AUC
0.7
0.6
0.90
0.5
0.4
batch
HAPT
Image
isolet
MicroV1 MicroV2
MNIST
musk
nomao
semeion
coil2000
dota
drive
madelon
Dataset
AE configuration
kNN
AEkNN PPL=0.75
AEkNN PPL=0.5
Fig. 9: AUC results for test data.
number of features in this dataset. Nonetheless, AEkNN gets better than kNN in
the rest of metrics for coil2000 dataset.
1500
1000
4
Time
Time
6
500
2
0
0
batch
coil2000
Image
madelon
MicroV1
MicroV2
musk
nomao
semeion
dota
drive
HAPT
isolet
MNIST
Dataset
AE configuration
kNN
AEkNN PPL=0.75
AEkNN PPL=0.5
Fig. 10: Time results for test data.
The running times for both algorithms are presented in Table 8 and Fig. 10. As
can be seen, the configuration that takes less time to classify is the one corresponding to AEkNN with PPL = (0.5), obtaining the lowest value for all datasets. This
is due to the higher compression of the data achieved by this configuration. In the
same way, AEkNN with PPL = (0.75) obtains better results than the algorithm
kNN in all cases.
Summarizing, it can be observed that the results obtained through AEkNN
improve those obtained with the original kNN algorithm for most of the datasets.
AEkNN, despite the transformation of the input space to reduce dimensionality,
the quality of the results in terms of classification performance are better than
those of kNN in most cases. In addition, in terms of classification time, it can be
noted how AEkNN with higher compression of information significantly reduces
the time spent on classification, without having a negative impact on the other
measures.
AEkNN: An AutoEncoder kNN-based classifier ...
23
To determine if there are statistically significant differences between the obtained results, the proper statistical test has been conducted. For this purpose,
the Wilcoxon test will be performed, comparing each version of AEkNN against
the results of the classical kNN algorithm. In Table 9, the results obtained for
Wilcoxon tests are shown.
Table 9: Result of Wilcoxon’s test (p-values ) comparing kNN vs AEkNN
AEkNN with PPL=(0.75)
AEkNN with PPL=(0.5)
Accuracy
0.0017
0.0023
F-Score
0.0003
0.0245
AUC
0.0085
0.0107
Time
0.0002
0.0001
As can be seen the p-values are rather low, so that statistical significant differences between the two AEkNN versions and the original kNN algorithm in
all considered measures exist can be concluded, considering the p-value threshold
within the usual [0.05, 0.1] range. On the one hand, taking into account Accuracy,
F-Score and AUC, the configuration with best results is that performing a 75% of
feature reduction. Therefore, it is the optimal solution from the point of view of
predictive performance. The reason for this might be that there is less compression of the data, therefore, there is less loss of information compared to the other
considered configuration (50%). On the other hand, considering running time, the
configuration with best results is that performing a 50% of feature reduction. It is
not surprising that having less features allows to compute distances in less time.
5.4 AEkNN vs PCA / LDA
The objective of this third part is to assess the competitiveness of AEkNN against
traditional dimensionality reduction algorithms. In particular, the algorithms used
will be PCA [63, 43] and LDA [32, 92], since they are traditional algorithms that
offer good results in this task [61]. To do so, a comparison will be made between
the results obtained with AEkNN, using the values of the PPL parameter selected
in section 5.3, and the results obtained with PCA and LDA algorithm on the
same datasets. It is important to note that the number of features selected with
these methods will be the same as with the AEkNN algorithm, so there are two
executions for each algorithm.
First, Table 10 shows the results for each one of the datasets and considered
measures. The results for the three algorithms are presented jointly, and the best
ones are highlighted in bold. One plot have been generated for each metric aiming
to optimize data visualization. In this case, the graphs represent the best value of
the two configurations for each algorithm in order to better visualize the differences
between the three methods. Fig. 11 represents the results for Accuracy, Fig. 12 for
F-Score and Fig. 13 for AUC.
The results shown in Table 10 indicate that AEkNN works better than PCA
and LDA for most datasets considering Accuracy. On the one hand, the version
of AEkNN with PPL = (0.75) improves PCA in in 12 out of 14 cases and LDA
in 9 out of 14 cases, obtaining the best overall results in 7 of them. On the other
hand, the version of AEkNN with PPL = (0.5) obtains better results than PCA
24
F. J. Pulgar et al.
1.00
Accuracy
0.75
0.50
0.25
batch
coil2000
dota
drive
HAPT
Image
isolet
madelon MicroV1 MicroV2 MNIST2
musk
nomao
semeion
Dataset
Algorithm
AEkNN
PCA
LDA
Fig. 11: Accuracy results for test data.
in 13 out of 14 cases and LDA in 8 out of 14 cases, being the best configuration
in 4 of them. In addition, LDA only obtains the best result in 4 cases. Fig. 11
confirms this trend. It can be observed that the bars, where AEkNN results are
represented, are higher in most datasets.
Analyzing the data corresponding to the metric F-Score, presented in Table
10, it can be observed that AEkNN produces an overall improvement over PCA
and LDA. The AEkNN version with PPL = (0.75) improves PCA in 12 out of 14
cases and LDA in 10 out of 14 cases, obtaining the best overall results in 6 of them.
The version of AEkNN with PPL = (0.5) obtains better results than PCA in all
cases and LDA in 8 out of 14 cases, being the best configuration in 4 of them.
PCA does not obtain any best result and LDA obtains the best result in 4 cases.
How the results of the versions corresponding to AEkNN show values higher than
those corresponding to kNN can be seen in Fig. 12.
1.00
F−Score
0.75
0.50
0.25
batch
coil2000
dota
drive
HAPT
Image
isolet
madelon MicroV1 MicroV2 MNIST2
musk
nomao
semeion
Dataset
Algorithm
AEkNN
PCA
LDA
Fig. 12: F-Score results for test data.
The data related to AUC, presented in Table 10, also show that AEkNN works
better than PCA and LDA. The AEkNN version with PPL = (0.75) improves PCA
in 13 out of 14 cases and LDA in 10 out of 14 cases, obtaining the best overall
results in 7 of them. The version of AEkNN with PPL = (0.5) obtains better
results than PCA in 13 out of 14 cases and LDA in 8 out of 14 cases, being the
best configuration in 4 of them. LDA only gets the best result in 4 cases. Fig. 13
confirms this trend.
0.862
0.092
0.899
0.517
0.869
0.206
0.931
0.706
0.511
0.443
0.590
0.640
0.135
0.082
0.676
0.091
0.895
0.519
0.818
0.189
0.982
0.907
0.501
0.959
0.886
0.949
0.818
0.649
0.866
0.091
0.895
0.520
0.889
0.152
0.982
0.890
0.512
0.960
0.902
0.943
0.861
0.690
0.952
0.863
0.946
0.486
0.893
0.995
0.990
0.910
0.567
0.842
0.883
0.959
0.867
0.625
0.945
0.796
0.947
0.487
0.897
0.995
0.988
0.906
0.542
0.829
0.887
0.960
0.872
0.636
0.580
0.023
0.861
0.486
0.821
0.176
0.962
0.624
0.519
0.180
0.504
0.704
0.116
0.005
0.867
0.030
0.946
0.486
0.871
0.190
0.945
0.718
0.503
0.478
0.638
0.704
0.116
0.037
0.672 0.866 0.951 0.943
0.015 0.015 0.922 0.889
0.944 0.944 0.541 0.543
0.489 0.493 0.515 0.514
0.851 0.891 0.890 0.894
0.053 0.044 0.997 0.997
0.989 0.989 0.966 0.958
0.906 0.893 0.928 0.927
0.503 0.491 0.547 0.540
0.841 0.841 0.917 0.898
0.888 0.903 0.942 0.946
0.949 0.943 0.975 0.972
0.823 0.871 0.931 0.934
0.644 0.695 0.887 0.897
0.734
0.500
0.523
0.513
0.797
0.502
0.912
0.732
0.523
0.553
0.660
0.828
0.532
0.500
0.893
0.491
0.532
0.515
0.869
0.510
0.930
0.792
0.510
0.690
0.748
0.798
0.532
0.537
0.780
0.500
0.545
0.517
0.817
0.500
0.965
0.926
0.505
0.903
0.951
0.966
0.898
0.891
0.862
0.500
0.549
0.519
0.889
0.500
0.965
0.913
0.505
0.899
0.957
0.954
0.929
0.916
image
drive
coil2000
dota
nomao
batch
musk
semeion
madelon
hapt
isolet
mnist
microv1
microv2
0.563
0.091
0.825
0.514
0.798
0.195
0.942
0.598
0.506
0.179
0.424
0.740
0.135
0.058
0.952
0.862
0.898
0.517
0.890
0.995
0.983
0.909
0.547
0.950
0.882
0.959
0.857
0.629
Dataset
0.945
0.791
0.900
0.516
0.894
0.995
0.979
0.905
0.540
0.946
0.885
0.960
0.863
0.638
Accuracy
F-Score
AUC
AEkNN
PCA
LDA
AEkNN
PCA
LDA
AEkNN
PCA
LDA
(0.75) (0.5) (0.75) (0.5) (0.75) (0.5) (0.75) (0.5) (0.75) (0.5) (0.75) (0.5) (0.75) (0.5) (0.75) (0.5) (0.75) (0.5)
Table 10: Accuracy, F-Score and AUC classification results of AEkNN (with different PPL), LDA and PCA for test data
AEkNN: An AutoEncoder kNN-based classifier ...
25
26
F. J. Pulgar et al.
1.0
AUC
0.9
0.8
0.7
0.6
0.5
batch
coil2000
dota
drive
HAPT
Image
isolet
madelon MicroV1 MicroV2 MNIST2
musk
nomao
semeion
Dataset
Algorithm
AEkNN
PCA
LDA
Fig. 13: AUC results for test data.
Summarizing, it can be observed that the results obtained through AEkNN improve those obtained with the PCA and LDA algorithms for most of the datasets.
The quality of the results with AEkNN in terms of classification performance are
better than those of PCA and LDA in most cases. This means that the high-level
features obtained by the AEkNN algorithm provide more relevant information than
those obtained by the PCA and LDA algorithms. Therefore, the results obtained
by classifying with the AEkNN algorithm improve those obtained with PCA and
LDA.
Previously, the data obtained in the experimentation have been presented and
a comparison between them is made. However, it is necessary to verify if there are
significant differences between the data corresponding to the different algorithms.
To do this, the Friedman test [31] will be applied. Average ranks obtained by
applying the Friedman test for Accuracy, F-Score and AUC measures are shown
in Table 11. In addition, Table 12 shows the different p-values obtained by the
Friedman test.
Table 11: Average rankings of the different PPL values by measure
Accuracy
Algorithm
Ranking
AEkNN PPL=0.75
AEkNN PPL=0.5
LDA PPL=0.5
LDA PPL=0.75
PCA PPL=0.5
PCA PPL=0.75
2.286
2.357
3.000
3.571
4.321
5.464
F-Score
Algorithm
Ranking
AEkNN PPL=0.75
AEkNN PPL=0.5
LDA PPL=0.75
LDA PPL=0.5
PCA PPL=0.5
PCA PPL=0.75
2.143
2.214
3.429
3.429
4.429
5.357
AUC
Algorithm
AEkNN PPL=0.75
AEkNN PPL=0.5
LDA PPL=0.5
LDA PPL=0.75
PCA PPL=0.5
PCA PPL=0.75
Table 12: Results of Friedman’s test (p-values )
Accuracy
F-Score
AUC
1.266e-05
7.832e-06
7.913e-07
Ranking
1.929
2.500
2.929
3.500
4.679
5.464
AEkNN: An AutoEncoder kNN-based classifier ...
27
As can be observed in Table 12, for Accuracy, F-Score and AUC there are
statistically significant differences between the different PPL values if we set the
p-value threshold to the usual range [0.05, 0.1]. It can be seen that AEkNN with
PPL = (0.75) offer better results than the remaining ones. In the three rankings
presented, the AEkNN configurations with PPL = (0.75) and PPL = (0.5) appear
first, clearly highlighted with respect to the other values. Therefore, it is considered that AEkNN obtains better predictive performance, since the reduction of
dimensionality generates more significant features.
5.5 General Guidelines on the use of AEkNN
AEkNN could be considered as a robust algorithm on the basis of the previous
analysis. The experimental work demonstrates that it has good performance with
the two PPL considered values. From the conducted experimentation some guidelines can also be extracted:
– When working with very high-dimensional datasets, it is recommended to use
AEkNN with the PPL = (0.5) configuration. In this study, this configuration
has obtained the best results for datasets having more than 600 features. The
reason is that the input data has a larger number of features and allows a
greater reduction without losing relevant information. Therefore, AEkNN can
compress more in these cases.
– When using binary datasets with a lower dimensionality, the AEkNN algorithm
with the PPL = (0.5) configuration continues to be the best choice. In our
experience, this configuration has shown to work better for binary datasets
with a number of features around 100. In these cases, the compression may be
higher since it is easier to discriminate by class.
– For all other datasets, the choice of configuration for AEkNN depends on the
indicator to be enhanced. On the one hand, if the goal is to achieve the best
possible predictive performance, the configuration with PPL = (0.75) must
be chosen. In these cases, AEkNN needs to use more original information.
On the other hand, when the interest is to optimize the running time, while
maintaining improvements in predictive performance with respect to kNN, the
configuration with PPL = (0.5) is the best selection. The reason is in the higher
compression of the data. AEkNN needs less time to classify lower-dimensional
data.
Summarizing, the configuration of AEkNN must be adapted to the data traits
to obtain optimal results. For this, a series of tips have been established.
6 Concluding remarks
In this paper, a new classification algorithm called AEkNN has been proposed.
This algorithm is based on kNN but aims to mitigate the problem that arises while
working with high-dimensional data. To do so, AEkNN internally incorporates a
model-building phase aimed to perform a reduction of the feature space, using
AEs for this purpose. The main reason that has led to the design of AEkNN are
the good results that have been obtained by AEs when they are used to generate
28
F. J. Pulgar et al.
higher-level features. AEkNN relies on an AE to extract a reduced representation
of a higher level that replaces the original data.
In order to determine if the proposed algorithm offers better results than the
kNN algorithm, a experimentation has been carried out. Firstly, the analysis of
these results have allowed to determine which AE structure works better. Furthermore, in the second part of the conducted experimentation, the results of the
best configurations have been compared with the results produced by kNN. As
has been stated, the results of AEkNN improve those obtained by kNN for all the
metrics. In addition, AEkNN offers a considerable improvement with respect to
the time invested in the classification.
In addition, a comparison has been made with other traditional methods applied to this problem, in order to verify that the AEkNN algorithm improves the
results when carrying out the dimensionality reduction task. For this, AEkNN
has been compared with LDA and PCA. The results show that the proposed
AEkNN algorithm improves the performance in classification for most of the
dataset used. This occurs because the features generated with the proposed algorithm are more significant and provide more relevant information to the classification using distance-based algorithms.
In conclusion, AEkNN is able to reduce the adverse effects of high-dimensional
data while performing instance-based classification, improving both running time
and classification performance. These results show that the use of AEs can be
helpful to solve this kind of obstacle, opening up new possibilities of future work in
which they are applied to help solve similar problems presented by other traditional
models.
Acknowledgment
The work of F. Pulgar was supported by the Spanish Ministry of Education under
the FPU National Program (Ref. FPU16/00324). This work was partially supported by the Spanish Ministry of Science and Technology under project TIN201568454-R.
References
1. Aha DW, Kibler D, Albert MK (1991) Instance-Based Learning Algorithms.
Machine Learning 6(1):37–66, DOI 10.1023/A:1022689900470
2. Aiello S, Kraljevic T, Maj P, with contributions from the H2Oai team (2016)
h2o: R Interface for H2O. URL https://CRAN.R-project.org/package=h2o, r
package version 3.8.1.3
3. Alpaydin E (2014) Introduction to machine learning. MIT press
4. Altman NS (1992) An Introduction to Kernel and Nearest-Neighbor Nonparametric Regression. The American Statistician 46(3):175–185, DOI 10.1080/
00031305.1992.10475879
5. Atkeson CG, Moorey AW, Schaalz S, Moore AW, Schaal S (1997) Locally Weighted Learning. Artificial Intelligence 11:11–73, DOI 10.1023/A:
1006559212014
AEkNN: An AutoEncoder kNN-based classifier ...
29
6. Bache K, Lichman M (2013) UCI Machine Learning Repository. URL http:
//www.ics.uci.edu/$\sim$mlearn/MLRepository.html
7. Belkin M, Niyogi P (2003) Laplacian eigenmaps for dimensionality reduction
and data representation. Neural Computation 15(6):1373–1396, DOI 10.1162/
089976603321780317
8. Bellman R (1957) Dynamic Programming. Princeton University Press
9. Bellman R (1961) Adaptive control processes: A guided tour. Princeton University Press
10. Bengio Y (2009) Learning Deep Architectures for AI. Foundations and Trends
in Machine Learning 2(1):1–127, DOI 10.1561/2200000006
11. Bengio Y (2013) Deep Learning of Representations: Looking Forward. In: International Conference on Statistical Language and Speech Processing, pp
1–37, DOI 10.1007/978-3-642-39593-2\ 1
12. Bengio Y, Courville A, Vincent P (2013) Representation Learning: A Review and New Perspectives. Pattern Analysis and Machine Intelligence, IEEE
Transactions on 35(8):1798–1828, DOI 10.1109/TPAMI.2013.50
13. Beyer K, Goldstein J, Ramakrishnan R, Shaft U (1999) When Is ”Nearest
Neighbor” Meaningful? In: International conference on database theory, pp
217–235, DOI 10.1007/3-540-49257-7\ 15
14. Bhatia V, Rani R (2018) Dfuzzy: a deep learning-based fuzzy clustering
model for large graphs. Knowledge and Information Systems DOI 10.1007/
s10115-018-1156-3, URL https://doi.org/10.1007/s10115-018-1156-3
15. Burges CJC (2005) Geometric Methods for Feature Extraction and Dimensional Reduction, Springer US, pp 59–91. DOI 10.1007/0-387-25465-X\ 4
16. Casillas J, Cordón O, Del Jesus MJ, Herrera F (2001) Genetic feature selection in a fuzzy rule-based classification system learning process for highdimensional problems. Information Sciences 136(1):135–157, DOI 10.1016/
S0020-0255(01)00147-5
17. Charte D, Charte F, Garca S, del Jesus MJ, Herrera F (2018) A practical tutorial on autoencoders for nonlinear feature fusion: Taxonomy, models, software
and guidelines. Information Fusion 44:78–96, DOI 10.1016/j.inffus.2017.12.007
18. Chung J, Gülçehre C, Cho K, Bengio Y (2015) Gated feedback recurrent
neural networks. In: Proceedings of the International Conference on Machine
Learning, pp 2067–2075
19. Ciresan D, Giusti A, Gambardella L, Schmidhuber J (2012) Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images. In: Advances in neural information processing systems, pp 2843–2851
20. Ciresan D, Meier U, Schmidhuber J (2012) Multi-column deep neural networks
for image classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 3642–3649, DOI 10.1109/CVPR.
2012.6248110
21. Coifman RR, Lafon S, Lee AB, Maggioni M, Nadler B, Warner F, Zucker
SW (2005) Geometric diffusions as a tool for harmonic analysis and structure
definition of data: Diffusion maps. Proceedings of the National Academy of
Sciences 102(21):7426–7431, DOI 10.1073/pnas.0500334102
22. Cole R, Fanty M (1990) Spoken letter recognition. In: Proceedings of the
workshop on Speech and Natural Language, pp 385–390, DOI 10.3115/116580.
116725
30
F. J. Pulgar et al.
23. Cover T, Hart P (1967) Nearest neighbor pattern classification. IEEE Transactions on Information Theory 13(1):21–27, DOI 10.1109/TIT.1967.1053964
24. Dasarathy BV (1991) Nearest neighbor norms: NN pattern classification techniques. IEEE Computer Society Press
25. Deng L (2014) Deep Learning: Methods and Applications. Foundations and
Trends in Signal Processing 7(3-4):197–387, DOI 10.1561/2000000039
26. Deng L, Yu D (2011) Deep convex net: A scalable architecture for speech
pattern classification. In: Proceedings of the Annual Conference of the International Speech Communication Association, pp 2285–2288
27. Deng Y, Ren Z, Kong Y, Bao F, Dai Q (2017) A hierarchical fused fuzzy deep
neural network for data classification. IEEE Transactions on Fuzzy Systems
25(4):1006–1012, DOI 10.1109/TFUZZ.2016.2574915
28. Deng Z, Zhu X, Cheng D, Zong M, Zhang S (2016) Efficient knn classification
algorithm for big data. Neurocomputing 195:143–148, DOI 10.1016/j.neucom.
2015.08.112
29. Duda RO, Hart PE, Stork DG (1973) Pattern Classification. Wiley New York
30. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S (2017)
Dermatologist-level classification of skin cancer with deep neural networks.
Nature 542(7639):115–118, DOI 10.1038/nature21056
31. Friedman M (1937) The use of ranks to avoid the assumption of normality
implicit in the analysis of variance. Journal of the American Statistical Association 32(200):675–701, DOI 10.1080/01621459.1937.10503522
32. Fukunaga K (2013) Introduction to statistical pattern recognition. Academic
Press
33. Goldberg DE, Holland JH (1988) Genetic Algorithms and Machine Learning.
Machine Learning 3(2):95–99, DOI 10.1023/A:1022602019183
34. Goodfellow I, Bengio Y, Courville A (2016) Deep Learning. The MIT Press
35. Guyon I, Gunn S, Ben-Hur A, Dror G (2004) Result Analysis of the NIPS 2003
Feature Selection Challenge. In: Proceedings of Neural Information Processing
Systems, vol 4, pp 545–552
36. Harsanyi JC, Chang CI (1994) Hyperspectral image classification and dimensionality reduction: An orthogonal subspace projection approach. IEEE Transactions on Geoscience and Remote Sensing 32(4):779–785, DOI 10.1109/36.
298007
37. Hinneburg A, Aggarwal CC, Keim DA (2000) What Is the Nearest Neighbor
in High Dimensional Spaces? In: Proceedings of the International Conference
on Very Large Databases, pp 506–515
38. Hinton G (2010) A practical guide to training restricted boltzmann machines.
Momentum 9(1):926, DOI 10.1007/978-3-642-35289-8\ 32
39. Hinton G, Salakhutdinov R (2012) A better way to pretrain deep Boltzmann
machines. Advances in Neural Information Processing Systems (3):2447–2455
40. Hinton G, Deng L, Yu D, Dahl GE, Mohamed AR, Jaitly N, Senior A, Vanhoucke V, Nguyen P, Sainath TN (2012) Deep neural networks for acoustic
modeling in speech recognition: The shared views of four research groups.
IEEE Signal Processing Magazine 29(6):82–97
41. Hinton GE, Salakhutdinov RR (2006) Reducing the Dimensionality of Data
with Neural Networks. Science 313(5786):504–507, DOI 10.1126/science.
1127647
AEkNN: An AutoEncoder kNN-based classifier ...
31
42. Hochreiter S, Schmidhuber J (1997) Long Short-Term Memory. Neural Computation 9(8):1735–1780, DOI 10.1162/neco.1997.9.8.1735
43. Hotelling H (1933) Analysis of a complex of statistical variables into principal
components. Journal of Educational Psychology 24(6):417–441, DOI 10.1037/
h0071325
44. Hughes GF (1968) On the Mean Accuracy of Statistical Pattern Recognizers.
IEEE Transactions on Information Theory 14(1):55–63, DOI 10.1109/TIT.
1968.1054102
45. Keyvanrad MA, Homayounpour MM (2017) Effective sparsity control in deep
belief networks using normal regularization term. Knowledge and Information
Systems 53(2):533–550, DOI 10.1007/s10115-017-1049-x, URL https://doi.
org/10.1007/s10115-017-1049-x
46. Kohavi R, Provost F (1998) Glossary of Terms. Machine Learning 30(2-3):271–
274, DOI 10.1023/A:1017181826899
47. Kotsiantis SB (2007) Supervised machine learning: A review of classification
techniques. Informatica 31:249–268
48. Kouiroukidis N, Evangelidis G (2011) The Effects of Dimensionality Curse
in High Dimensional kNN Search. In: Proceedings of the 15th Panhellenic
Conference on Informatics, IEEE, pp 41–45, DOI 10.1109/PCI.2011.45
49. Krizhevsky A, Sutskever I, Geoffrey E H (2012) ImageNet Classification with
Deep Convolutional Neural Networks. In: Advances in neural information processing systems, pp 1097–1105
50. LeCun Y, Bengio Y (1995) Convolutional networks for images, speech, and
time series. The handbook of brain theory and neural networks 3361:255–258,
DOI 10.1109/IJCNN.2004.1381049
51. Lecun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11):2278–2324,
DOI 10.1109/5.726791
52. LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521(7553):436–
444, DOI 10.1038/nature14539
53. Lee H, Grosse R, Ranganath R, Ng AY (2009) Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations. In:
Proceedings of the 26th Annual International Conference on Machine Learning, ACM Press, vol 2008, pp 609–616, DOI 10.1145/1553374.1553453
54. Lee JA, Verleysen M (2007) Nonlinear Dimensionality Reduction. Springer
Science & Business Media
55. Lin Y, Zhang T, Zhu S, Yu K (2010) Deep coding network. In: Advances in
Neural Information Processing Systems, pp 1405–1413
56. Liou CY, Cheng WC, Liou JW, Liou DR (2014) Autoencoder for words. Neurocomputing 139:84–96, DOI 10.1016/j.neucom.2013.09.055
57. Liu H, Motoda H (1998) Feature extraction, construction and selection: A
data mining perspective, vol 453. Springer Science & Business Media, DOI
10.1007/978-1-4615-5725-8
58. Liu H, Motoda H (2007) Computational methods of feature selection. CRC
Press
59. Maillo J, Luengo J, Garcı́a S, Herrera F, Triguero I (2017) Exact fuzzy knearest neighbor classification for big datasets. In: IEEE International Conference on Fuzzy Systems, IEEE, pp 1–6, DOI 10.1109/FUZZ-IEEE.2017.
8015686
32
F. J. Pulgar et al.
60. Maillo J, Ramı́rez S, Triguero I, Herrera F (2017) knn-is: An iterative sparkbased design of the k-nearest neighbors classifier for big data. KnowledgeBased Systems 117:3–15, DOI 10.1016/j.knosys.2016.06.012
61. Martı́nez AM, Kak AC (2001) PCA versus LDA. IEEE Transactions on Pattern Analysis and Machine Intelligence 23(2):228–233, DOI 10.1109/34.908974
62. Min R, Stanley DA, Yuan Z, Bonner A, Zhang Z (2009) A Deep Non-linear
Feature Mapping for Large-Margin kNN Classification. In: Proceedings of the
International Conference on Data Mining, IEEE, pp 357–366, DOI 10.1109/
ICDM.2009.27
63. Pearson K (1901) LIII. On lines and planes of closest fit to systems of points
in space. The London, Edinburgh, and Dublin Philosophical Magazine and
Journal of Science 2(11):559–572, DOI 10.1080/14786440109462720
64. R Core Team (2016) R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, URL https:
//www.R-project.org/
65. Radovanović M, Nanopoulos A, Ivanović M (2010) Hubs in Space : Popular
Nearest Neighbors in High-Dimensional Data. Journal of Machine Learning
Research 11:2487–2531
66. Reyes-Ortiz JL, Oneto L, Samà A, Parra X, Anguita D (2016) TransitionAware Human Activity Recognition Using Smartphones. Neurocomputing
171:754–767, DOI 10.1016/j.neucom.2015.07.085
67. Rifai S, Muller X (2011) Contractive Auto-Encoders: Explicit Invariance During Feature Extraction. In: Proceedings of the 28th International Conference
on Machine Learning, vol 85, pp 833–840
68. Roweis ST, Saul LK (2000) Nonlinear dimensionality reduction by locally linear embedding. Science 290(5500):2323–2326, DOI 10.1126/science.290.5500.
2323
69. Sak H, Senior A, Beaufays F (2014) Long Short-Term Memory Recurrent Neural Network Architectures for Large Scale Acoustic Modeling. In: Proceedings
of the Annual Conference of the International Speech Communication Association, September, pp 338–342
70. Salakhutdinov R, Mnih A, Hinton G (2007) Restricted boltzmann machines
for collaborative filtering. In: Proceedings of the 24th International Conference
on Machine learning, ACM, pp 791–798, DOI 10.1145/1273496.1273596
71. Saul LK, Weinberger KQ, Ham JH, Sha F, Lee DD (2006) Spectral Methods
for Dimensionality Reduction. The MIT Press, pp 292–308, DOI 10.7551/
mitpress/9780262033589.003.0016
72. Sheskin DJ (2004) Handbook of parametric and nonparametric statistical procedures. Technometrics 46:1193, DOI 10.1198/tech.2004.s209
73. Snoek J, Adams RP, Larochelle H (2012) Nonparametric guidance of autoencoder representations using label information. Journal of Machine Learning
Research 13(Sep):2567–2588
74. Sordoni A, Bengio Y, Vahabi H, Lioma C, Grue Simonsen J, Nie JY (2015)
A hierarchical recurrent encoder-decoder for generative context-aware query
suggestion. In: Proceedings of the 24th ACM International on Conference on
Information and Knowledge Management, ACM, pp 553–562
75. Spearman C (1904) ”General Intelligence,” Objectively Determined and
Measured. The American Journal of Psychology (2):201–292, DOI 10.2307/
1412107
AEkNN: An AutoEncoder kNN-based classifier ...
33
76. Tang P, Zhang J, Wang X, Feng B, Roli F, Liu W (2017) Learning extremely
shared middle-level image representation for scene classification. Knowledge
and Information Systems 52(2):509–530, DOI 10.1007/s10115-016-1015-z,
URL https://doi.org/10.1007/s10115-016-1015-z
77. Tenenbaum JB (2000) A Global Geometric Framework for Nonlinear Dimensionality Reduction. Science 290(5500):2319–2323, DOI 10.1126/science.290.
5500.2319
78. Torgerson W (1952) Multidimensional scaling: I. Theory and method. Psychometrika 17(4):401–419, DOI 10.1007/BF02288916
79. Van Der Maaten LJP, Postma EO, Van Den Herik HJ (2009) Dimensionality
Reduction: A Comparative Review. Journal of Machine Learning Research
10:1–41, DOI 10.1.1.112.5472
80. Van Der Putten P, Van Someren M (2004) A bias-variance analysis of a real
world learning problem: The coil challenge 2000. Machine learning 57(1):177–
195
81. Venna J (2007) Dimensionality reduction for visual exploration of similarity
structures. Helsinki University of Technology
82. Vergara A, Vembu S, Ayhan T, Ryan MA, Homer ML, Huerta R (2012) Chemical gas sensor drift compensation using classifier ensembles. Sensors and Actuators B: Chemical 166:320–329, DOI 10.1016/j.snb.2012.01.074
83. Vincent P, Larochelle H, Bengio Y, Manzagol PA (2008) Extracting and composing robust features with denoising autoencoders. In: Proceedings of the
25th International Conference on Machine Learning, ACM, pp 1096–1103,
DOI 10.1145/1390156.1390294
84. Wang W, Huang Y, Wang Y, Wang L (2014) Generalized autoencoder: A
neural network framework for dimensionality reduction. In: Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition Workshops,
pp 496–503, DOI 10.1109/CVPRW.2014.79
85. Wang X (2011) A fast exact k-nearest neighbors algorithm for high dimensional
search using k-means clustering and triangle inequality. In: Proceedings of
the International Joint Conference on Neural Networks, IEEE, pp 1293–1299,
DOI 10.1016/j.patcog.2010.01.003
86. Weinberger KQ, Saul LK (2006) An introduction to nonlinear dimensionality reduction by maximum variance unfolding. In: Proceedings of the 21st
National Conference on Artificial Intelligence, vol 2, pp 1683–1686
87. Wiley JF (2016) R Deep Learning Essentials. Packt Publishing Ltd.
88. Xia Y, Peng Y, Zhang X, Bae H (2017) Demst-knn: A novel classification
framework to solve imbalanced multi-class problem. In: Computer Science Online Conference, Springer, pp 291–301, DOI 10.1007/978-3-319-57261-1\ 29
89. Yang M, Zhang L, Yang J, Zhang D (2011) Robust sparse coding for face
recognition. In: IEEE Conference on Computer Vision and Pattern Recognition, IEEE, pp 625–632, DOI 10.1109/CVPR.2011.5995393
90. Ye Y, Chen L, Hou S, Hardy W, Li X (2018) Deepam: a heterogeneous
deep learning framework for intelligent malware detection. Knowledge and
Information Systems 54(2):265–285, DOI 10.1007/s10115-017-1058-9, URL
https://doi.org/10.1007/s10115-017-1058-9
91. Yu C, Ooi BC, Tan KL, Jagadish HV (2001) Indexing the Distance: An Efficient Method to KNN Processing. In: International Conference on Very Large
Databases, pp 421–430
34
F. J. Pulgar et al.
92. Yu H, Yang J (2001) A direct lda algorithm for high-dimensional data-with
application to face recognition. Pattern Recognition 34(10):2067–2070, DOI
10.1016/S0031-3203(00)00162-X
93. Zhang J, Tian G, Mu Y, Fan W (2014) Supervised deep learning with auxiliary
networks. In: Proceedings of the 20th ACM SIGKDD international conference
on Knowledge discovery and data mining, ACM, pp 353–361
94. Zhou G, Zhu Z, He T, Hu XT (2016) Cross-lingual sentiment classification with stacked autoencoders. Knowledge and Information Systems
47(1):27–44, DOI 10.1007/s10115-015-0849-0, URL https://doi.org/10.
1007/s10115-015-0849-0
AEkNN: An AutoEncoder kNN-based classifier ...
35
About the author – FRANCISCO JAVIER PULGAR RUBIO received his
B.Sc. degree in Computer Science from the University of Jaén in 2013. He is a PhD
student in the Computer Science Department at the University of Jaén (Spain).
His research interests include areas such as classification, imbalance problems, high
dimensionality problem, neural network design, subgroup discovery, big data and
parallelization techniques.
About the author – FRANCISCO CHARTE received his B.Eng. degree in
Computer Science from the University of Jaén in 2010 and his M.Sc. and Ph.D. in
Computer Science from the University of Granada in 2011 and 2015, respectively.
He is currently a researcher at the University of Jaén (Spain). His main research
interests include machine learning with applications to multi-label classification,
high dimensionality and imbalance problems, and deep learning related techniques.
About the author – ANTONIO J. RIVERA received his B.Sc. degree and
his Ph.D. in Computer Science from the University of Granada in 1995 and 2003,
respectively. He is a lecturer of Computer Architecture and Computer Technology
with the Computer Science Department at the University of Jaén (Spain). His
research interests include areas such as multilabel classification, imbalance problems, evolutionary computation, neural network design, time series prediction and
regression tasks.
About the author – MARÍA J. DEL JESUS received the M.Sc. and Ph.D.
degrees in Computer Science from the University of Granada, Granada, Spain, in
1994 and 1999, respectively. She is an Associate Professor with the Department
of Computer Science, University of Jaén, Spain. Her current research interests include fuzzy rule-based systems, genetic fuzzy systems, subgroup discovery, data
preparation, feature selection, evolutionary radial basis neural networks, knowledge extraction based on evolutionary algorithms, and data mining.
| 9cs.NE
|
arXiv:1701.07535v2 [math.ST] 14 Dec 2017
Stratified Splitting for Efficient Monte Carlo
Integration
Radislav Vaisman, Robert Salomone, and Dirk P. Kroese
School of Mathematics and Physics, The University of Queensland
Brisbane, Australia
[email protected]
[email protected]
[email protected]
Abstract
The efficient evaluation of high-dimensional integrals is of importance in
both theoretical and practical fields of science, such as data science, statistical
physics, and machine learning. However, due to the curse of dimensionality,
deterministic numerical methods are inefficient in high-dimensional settings.
Consequentially, for many practical problems one must resort to Monte Carlo
estimation. In this paper, we introduce a novel Sequential Monte Carlo technique called Stratified Splitting. The method provides unbiased estimates and
can handle various integrand types including indicator functions, which are
used in rare-event probability estimation problems. Moreover, we demonstrate
that a variant of the algorithm can achieve polynomial complexity. The results
of our numerical experiments suggest that the Stratified Splitting method is
capable of delivering accurate results for a variety of integration problems.
1
Introduction
We consider the evaluation of expectations and integrals of the form
Z
X
Ef [ϕ(X)] =
ϕ(x)f (x) or Ef [ϕ(X)] =
ϕ(x)f (x) dx,
X
X
1
where X ∼ f is a random variable taking values in a set X ⊆ Rd , f is a probability
density function (pdf) with respect to the Lebesgue or counting measure, and ϕ :
X → R is a real-valued function.
The evaluation of such high-dimensional integrals is of critical importance in
many scientific areas, including statistical inference [13], rare-event probability estimation [1], machine learning [25, 41], and cryptography [28]. An important application is the calculation of the normalizing constant of a probability distribution,
such as the marginal likelihood (model evidence). However, often obtaining even a
reasonably accurate estimate of Ef [ϕ(X)] can be hard [39]. Deterministic computation methods that use Fubini’s theorem [12] and quadrature rules or extrapolations
[11] suffer from the curse of dimensionality, with the number of required function
evaluations growing exponentially with the dimension. Many other deterministic
and randomized methods have been proposed to estimate high-dimensional integrals.
Examples include Bayesian quadrature, sparse grids, and various Monte Carlo, quasiMonte Carlo, Nested Sampling, and Markov Chain Monte Carlo (MCMC) algorithms
[19, 26, 30, 32, 35, 42]. There are also procedures based on the sequential Monte
Carlo (SMC) approach [8] that provide consistent and unbiased estimators that have
asymptotic normality. However, many popular methods are not unbiased and some
are not known to be consistent. For example, Nested Sampling produces biased
estimates and even its consistency, when Markov Chain Monte Carlo (MCMC) is
used, remains an open problem [7]. An alternative criterion that one might use to
describe the efficiency of algorithms is computational complexity, which often used
in a computer science context. Here, an algorithm is considered to be efficient if it
scales polynomially (rather than exponentially) in the size of the problem.
In this paper, we propose a novel SMC approach for reliable and fast estimation
of high-dimensional integrals. Our method extends the Generalized Splitting (GS)
algorithm of [4], to allow the estimation of quite general integrals. In addition,
our algorithm is specifically designed to perform efficient sampling in regions of X
where f takes small values and ϕ takes large values. In particular, we present a way
of implementing stratification for variance reduction in the absence of knowing the
strata probabilities. A benefit of the proposed Stratified Splitting algorithm (SSA) is
that it provides an unbiased estimator of Ef [ϕ(X)], and that it can be analyzed in a
non-asymptotic setting. In particular, we prove that a simplified version of SSA can
provide polynomial complexity, under certain conditions. We give a specific example
of a #P complete problem where polynomial efficiency can be achieved, providing a
polynomial bound on the number of samples required to achieve a predefined error
bound.
The SSA uses a stratified sampling scheme (see, e.g., [39], Chapter 5), defining
2
a partition of the state space into strata, and using the law of total probability to
deliver an estimator of the value of the integral. To do so, one needs to obtain a
sample population from each strata and know the exact probability of each such
strata. Under the classical stratified sampling framework, it is assumed that the
former is easy to achieve and the latter is known in advance. However, such favorable
scenarios are rarely seen in practice. In particular, obtaining samples from within a
stratum and estimating the associated probability that a sample will be within this
stratum is hard in general [23]. To resolve this issue, the SSA incorporates a multilevel splitting mechanism [4, 9, 24, 38] and uses an appropriate MCMC method to
sample from conditional densities associated with a particular stratum.
The rest of the paper is organized as follows. In Section 2 we introduce the
SSA, explain its correspondence to a generic multi-level sampling framework, and
prove that the SSA delivers an unbiased estimator of the expectation of interest. In
Section 3, we provide a rigorous complexity analysis of the approximation error of the
proposed method, under a simplified setting. In Section 4, we introduce a challenging
estimation problem called the weighted component model, and demonstrate that
SSA can provide an arbitrary level of precision for this problem with polynomial
complexity in the size of the problem. In Section 5, we report our numerical findings
on various test cases that typify classes of problems for which the SSA is of practical
interest. Finally, in Section 6 we summarize the results and discuss possible directions
for future research. Detailed proofs are given in the appendix.
2
2.1
Stratified splitting algorithm
Generic multilevel splitting framework
We begin by considering a very generic multilevel splitting framework, similar to
[14]. Let X ∼ f be a random variable taking values in a set X . Consider a
decreasing sequence of sets X = XS
0 ⊇ · · · ⊇ Xn = ∅ and define Zt = Xt−1 \ Xt ,
for t = 1, . . . , n. Note that Xt−1 = ni=t Zi and that {Zt } yields a partition of X ;
that is,
n
[
X =
Zt , Zt1 ∩ Zt2 = ∅ for 1 ≤ t1 < t2 ≤ n.
(1)
t=1
We can define a sequence of conditional pdfs
ft (x) = f (x | x ∈ Xt−1 ) =
f (x)1{x ∈ Xt−1 }
Pf (X ∈ Xt−1 )
3
for t = 1, . . . , n,
(2)
where 1 denotes the indicator function. Also, define
gt (x) = f (x | x ∈ Zt ) =
f (x)1{x ∈ Zt }
Pf (X ∈ Zt )
for t = 1, . . . , n.
(3)
Our main objective is to sample from the pdfs ft and gt in (2) and (3), respectively. To do so, we first formulate a generic multilevel splitting framework, given in
Algorithm 1.
Algorithm 1: Generic multilevel splitting framework
input : X0 , . . . , Xn and {ft , gt }1≤t≤n .
output: Samples from ft and gt for 1 ≤ t ≤ n.
Create a multi-set X1 of samples from f1 .
for t = 1 to n do
Set Zt ← Xt ∩ Zt .
Set Yt ← Xt \ Zt .
if t < n then
Create a multi-set Xt+1 of samples (particles) from ft+1 , (possibly)
using elements of the Yt set. /* This step is called the
splitting or the rejuvenation step.
*/
return multi-sets {Xt }1≤t≤n , and {Zt }1≤t≤n .
Note that the samples in {Xt }1≤t≤n and {Zt }1≤t≤n are distributed according to
ft and gt , respectively, and these samples can be used to handle several tasks. In
particular, the {Xt }1≤t≤n sets allow one to handle the general non-linear Bayesian
filtering problem [17]. Moreover, by tracking the cardinalities of the sets {Xt }1≤t≤n
and {Zt }1≤t≤n , one is able to tackle hard rare-event probability estimation problems,
such as delivering estimates of Pf (X ∈ Zn ) [4, 26, 40]. Finally, it was recently shown
by [43] that Algorithm 1 can be used as a powerful variance minimization technique
for any general SMC procedure. In light of the above, we propose taking further
advantage of the sets {Xt }1≤t≤n and {Zt }1≤t≤n , to obtain an estimation method
suitable for general integration problems.
2.2
The SSA set-up
Following the above multilevel splitting framework, it is convenient to construct
the sequence of sets {Xt }0≤t≤n by using a performance function S : X → R, in
such a way that {Xt }0≤t≤n can be written as super level-sets of S for chosen levels
4
γ0 , . . . , γn , where γ0 and γn are equal to inf x∈X S(x) and supx∈X S(x), respectively.
In particular, Xt = {x ∈ X : S(x) ≥ γt } for t = 0, . . . , n. The partition {Zt }1≤t≤n ,
and the densities {ft }1≤t≤n and {gt }1≤t≤n , are defined as before via (1), (2), and (3),
respectively. Similarly, one can define a sequence of sub level-sets of S; in this paper
we use the latter for some cases and whenever appropriate.
def
Letting zt = Ef [ϕ (X) | X ∈ Zt ] Pf (X ∈ Zt ) for t = 1, . . . , n, and combining (1)
with the law of total probability, we arrive at
def
z = Ef [ϕ (X)] =
n
X
t=1
Ef [ϕ (X) | X ∈ Zt ] Pf (X ∈ Zt ) =
n
X
zt .
(4)
t=1
The SSA proceeds with the construction of estimators Zbt for zt for t = 1, . . . , n
and, as soon as these are available, we can use (4) to deliver the SSA estimator for
P
z, namely Zb = nt=1 Zbt .
def
def
b t and Pbt
For 1 ≤ t ≤ n, let ϕt = Ef [ϕ (X) | X ∈ Zt ], pt = Pf (X ∈ Zt ), and let Φ
b t Pbt , and recall that, under
be estimators of ϕt and pt , respectively. We define Zbt = Φ
the multilevel splitting framework, we obtain the sets {Xt }1≤t≤n , and {Zt }1≤t≤n .
b t }1≤t≤n and {Pbt }1≤t≤n , in
These sets are sufficient to obtain unbiased estimators {Φ
the following way.
b t to be the (unbiased) Crude Monte Carlo (CMC) estimator of ϕt ,
1. We define Φ
that is,
X
bt = 1
ϕ(Z) for all t = 1, . . . , n.
Φ
|Zt | Z∈Z
t
2. The estimator Pbt is defined similar to the one used in the Generalized Splitting
(GS) algorithm of [4]. In particular, the GS product estimator is defined as foldef
def
lows. Define the level entrance probabilities r0 =Q1, rt = Pf (X ∈ Xt | X ∈ Xt−1 )
for t = 1, . . . , n, and note that Pf (X ∈ Xt ) = ti=0 ri . Then, for t = 1, . . . , n,
it holds that
pt = Pf (X ∈ Zt ) = Pf (X ∈ Xt−1 ) − Pf (X ∈ Xt )
=
t−1
Y
i=0
ri −
t
Y
i=0
ri = (1 − rt )
t−1
Y
ri .
i=0
bt ) Qt−1 R
b
b def
This suggests the estimator Pbt = (1 − R
i=0 i for pt , where R0 = 1 and
bt = |Yt | = |Xt \Zt | for all t = 1, . . . , n.
R
|Xt |
|Xt |
5
In practice, obtaining the {Xt }1≤t≤n and {Zt }1≤t≤n sets requires sampling from the
conditional pdfs in (2) and (3). However, for many specific applications, designing
such a procedure can be extremely challenging. Nevertheless, we can use the Yt =
Xt \ Zt set from iteration t to sample Xt+1 from ft+1 for each t = 1, . . . , n − 1,
via MCMC. In particular, the particles from the Yt set can be “split”, in order
to construct the desired set Xt+1 for the next iteration, using a Markov transition
kernel κt+1 (· | ·) whose stationary pdf is ft+1 , for each t = 1, . . . , n − 1. Algorithm 2
summarizes the general procedure for the SSA. The particular splitting step described
in Algorithm 2 is a popular choice [3, 4, 37], especially for hard problems with
unknown convergence behavior of the corresponding Markov chain.
Algorithm 2: SSA for estimating z = Ef [ϕ (X)]
input : A set X , a pdf f , the functions ϕ : X → R and S : X → R, a sequence of
levels γ0 , . . . , γn , and the sample size N ∈ N.
b — an estimator of z = Ef [ϕ(X)].
output: Z
b
Set R0 ← 1, X1 ← ∅, and f1 ← f .
for i = 1 to N do
draw X ∼ f1 (x) and add X to X1 .
for t = 1 to n do
Set Zt ← {X ∈ Xt : X ∈ Zt } and Yt ← Xt \ Zt .
bt ← 1 P
Set Φ
X∈Zt ϕ(X).
|Zt |
|Y
|
t
bt ←
bt ← 1 − R
bt Qt−1 R
b
Set R
,
and
P
j=0 j .
N
b t Pbt .
Set Zbt ← Φ
if t < n then
/* Performing splitting to obtain Xt+1 .
Set Xt+1 ← ∅ and draw Ki ∼ Bernoulli(0.5), for i = 1, . . . , |Yt |, such that
P|Yt |
i=1 Ki = N mod |Yt |.
for Y ∈ Yt doj
k
Set Mi ←
N
|Yt |
*/
+ Ki and Xi,0 ← Y.
for j = 1 to Mi do
Draw Xi,j ∼ κτt+1 (· | Xi,j−1 ) (where κτt+1 (· | ·) is a τ -step transition kernel
using κt+1 (· | ·)), and add Xi,j to Xt+1 .
Pn
return Zb = t=1 Zbt .
Remark 1 (Independent SSA) Obviously, the SSA above produces dependent
samples in sets Xt and Zt . It will be convenient to also consider a version of SSA
where the samples are independent. This can for example be achieved by simulating
multiple independent runs of SSA, until N realizations X ∈ Xt have been produced.
Of course this Independent SSA (ISSA) variant is much less efficient than the original
algorithm, but the point is that the computational effort (the run time complexity)
6
remains polynomial — O(C 2 n) versus O(Cn), where C is the complexity of SSA
for each level. Hence, from a complexity point of view we may as well consider the
independent SSA variant. This is what will be done in Theorem 2.
Theorem 1 (Unbiased
h iestimator) Algorithm 2 outputs an unbiased estimator;
that is, it holds that E Zb = Ef [ϕ (X)] = z.
Proof 1 See Appendix A.
We next proceed with a clarification for a few remaining practical issues regarding
the SSA.
Determining the SSA levels. It is often difficult to make an educated guess how
to set the values of the level thresholds. However, the SSA requires the values of
{γt }1≤t≤n to be known in advance, in order to ensure that the estimator is unbiased.
To resolve this issue, we perform a single pilot run of Algorithm 2 using a so-called
rarity parameter 0 < ρ < 1. In particular, given samples from an Xt set, we take the
ρ |Xt | performance quantile as the value of the corresponding level γt , and form the
next level set. Such a pilot run helps to establish a set of threshold values adapted
to the specific problem. After the completion of the pilot run we simply continue
with a regular execution of Algorithm 2 using the level threshold values observed in
the pilot run.
Controlling the SSA error. A common practice when working with a Monte
Carlo algorithm that outputs an unbiased estimator, is to run it for R independent
replications to obtain Zb(1) , . . . , Zb(R) , and report the average value. Thus, for a final
estimator, we take
R
X
−1
b
Z=R
Zb(j) .
j=1
To measure the quality of the SSA output, we use the estimator’s relative error (RE),
which is defined by
r
Var Zb
RE = h i √ .
E Zb R
As the variance and expectation of the estimator are not known explicitly, we
report an estimate of the relative error by estimating both terms from the result of
the R runs.
7
3
Complexity
In this section, we present a theoretical complexity analysis of the Independent SSA
variant, discussed in Remark 1. Our focus is on establishing time-complexity results,
and thus our style of analysis is similar to that used for approximate counting algorithms. For an extensive overview, we refer to [29, Chapter 10]. We begin with a
definition of a randomized algorithm’s efficiency.
Definition 1 ([29]) A randomized algorithm gives an (ε, δ)-approximation for the
value z if the output Zb of the algorithm satisfies
b
P z(1 − ε) ≤ Z ≤ z(1 + ε) ≥ 1 − δ.
With the above definition in mind, we now aim to specify the sufficient conditions for
the SSA to provide an (ε, δ)-approximation
n
oto z. A key component in our analysis is
(m)
to construct a Markov chain Xt , m ≥ 0 with stationary pdf ft (defined in (2)),
for all 1 ≤ t ≤ n, and to consider the speed of convergence of the distribution of
(m)
Xt as m increases. Let µt be the probability distribution corresponding to ft , so
Z
ft (u)λ(du),
µt (A) =
A
for all Borel sets A, where λ is some base measure, such as the Lebesgue or counting
measure. To proceed, we have
(τ )
(0)
τ
κt (A | x) = P Xt ∈ A | Xt = x ,
for the τ -step transition law of the Markov chain. Consider the total variation
distance between κτt (· | x) and µt , defined as:
kκτt (· | x) − µt kTV = sup |κτt (A | x) − µt (A)|.
A
An essential ingredient of our analysis is the so-called mixing time (see [36] and [27]
for an extensive overview), which is defined as
τmix (ε, x) = min {τ : kκτt (· | x) − µt kTV ≤ ε}.
Let µ
bt = κτt (· | x ) be the SSA sampling distribution at steps 1 ≤ t ≤ n where, for
simplicity, we suppress x in the notation of µ
bt .
8
Finally, similar to µt and µ
bt , let νt be the probability distribution corresponding
to the pdf gt (defined in (3)), and let νbt be the SSA sampling distribution, for all
1 ≤ t ≤ n.
Theorem 2 details the main efficiency result for the ISSA. We reiterate that from
a complexity point of view the independence setting does not impose a theoretical
limitation, as the run time complexity remains polynomial. An advantage is that by
using ISSA we can engage powerful concentration inequalities [6, 20].
Theorem 2 (Complexity of the ISSA) Let ϕ be a strictly positive real-valued
function, at = minx∈Zt {ϕ(x)}, bt = maxx∈Zt {ϕ(x)}, and rt = min {rt , 1 − rt } for
1 ≤ t ≤ n. Let µ
bt and νbt be the ISSA sampling distributions at steps 1 ≤ t ≤ n, for Xt
and Zt , respectively. Then, the ISSA gives an (ε, δ)-approximation to z = Ef [ϕ(X)],
provided that for all 1 ≤ t ≤ n the following holds.
1.
kb
µt − µt kTV ≤
2.
kb
νt − νt kTV ≤
ε rt
32n
and
ε at
16(bt −at )
|Xt | ≥
and
3072 n2 ln(4n2 /δ)
.
ε2 r2t
|Zt | ≥
128(bt −at )2 ln(4n/δ)
.
ε2 a2t
Proof 2 See Appendix A.
In some cases, the distributions of the states in Xt and Zt generated by Markov
chain defined by the kernel κτt , approach the target distributions µt and νt very fast.
This occurs for example when there exists a polynomial in n (denoted by P(n)),
such that the mixing time [27] is bounded by O(P(n)), (bt − at )2 /a2 = O(P(n)), and
rt = O(1/P(n)) for all 1 ≤ t ≤ n. In this case, the ISSA becomes a fully polynomial
randomized approximation scheme (FPRAS) [29]. In particular, the ISSA results in
a desired (ε, δ)-approximation to z = Ef [ϕ(X)] with running time bounded by a
polynomial in n, ε−1 , and ln(δ −1 ). Finally, it is important to note that an FPRAS
algorithm for such problems is essentially the best result one can hope to achieve
[22].
We next illustrate the use of Theorem 2 with an example of a difficult problem
for which the ISSA provides an FPRAS.
4
FPRAS for the weighted component model
Consider a system of k components. Each component i generates a specific amount
of benefit, which is given by a positive real number wi , i = 1, . . . , k. In addition,
each component can be operational or not.
9
Let w = (w1 , . . . , wk )> be the column vector of component weights (benefits), and
x = (x1 , . . . , xk )> be a binary column vector, where xi indicates the ith component’s
operational status for 1 ≤ i ≤ k. That is, if the component i is operational xi = 1,
and xi = 0 if it is not. Under this setting, we define the system performance as
S(w, x) =
k
X
wi xi = w> x.
i=1
We further assume that all elements are independent of each other, and that each
element is operational with probability 1/2 at any given time. For the above system
definition, we might be interested in the following questions.
1. Conditional
expectation estimation. Given a minimal threshold performance
Pk
γ ≤ i=1 wi , what is the expected system performance? That is to say, we are
interested in the calculation of
E [S(w, X) | S(w, X) ≤ γ] ,
(5)
where X is a k-dimensional binary vector generated uniformly at random from
the {0, 1}k set. This setting appears (in a more general form), in a portfolio
credit risk analysis [16], and will be discussed in Section 5.
2. Tail probability estimation [1]. Given the minimal threshold performance γ,
what is the probability that the overall system performance is smaller than γ?
In other words, we are interested in calculating
P(S(X) ≤ γ) = E [1{S(X) ≤ γ}] .
(6)
The above problems are both difficult, since a uniform generation of X ∈ {0, 1}k ,
such that w> X ≤ γ, corresponds to the knapsack problem, which belongs to #P
complexity class [31, 44]. In this section, we show how one can construct an FPRAS
for both problems under the mild condition that the difference between the minimal
and the maximal weight in the w vector is not large. This section’s main result is
summarized next.
Proposition 1 Given a weighted component model with k weights, w = (w1 , . . . , wk )
and a threshold γ, let w = min{w}, w = max{w}. Then, provided that w =
O (P(k)) w, there exists an FPRAS for the estimation of both (5) and (6).
10
Prior to stating the proof of Proposition 1, define
(
)
k
X
Xb = x ∈ {0, 1}k :
wi x i ≤ b
i=1
for b ∈ R,
(7)
and let µb be the uniform distribution on the Xb set. [31] introduce an MCMC
algorithm that is capable of sampling from the Xb set almost uniformly at random. In
particular, this algorithm can sample X ∼ µ
bb , such that kb
µb − µb kTV ≤ ε. Moreover,
the authors show that their Markov chain mixes rapidly, and in particular, that its
mixing time is polynomial in k and is given by τmix (ε) = O k 9/2+ε . Consequentially,
the sampling from µ
bb can be performed in O(P(k)) time for any ε > 0.
The proof of Proposition 1 depends on the following technical lemma.
Lemma 1 Let X ∼ π
b be a strictly positive univariate random variable such that
a ≤ X ≤ b, and let X1 , . . . , Xm be its independent realizations. Then, provided that
kb
π − πkTV
εa
≤
,
4(b − a)
and
(b − a2 ) ln(2/δ)
m≥
,
2(ε/4)2 a2
it holds that:
1 X
P (1 − ε)Eπ [X] ≤
Xi ≤ (1 + ε)Eπ [X]
m i=1
m
!
≥ 1 − δ.
Proof 3 See Appendix A.
We next proceed with the proof of Proposition 1 which is divided into two parts.
The first for the conditional expectation estimation and the second for the tail probability evaluation.
Proof 4 (Proposition 1: FPRAS for (5)) With the powerful result of [31] in hand,
one can achieve a straightforward development of an FPRAS for the conditional expectation estimation problem. The proof follows immediately from Lemma 1. In
particular, all we need to do in order to achieve an (ε, δ) approximation to (5) is to
generate
(w − w)2 ln(2/δ)
m=
2(ε/4)2 w2
samples from µ
bγ , such that
kb
µγ − µγ kTV ≤
11
εw
.
4(w − w)
Recall that the mixing time is polynomial in k, and note that the number of samples
m is also polynomial in k, thus the proof is complete, since
m=
(w − w)2 ln(2/δ)
(ε/4)2 w2
=
|{z}
w=O(P(k))w
O(P(k))
ln(2/δ)
.
ε2
Proof 5 (Proposition 1: FPRAS for (6)) In order to put this problem into the
setting of Theorem 2 and achieve an FPRAS, a careful definition of the corresponding
level sets is essential. In particular, the number of levels should be polynomial in k,
and the level entrance probabilities {rt }1≤t≤n , should not be too small. Fix
P
k
w
−
γ
i=1 i
,
n=
w
to be the number of levels, and set γt = γ + (n − t) w for t = 0, . . . , n. For general γ
it holds that
E [1{S(X) ≤ γ}] = E [1{S(X) ≤ γ + n w} | S(X) ≤ γ + n w] P (S(X) ≤ γ + n w)
"
#
!
k
k
X
X
+ E 1{S(X) ≤ γ} | γ + n w < S(X) ≤
wi P γ + n w < S(X) ≤
wi
i=1
= E [1{S(X) ≤ γ} | S(X) ≤ γ + n w]
|
{z
}
k
2 −1
+
2k
(∗)
k
X
i=1
wi
!
i=1
1
,
2k
where the last equality follows from the fact
Pk that there is only one vector x =
(1, 1, . . . , 1) for which γ + n w < S(x) ≤ i=1 wi . That is, it is sufficient to develop an efficient approximation to (∗) only, since the rest are constants.
We continue by defining the sets X = Xγ0 ⊇ · · · ⊇ Xγn via (7), and by noting
that for this particular problem, our aim is to find P(X ∈ Xγn ), so the corresponding
estimator simplifies to (see Section 2.2 (2)),
n
Y
Zb =
t=0
bt .
R
In order to show that the algorithm provides an FPRAS, we will need to justify only
condition (1) of Theorem 2, which is sufficient in our case because we are dealing
with an indicator integrand. Recall that the formal requirement is
kb
µt − µt kTV ≤ ε rt /32n,
and
12
|Xt | ≥ 3072 n2 ln(4n2 /δ)/ε2 r2t ,
where µt is the uniform distribution on Xγt for t = 0, . . . , n, and each sample in Xt
is distributed according to µ
bt . Finally, the FPRAS result is established by noting that
the following holds.
1. From Lemma 6 in the Appendix, we have that rt ≥
1
k+1
for 1 ≤ t ≤ n.
2. The sampling from µ
bt can be performed in polynomial (in k) time [31].
3. The number of levels n (and thus the required sample size {|Xt |}1≤t≤n ) is polynomial in k, since
P
k
w
−
τ
i
kw
i=1
≤
=
O(P(k)).
|{z}
w
w
w=O(P(k))w
Unfortunately, for many problems, an analytical result such as the one obtained
in this section is not always possible to achieve. The aim of the following numerical
section is to demonstrate that the SSA is capable of handling hard problems in the
absence of theoretical performance.
5
Numerical experiments
5.1
Portfolio credit risk
We consider a portfolio credit risk setting [16]. Given a portfolio of k assets, the
portfolio loss L is the random variable
L=
k
X
li Xi ,
(8)
i=1
where li is the risk of asset i ∈ {1, . . . , k}, and Xi is an indicator random variable
that models the default of asset i. Under this setting (and similar to Section 4), one
is generally interested in the following.
1. Conditional Value at Risk (CVaR). Given a threshold (value at risk) v, calculate
the conditional value at risk c = E[L | L ≥ v].
2. Tail probability estimation. Given the value at risk, calculate the tail probability P(L ≥ v) = E [1{L ≥ v}].
13
The SSA can be applied to both
Pk problems as follows. For tail probability estimation, we simply set ϕ(x) = 1{ i=1 li xi ≥ v}. For conditional expectation, we set
P
ϕ(x) = ki=1 li xi .
Note that the tail probability estimation (for which the integrand is the indicator
function), is a special case of a general integration. Recall that the GS algorithm of [4]
works on indicator integrands, and thus GS is a special case of the SSA. Consequently,
in this section we will investigate the more interesting (and more general) scenario
of estimating an expectation conditional on a rare event.
As our working example, we consider a credit risk in a Normal Copula model and,
in particular, a 21 factor model from [16] with 1, 000 obligors. Thus, the integrals of
interest are 1021 dimensional.
The SSA setting is similar to the weighted component model from Section 4. We
define a k-dimensional binary vector x = (x1 , . . . , xk ), for which xi stands for the
ith asset default (xi = 1 for default, and 0 otherwise). We take the performance
function S(x) to be the loss function (8). Then, the level sets are defined naturally
by Xt = {x : P
S(x) ≥ γt }, (see also Section 2.2). In our experiment, we set γ0 = 0
and γn = 1 + ki=1 li . In order to determine the remaining levels γ1 , . . . , γn−1 , we
execute a pilot run of Algorithm 2 with N = 1, 000 and ρ = 0.1. As an MCMC
sampler, we use a Hit-and-Run algorithm [26, Chapter 10, Algorithm 10.10], taking
a new sample after 50 transitions.
It is important to note that despite the existence of several algorithms for estimating c, the SSA has an interesting feature, that (to the best of our knowledge) is
not present in other methods. Namely, one is able to obtain an estimator for several
CVaRs via a single SSA run. To see this, consider the estimation of c1 , . . . , cs for
s ≥ 1. Suppose that v1 ≤ · · · ≤ vs and note that it will be sufficient to add these
values to the {γt } (as additional levels), and retain s copies of Pbt and Zbt . In particular, during the SSA execution, we will need to closely follow the γ levels, and as
soon as we encounter a certain vj for 1 ≤ j ≤ s, we will start to update the corre(j)
(j)
sponding values of Pbt and Zbt , in order to allow the corresponding estimation of
cj = E[L | L ≥ vj ]. Despite that such a procedure introduces a dependence between
the obtained estimators, they still remain unbiased.
To test the above setting, we perform the experiment with a view to estimate
{cj }1≤j≤13 using the following values at risk:
{10000, 14000, 18000, 22000, 24000, 28000, 30000,
34000, 38000, 40000, 44000, 48000, 50000}.
(9)
The execution of the SSA pilot run with the addition of the desired VaRs (levels)
14
from (9) (marked in bold), yields the following level values of (γ0 , . . . , γ21 ):
(0, 788.3, 9616.7, 10000, 14000, 18000, 22000, 24000, 28000,
30000, 34000, 38000, 40000, 44000, 47557.6, 48000, 49347.8,
50000, 50320.6, 50477.4, 50500, ∞).
Table 1 summarizes the results obtained by executing 1 pilot and 5 regular independent runs of the SSA. For each run, we use the parameter set that was specified
for the pilot run (N = 1, 000 and burn-in of 50). The overall execution time (for all
these R = 1 + 5 = 6 independent runs) is 454 seconds. The SSA is very accurate.
In particular, we obtain an RE of less than 1% for each b
c while employing a very
modest effort.
v
b
c
10000 1.68 × 104
14000 2.09 × 104
18000 2.46 × 104
22000 2.82 × 104
24000 2.99 × 104
28000 3.32 × 104
30000 3.48 × 104
RE
0.67 %
0.51 %
0.21 %
0.19 %
0.23 %
0.21 %
0.14 %
v
34000
38000
40000
44000
48000
50000
b
c
3.80 × 104
4.11 × 104
4.26 × 104
4.56 × 104
4.86 × 104
5.01 × 104
RE
0.05 %
0.05 %
0.08 %
0.08 %
0.02 %
0.02 %
Table 1: The SSA results for the Normal copula credit risk model with 21 factors
and 1, 000 obligors.
The obtained result is especially appealing, since the corresponding estimation
problem falls into rare-event setting [15]. That is, a CMC estimator will not be
applicable in this case.
5.2
Self-avoiding walks
In this section, we consider random walks of length n on the two-dimensional lattice
of integers, starting from the origin. In particular, we are interested in estimating
the following quantities:
15
1. cn : the number of SAWs of length n,
2. ∆n : the expected distance of the final SAW coordinate to the origin.
To put these SAW problems into the SSA framework, define the set of directions,
X = {Left, Right, Up, Down}n , and let f be the uniform pdf on X . Let ξ(x) denote
the final coordinate of the random walk represented
by the directions
vector x. We
have cn = Ef [1{X is SAW}] and ∆n = Ef kξ(X)k 1{X is SAW} .
Next, we let Xt ⊆ X be the set of all directions vectors that yield a valid selfavoiding walk of length at least t, for 0 ≤ t ≤ n. In addition, we define Zt to be
the set of all directions vectors that yield a self-avoiding walk of length (exactly)
t, for 1 ≤ t ≤ n. The above gives the required partition of X . Moreover, the
simulation from ft (x) = f (x | x ∈ Xt−1 ), reduces to the uniform selection of the
SAW’s direction at time 1 ≤ t ≤ n.
Our experimental setting for SAWs of lengths n is as follows. We set the sample
size of the SSA to be Nt = 1000 for all t = 1, . . . , n. In this experiment, we are
interested in both the probability that X lies in Zn , and the expected distance of
X ∈ Zn (uniformly selected) to the origin. These give us the required estimators
of cn and ∆n , respectively. The leftmost plot of Fig. 1 summarizes a percent error
(PE), which is defined by
b
cn − cn
,
PE = 100
cn
where b
cn stands for the SSA’s estimator of cn .
10
200
0
R/n
PE (%)
5
−5
−10
−15
10
20
30
40
50
3%
1%
150
100
50
3% RE
1% RE
60 70
0
10
20
30
40
50
60
70
n
n
Figure 1: The PE (leftmost plot) and the number of independent runs divided by
SAW’s length (rightmost plot) of the SSA as a function of SAW length n for 3% and
1% RE.
In order to explore the convergence of the SSA estimates to the true quantity of
interest, the SSA was executed for a sufficient number of times to obtain 3% and
16
1% relative error (RE) [39], respectively. The exact cn values for n = 1, . . . , 71 were
taken from [18, 21]); naturally, when we allow a smaller RE, that is, when we increase
R, the estimator converges to the true value cn , as can be observed in leftmost plot
of Fig. 1. In addition, the rightmost plot of Fig. 1 shows that regardless of the RE,
the required number of independent SSA runs (R) divided by SAW’s length (n), is
growing linearly with n.
To further investigate the SSA convergence, we consider the two properties of
SAW’s. In particular, the following holds [2, 10, 33, 34]
1
1. µ = limn→∞ cnn ∈ [µ, µ] = [2.62002, 2.679192495].
e n n = ∆n .
2. ∆n = n 4 ≤ ∆n ≈ n3/4 = ∆
1
Fig. 2 summarizes our results compared to these bounds. In particular, we run the
SSA to achieve the 3% RE (for ĉn ) for 1 ≤ n ≤ 200. It can be clearly observed, that
1/n
the estimator b
cn converges toward the [µ, µ] interval as n grows. It is interesting to
note that, at least for small n, ∆n seems to grow at a rate smaller than the suggested
3/4.
µ
1/n
b
cn
µ
2.9
2.8
50
40
30
en
∆
bn
∆
∆n
20
2.7
10
2.6
50
100
n
150
1/n
6
Figure 2: The b
cn
Discussion
0
200
50
100
n
150
200
b n as a function of the SAW’s length n.
and the ∆
In this paper we described a general procedure for multi-dimensional integration, the
SSA, and applied it to various problems from different research domains. We showed
that this method belongs to a very general class of SMC algorithms and developed
its theoretical foundation. The proposed SSA is relatively easy to implement and
17
our numerical study indicates that the SSA yields good performance in practice.
However, it is important to note that generally speaking, the efficiency of the SSA
and similar sequential algorithms is heavily dependent on the mixing time of the
corresponding Markov chains that are used for sampling. A rigorous analysis of the
mixing time for different problems is thus of great interest. Finally, based on our
numerical study, it will be interesting to apply the SSA to other practical problems.
7
Acknowledgements
This work was supported by the Australian Research Council Centre of Excellence
for Mathematical & Statistical Frontiers, under grant number CE140100049.
References
[1] Asmussen, S., and Glynn, P. W. Stochastic Simulation: Algorithms and
Analysis. Applications of Mathematics. Springer Science and Business Media,
LLC, 2007.
[2] Beyer, W., and Wells, M. Lower bound for the connective constant of a
self-avoiding walk on a square lattice. Journal of Combinatorial Theory, Series
A 13, 2 (1972), 176–182.
[3] Botev, Z., and Kroese, D. P. An efficient algorithm for rare-event probability estimation, combinatorial optimization, and counting. Methodology and
Computing in Applied Probability 10, 4 (December 2008), 471–505.
[4] Botev, Z. I., and Kroese, D. P. Efficient Monte Carlo simulation via the
Generalized Splitting method. Statistics and Computing 22 (2012), 1–16.
[5] Bullen, P. A Dictionary of Inequalities. Monographs and Research Notes in
Mathematics. Taylor & Francis, Oxfordshire, 1998.
[6] Chernoff, H. A Measure of Asymptotic Efficiency for Tests of a Hypothesis
Based on the sum of Observations. Ann. Math. Statist. 23, 4 (12 1952), 493–507.
[7] Chopin, N., and Robert, C. P. Properties of nested sampling. Biometrika
97, 3 (2010), 741–755.
18
[8] Del Moral, P., Doucet, A., and Jasra, A. Sequential Monte Carlo samplers. Journal of the Royal Statistical Society: Series B (Statistical Methodology)
68, 3 (2006), 411–436.
[9] Duan, Q., and Kroese, D. P. Splitting for multi-objective optimization.
Methodology and Computing in Applied Probability (June 2017), 1–17.
[10] Duminil–Copin, H., and Hammond, A. Self-avoiding walk is sub-ballistic.
Communications in Mathematical Physics 324, 2 (2013), 401–423.
[11] Forsythe, G. E., Malcolm, M. A., and Moler, C. B. Computer methods
for mathematical computations. Prentice-Hall series in automatic computation.
Prentice-Hall, Englewood Cliffs (N.J.), 1977.
[12] Friedman, H. A consistent Fubini-Tonelli theorem for nonmeasurable functions. Illinois J. Math. 24, 3 (1980), 390–395.
[13] Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. Bayesian
Data Analysis, 3 ed. Taylor & Francis, Oxfordshire, July 2003.
[14] Gilks, W. R., and Berzuini, C. Following a moving target - Monte Carlo
inference for dynamic Bayesian models. Journal of the Royal Statistical Society:
Series B (Statistical Methodology) 63, 1 (2001), 127–146.
[15] Glasserman, P. Monte Carlo methods in financial engineering. Applications
of mathematics. Springer, New York, 2004. Permire parution en dition broche
2010.
[16] Glasserman, P., and Li, J. Importance sampling for portfolio credit risk.
Management Science 51, 11 (2005), 1643–1656.
[17] Gordon, N. J., Salmond, D. J., and Smith, A. F. M. Novel approach to
nonlinear/non-Gaussian Bayesian state estimation. Radar and Signal Processing, IEE Proceedings F 140, 2 (Apr. 1993), 107–113.
[18] Guttmann, A. J., and Conway, A. R. Square lattice self-avoiding walks
and polygons. Annals of Combinatorics 5, 3 (2001), 319–345.
[19] Heiss, F., and Winschel, V. Likelihood approximation by numerical integration on sparse grids. Journal of Econometrics 144, 1 (2008), 62–80.
[20] Hoeffding, W. Probability inequalities for sums of bounded random variables.
Journal of the American Statistical Association 58, 301 (1963), 13–30.
19
[21] Jensen, I. Enumeration of self-avoiding walks on the square lattice. Journal
of Physics A: Mathematical and General 37, 21 (2004), 5503.
[22] Jerrum, M., and Sinclair, A. The Markov chain Monte Carlo method:
an approach to approximate counting and integration. In Approximation Algorithms for NP-hard Problems (Boston, 1996), D. Hochbaum, Ed., PWS Publishing, pp. 482–520.
[23] Jerrum, M., Valiant, L. G., and Vazirani, V. V. Random Generation
of Combinatorial Structures from a Uniform Distribution. Theor. Comput. Sci.
43 (1986), 169–188.
[24] Kahn, H., and Harris, T. E. Estimation of particle Transmission by Random Sampling. National Bureau of Standards Applied Mathematics Series 12
(1951), 27–30.
[25] Koller, D., and Friedman, N. Probabilistic Graphical Models: Principles
and Techniques - Adaptive Computation and Machine Learning. The MIT Press,
Cambridge, Massachusetts, 2009.
[26] Kroese, D. P., Taimre, T., and Botev, Z. I. Handbook of Monte Carlo
methods. John Wiley and Sons, New York, 2011.
[27] Levin, D. A., Peres, Y., and Wilmer, E. L. Markov chains and mixing
times. Providence, R.I. American Mathematical Society, 2009. With a chapter
on coupling from the past by James G. Propp and David B. Wilson.
[28] McGrayne, S. The Theory that Would Not Die: How Bayes’ Rule Cracked
the Enigma Code, Hunted Down Russian Submarines, & Emerged Triumphant
from Two Centuries of Controversy. Yale University Press, London, 2011.
[29] Mitzenmacher, M., and Upfal, E. Probability and computing : randomized
algorithms and probabilistic analysis. Cambridge University Press, New York,
2005.
[30] Morokoff, W. J., and Caflisch, R. E. Quasi-Monte Carlo Integration.
Journal of Computational Physics 122, 2 (1995), 218–230.
[31] Morris, B., and Sinclair, A. Random walks on truncated cubes and sampling 0-1 knapsack solutions. SIAM Journal on Computing 34, 1 (2004), 195–
226.
20
[32] Newman, M., and Barkema, G. Monte Carlo Methods in Statistical Physics.
Clarendon Press, Oxford, New York, 1999.
[33] Nienhuis, B. Exact Critical Point and Critical Exponents of O(n) Models in
Two Dimensions. Phys. Rev. Lett. 49 (Oct 1982), 1062–1065.
[34] Noonan, J. New upper bounds for the connective constants of self-avoiding
walks. Journal of Statistical Physics 91, 5 (1998), 871–888.
[35] O’Hagan, A. Bayes-Hermite quadrature. Journal of Statistical Planning and
Inference 29, 3 (1991), 245–260.
[36] Roberts, G. O., Rosenthal, J. S., et al. General state space Markov
chains and MCMC algorithms. Probability Surveys 1 (2004), 20–71.
[37] Rubinstein, R. Y. The gibbs cloner for combinatorial optimization, counting and sampling. Methodology and Computing in Applied Probability 11, 4
(December 2009), 491–549.
[38] Rubinstein, R. Y. Randomized algorithms with splitting: Why the classic
randomized algorithms do not work and how to make them work. Methodology
and Computing in Applied Probability 12, 1 (March 2010), 1–50.
[39] Rubinstein, R. Y., and Kroese, D. P. Simulation and the Monte Carlo
Method, 3 ed. John Wiley & Sons, New York, 2017.
[40] Rubinstein, R. Y., Ridder, A., and Vaisman, R. Fast Sequential Monte
Carlo Methods for Counting and Optimization. John Wiley & Sons, New York,
2013.
[41] Russell, S., and Norvig, P. Artificial Intelligence: A Modern Approach,
3 ed. Prentice Hall, Englewood Cliffs (N.J.), 2009.
[42] Skilling, J. Nested sampling for general Bayesian computation. Bayesian
Anal. 1, 4 (12 2006), 833–859.
[43] Vaisman, R., Kroese, D. P., and Gertsbakh, I. B. Splitting sequential
Monte Carlo for efficient unreliability estimation of highly reliable networks.
Structural Safety 63 (2016), 1 – 10.
[44] Valiant, L. G. The complexity of enumeration and reliability problems. SIAM
Journal on Computing 8, 3 (1979), 410–421.
21
Appendix
A
Technical arguments
Proof of Theorem 1. Recall that Z1 , . . . , Zn is a partition of the set X , so from the law
of total probability we have
Ef [ϕ (X)] =
n
X
t=1
Ef [ϕ (X) | X ∈ Zt ] Pf (X ∈ Zt ).
b = Pn Z
b
By the linearity of expectation, and since Z
t=1 t , it will be sufficient to show that
for all t = 1, . . . , n, it holds that
h i
E Zbt = Ef [ϕ (X) | X ∈ Zt ] Pf (X ∈ Zt ) .
To see this, we need the following.
1. Although that the samples in the Zt set for 1 ≤ t ≤ n are not independent due to
MCMC and splitting usage, they still have the same distribution; that is, for all t,
it holds that
"
#
X
ϕ(X) |Zt | = |Zt | Ef [ϕ (X) | X ∈ Zt ] .
(10)
E
X∈Zt
2. From the unbiasedness of multilevel splitting [4, 26], it holds for all 1 ≤ t ≤ n that
t−1
h i
Y
bt
bj = Pf (X ∈ Zt ) .
E Pbt = E 1 − R
R
(11)
j=0
Combining (10) and (11) with a conditioning on the cardinalities of the Zt sets, we complete
the proof with:
22
"
#
h i
h
i
X
1
b t Pbt = E Pbt
E Zbt = E Φ
ϕ(X)
|Zt |
X∈Zt
##
" "
X
1
ϕ(X) |Z0 |, . . . , |Zt |
= E E Pbt
|Zt |
X∈Zt
"
"
##
X
1
= E Pbt
E
ϕ(X) |Z0 |, . . . , |Zt |
|Zt |
X∈Zt
1
b
= E Pt
|Zt | Ef [ϕ (X) | X ∈ Zt ]
|{z}
|Zt |
(10)
h i
= Ef [H (X) | X ∈ Zt ] E Pbt |{z}
= Ef [H (X) | X ∈ Zt ] Pf (X ∈ Zt ) .
(11)
Proof of Lemma 1. Recall that
kb
π − πkTV
1
sup
=
b − a ϕ : R→[a,b]
Z
ϕ(x) π
b(dx) −
Z
ϕ(x) π(dx) ,
for any function ϕ : R → [a, b], (Proposition 3 in [36]). Hence,
Eπ [X] − (b − a)
εa
εa
≤ Eπb [X] ≤ Eπ [X] + (b − a)
.
4(b − a)
4(b − a)
Combining this with the fact that X ≥ a, we arrive at
1−
ε
εa
E [X]
εa
ε
≤1−
≤ πb
≤1+
≤1+ .
4
4Eπ [X]
Eπ [X]
4Eπ [X]
4
Next, since
Eπb
"
#
m
1 X
Xi = Eπb [X] ,
m
i=1
we can apply the [20] inequality, to obtain
!
1 Pm
X
ε
ε
i
≤1+
≥ 1 − δ,
P 1 − ≤ m i=1
4
Eπb [X]
4
for
m=
(12)
(b − a)2 ln(2/δ)
(b − a)2 ln(2/δ)
≥
.
2
2(ε/4)2 a2
2(ε/4)2 (Eπb [X])
23
(13)
Finally, we complete the proof by combining (12) and (13), to obtain:
!
2
1 Pm
X
ε/2 2 Eπb [X] m
ε/2
i
i=1
P 1+ε≤ 1−
≤
≤1+ε
≤ 1+
2
Eπ [X] Eπb [X]
2
!
m
1 X
= P (1 − ε)Eπ [X] ≤
Xi ≤ (1 + ε)Eπ [X] ≥ 1 − δ.
m
i=1
Proof of Theorem 2. The proof of this theorem consists of the following steps.
1. In Lemma 2, we prove that an existence of an ε, nδ P
-approximation to {zt }1≤t≤n
implies an existence of an (ε, δ)-approximation to z = nt=1 zt .
δ
-approximation to {ϕt }1≤t≤n
2. In Lemma 3, we prove that an existence of an 4ε , 2n
δ
and {pt }1≤t≤n implies an ε, n -approximation existence to {zt }1≤t≤n .
δ
-approximations to ϕt and pt
3. In Lemmas 1, 4 and 5, we provide the required 4ε , 2n
for 1 ≤ t ≤ n.
Lemma 2 Suppose that for all t = 1, . . . , n, an ε, nδ -approximation to zt exists. Then,
P z(1 − ε) ≤ Zb ≤ z(1 + ε) ≥ 1 − δ.
Proof 6 From the assumption of the existence of the ε, nδ -approximation to zt for each
1 ≤ t ≤ n, we have
δ
δ
P Zbt − zt ≤ εzt ≥ 1 − , and P Zbt − zt > εzt < .
n
n
By using Boole’s inequality (union bound), we arrive at
n
X
bt − zt > εzt < n δ = δ,
P ∃ t : Zbt − zt > εzt ≤
P Z
n
t=1
that is, it holds for all t = 1, . . . , n, that
bt − zt ≤ εzt = 1 − P ∃ t : Zbt − zt > εct ≥ 1 − δ,
P ∀t : Z
and hence,
P (1 − ε)
n
X
t=1
zt ≤
n
X
t=1
Zbt ≤ (1 + ε)
n
X
t=1
24
zt
!
b ∈≤ z(1 ± ε) ≥ 1 − δ.
=P Z
δ
Lemma 3 Suppose that for all t = 1, . . . , n, there exists an 4ε , 2n
-approximation to ϕt
and pt . Then,
δ
P zt (1 − ε) ≤ Zbt ≤ zt (1 + ε) ≥ 1 −
for all t = 1, . . . , n.
n
δ
Proof 7 By assuming an existence of 4ε , 2n
-approximation to ϕt and pt , namely:
!
!
bt
Pbt
Φ
− 1 ≤ ε/4 ≥ 1 − δ/2n, and P
− 1 ≤ ε/4 ≥ 1 − δ/2n,
P
ht
pt
and combining it with the union bound, we arrive at
2
2
b t Pbt
ε/2
Φ
ε/2
P 1 − ε ≤ 1 −
≤
≤ 1+
≤ 1 + ε ≥ 1 − δ/n,
|{z}
|{z}
2
ϕt pt
2
(∗)
(∗)
where (∗) follows from the fact that for any 0 < |ε| < 1 and n ∈ N we have
ε/2 n
ε/2 n
1−ε≤ 1−
and
1+
≤ 1 + ε.
n
n
(14)
To see that
n (14) holds, note that by using exponential inequalities from [5], we have that
ε/2
1− n
≥ 1 − 2ε ≥ 1 − ε. In addition, it holds that |eε − 1| < 7ε/4, and hence:
ε/2 n
7(ε/2)
1+
≤ eε/2 ≤ 1 +
≤ 1 + ε.
n
4
δ
To complete the proof of Theorem 2, we need to provide 4ε , 2n
-approximations to
both ϕt and pt . However, at this stage we have to take into account a specific splitting
strategy, since the SSA sample size bounds depend on the latter. Here we examine the
independent setting, for which the samples in each {Xt }1≤t≤n set are independent. That
is, we use multiple runs of the SSA at each stage (t = 1, . . . , n) of the algorithm execution.
See Remark 1 for further details.
Remark 2 (Lemma 1 for binary random variables) For a binary random variable
X ∈ {0, 1}, with a known lower bound on its mean, Lemma 1 can be strengthened via the
usage of [6] bound instead of the [20] inequality. In particular, the following holds.
Let X ∼ π
b(x) be a binary random variable and let X1 , . . . , Xm be its independent
realizations. Then, provided that Eπb [X] ≥ E0πb [X],
kb
π − πkTV ≤
ε E0πb [X]
,
4
and m ≥
25
3 ln(2/δ)
,
(ε/4)2 (E0πb [X])2
it holds that:
m
1 X
P (1 − ε)Eπ [X] ≤
Xi ≤ (1 + ε)Eπ [X]
m
i=1
!
≥ 1 − δ.
The corresponding proof is almost identical to the one presented in Lemma 1. The major
difference is the bound on the sample size in (13), which is achieved via the Chernoff bound
from [29, Theorem 10.1] instead of Hoeffding’s inequality.
Lemma 4 Suppose that at = minx∈Zt {ϕ(x)}, bt = maxx∈Zt {ϕ(x)} for all t = 1, . . . , n.
Then, provided that the samples in the Zt set are independent, and are distributed according
to νbt such that
kb
νt − νt kTV ≤
ε at
,
16(bt − at )
b t = |Zt |−1 P
then Φ
X∈Zt ϕ(X) is an
and
ε δ
4 , 2n
|Zt | ≥
128(bt − at )2 ln(4n/δ)
,
ε2 a2t
-approximation to ϕt .
Proof 8 The proof is an immediate consequence of Lemma 1. In particular, note that
kb
νt − νt kTV ≤
and that
|Zt | ≥
ε
4
at
ε at
=
,
4(bt − at )
16(bt − at )
δ
)
(bt − at )2 ) ln(2/ 2n
128(bt − at )2 ln(4n/δ)
=
.
2( 4ε /4)2 a2
ε2 a2t
Lemma 5 Suppose that the samples in the Xt set are independent, and are distributed
according to µ
bt , such that
3072 n2 ln(4n2 /δ)
,
ε2 r2t
δ
where rt = min {rt , 1 − rt } for 1 ≤ t ≤ n. Then, Pbt is an 4ε , 2n
-approximation to pt .
kb
µt − µt kTV ≤
ε rt
,
32n
and
|Xt | ≥
bt Qt−1 R
b
Proof 9 Recall that Pbt = 1 − R
j=0 j for t = 1, . . . , n. Again, by combining the
union bound with (14),
we conclude that the desired approximation to pt can be obtained by
ε
deriving the 8n
, 2nδ 2 -approximations for each rt and 1 − rt . In this case, the probability
bt /rt satisfies 1 − ε/8n ≤ R
bt /rt ≤ 1 + ε/8n is at least 1 − δ/2n2 .
that for all t = 1, . . . , n, R
bt )/(1 − rt ), and thus, we arrive at:
The same holds for (1 − R
!
n
n
ε
ε
bt
/2
/2
P
P 1 − ε/4 ≤ 1 − 4
≤
≤ 1+ 4
≤ 1 + ε/4 ≥ 1 − δ/2n.
n
pt
n
26
bt and (1 − R
bt ) are easily achieved via Remark 2. In particular,
The bounds for each R
ε
it is not very hard to verify that in order to get an 8n
, 2nδ 2 -approximation, it is sufficient
to take
ε
r
ε rt
,
kb
µt − µt kTV ≤ 8n t =
4
32n
and
3 ln 2/ 2nδ 2
3072 n2 ln(4n2 /δ)
|Xt | ≥
=
.
ε
( 8n /4)2
ε2 r2t
Lemma 6 Suppose without loss of generality that w = (w1 , . . . , wk ) satisfies w1 ≤ w2 ≤
· · · ≤ wk , that is w = w1 . Then, for Xb and Xb−w1 sets defined via (7), it holds that:
r=
|Xb−w1 |
1
≥
.
|Xb |
k+1
Proof 10 For any b ∈ R and x = (x1 , . . . , xk ), define a partition of Xb via
(w1 )
Xb
= {x ∈ Xb : x1 = 1},
and
(−w1 )
Xb
= {x ∈ Xb : x1 = 0}.
Then, the following holds.
(w1 )
1. For any x ∈ Xb
in
(−w )
Xb−w11
, replace x1 = 1 with x1 = 0, and note that the resulting vector is
(w1 )
(−w )
set, since its performance is at most b − w1 , that is Xb
≤ Xb−w11 .
(−w )
Similarly, for any x ∈ Xb−w11 , setting x1 = 1 instead of x1 = 0, results in a vector
(w1 )
which belongs to the Xb
set. That is:
(w1 )
Xb
(−w )
= Xb−w11 .
(15)
(w )
2. For any x ∈ Xb 1 , replace x1 = 1 with x1 = 0 and note that the resulting vector is
(−w )
(w )
(−w )
(−w )
now in the Xb 1 set, that is |Xb 1 | ≤ |Xb 1 |. In addition, for any x ∈ Xb 1 ,
there are at most k − 1 possibilities to replace x’s non-zero entry with zero and
(w )
(−w )
set x1 = 1, such that the result will be in the Xb 1 set. That is, Xb 1 ≤
(w1 )
(k − 1) Xb
+ 1, (where +1 stands for the vector of zeros), and we arrive at
(w1 )
Xb
(−w1 )
≤ Xb
(w1 )
≤ (k − 1) Xb
27
(w1 )
+ 1 ≤ k Xb
.
(16)
Combining (15) and (16), we complete the proof by noting that
(w )
(−w )
(−w )
Xb−w11
Xb−w1 1 + Xb−w11
|Xb−w1 |
≥
=
(w )
(−w )
(w )
(−w )
|Xb |
Xb 1 + Xb 1
Xb 1 + X b 1
(w1 )
=
|{z}
Xb
(w1 )
(15),(16) Xb
+k
(w )
Xb 1
=
1
.
k+1
28
| 10math.ST
|
Cloud-aided collaborative estimation by ADMM-RLS
algorithms for connected vehicle prognostics
Technical Report TR-2017-01
arXiv:1709.07972v1 [cs.SY] 22 Sep 2017
Valentina Breschi∗, Ilya Kolmanovsky†, Alberto Bemporad∗
September 26, 2017
Abstract
As the connectivity of consumer devices is rapidly growing and cloud
computing technologies are becoming more widespread, cloud-aided techniques for parameter estimation can be designed to exploit the theoretically unlimited storage memory and computational power of the “cloud”,
while relying on information provided by multiple sources.
With the ultimate goal of developing monitoring and diagnostic strategies, this report focuses on the design of a Recursive Least-Squares (RLS)
based estimator for identification over a group of devices connected to the
“cloud”. The proposed approach, that relies on Node-to-Cloud-to-Node
(N2C2N) transmissions, is designed so that: (i) estimates of the unknown
parameters are computed locally and (ii) the local estimates are refined
on the cloud. The proposed approach requires minimal changes to local
(pre-existing) RLS estimators.
1
Introduction
With the increasing connectivity between devices, the interest in distributed solutions for estimation [13], control [5] and machine learning [4] has been rapidly
growing. In particular, the problem of parameter estimation over networks has
been extensively studied, especially in the context of Wireless Sensor Networks
(WSNs). The methods designed to solve this identification problem can be divided into three groups: incremental approaches [10], diffusion approaches [3]
and consensus-based distributed strategies [11]. Due to the low communication power of the nodes in WSNs, research has mainly been devoted to obtain
fully distributed approaches, i.e. methods that allow exchanges of information
between neighbor nodes only. Even though such a choice enables to reduce
multi-hop transmissions and improve robustness to node failures, these strategies allows only neighbor nodes to communicate and thus to reach consensus.
The results in the report have been partially presented in a paper submitted to ACC 2018.
∗ Valentina Breschi and Alberto Bemporad are with the IMT School for Advanced Studies
Lucca, Piazza San Francesco 19, 55100 Lucca, Italy. [email protected];
[email protected]
† Ilya Kolmanovsky is with the Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI 48109, USA. [email protected]
1
Global Updates
CLOUD
Communication Layer
Broadcast
Global Updates
Collect information
······
······
Local Updates
Figure 1: Cloud-connected vehicles.
As a consequence, to attain consensus on the overall network, its topology has
to be chosen to enable exchanges of information between the different groups of
neighbor nodes.
At the same time, with recent advances in cloud computing [12] it has now
become possible to acquire and release resources with minimum effort so that
each node can have on-demand access to shared resources, theoretically characterized by unlimited storage space and computational power. This motivates to
reconsider the approach towards a more centralized strategy where some computations are performed at the node level, while the most time and memory
consuming ones are executed “on the cloud”. This requires the communication
between the nodes and a fusion center, i.e. the “cloud”, where the data gathered
from the nodes are properly merged.
Cloud computing has been considered for automotive vehicle applications in [7][8] and [14]. As motivating example for another possible automotive application,
consider a vehicle fleet with vehicles connected to the “cloud” (see Figure 1).
In such a setting, measurements taken on-board of the vehicles can be used for
cloud-based diagnostics and prognostics purposes. In particular, the measurements can be used to estimate parameters that may be common to all vehicles,
such as parameters in components wear models or fuel consumption models, and
parameters that may be specific to individual vehicles. References [15] and [6]
suggest potential applications of such approaches for prognostics of automotive
fuel pumps and brake pads. Specifically, the component wear rate as a function
of the workload (cumulative fuel flow or energy dissipated in the brakes) can be
common to all vehicles or at least to all vehicles in the same class.
A related distributed diagnostic technique has been proposed in [1]. However it
relies on a fully-distributed scheme, introduced to reduce long distance transmissions and to avoid the presence of a “critic” node in the network, i.e. a node
whose failure causes the entire diagnostic strategy to fail.
2
In this report a centralized approach for recursive estimation of parameters
in the least-squares sense is presented. The method has been designed under
the hypothesis of (i) ideal transmission, i.e. the information exchanged between
the cloud and the nodes is not corrupted by noise, and the assumption that (ii)
all the nodes are described by the same model, which is supposed to be known
a priori. Differently from what is done in many distributed estimation methods (e.g. see [11]), where the nodes estimate common unknown parameters, the
strategy we propose allows to account for more general consensus constraint. As
a consequence, for example, the method can be applied to problems where only
a subset of the unknowns is common to all the nodes, while other parameters
are purely local, i.e. they are different for each node.
Our estimation approach is based on defining a separable optimization problem
which is then solved through the Alternating Direction Method of Multipliers
(ADMM), similarly to what has been done in [11] but in a somewhat different
setting. As shown in [11], the use of ADMM leads to the introduction of two
time scales based on which the computations have to be performed. In particular, the local time scale is determined by the nodes’ clocks, while the cloud
time scale depends on the characteristics of the resources available in the center
of fusion and on the selected stopping criteria, used to terminate the ADMM
iterations.
The estimation problem is thus solved through a two-step strategy. In particular: (i) local estimates are recursively retrieved by each node using the measurements acquired from the sensors available locally; (ii) global computations are
performed to refine the local estimates, which are supposed to be transmitted to
the cloud by each node. Note that, based on the aforementioned characteristics,
back and forth transmissions to the cloud are required. A transmission scheme
referred to as Node-to-Cloud-to-Node (N2C2N) is thus employed.
The main features of the proposed strategies are: (i) the use of recursive formulas to update the local estimates of the unknown parameters; (ii) the possibility
to account for the presence of both purely local and global parameters, that can
be estimated in parallel; (iii) the straightforward integration of the proposed
techniques with pre-existing Recursive Least-Squares (RLS) estimators already
running on board of the nodes.
The report is organized as follows. In Section 2 ADMM is introduced, while
in Section 3 is devoted to the statement of the considered problem. The approach for collaborative estimation with full consensus is presented in Section 4,
along with the results of simulation examples that show the effectiveness of the
approach and its performance in different scenarios. In Section 5 and Section 6
the methods for collaborative estimation with partial consensus and for constrained collaborative estimation with partial consensus are described, respectively. Results of simulation examples are also reported. Concluding remarks
and directions for future research are summarized in Section 7.
1.1
Notation
Let Rn be the set of real vectors of dimension n and R+ be the set of positive
real number, excluding zero. Given a set A, let Ă be the complement of A.
Given a vector a ∈ Rn , kak2 is the Euclidean norm of a. Given a matrix
3
A ∈ Rn×p , A0 denotes the transpose of A. Given a set A, let PA denote the
Euclidean projection onto A. Let In be the identity matrix of size n and 0n be
an n-dimensional column vector of ones.
2
Alternating Direction Method of Multipliers
The Alternating Direction Method of Multipliers (ADMM) [2] is an algorithm
tailored for problems in the form
minimize
f (θ)+
subject to
Aθ + Bz = c,
(1)
where θ ∈ Rnθ , z ∈ Rnz , f : Rnθ → R ∪ {+∞} and g : Rnz → R ∪ {+∞} are
closed, proper, convex functions and A ∈ Rp×nθ , B ∈ Rp×nz , c ∈ Rp .
To solve Problem (1), the ADMM iterations to be performed are
θ(k+1) = argmin L(θ, z (k) , δ (k) ),
(2)
θ
z (k+1) = argmin L(θ(k+1) , z, δ (k) ),
δ (k+1) = δ
z
(k)
(3)
+ ρ(Aθ(k+1) + Bz (k+1) − c),
(4)
where k ∈ N indicates the ADMM iteration, L is the augmented Lagrangian
associated to (1), i.e.
L(θ, z, δ) = f (θ) + g(z) + δ 0 (Aθ + Bz − c) +
ρ
2
kAθ + Bz − ck2 ,
2
(5)
δ ∈ Rp is the Lagrange multiplier and ρ ∈ R+ is a tunable parameter (see [2]
for possible tuning strategies). Iterations (2)-(4) have to be run until a stopping
criteria is satisfied, e.g. the maximum number of iterations is attained.
It has to be remarked that the convergence of ADMM to high accuracy
results might be slow (see [2] and references therein). However, the results
obtained with a few tens of iterations are usually accurate enough for most of
applications. For further details, the reader is referred to [2].
2.1
ADMM for constrained convex optimization
Suppose that the problem to be addressed is
min
f (θ)
θ
s.t.
(6)
θ ∈ C,
with θ ∈ Rnθ , f : Rnθ → R ∪ {+∞} being a closed, proper, convex function and
C being a convex set, representing constraints on the parameter value.
As explained in [2], (6) can be recast in the same form as (1) through the
introduction of the auxiliary variable z ∈ Rnθ and the indicator function of set
C, i.e.
(
0
if z ∈ C
g(z) =
.
(7)
+∞ otherwise
4
In particular, (6) can be equivalently stated as
min f (θ) + g(z)
θ,z
(8)
θ − z = 0.
s.t.
Then, the ADMM scheme to solve (8) is
θ(k+1) = argmin L(θ, z (k) , δ (k) ),
(9)
θ
z (k+1) = PC (θ(k+1) + δ (k) ),
(10)
δ (k+1) = δ (k) + ρ(θ(k+1) − z (k+1) )
(11)
with L equal to
ρ
L(θ, z, δ) = f (θ) + g(z) + δ 0 (θ − z) + kθ − zk22
2
2.2
ADMM for consensus problems
Consider the optimization problem given by
min
g
θ
N
X
fn (θg ),
(12)
n=1
where θg ∈ Rnθ and each term of the objective, i.e. fn : Rnθ → R ∪ {+∞}, is a
proper, closed, convex function.
Suppose that N processors are available to solve (12) and that, consequently,
we are not interested in a centralized solution of the consensus problem. As
explained in [2], ADMM can be used to reformulate the problem so that each
term of the cost function in (12) is handled by its own.
In particular, (12) can be reformulated as
N
X
minimize
fn (θn )
(13)
n=1
θn − θ g = 0
subject to
n = 1, . . . , N.
Note that, thanks to the introduction of the consensus constraint, the cost
function in (13) is now separable.
The augmented Lagrangian correspondent to (13) is given by
g
N
L({θn }N
n=1 , θ , {δn }n=1 ) =
N
X
fn (θn ) + δn0 (θn − θg ) +
n=1
ρ
2
kθn − θg k2 , (14)
2
and the ADMM iterations are
θn(k+1) = argmin Ln (θn , δn(k) , θg,(k) ), n = 1, . . . , N
(15)
θn
N
1 X (k+1) 1 (k)
θn
+ δn
,
N n=1
ρ
= δn(k) + ρ θn(k+1) − θg,(k+1) , n = 1, . . . , N
θg,(k+1) =
(16)
δn(k+1)
(17)
5
with
ρ
Ln = fn (θn ) + (δn )0 (θn − θg ) + kθn − θg k22 .
2
Note that, on the one hand (15) and (17) can be carried out independently by
each agent n ∈ {1, . . . , N }, (16) depends on all the updated local estimates.
The global estimate should thus be updated in a “fusion center”, where all the
local estimates are collected and merged.
3
Problem statement
Assume that (i) measurements acquired by N agents are available and that
(ii) the behavior of the N data-generating systems is described by the same
known model. Suppose that some parameters of the model, θn ∈ Rnθ with
n = 1, . . . , N , are unknown and that their value has to to be retrieved from
data. As the agents share the same model, it is also legitimate to assume that
(iii) there exist a set of parameters θg ∈ Rng , with ng ≤ nθ , common to all the
agents.
We aim at (i) retrieving local estimates of {θn }N
n=1 , employing information
available at the local level only, and (ii) identifying the global parameter θg at
the “cloud” level, using the data collected from all the available sources. To
accomplish these tasks (i) N local processors and (ii) and a “cloud”, where the
data are merged are needed.
The considered estimation problem can be cast into a separable optimization
problem, given by
N
X
min
fn (θn )
θn
s.t.
n=1
F (θn ) = θg ,
(18)
θn ∈ Cn , n = 1, . . . , N
nθ
where fn : R → R∪{+∞} is a closed, proper, convex function, F : Rnθ → Rng
is a nonlinear operator and Cn ⊂ Rnθ is a convex set representing constraints
on the parameter values. Note that, constraints on the value of the global pa˘ with θ ∈ C.
rameter can be enforced if Cn = C ∪ {Cn ∩ C},
Assume that the available data are the output/regressor pairs collected from
each agent n ∈ {1, . . . , N } over an horizon of length T ∈ N, i.e. {yn (t), Xn (t)}Tt=1 .
Relying on the hypothesis that the regressor/output relationship is well modelled as
yn (t) = Xn (t)0 θn + en (t),
(19)
with en (t) ∈ Rny being a zero-mean additive noise independent of the regressor
Xn (t) ∈ Rnθ ×ny , we will focus on developing a recursive algorithm to solve (18)
with the local cost functions given by
T
fn (θn ) =
1 X T −t
2
λ
kyn (t) − Xn (t)0 θn k2 .
2 t=1 n
6
(20)
The forgetting factor λn ∈ (0, 1] is introduced to be able to estimate timevarying parameters. Note that different forgetting factors can be chosen for
different agents.
Remark 1 ARX models
Suppose that an AutoRegressive model with eXogenous inputs (ARX) has to be
identified from data. The input/output relationship is thus given by
y(t) = θ1 y(t − 1) + . . . + θna y(t − na )+
+ θna +1 u(t − nk − 1) + . . . + θna +nb u(t − nk − nb ) + e(t)
(21)
where u is the deterministic input, {na , nb } indicate the order of the system, nk
is the input/output delay.
Note that (21) can be recast as the output/regressor relationship with the regressor defined as
0
X(t) = y(t − 1)0 . . . y 0 (t − na ) u(t − nk − 1)0 . . . u(t − nk − nb )0
(22)
It is worth to point out that, in the considered framework, the parameters na , nb
and nk are the same for all the N agents, as they are supposed to be described
by the same model.
4
Collaborative estimation for full consensus
Suppose that the problem to be solve is (12), i.e. we are aiming at achieving
full consensus among N agents. Consequently, the consensus constraint in (18)
has to be modified as
F (θn ) = θg → θn = θg
and Cn = Rnθ , so that θn ∈ Cn can be neglected for n = 1, . . . , N . Moreover,
as we are focusing on the problem of collaborative least-squares estimation, we
are interested in the particular case in which the local cost functions in (13) are
equal to (20) .
Even if the considered problem can be solved in a centralized fashion, our
goal is to obtain estimates of the unknown parameters both (i) at a local level
and (ii) on the “cloud”. With the objective of distributing the computation
among the local processors and the “cloud”, we propose 5 approaches to address
(13).
4.1
Greedy approaches
All the proposed ‘greedy’ approaches rely on the use, by each local processor,
of the standard Recursive Least-Squares (RLS) method (see [9]) to update the
N
local estimates, {θ̂n }N
n=1 . Depending on the approach, {θ̂n }n=1 are then combined on the “cloud” to update the estimate of the global parameter.
The first two methods that are used to compute the estimates of the unknown
parameters both (i) locally and (ii) on the “cloud” are:
7
{θ̂1 (0), φ1 (0)}
#1
φ1 (0)
{θ̂N (0), φN (0)}
···
θ̂1 (t)
or
{θ̂1 (t), φ1 (t)}
···
#1
#N
θ̂N (t)
or
{θ̂N (t), φN (t)}
θ̂g
φN (0)
θ̂1 (t)
or
{θ̂1 (t), φ1 (t)}
CLOUD
θ̂og
θ̂g
#N
θ̂N (t)
or
{θ̂N (t), φN (t)}
θ̂g
CLOUD
θ̂g
(a) S-RLS and SW-RLS
(b) M-RLS and MW-RLS
Figure 2: Greedy approaches. Schematic of the information exchanges between
the agents and the “cloud”.
1. Static RLS (S-RLS) The estimate of the global parameter is computed
as
N
1 X
g
θ̂ =
θ̂n (t).
(23)
N n=1
2. Static Weighted RLS (SW-RLS) Consider the matrices {φn }N
n=1 , obtained applying standard RLS at each node (see [9]), and assume that
g
{φn }N
n=1 are always invertible. The estimate θ̂ is computed as the weighted
average of the local estimates
θ̂g =
N
X
!−1
φn (t)−1
n=1
N
X
!
φn (t)−1 θ̂n (t) .
(24)
n=1
Considering that φn is an indicator of the accuracy of the nth local estimate, (24) allows to weight more the “accurate ” estimates then the
“inaccurate” ones.
S-RLS and SW-RLS allow to achieve our goal, i.e. (i) obtain a local estimate of
the unknowns and (ii) compute θ̂g using all the information available. However,
looking at the scheme in Figure 2(a) and at Algorithm 1, it can be noticed that
the global estimate is not used at a local level.
Thanks to the dependence of θ̂g on all the available information, the local use
of the global estimate might enhance the accuracy of {θ̂n }N
n=1 . Motivated by
this observation, we introduce two additional methods:
4. Mixed RLS (M-RLS)
5. Mixed Weighted RLS (MW-RLS)
While M-RLS relies on (23), in MW-RLS the local estimates are combined as in
(24). However, as shown in Figure 2(b) and outlined in Algorithm 2, the global
estimate θ̂g is fed to the each local processor and used to update the local
estimates instead of their values at the previous step. Note that, especially at
8
Algorithm 1 S-RLS and SW-RLS
Input: Sequence of observations {Xn (t), yn (t)}Tt=1 , initial matrices φn (0) ∈
nθ ×nθ
, initial estimates θ̂n (0) ∈ Rnθ , n = 1, . . . , N
R
1. for t = 1, . . . , T do
Local
1.1. for n = 1, . . . , N do
1.1.1. compute Kn (t), φn (t) and θ̂n (t) with standard RLS [9];
1.2. end for;
Global
1.1. compute θ̂g ;
2. end.
Output: Local estimates {θ̂n (t)}Tt=1 , n = 1, . . . , N , estimated global parameters {θ̂g (t)}Tt=1 .
Algorithm 2 M-RLS and MW-RLS
Input: Sequence of observations {Xn (t), yn (t)}Tt=1 , initial matrices φn (0) ∈
nθ ×nθ
R
, n = 1, . . . , N , initial estimate θ̂og .
1. for t = 1, . . . , T do
Local
1.1. for n = 1, . . . , N do
1.1.1. set θ̂n (t − 1) = θ̂g (t − 1);
1.1.2. compute Kn (t), φn (t) and θ̂n (t) with standard RLS [9];
1.2. end for;
Global
1.1. compute θ̂g ;
2. end.
Output: Local estimates {θ̂n (t)}Tt=1 , n = 1, . . . , N , estimated global parameters {θ̂g (t)}Tt=1 .
the beginning of the estimation horizon, the approximation made in M-RLS
and MW-RLS might affect negatively some of the local estimates, e.g. the ones
obtained by the agents characterized by a relatively small level of noise.
Remark 2 While S-RLS and M-RLS require the local processors to transmit
N
to the “cloud” only {θ̂n }N
n=1 , the pairs {θ̂n , φn }n=1 have to be communicated to
the “cloud” with both SW-RLS and MW-RLS (see (23) and (24), respectively).
Moreover, as shown in Figure 2, while S-RLS and SW-RLS require Node-toCloud-to-Node (N2C2N) transmissions, M-RLS and MW-RLS are based on a
9
Node-to-Cloud (N2C) communication policy.
4.2
ADMM-based RLS (ADMM-RLS) for full consensus
Instead of resorting to greedy methods, we propose to solve (12) with ADMM.
Note that the same approach has been used to develop a fully distributed scheme
for consensus-based estimation over Wireless Sensor Networks (WSNs) in [11].
However, our approach differs from the one introduced in [11] as we aim at
exploiting the “cloud” to attain consensus and, at the same time, we want local
estimates to be computed by each node.
As the problem to be solved is equal to (13), the ADMM iterations to be performed are (15)-(17), i.e.
o
n
ρ
θ̂n (T )(k+1) = argmin fn (θn ) + (δn(k) )0 (θn − θ̂g,(k) ) + kθn − θ̂g,(k) k22 ,
2
θn
N
1 X (k+1) 1 (k)
θ
+ δn
,
θ̂g,(k+1) =
N n=1 n
ρ
δn(k+1) = δn(k) + ρ θ̂n(k+1) (T ) − θ̂g,(k+1) , n = 1, . . . , N
with the cost functions fn defined as in (14) and where the dependence on T
of the local estimates is stressed to underline that only the updates of θ̂n are
directly influenced by the current measurements. Note that the update for θ̂g
is a combination of the mean of the local estimates, i.e. (23), and the mean of
the Lagrange multipliers.
As (16)-(17) are independent from the specific choice of fn (θn ), we focus on
the update of the local estimates, i.e. (15), with the ultimate goal of finding
recursive updates for θ̂n .
Thanks to the characteristics of the chosen local cost functions, the closedform solution for the problem in (15) is given by
θ̂n(k+1) (T ) = φn (T ) Yn (T ) − δn(k) + ρθ̂g,(k) ,
(25)
Yn (t) =
t
X
λt−τ
n Xn (τ )yn (τ ), t = 1, . . . , T,
(26)
τ =1
φn (t) =
t
X
!−1
0
λt−τ
n Xn (τ )(Xn (τ ))
+ ρInθ
, t = 1, . . . , T.
(27)
τ =1
With the aim of obtaining recursive formulas to update θ̂n , consider the local
estimate obtained at T − 1, which is given by
θ̂n (T − 1) = φn (T − 1) Yn (T − 1) + ρθ̂g (T − 1) − δn (T − 1) ,
(28)
with δn (T −1) and θ̂g (T −1) denoting the Lagrange multiplier and the global es(k)
timate computed at T −1, respectively. It has then to be proven that θ̂n (T −1)
can be computed as a function of θ̂(T − 1), yn (T ) and Xn (T ).
10
Consider the inverse matrix φn (27), given by
φn (T )−1 = Xn (T ) + ρInθ ,
Xn (t) =
t
X
0
λt−τ
n Xn (τ )(Xn (τ )) .
τ =1
Based on (27), it can be proven that φn (T )−1 can be computed as a function of
φn (T − 1)−1 . In particular:
φn (T )−1 = Xn (T ) + ρInθ =
= λn Xn (T − 1) + Xn (T )(Xn (T ))0 + ρInθ =
= λn [Xn (T − 1) + ρInθ ] + Xn (T )(Xn (T ))0 + (1 − λn )ρInθ =
= λn φn (T − 1)−1 + Xn (T )(Xn (T ))0 + (1 − λn )ρInθ .
Introducing the extended regressor vector X̃n (T )
p
X̃n (T ) = Xn (T )
(1 − λn )ρInθ ∈ Rnθ ×(ny +nθ ) ,
(29)
(30)
(29) can then be further simplified as
φn (T )−1 = λn φn (T − 1)−1 + X̃n (T )(X̃n (T ))0 .
Applying the matrix inversion lemma, the resulting recursive formulas to update
φn are
Rn (T ) = λn I(ny +nθ ) + (X̃n (T ))0 φn (T − 1)X̃n (T ),
−1
Kn (T ) = φn (T − 1)X̃n (T )(Rn (T )) ,
0
φn (T ) = λ−1
I
−
K
(T
)(
X̃
(T
))
φn (T − 1).
n
n
n
n
θ
(31)
(32)
(33)
Note that the gain Kn and matrix φn are updated as in standard RLS (see [9]),
with the exceptions of the increased dimension of the identity matrix in (31)
and the substitution of the regressor with X̃n . Only when λn = 1 the regressor
Xn and X̃n are equal. Moreover, observe that (31)-(33) are independent from
k and, consequently, {Rn , Kn , φn }N
n=1 can be updated once fer step t.
Consider again (25). Adding and subtracting
h
i
λn φn (T ) ρθ̂g (T − 1) − δn (T − 1)
to (25), the solution of (15) corresponds to
h
θ̂n(k+1) (T ) = φn (T ) λn Yn (T − 1) − δn (T − 1) + ρθ̂g (T − 1) +
i
+Xn (T )yn (T ) − δn(k) − λn δn (T − 1) + ρ θ̂g,(k) − λn θ̂g (T − 1) =
= θ̂nRLS (T ) + θ̂nADM M,(k+1) (T ),
(34)
with
n
θ̂nRLS (T ) = φn (T ) λn Yn (T − 1) + ρθ̂g (T − 1) − δn (T − 1) +
+Xn (T )yn (T )} ,
i
h
(k+1)
(k+1)
= φn (T ) ρ∆g,λn (T ) − ∆λn (T ) ,
θ̂nADM M,(k+1) (T )
11
(35)
(36)
and
g,(k)
∆k+1
− λn θ̂g (T − 1),
g,λn (T ) = θ̂
(k+1)
∆λn (T )
= δn(k) − λn δn (T − 1).
(37)
(38)
Observe that (36) is independent from the past data-pairs {yn (t), Xn (t)}Tt=1 ,
while (35) depends on Yn (T − 1). Aiming at obtaining recursive formulas to
update θ̂n , the dependence of (35) should be eliminated.
Consider (35). Exploiting (33) and (28), θ̂nRLS (T ) is given by
n
o
θ̂nRLS (T ) = φn (T − 1) Yn (T − 1) + ρθ̂g (T − 1) − δn (T − 1) +
n
o
− Kn (T )(X̃n (T )0 )φn (T − 1) Yn (T − 1) + ρθ̂g (T − 1) − δn (T − 1) +
+ φn (T )Xn (T )yn (T ) =
= θ̂n (T − 1) − Kn (T )(X̃n (T ))0 θ̂n (T − 1) + φn (T )Xn (T )yn (T ).
(39)
For (39) to be dependent on the extended regressor only, we define the extended
measurement vector
0
ỹn (T ) = (yn (T ))0 01×ng .
The introduction of ỹn yields (39) can be modified as
θ̂nRLS (T ) = θ̂n (T − 1) − Kn (T )(X̃n (T ))0 θ̂n (T − 1) + φn (T )X̃n (T )ỹn (T ).
Notice that the equality φn (T )X̃n (T ) = Kn (T ) holds and it can be proven as
follows
0
φn (T )X̃n (T ) = λ−1
I
−
K
(T
)(
X̃
(T
))
φn (T − 1)X̃n (T ) =
n
n
n
n
θ
= λ−1
Inθ − φn (T − 1)X̃n (T )(Rn (T ))−1 (X̃n (T ))0 φn (T − 1)X̃n (T ) =
n
−1
−1
0
= φn (T − 1)X̃n (T ) λ−1
I
−
λ
(R
(T
))
(
X̃
(T
))
φ
(T
−
1)
X̃
(T
)
=
n
n
n
n
n
n
n
θ
= φn (T − 1)X̃n (T ) λ−1
n Inθ +
0
−1
−λ−1
(X̃n (T ))0 φn (T − 1)X̃n (T ) =
n (λn I(ny +nθ ) + (X̃n (T )) φn (T − 1)X̃n (T ))
−1
−1
0
−1
= φn (T − 1)X̃n (T ) λ−1
n Inθ − λn (I(ny +nθ ) + λn (X̃n (T )) φn (T − 1)X̃n (T )) ·
·(X̃n (T ))0 φn (T − 1)X̃n (T )λn−1 =
−1
= φn (T − 1)X̃n (T ) λn Inθ + (X̃n (T ))0 φn (T − 1)X̃n (T )
= Kn (T ),
where the matrix inversion lemma and (32)-(33) are used.
It can thus be proven that θ̂nRLS can be updated as
θ̂nRLS (T ) = θ̂n (T − 1) + Kn (T )(ỹn (T ) − X̃n (T )0 θ̂n (T − 1)).
(40)
While the update for θ̂nADM M (36) depends on both the values of the Lagrange multipliers and the global estimates, θ̂nRLS (40) is computed on the basis
12
φ1 (0)
···
#1
θ̂1 (t)
φN (0)
#N
{θ̂1RLS (t), φ1 (t)}
RLS
{θ̂N
(t), φN (t)}
CLOUD
θ̂N (t)
{θ̂nADMM }N
n=1
θ̂g , δn , {θ̂n (t)}N
n=1
Figure 3: ADMM-RLS. Schematic of the information exchanges between the
agents and the “cloud”when using a N2C2N communication scheme.
of the previous local estimate and the current measurements. Consequently,
θ̂nRLS is updated recursively.
Under the hypothesis that both θ̂g and δn are stored on the “cloud”, it does
seems legitimate to update θ̂g and δn on the “cloud”, along with θ̂nADM M . Instead, the partial estimates θ̂nRLS , n = 1, . . . , N , can be updated by the local
processors. Thanks to this choice, the proposed method, summarized in Algorithm 3 and Figure 3, allows to obtain estimates both at the (i) agent and (ii)
“cloud” level.
Observe that, thanks to the independence of (40) from k, θ̂nRLS can be updated
once per step t. The local updates are thus regulated by a local clock and not
by the one controlling the ADMM iterations on the “cloud”.
Looking at (31)-(33) and (40), it can be noticed that θ̂nRLS is updated through
standard RLS, with the exceptions that, at step t ∈ {1, . . . , T }, the update depends on the previous local estimate θ̂n (t−1) instead of depending on θ̂nRLS (t−1)
and that the output/regressor pair {yn (t), Xn (t)} is replaced with {ỹn (t), X̃n (t)}.
As a consequence, the proposed method can be easily integrated with preexisting RLS estimators already available locally.
Remark 3 Algorithm 1 requires the initialization of the local and global estimates. If some data are available to be processed in a batch mode, θ̂n (0) can be
chosen as the best linear model, i.e.
θ̂n (0) = argmin
θn
τ
X
kyn (t) − Xn (t)0 θk22
t=1
and θ̂g (0) can be computed as the mean of {P θ̂n (0)}N
n=1 . Moreover, the matrices
φn , n = 1, . . . , N , can be initialized as φn (0) = γInθ , with γ > 0.
Remark 4 The chosen implementation requires θ̂nRLS and φn to be transmitted
from the local processors to the “cloud” at each step, while the “cloud”has to
communicate θ̂n to all the agents. As a consequence, the proposed approach is
based on N2C2N transmissions.
13
Algorithm 3 ADMM-RLS for full consensus (N2C2N)
Input: Sequence of observations {Xn (t), yn (t)}Tt=1 , initial matrices φn (0) ∈
Rnθ ×nθ , initial local estimates θ̂n (0), initial dual variables δn,o , n = 1, . . . , N ,
initial global estimate θ̂og , parameter ρ ∈ R+ .
1. for t = 1, . . . , T do
Local
1.1. for n = 1, . . . , N do
1.1.1. compute X̃n (t) as in (30);
1.1.2. compute Kn (t) and φn (t) with (32) - (33);
1.1.3. compute θ̂nRLS (t) with (40);
1.2. end for;
Global
1.1. do
ADM M,(k+1)
(t) with (36), n = 1, . . . , N ;
1.1.1. compute θ̂n
1.1.2. compute θ̂g,(k+1) (t) with (16);
(k+1)
with (17), n = 1, . . . , N ;
1.1.3. compute δn
1.2. until a stopping criteria is satisfied (e.g. maximum number of
iterations attained);
2. end.
Output: Estimated global parameters {θ̂g (t)}Tt=1 , estimated local parameters
{θ̂n (t)}Tt=1 , n = 1, . . . , N .
4.3
Example 1. Static parameters
Suppose that N data-generating systems are described by the following models
yn (t) = 0.9yn (t − 1) + 0.4un (t − 1) + en (t),
(41)
0
where yn (t) ∈ R, Xn (t) = [ yn (t−1) un (t−1) ] , un is known and is generated in this
example as a sequence of i.i.d. elements uniformly distributed in the interval
[ 2 3 ] and en ∼ N (0, Rn ) is a white noise sequence, with {Rn ∈ N}N
n=1 randomly
chosen in the interval [ 1 30 ]. Evaluating the effect of the noise on the output yn
through the Signal-to-Noise Ratio SN Rn , i.e.
PT
2
(yn (t) − en (t))
dB
(42)
SNRn = 10 log t=1PT
2
t=1 en (t)
the chosen covariance matrices yield SNRn ∈ [7.8 20.8] dB, n = 1, . . . , N . Note
that (41) can be equivalently written as
0
yn (t) = (Xn (t))0 θg + en (t) with θg = [ 0.9 0.4 ]
and the regressor Xn (t) is defined as in (22), i.e. Xn = [ yn (t−1) un (t−1) ].
14
1.2
1.2
1
1
0.8
0.8
0.6
0.4
0.6
0.2
0
200
400
600
800
1000
(a) θ1g vs θ̂1g
0
200
400
600
800
1000
(b) θ2g vs θ̂2g
Figure 4: Example 1. True vs estimated parameters. Black : true, red : C-RLS,
blue : S-RLS, cyan : SW-RLS, magenta : M-RLS, green : MW-RLS.
Observe that the deterministic input sequences {un (t)}Tt=1 are all different.
However, they are all generated accordingly to the same distribution, as it seems
reasonable to assume that systems described by the same model are characterized by similar inputs.
Initializing φn as φn (0) = 0.1Inθ , while θ̂n (0) and θ̂og are sampled from the
distributions N (θ̂g , 2Inθ ) and N (θ̂g , Inθ ), respectively, and {λn = Λ}N
n=1 , with
Λ = 1, we first evaluate the performance of the greedy approaches. The actual
parameter θg and the estimate obtained with the different greedy approaches
are reported in Figure 4.
Despite the slight difference performances in the first 300 steps, which seems
to be legitimate, the estimates obtained with SW-RLS, M-RLS and MW-RLS
are similar. Moreover, θ̂g obtained with the different methods are comparable
with respect with the estimate computed with C-RLS.
In particular, the similarities between the estimates obtained with M-RLS, MWRLS and C-RLS prove that, in the considered case, the choice of the “mixed”
strategy allows to enhance the accuracy of θ̂g . Comparing the estimates obtained with S-RLS and SW-RLS, observe that the convergence of the estimate
to the actual value of θg tends to be faster if θ̂g is computed as in (24).
Setting ρ = 0.1, the performance of the ADMM-RLS are assessed for different values of N and T . Moreover, the retrieved estimates are compared to the
ones obtained with C-RLS and the greedy approaches.
The accuracy of the estimate θ̂g is assessed through the Root Mean Square Error
(RMSE), i.e.
v
2
uP
u T
θig − θ̂ig (t)
t
t=1
, i = 1, . . . , ng .
(43)
RMSEgi =
T
As expected (see Table 1), the accuracy of the estimates tends to increase if
the number of local processors N and the estimation horizon T increase. In the
case N = 100 and T = 1000, the estimates obtained with both C-RLS and the
SW-RLS and MW-RLS have comparable accuracy. See Table 2.
The estimates obtained with ADMM-RLS, C-RLS and MW-RLS are further
compared in Figure 5 and, as expected the three estimates are barely distinguishable. Thus the proposed ADMM-RLS algorithm, which uses local esti-
15
Table 1: ADMM-RLS: kRMSEg k2
T
N
2
10
102
10
102
103
104
1.07
0.55
0.39
0.33
0.22
0.11
0.16
0.09
0.03
0.10
0.03
0.01
Table 2: kRMSEg k2 : C-RLS and greedy methods vs ADMM-RLS
kRMSEg k2
C-RLS
0.03
S-RLS
0.05
SW-RLS
0.03
1.2
Method
M-RLS
0.04
MW-RLS
0.03
ADMM-RLS
0.03
0.2
1.1
0.1
1
0.9
0
0.8
0.7
-0.1
0
200
400
600
800
1000
0
200
(a) θ1g vs θ̂1g
400
600
800
1000
800
1000
(b) |θ̂1g − θ1g |
1
0.6
0.8
0.4
0.6
0.2
0.4
0
0.2
0
200
400
600
800
1000
(c) θ2g vs θ̂2g
0
200
400
600
(d) |θ̂2g − θ2g |
Figure 5: Example 1. Model parameters. Black : true, red : C-RLS, green :
MW-RLS, blue : ADMM-RLS.
mates and the cloud, is able to obtain good accuracy versus the fully centralized
approach. Moreover, ADMM-RLS allows to retrieve estimates as accurate as
the ones obtained with the MW-RLS, i.e. the greedy approach associated with
the least RMSE.
4.3.1
Non-informative agents
Using the previously introduced initial setting and parameters, lets assume that
some of the available data sources are non-informative, i.e. some systems are
not excited enough to be able to retrieve locally an accurate estimate of all
the unknown parameters [9]. Null input sequences and white noise sequences
characterized by Rn = 10−8 are used to simulate the behavior of the Nni ≤ N
non-informative agents.
16
Table 3: Example 1. kRMSEg k2 vs Nni
kRMSEg k2
Nni
10
20
0.02
0.02
1
0.02
50
0.03
Table 4: Example 1. kRMSEg k2 : 20% of non-informative agents
kRMSEg k2
C-RLS
0.02
S-RLS
0.03
SW-RLS
0.02
Method
M-RLS
0.07
MW-RLS
0.03
ADMM-RLS
0.02
1.5
1.6
1.4
1
1.2
1
0.5
0.8
0
1000
2000
(a)
θ1g
3000
vs
4000
5000
θ̂1g
0
1000
2000
(b)
θ2g
3000
vs
4000
5000
θ̂2g
Figure 6: Example 1. Model parameters vs Nni . Black : true, red : Nni = 1,
blue : Nni = 10, cyan : Nni = 20, magenta : Nni = 50.
Consider the case N = 100 and T = 5000. The performance of ADMM-RLS
are studied under the hypothesis that an increasing number Nni of systems is
non-informative. Looking at the RMSEs in Table 3 and the estimates reported
in Figure 6, it can be noticed that the quality of the estimate starts to decrease
only when half of the available systems are non-informative. In case of Nni = 20,
the estimates obtained with ADMM-RLS are then compared with the ones
computed with C-RLS and the greedy approaches. As it can be noticed from
the RMSEs reported in Table 4, in presence of non-informative agents SW-RLS
tends to perform better than the other greedy approaches and the accuracy of
the estimates obtained with C-RLS, SW-RLS and ADMM-RLS are comparable.
4.3.2
Agents failure
Consider again N = 100 and T = 5000 and suppose that, due to a change in the
behavior of Nf local agents the parameters of their models suddenly assume different values with respect to [ 0.9 0.4 ]. We study the performance of ADMM-RLS
under the hypothesis that the change in the value of the parameters happens
at an unknown instant tn , randomly chosen in the interval [1875, 3750] samples,
and simulating the change in the local parameters using θn,1 sampled from the
distribution U[ 0.2 0.21 ] and θn,2 sampled from U[ 1.4 1.43 ] after tn .
Observe that it might be beneficial to use a non-unitary forgetting factor, due
to the change in the local parameters. Consequently, λn , n = 1, . . . , N , is set to
0.99 for all the N agents.
The performance of ADMM-RLS are initially assessed considering an increasing
number of systems subject to failure. See Table 5 and Figure 7. Observe that
the failure of the agents seems not to influence the accuracy of the obtained
17
Table 5: Example 1. ADMM-RLS: kRMSEg k2 vs Nf
Nf
kRMSEg k2
1
0.03
10
0.03
20
0.03
50
0.04
1.8
0.8
1.6
0.6
1.4
1.2
0.4
1
0.2
0
0.8
0
1000
2000
3000
4000
5000
0
1000
(a) θ1g vs θ̂1g
2000
3000
4000
5000
4000
5000
(b) |θ̂1g − θ1g |
1.5
1
0.8
1
0.6
0.4
0.2
0.5
0
0
1000
2000
3000
4000
5000
(c) θ2g vs θ̂2g
0
1000
2000
3000
(d) |θ̂2g − θ2g |
Figure 7: Example 1. Model parameters vs Nf . Black : true, red : Nf = 1,
blue : Nf = 10, cyan : Nf = 20, magenta : Nf = 50.
estimates if Nf 6= 50. The use of ADMM-RLS thus allows to compute accurate
global estimates even when some of the agent experience a failure.
4.4
Example 2. Time-varying parameters
The presence of the forgetting factor in the cost functions fn (see (20)) allows
to estimate time-varying parameters, as it enables to weight differently past and
currently collected data.
Suppose that the behavior of N systems is described by the ARX model
yn (t + 1) = θ1g (t)yn (t − 1) + θ2g (t)un (t − 1) + en (t)
(44)
where θ1g = 0.9 sin (x) and θ2g = 0.4 cos (x), with x ∈ [0, 2π], and un ∼ U[ 2 3 ] .
The white noise sequences en ∼ N (0, Rn ), n = 1, . . . , N , have covariances Rn
randomly selected in the interval [ 1 30 ] yielding to SN Rn ∈ [ 2.4 6.5 ] dB.
Considering an estimation horizon T = 1000, imposing φn as φn (0) = 0.1Inθ ,
while θ̂n (0) and θ̂og are sampled from the distributions N (θ̂g , 2Inθ ) and N (θ̂g , Inθ ),
respectively, ρ = 0.1 and setting {λn = Λ}N
n=1 , with Λ = 0.95, the performances
of ADMM-RLS are compared with the ones of C-RLS and the four greedy approaches. See Table 6. As for the case where time-invariant parameters have to
be estimated (see Example 1), SW-RLS and MW-RLS tend to perform slightly
better than the other greedy approaches. Note that the accuracy of the estimates C-RLS, SW-RLS and MW-RLS is comparable.
Figure 8 reports the actual global parameters and the estimates obtained
18
Table 6: Example 2. kRMSEg k2 vs Method
kRMSEg k2
C-RLS
0.08
S-RLS
0.10
SW-RLS
0.08
Method
M-RLS
0.09
MW-RLS
0.08
ADMM-RLS
0.08
1
0.2
0.5
0.1
0
0
-0.5
-1
-0.1
0
200
400
600
800
1000
0
200
(a) θ1g vs θ̂1g
400
600
800
1000
800
1000
(b) |θ̂1g − θ1g |
1
0.6
0.5
0.4
0
0.2
0
-0.5
0
200
400
600
800
1000
0
200
(c) θ2g vs θ̂2g
400
600
(d) |θ̂2g − θ2g |
Figure 8: Example 2. True vs estimated model parameters. Black : true, red :
C-RLS, blue : ADMM-RLS.
with C-RLS and ADMM-RLS, along with the estimation errors. As already observed, the accuracy of the estimates computed with C-RLS and ADMM-RLS
is comparable.
5
Collaborative estimation for partial consensus
Consider the more general hypothesis that there exist a parameter vector θg ∈
Rng , with ng ≤ nθ such that:
P θn = θg ∀n ∈ {1, . . . , N },
(45)
where P ∈ Rng ×nθ is a matrix assumed to be known a priori. The problem that
we want to solve is then given by
min
{θn }N
n=1
s.t.
N
X
fn (θn )
n=1
(46)
g
P θn = θ , n = 1, . . . , N,
with fn defined as in (20). Note that (46) corresponds to (18) with the consensus
constraint modified as
F (θn ) = θg → P θn = θg .
The considered consensus constraint allows to enforce consensus over a linear
combination of the components of θn . Note that, through proper choices of P ,
19
different settings can be considered, e.g. if P = Inθ then θn = θg and thus (46)
is equal to (12). We can also enforce consensus only over some components of
θn , so that some of the unknowns are assumed to be global while others are
supposed to assume a different value for each agent.
g
As we are interested in obtaining an estimate for both {θn }N
n=1 and θ , note
that (46) cannot be solved resorting to a strategy similar to C-RLS (see Appendix A). In particular, even if properly modified, a method as C-RLS would
allow to compute an estimate for the global parameter only.
The ADMM iterations to solve problem (46) are given by
θ̂n(k+1) (T ) = argmin L(θn , θ̂g,(k) , δn(k) ),
(47)
g
(k) N
θ̂g,(k+1) = argmin L({θ̂n(k+1) (T )}N
n=1 , θ , {δn }n=1 ),
(48)
δn(k+1) = δn(k) + ρ(P θ̂n(k+1) (T ) − θ̂g,(k+1) ),
(49)
θn
θg
with k ∈ N indicating the ADMM iteration, ρ ∈ R+ being a tunable parameter,
δn ∈ Rng representing the Lagrange multiplier and the augmented Lagrangian
L given by
L=
N n
o
X
ρ
2
fn (θn ) + δn0 (P θn − θg ) + kP θn − θg k2 .
2
n=1
(50)
Note that the dependence on T is explicitly indicated only for the local estimates
θ̂n , as they are the only quantities directly affected by the measurement and the
regressor at T .
Consider the update of the estimate θ̂g . The closed form solution for (48) is
θ̂g,(k+1) =
N
1
1 X
P θ̂n(k+1) (T ) + δn(k) .
N n=1
ρ
(51)
The estimate of the global parameter is thus updated through the combination
(k+1)
of the mean of {δn }N
(T )}N
n=1 and the mean of {P θ̂n
n=1 . As expected, (51)
resembles (16), where the local estimates are replaced by a linear combination
of their components.
Consider the update for the estimate of the local parameters. The close form
solution for (47) is given by:
n
o
θ̂n(k+1) (T ) = φn (T ) Yn (T ) + P 0 (ρθ̂g,(k) − δn(k) ) ,
(52)
Yn (t) =
t
X
λt−τ
n Xn (τ )yn (τ ), t = 1, . . . , T,
(53)
τ =1
"
φn (t) =
t
X
#
0
λt−τ
+ ρP 0 P
n Xn (τ )Xn (τ )
!−1
, t = 1, . . . , T.
(54)
τ =1
As also in this case we are interested in obtaining recursive formulas for the
local updates, consider θ̂n (T − 1), defined as
θ̂n (T − 1) = φn (T − 1) Yn (T − 1) + P 0 (ρθ̂g (T − 1) − δn (T − 1)) ,
(55)
20
where φn (T − 1) is equal to (54), and θ̂g (T − 1) and δn (T − 1) are the global
estimate and the Lagrange multiplier obtained at T − 1, respectively.
Observe that the following equalities hold
−1
φn (T ) = (Xn (T ) + ρP 0 P )
=
−1
= (λn Xn (T − 1) + Xn (T )Xn (T )0 + ρP 0 P )
0
=
−1
= (λn (Xn (T − 1) + ρP P ) + Xn (T )Xn (T ) + ρ(1 − λn )P 0 P )
−1
,
= λn φn (T − 1)−1 + Xn (T )Xn (T )0 + ρ(1 − λn )P 0 P
with
Xn (t) =
t
X
0
=
0
λt−τ
n Xn (τ )(Xn (τ )) , t = 1, . . . , T.
τ =1
Introducing the extended regressor
p
X̃n (T ) = Xn (T )
ρ(1 − λn )P 0 ∈ Rnθ ×(ny +ng )
(56)
and applying the matrix inversion lemma, it can be proven that φn can be
updated as
Rn (T ) = λn I(ny +ng ) + (X̃n (T ))0 φn (T − 1)X̃n (T ),
Kn (T ) = φn (T − 1)X̃n (T ) (Rn (T ))
φn (T ) =
λ−1
n (Inθ
−1
,
0
− Kn (T )(X̃n (T )) )φn (T − 1).
(57)
(58)
(59)
Note that (57)-(59) are similar to (31)-(33), with differences due to the new
definition of the extended regressor.
Consider again (52). Adding and subtracting
λn φn (T )P 0 ρθ̂g (T − 1) − δn (T − 1)
(k+1)
can be computed as
h
θ̂n(k+1) (T ) = φn (T ) λn Yn (T − 1) + P 0 (ρθ̂g (T − 1) − δn (T − 1) )+
i
+Xn (T )yn (T ) − P 0 δn(k) − λn δn (T − 1) + P 0 ρ θ̂g,(k) − λn θ̂g (T − 1) =
to (52), θ̂n
= θ̂nRLS (T ) + θ̂nADM M,(k+1) (T ).
(60)
In particular,
n
o
θ̂nRLS (T ) = φn (T )λn Yn (T − 1) + ρP 0 θ̂(T − 1) − P 0 δn (T − 1) +
+ φn (T )X̃n (T )yn (T ),
and
(k+1)
(k+1)
θ̂nADM M,(k+1) (T ) = φn (T )P 0 ρ∆g,λn (T ) − ∆λn
,
with
g,(k)
∆k+1
− λn θ̂g (T − 1),
g,λn (T ) = θ̂
(k+1)
∆λn
(T ) = δn(k) − λn δn (T − 1).
21
(61)
(62)
Observe that, as for (16) and (51), (62) differs from (36) because of the presence
of P .
Note that, accounting for the definition of φn (T − 1), exploiting the equality Kn (T ) = φn (T )X̃n (T ) (see Section 4 for the proof) and introducing the
extended measurement vector
0
ỹn (T ) = yn (T )0 O1×ng ,
the formula to update θ̂nRLS in (62) can be further simplified as
n
o
θ̂nRLS (T ) = φn (T − 1) Yn (T − 1) + P 0 (ρθ̂g (T − 1) − δn (T − 1)) +
n
o
− Kn (T )(X̃n (T )0 )φn (T − 1) Yn (T − 1) + P 0 (ρθ̂g (T − 1) − δn (T − 1)) +
+ φn (T )Xn (T )yn (T ) =
= θ̂n (T − 1) − Kn (T )(X̃n (T ))0 θ̂n (T − 1) + φn (T )X̃n (T )ỹn (T ) =
= θ̂n (T − 1) + Kn (T )(ỹn (T ) − (X̃n (T ))0 θ̂n (T − 1)).
(63)
As the method tailored to attain full consensus (see Section 4), note that both
θ̂g and δn should be updated on the “cloud”. As a consequence, also θ̂nADM M
should be updated on the “cloud”, due to its dependence on both θ̂g and δn . On
the other hand, θ̂nRLS can be updated by the local processors. As for the case
considered in Section 4, note that (63) is independent from k and, consequently,
the synchronization between the local clock and the one on the “cloud”is not
required.
The approach is outlined in Algorithm 4 and the transmissions characterizing each iteration is still the one reported in the scheme in Figure 3. As a
consequence, the observations made in Section 4 with respect to the information exchange between the nodes and the “cloud” hold also in this case.
5.1
Example 3
Assume to collect data for T = 1000 from a set of N = 100 dynamical systems
modelled as
yn (t) = θ1g yn (t − 1) + θn,2 yn (t − 2) + θ2g un (t − 1) + en (t),
0
(64)
where θg = [ 0.2 0.8 ] and θn,2 is sampled from a normal distribution N (0.4, 0.0025),
so that it is different for the N systems. The white noise sequence en ∼
N (0, Rn ), where, for the “informative’ systems, Rn ∈ [1 20] yields SN R ∈
[3.1, 14.6] dB (see (42)).
Initializing φn as φn (0) = 0.1Inθ , while θ̂n (0) and θ̂og are sampled from the distributions N (θ̂g , 2Inθ ) and N (θ̂g , Inθ ), respectively, {λn = Λ}N
n=1 , with Λ = 1,
and ρ = 0.1, the performance of the proposed approach are evaluated. Figure 9
shows θ̂g obtained with ADMM-RLS, along with the estimation error. Observe
that the estimates tends to converge to the actual value of the global parameters.
To further assess the performances of ADMM-RLS, θn , θ̂n and θ̂nRLS obtained
for the 5th system, i.e. n = 5, are compared in Figure 10. It can thus be seen
22
Algorithm 4 ADMM-RLS for partial consensus (N2C2N)
Input: Sequence of observations {Xn (t), yn (t)}Tt=1 , initial matrices φn (0) ∈
Rnθ ×nθ , initial local estimates θ̂n (0), initial dual variables δn,o , forgetting factors
λn , n = 1, . . . , N , initial global estimate θ̂og , parameter ρ ∈ R+ .
1. for t = 1, . . . , T do
Local
1.1. for n = 1, . . . , N do
1.1.1. compute X̃n (t) with (56);
1.1.2. compute Kn (t) and φn (t) with (58) - (59);
1.1.3. compute θ̂nRLS (t) with (63);
1.2. end for;
Global
1.1. do
ADM M,(k+1)
(t) with (62), n = 1, . . . , N ;
1.1.1. compute θ̂n
(k+1)
1.1.2. compute θ̂n
1.1.3. compute θ̂
(t) with (60), n = 1, . . . , N ;
g,(k+1)
with (51);
(k+1)
δn
with (49), n = 1, . . . , N ;
1.1.4. compute
1.2. until a stopping criteria is satisfied (e.g. maximum number of
iterations attained);
2. end.
Output: Estimated global parameters {θ̂g (t)}Tt=1 , estimated local parameters
{θ̂n (t)}Tt=1 , n = 1, . . . , N .
that the difference between θ̂nRLS and θ̂n is mainly noticeable at the beginning of
the estimation horizon, but then θ̂nRLS and θ̂n are barely distinguishable. Note
that SNR5 = 8.9 dB.
5.1.1
Non-informative agents
Suppose that among the N = 100 systems described by the model in (64),
Nni = 20 randomly chosen agents are non-informative, i.e. their input sequences un are null and Rn = 10−8 .
As it can be observed from the estimates reported in Figure 11, {θ̂ig }2i=1 converge to the actual values of the global parameters even if 20% of the systems
provide non-informative data.
The local estimates θ̂n,2 for the 8th and 65th system (SN R65 ≈ 6 dB) are
reported in Figure 12. As, the 8th system is among the ones with a non exciting input, θ̂8,2 = θ̂8,2 (0) over the estimation horizon. Instead, θ̂65,2 tends
to converge to the actual value of θ65,2 . Even if the purely local parameter is
not retrieved from the data, using the proposed collaborative approach θ8,1 and
θ8,3 are accurately estimated (see Figure 13). We can thus conclude that the
proposed estimation method “forces” the estimates of the global components of
23
0.4
0.3
0.2
0.2
0.1
0
0
-0.2
-0.1
0
200
400
(a)
θ1g
600
vs
800
1000
0
200
θ̂1g
400
(b)
1.4
0.6
1.2
0.4
|θ̂1g
600
−
800
1000
θ1g |
1
0.2
0.8
0
0.6
0
200
400
600
800
1000
0
200
(c) θ2g vs θ̂2g
400
600
800
1000
(d) |θ̂2g − θ2g |
Figure 9: Example 3. True vs estimated global parameters. Black : true, blue
: ADMM-RLS.
0.5
0.5
0
0
-0.5
-0.5
-1.5
-1
-2
0
200
400
600
800
1000
0
RLS
(a) θ5,1 vs θ̂5,1 and θ̂5,1
200
400
600
800
1000
RLS
(b) θ5,2 vs θ̂5,2 and θ̂5,2
2.5
2
1.5
1
0
200
400
600
800
1000
RLS
(c) θ5,3 vs θ̂5,3 and θ̂5,3
Figure 10: Example 3. Local parameter θn,2 , n = 5. Black : true, blue : θ̂5 ,
red: θ̂5RLS .
θn to follow θ̂g , which is estimated automatically discarding the contributions
from the systems that lacked excitation.
24
1.4
0.4
1.2
0.2
1
0
0.8
-0.2
0.6
0
200
400
600
800
1000
0
200
(a) θ1g vs θ̂1g
400
600
800
1000
(b) θ2g vs θ̂2g
Figure 11: Example 3. True vs estimated global parameters. Black : true, blue
: ADMM-RLS.
1
0.4
0
0.2
-1
0
-2
-3
-0.2
-4
-0.4
-5
0
200
400
600
800
0
1000
200
(a) θ8,2 vs θ̂8,2
400
600
800
1000
(b) θ65,2 vs θ̂65,2
Figure 12: Example 3. Local parameters θn,2 , n = 8, 65. Black : true, blue :
ADMM-RLS.
0.35
1.6
0.3
1.4
0.25
1.2
1
0.2
0.8
0.15
0.6
0
200
400
600
800
1000
0
RLS
(a) θ8,1 vs θ̂8,1 and θ̂8,1
200
400
600
800
1000
RLS
(b) θ8,3 vs θ̂8,3 and θ̂8,3
Figure 13: Example 3. Local parameters θ8,i , i = 1, 3. Black : true, blue :
RLS
θ̂8,i
, red: θ̂8,i .
6
Constrained Collaborative estimation for partial consensus
Suppose that the value of the local parameter θn is constrained to a set Cn and
that this hypothesis holds for all the agents n ∈ {1, . . . , N }. With the objective
of reaching partial consensus among the agents, the problem to be solved can
thus be formulated as
minimize
N
X
fn (θn )
n=1
s.t.
(65)
P θn = θ, n = 1, . . . , N,
θn ∈ Cn , n = 1, . . . , N.
25
Observe that (65) corresponds to (18) if the nonlinear consensus constraint is
replaced with (45).
To use ADMM to solve (65), the problem has to be modified as
minimize
N
X
{fn (θn ) + gn (zn )}
n=1
P θn = θg n = 1, . . . , N
s.t.
(66)
θn = zn , n = 1, . . . , N
N
where {gn }N
n=1 are the indicator functions of the sets {Cn }n=1 (defined as in (7))
nθ N
and {zn ∈ R }n=1 are auxiliary variables. Observe that (66) can be solved with
ADMM. Given the augmented Lagrangian associated with (66), i.e.
L=
N
X
0
0
{fn (θn ) + gn (zn ) + δn,1
(θn − zn ) + δn,2
(P θn − θg )+
n=1
+
ρ2
ρ1
kθn − zn k22 + kP θn − θg k22 },
2
2
(67)
the iterations that have to be performed to solve the addressed problem with
ADMM are
θ̂n(k+1) (T ) = argmin L(θn , θ̂g,(k) , zn(k) , δn(k) ),
(68)
zn(k+1) = argmin L(θ̂n,(k+1) (T ), θ̂g,(k) , zn , δn(k) ),
(69)
g
(k+1) (k) N
θ̂g,(k+1) = argmin L({θ̂n(k+1) }N
, δn }n=1 ),
n=1 , θ , {zn
(70)
θn
zn
θg
(k+1)
= δn,1 + ρ1 (θ̂n(k+1) (T ) − z (k+1) ),
(k+1)
= δn,2 + ρ2 (P θ̂n(k+1) (T ) − θ̂g,(k+1) ).
δn,1
δn,2
(k)
(71)
(k)
(72)
N
Note that two sets of Lagrangian multipliers, {δn,1 }N
n=1 and {δn,2 }n=1 , have
ng
been introduced. While δn,1 ∈ R is associated with the partial consensus
constraint, δn,2 ∈ Rnθ is related to the constraint θn ∈ Cn , n = 1, . . . , N .
Solving (69)-(70), the resulting updates for the auxiliary variables and the
global estimates are
1 (k)
(k+1)
(k+1)
(73)
zn
=PCn θ̂n
(T ) + δn,1 , n = 1, . . . , N,
ρ1
N
1 X
1 (k)
θ̂g,(k+1) =
P θ̂n(k+1) (T ) + δn,2 .
(74)
N n=1
ρ2
Observe that z-update is performed projecting onto the set Cn a combination of
(k)
the updated local estimate and δn,1 , while θ̂g,(k+1) is computed as in Section 5,
with δn replaced by δn,2 .
26
Consider the close form solution of (68), which is given by
o
n
(k)
(k)
θ̂n(k+1) (T ) = φn (T ) Yn (T ) − δn,1 − P 0 δn,2 + ρ1 zn(k) + ρ2 P 0 θ̂g,(k) ,
Yn (t) =
t
X
λt−τ
n Xn (τ )yn (τ ),
(75)
(76)
τ =1
"
φn (t) =
t
X
!−1
#
0
λt−τ
+ ρ1 Inθ + ρ2 P 0 P
n Xn (τ )Xn (τ )
.
(77)
τ =1
Aiming at finding recursive formulas to update the estimates of the local parameters, we introduce the nth local estimate obtained at T − 1, i.e.
θ̂n (T − 1) = φn (T − 1) {Yn (T − 1) − δn,1 (T − 1) − P 0 δn,2 (T − 1)+
o
+ρ1 zn (T − 1) + ρ2 P 0 θ̂g (T − 1)
(78)
with δn,1 (T − 1), δn,2 (T − 1), zn (T − 1) and θ̂g (T − 1) being the Lagrange multipliers and the global estimate obtained at T − 1, respectively.
(k+1)
To obtain recursive formulas to compute θ̂n
, we start proving that φn (T )
can be computed as a function of φn (T − 1). in particular, introducing
Xn (t) =
t
X
0
λt−τ
n Xn (τ )(Xn (τ )) ,
τ =1
note that
φn (T )−1 = Xn (T ) + ρ1 Inθ + ρ2 P 0 P =
= λn Xn (T − 1) + Xn (T )Xn (T )0 + ρ1 Inθ + ρ2 P 0 P =
= λn [Xn (T − 1) + ρ1 Inθ + ρ2 P 0 P ] + Xn (T )Xn (T )0 + (1 − λn )ρ1 + (1 − λn )ρ2 P 0 P =
= λn φn (T − 1)−1 + Xn (T )Xn (T )0 + (1 − λn )ρ1 + (1 − λn )ρ2 P 0 P.
Defining the extended regressor as
p
X̃n (T ) = Xn (T )
(1 − λn )ρ1 Inθ
(1 − λn )ρ2 P 0 ∈ Rnθ ×(ny +nθ +ng ) ,
(79)
and applying the matrix inversion lemma, it can be easily proven that φn (T )
can then be computed as:
p
Rn (T ) = λn I(ny +nθ +ng ) + X̃n (T )0 φn (T )X̃n (T ),
Kn (T ) = φn (T − 1)X̃n (T )(Rn (T ))
φn (T ) =
λ−1
n (Inθ
−1
,
0
− Kn (T )X̃n (T ) )φn (T − 1).
(80)
(81)
(82)
The same observations relative to the update of φn made in Section 5 holds also
in the considered case.
Consider (75). Adding and subtracting
h
i
λn −δn,1 (T − 1) − P 0 δn,2 (T − 1) + ρ1 zn (T − 1) + ρ2 P 0 θ̂g (T − 1)
27
to (75) and considering the definition of φn (T − 1) (see (77)), the formula to
update θ̂n can be further simplified as
θ̂n(k+1) (T ) = φn (T ){λn (Yn (T − 1) − δn,1 (T − 1) − P 0 δn,2 (T − 1)+
+ρ1 zn (T − 1) + ρ2 P 0 θ̂g (T − 1) + Xn (T )yn (T ) + ρ1 (zn(k) − λn zn (T − 1))+
(k)
+ ρ2 P 0 (θ̂g,(k) − λn θ̂g (T − 1)) − (δn,1 − λn δn,1 (T − 1))+
(k)
− P 0 (δn,2 − λn δn,2 (T − 1))} =
= θ̂n (T − 1) − Kn (T )X̃n (T )θ̂n (T − 1) + φn (T ){Xn (T )yn (T )+
+ ρ1 (zn(k) − λn zn (T − 1)) + ρ2 P 0 (θ̂g,(k) − λn θ̂g (T − 1))}+
(k)
(k)
− (δn,1 − λn δn,1 (T − 1)) − P 0 (δn,2 − λn δn,2 (T − 1))} =
= θ̂nRLS (T ) + θ̂nADM M,(k+1) (T ).
(83)
In particular,
θ̂nRLS = φn (T )λn (Yn (T − 1) − δn,1 (T − 1) − P 0 δn,2 (T − 1) + ρ1 zn (T − 1)+
+ρ2 P 0 θ̂g (T − 1) + φn (T )Xn (T )yn (T ),
(84)
while
i
h
(k+1)
(k+1)
(k+1)
(k+1)
θ̂nADM M,(k+1) (T ) = φn (T ) ρ1 ∆z,λn (T ) + ρ2 P 0 ∆g,λn (T ) − ∆1,λn − P 0 ∆2,λn .
(85)
with
(k+1)
∆z,λn (T ) = zn(k) − λn zn (T − 1),
(k+1)
∆g,λn (T ) = θ̂g,(k) − λn θ̂g (T − 1),
(k+1)
(k)
(k+1)
(k)
∆1,λn = δn,1 − λn δn,1 (T − 1),
∆2,λn = δn,2 − λn δn,2 (T − 1).
Note that (85) differs from (62) because of the introduction of the additional
terms ∆z,λn and ∆1,λn .
Similarly to what is presented in Section 5, thanks to (82) the formula to update
θ̂nRLS can be further reduced as
θ̂nRLS = θ̂n (T − 1) − Kn (T )(X̃n (T ))0 θ̂n (T − 1) + φn (T )Xn (T )yn (T ) =
= θ̂n (T − 1) − Kn (T )(X̃n (T ))0 θ̂n (T − 1) + φn (T )X̃n (T )ỹn (T ),
with the extended measurement vector ỹn (T ) is defined as
0
ỹn (T ) = yn (T )0 O1×nθ O1×nng .
Exploiting the equality Kn (T ) = φn (T )X̃n (T ) (the proof can be found in (4)),
it can thus be proven that
θ̂nRLS = hatθn (T − 1) + Kn (T )(ỹn (T ) − (X̃n (T ))0 θ̂n (T − 1)).
(86)
It is worth remarking that θ̂nRLS can be updated (i) locally, (ii) recursively and
(iii) once per step t.
28
Remark 5 The proposed method, summarized in Algorithm 5 and in Figure 3,
requires the agents to transmit {θ̂nRLS , φn } to the “cloud”, while the “cloud” has
to communicate θ̂n to each node once it has been computed. As a consequence,
a N2C2N transmission scheme is required.
6.1
Example 4
Suppose that the data are gathered from N = 100 systems, described by (64)
and collected over an estimation horizon T = 5000. Moreover, assume that the
a priori information constraints parameter estimates to the following ranges:
`n,1 ≤ θ̂n,1 ≤ upn,1 `n,2 ≤ θ̂n,2 ≤ upn,2
`n,3 ≤ θ̂n,3 ≤ upn,3 .
(87)
Algorithm 5 ADMM-RLS algorithm for constrained consensus
Input: Sequence of observations {Xn (t), yn (t)}Tt=1 , initial matrices φn (0) ∈
nθ ×nθ
o
o
R
, initial local estimates θ̂n (0), initial dual variables δn,1
and δn,2
, initial auxiliary variables ẑn,o , forgetting factors λn , n = 1, . . . , N , initial global
estimate θ̂og , parameters ρ1 , ρ2 ∈ R+ .
1. for t = 1, . . . , T do
Local
1.1. for n = 1, . . . , N do
1.1.1. compute X̃n (t) with (79);
1.1.2. compute Kn (t) and φn (t) with (81) - (82);
1.1.3. compute θ̂nRLS (t) with (86);
1.2. end for;
Global
1.1. do
ADM M,(k+1)
(t) with (85), n = 1, . . . , N ;
1.1.1. compute θ̂n
(k+1)
1.1.2. compute θ̂n
1.1.3. compute
(t) with (83), n = 1, . . . , N ;
(k+1)
zn
(t)
g,(k+1)
1.1.4. compute θ̂
with (73), n = 1, . . . , N ;
with (74);
(k+1)
1.1.5. compute δn,1
with (71), n = 1, . . . , N ;
(k+1)
δn,2
with (72), n = 1, . . . , N ;
1.1.6. compute
1.2. until a stopping criteria is satisfied (e.g. maximum number of
iterations attained);
2. end.
Output: Estimated global parameters {θ̂g (t)}Tt=1 , estimated local parameters
{θ̂n (t)}Tt=1 , n = 1, . . . , N .
29
100
10-2
-4
10
-4
10
-2
10
10
0
10
2
Figure 14: Example 4. N̄ b vs ρ1 /ρ2 : black = N̄1b , red = N̄2b , blue = N̄3b .
Observe that the parameters ρ1 , ρ2 ∈ R+ have to be tuned. To assess how
the choice of these two parameters affects the satisfaction of (87), consider the
number of steps the local estimates violate the constraints over the estimation
horizon T , {Nib }3i=1 . Assuming that “negligible” violations of the constraints
are allowed, (87) are supposed to be violated if the estimated parameters fall
outside the interval Bn = [ `n −10−4 un +10−4 ]. Considering the set of constraints
S2 = {`n = [ 0.19 θn,2 −0.1 0.79 ] , upn = [ 0.21 θn,2 +0.1 0.81 ]},
Figure 14 shows the average percentage of violations over the N agents obtained
fixing ρ2 = 0.1 and choosing
ρ1 = {10−5 , 10−4 , 10−3 , 10−2 .10−1 , 1, 10, 20}.
Observe that if ρ1 dominates over ρ2 the number of violations tends to decrease,
as in the augmented Lagrangian (87) are weighted more than the consensus
constraint. However, if ρ1 /ρ2 > 100, {N̄ib }3i=1 tend to slightly increase. It
is thus important to trade-off between the weights attributed to (87) and the
consensus constraint. To evaluate how the stiffness of the constraints affects the
choice of the parameters, {Nib }3i=1 are computed considering three different sets
of box constraints
S1 = {`n = [ 0.195 θn,2 −0.05 0.795 ] , upn = [ 0.205 θn,2 +0.05 0.805 ]},
S2 = {`n = [ 0.19 θn,2 −0.1 0.79 ] , upn = [ 0.21 θn,2 +0.1 0.81 ]},
S3 = {`n = [ 0.15 θn,2 −0.5 0.75 ] , upn = [ 0.25 θn,2 +0.5 0.85 ]}.
The resulting {N̄ib }3i=1 are reported in Figure 15. Note that also in this case the
higher the ratio ρ1 /ρ2 is, the smaller {N̄ib }3i=1 are. However, also in this case,
the constraint violations tend to increase for ρ1 /ρ2 > 100.
Focusing on the assessment of ADMM-RLS performances when the set of
constraints is S2 , Figure 16 shows the global estimates obtained using the same
initial conditions and forgetting factors as in Section 6, with ρ1 = 10 and ρ2 =
0.1. Note that the global estimates satisfy (87), showing that the constraints on
the global estimate are automatically enforced imposing θn ∈ Cn . As it concerns
the RMSEs for θ̂g (43), they are equal to:
RM SE1g = 0.001 and RM SE2g = 0.006,
and their relatively small values can be related to the introduction of the additional constraints, that allow to limit the resulting estimation error.
30
100
100
-10
-10
10
10
10
-4
-2
10
10
0
10
2
10
-4
-2
10
(a) N̄1b
10
0
10
2
(b) N̄2b
100
10-2
-4
10
-4
-2
10
10
10
0
10
2
(c) N̄3b
Figure 15: Example 4. Average percentage of constraint violations N̄ib %, i =
1, 2, 3, vs ρ1 /ρ2 . Black : S1 , red : S2 , blue : S3 .
0.22
0.85
0.21
0.2
0.8
0.19
0.18
0.75
0
1000
2000
(a)
θ1g
3000
vs
4000
5000
θ̂1g
0
1000
2000
(b)
θ2g
3000
vs
4000
5000
θ̂2g
Figure 16: Example 4.Global model parameters: black = true, blue = ADMMRLS, red = upper and lower bounds.
Figure 17 show the estimate θ̂n for n = 11, with SN R11 = 10.6 dB. Note
that the estimated parameters tend to satisfy the constraints. In Figure 18 θ̂n
and θ̂nRLS , with n = 11, are compared. As it can be noticed, while θ̂11 satisfied
RLS
the imposed constraints on its values, the effect of using θ̂11 to update θ̂11
(see (86)) is not strong enough to enfoce also the estimates computed locally
to satisfy the contraints. To further assess the performance of the proposed
approach, the RMSE for the local estimates
v
2
uP
u T
t t=1 θn,i − θ̂n,i (t)
.
(88)
RMSEn,i =
T
is also considered. RM SEn,2 obtained for each of the N systems is reported in
Figure 19 and, as it can be noticed, RM SEn,2 is relatively small. As for the
global parameters’ estimates, this result can be related to the introduction of
the additional constraints.
31
0.22
0.6
0.21
0.5
0.2
0.4
0.19
0.18
0.3
0
1000
2000
3000
4000
5000
0
1000
(a) θ11,1 vs θ̂11,1
2000
3000
4000
5000
(b) θ11,2 vs θ̂11,2
0.85
0.8
0.75
0
1000
2000
3000
4000
5000
(c) θ11,3 vs θ̂11,3
Figure 17: Local parameter θn , n = 11. Black : true, blue : ADMM-RLS, red :
upper and lower bounds
0.8
0.3
0.2
0.6
0.1
0.4
0
-0.1
0.2
0
200
400
600
800
1000
0
RLS
(a) θ11,1 vs θ̂11,1 and θ̂11,1
200
400
600
800
1000
RLS
(b) θ11,2 vs θ̂11,2 and θ̂11,2
1.2
1
0.8
0.6
0
1000
2000
3000
4000
5000
RLS
(c) θ11,3 vs θ̂11,3 and θ̂11,3
Figure 18: Local parameter θn , n = 11, for t ∈ [1 1000]. Black : true, blue :
RLS
θ̂11
, cyan : θ̂11 , red : upper and lower bounds
7
Concluding Remarks and Future Work
In this report a method for collaborative least-squares parameter estimation
is presented based on output measurements from multiple systems which can
perform local computations and are also connected to a centralized resource in
the “cloud”. The approach includes two stages: (i) a local step, where estimates
of the unkown parameters are obtained using the locally available data, and (ii)
32
0.04
0.03
0.02
0.01
0
10
20
30
40
50
60
70
80
90
100
Figure 19: RM SE2 for each agent n, n = 1, . . . , N .
a global stage, performed on the cloud, where the local estimates are fused.
Future research will address extentions of the method to the nonlinear and
multi-class consensus cases. Moreover, an alternative solution of the problem
will be studied so to replace the transmission policy required now, i.e. N2C2N,
with a Node-to-Cloud (N2C) communication scheme. This change should allow
to alleviate problems associated with the communication latency between the
cloud and the nodes. Moreover, it should enable to obtain local estimators that
run independently from the data transmitted by the cloud, and not requiring
synchronous processing by the nodes and “cloud”. Other, solutions to further
reduce the trasmission complexity and to obtain an asynchronous scheme with
the same characteristics as the one presented in this report will be investigated.
A
Centralized RLS
Consider problem (12), with the cost functions given by
T
1X
kyn (t) − (Xn (t))0 θn k22 .
fn (θn ) =
2 t=1
The addressed problem can be solved in a fully centralized fashion, if at each step
t all the agents transmit the collected data pairs {yn (t), Xn (t)}, n = 1, . . . , N ,
to the “cloud”. This allows the creation of the lumped measurement vector and
regressor, given by
0
y̌(t) = y1 (t)0 . . . yN (t)0 ∈ RN ·ny ×1 ,
(89)
0
X̌(t) = X1 (t)0 . . . XN (t)0 ∈ Rnθ ×ny ·N .
Through the introduction of the lumped vectors, (12) with fn as in (20) is
equivalent to
T
1X
2
y̌(t) − (X̌(t))0 θg 2 .
(90)
min
θg 2
t=1
The estimate for the unknown parameters θ̂g can thus be retrieved applying
standard RLS (see [9]), i.e. performing at each step t the following iterations
−1
K(t) = φ(t − 1)X̌(t) ID + (X̌(t))0 φ(t − 1)X̌(t)
,
(91)
0
φ(t) = Inθ − K(t)(X̌(t)) φ(t − 1),
(92)
g
g
0 g
θ̂ (t) = θ̂ (t − 1) + K(t) y̌(t) − (X̌(t)) θ̂ (t − 1) ,
(93)
with
D = N · ny × 1.
33
References
[1] F. Boem, Y. Xu, C. Fischione, and T. Parisini. A distributed estimation
method for sensor networks based on pareto optimization. In 2012 IEEE
51st IEEE Conference on Decision and Control (CDC), pages 775–781, Dec
2012.
[2] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein. Distributed
optimization and statistical learning via the alternating direction method
of multipliers. Found. Trends Mach. Learn., 3(1):1–122, January 2011.
[3] F. S. Cattivelli, C. G. Lopes, and A. H. Sayed. Diffusion recursive leastsquares for distributed estimation over adaptive networks. IEEE Transactions on Signal Processing, 56(5):1865–1877, May 2008.
[4] Pedro A. Forero, Alfonso Cano, and Georgios B. Giannakis. Consensusbased distributed support vector machines. The Journal of Machine Learning Research, 11:1663–1707, Aug 2010.
[5] F. Garin and L. Schenato. A Survey on Distributed Estimation and Control
Applications Using Linear Consensus Algorithms, pages 75–107. Springer
London, London, 2010.
[6] M.N. Howell, J.P. Whaite, P. Amatyakul, Y.K. Chin, M.A. Salman, C.H.
Yen, and M.T. Riefe. Brake pad prognosis system, Apr 2010. US Patent
7,694,555.
[7] Z. Li, I. Kolmanovsky, E. Atkins, J. Lu, D. P. Filev, and J. Michelini.
Road risk modeling and cloud-aided safety-based route planning. IEEE
Transactions on Cybernetics, 46(11):2473–2483, Nov 2016.
[8] Z. Li, I. Kolmanovsky, E. M. Atkins, J. Lu, D. P. Filev, and Y. Bai.
Road disturbance estimation and cloud-aided comfort-based route planning. IEEE Transactions on Cybernetics, PP(99):1–13, 2017.
[9] L. Ljung. System identification: theory for the user. Prentice-Hall Englewood Cliffs, NJ, 1999.
[10] C. G. Lopes and A. H. Sayed. Incremental adaptive strategies over distributed networks. IEEE Transactions on Signal Processing, 55(8):4064–
4077, Aug 2007.
[11] G. Mateos, I. D. Schizas, and G. B. Giannakis. Distributed recursive leastsquares for consensus-based in-network adaptive estimation. IEEE Transactions on Signal Processing, 57(11):4583–4588, Nov 2009.
[12] Peter M. Mell and Timothy Grance. Sp 800-145. the nist definition of cloud
computing. Technical report, Gaithersburg, MD, United States, 2011.
[13] R. Olfati-Saber. Distributed kalman filtering for sensor networks. In 2007
46th IEEE Conference on Decision and Control, pages 5492–5498, Dec
2007.
34
[14] E. Ozatay, S. Onori, J. Wollaeger, U. Ozguner, G. Rizzoni, D. Filev,
J. Michelini, and S. Di Cairano. Cloud-based velocity profile optimization for everyday driving: A dynamic-programming-based solution. IEEE
Transactions on Intelligent Transportation Systems, 15(6):2491–2505, Dec
2014.
[15] E. Taheri, O. Gusikhin, and I. Kolmanovsky. Failure prognostics for intank fuel pumps of the returnless fuel systems. In Dynamic Systems and
Control Conference, Oct 2016.
35
| 3cs.SY
|
Automatic Backward Differentiation for American
Monte-Carlo Algorithms (Conditional Expectation)
Christian P. Fries
arXiv:1707.04942v1 [q-fin.CP] 16 Jul 2017
[email protected]
June 27, 2017
Version 0.4.0
Abstract
In this note we derive the backward (automatic) differentiation (adjoint [automatic]
differentiation) for an algorithm containing a conditional expectation operator. As
an example we consider the backward algorithm as it is used in Bermudan product
valuation, but the method is applicable in full generality.
The method relies on three simple properties:
1. a forward or backward (automatic) differentiation of an algorithm containing
a conditional expectation operator results in a linear combination of the
conditional expectation operators;
d
E(Y ) =
2. the differential of an expectation is the expectation of the differential dx
d
E( dx
Y );
3. if we are only interested in the expectation of the final result (as we are in all
valuation problems), we may use E(A · E(B|F)) = E(E(A|F) · B), i.e.,
instead of applying the (conditional) expectation operator to a function of the
underlying random variable (continuation values), it may be applied to the
adjoint differential.
The methodology not only allows for a very clean and simple implementation,
but also offers the ability to use different conditional expectation estimators in the
valuation and the differentiation.
Disclaimer: The views expressed in this work are the personal views of the authors and do not
necessarily reflect the views or policies of current or previous employers.
Feedback welcomed at [email protected].
1
Automatic Backward Differentiation for American Monte-Carlo
Christian Fries
Contents
1
Introduction
3
2
American Monte-Carlo and Bermudan Option Valuation
2.1 Bermudan Option Valuation . . . . . . . . . . . . . . . . . . . .
2.2 Bermudan Digital Option Valuation . . . . . . . . . . . . . . . .
3
4
4
3
Automatic Differentiation of the American-Monte-Carlo Backward Algorithm
5
3.1 Forward Differentiation . . . . . . . . . . . . . . . . . . . . . . .
5
3.2 Backward Differentiation . . . . . . . . . . . . . . . . . . . . . .
6
4
Numerical Results
7
5
Implementation Design
8
6
Conclusion
8
References
c 2017 Christian Fries
10
2
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
1
Christian Fries
Introduction
Given a filtered probability space (Ω, Q, {Ft }) we consider a Monte-Carlo simulation of a (time-discretised) stochastic process, i.e., we simulate sample path ωi
of sequences of random variables. Assuming that the drawings are uniform with
respect to the measure Q, this allows to approximate the unconditional expectation
EQ (X) of a random variable X via
n−1
1X
E (X) ≈
X(ωl ).
n
Q
l=0
However, the calculation of conditional expectations EQ (X|FTi ) is known to be
involved, since the Monte-Carlo simulation does not provide a discretization of
the filtration {Ft }. For the calculation (estimation) of conditional expectations
numerical approximations, like least-square regressions can be used. These methods
are often referred to as American Monte-Carlo, c.f. [6].
Let x denote a given model parameter used in the generation of the MonteCarlo simulation. Given a valuation algorithm which calculates the unconditional
expectation EQ (V ) of a random variable V , where the calculation of V involves
one or more conditional expectations, we consider the calculation of the derivative
d Q
dx E (V ).
Using automatic differentiation (i.e., analytic differentiation and the chain rule),
see [2, 8], this would require a differentiation of the conditional expectation operator
or - in practice - the approximation of the the conditional expectation operator.
Hence, direct application of automatic differentiation would result in the differentiation of the approximation of the the conditional expectation operator, e.g., the
regression, see [1, 3].
First of all, this is know to be difficult due to the cross-sectional character of
such approximations. In addition, it is in general questionable to differentiate an
approximation, since it is not guaranteed that the differentiation of an approximation
is a good approximation of the differential.
However, we can easily avoid this issue: First the differentiation of the conditional expectation can be replaced by a conditional expectation of a differentiation.
While this alone would present an improvement, it seems to make (backward) automatic differentiation much more complicated, since we would have to calculate a
derivative for each argument of a conditional expectation. In the next section we
show that we can simplify the algorithm by taking the conditional expectation of
the adjoint.
2
American Monte-Carlo and Bermudan Option Valuation
We shortly given the definition of the backward algorithm. For details see [6], we
use a similar notation here.
c 2017 Christian Fries
3
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
Christian Fries
Let 0 = T0 < T1 < . . . < Tn denote a given time discretization. Let Vi ,
i = 1, . . . , n denote the time-Ti numéraire relative values of given underlyings.
Here Vi are FTi -measurable random variables. Then let Ui be defined as follows:
Un+1 := 0
Ui := Bi EQ (Ui+1 |FTi ) , Ui+1 , Vi ,
where Bi are arbitrary function.
2.1
Bermudan Option Valuation
For a Bermudan option Bi is given by the optimal exercise criteria, i.e., Bi (x, u, v) :=
G(x − v, u, v) with
u if y>0
G(y, u, v) :=
.
v else
This defines a backward induction i = n, n − 1, . . . , 1 for Ui . For a Bermudan
option we have that the unconditional expectation (assuming that we have N (T0 ) =
1 for the numéraire)
EQ (U1 )
is the (risk) neutral value of the Bermudan option with exercise dates T1 < . . . < Tn
and exercise values Vi .
2.2
Bermudan Digital Option Valuation
For a Bermudan option the exercise is optimal. This allows for a popular “trick”
to ignore the first order derivative of G(y, u, v) with respect to y. Hence, for a
Bermudan option it is not necessary to differentiate Bi with respect to x.
However, this is just a special property of the function G and only holds for
first order derivatives. For example, the trick cannot applied to a Bermudan digital
option, i.e., Bi (x, u, v) := H(x − v, u, v) with
u if y>0
H(y, u, v) :=
.
1 else
We will use this product as a test case of the methodology in Section 4.
Remarks: The Bermudan digital option shows a dependency on the first order
derivative with respect to the indicator condition y, since the expectation of the two
outcomes u and 1 differs, i.e., we have a jump at y = 0. For the classic Bermudan
option we have that the expectation of u and v agree at y = 0.
The (automatic) differentiation of a jump may appear as an issue, but it is
possible to solve this elegantly in an stochastic automatic differentiation [7] and
in this note we focus on the treatment of the conditional expectation (however,
our numerical results demonstrate that the algorithm [5] treats the discontinuity
correctly, c.f. [7]).
c 2017 Christian Fries
4
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
3
Christian Fries
Automatic Differentiation of the American-Monte-Carlo
Backward Algorithm
Let z denote an arbitrary model parameter. We assume that the underlying values
Vi depend on z.
For i = 1, . . . , n we have
d
d
Ui =
Bi EQ (Ui+1 |FTi ) , Ui+1 , Vi
dz
dz
and writing Bi = Bi EQ (Ui+1 |FTi ) , Ui+1 , Vi to avoid the lengthy argument list
dBi
d
Ui =
·
dz
dx
and using
d Q
dz E (Ui )
d Q
dBi dUi+1 dBi dVi
E (Ui+1 |FTi ) +
·
+
·
dz
du
dz
dv dz
d
dBi
Ui =
· EQ
dz
dx
3.1
d
dz Ui
= EQ
dUi+1
|FTi
dz
+
dBi dUi+1 dBi dVi
·
+
·
du
dz
dv dz
(1)
Forward Differentiation
Applying this relation iteratively (plugging the expression of
i
of dU
dz for j = i, . . . , n − 1) gives
n
X
d
Ui =
dz
j=i
j−1
Y
k=i
dBk
du
!
dBj
· EQ
dx
dUj+1
|FTj
dz
+
dUj+1
dz
j−1
Y
k=i
into the equation
dBk
du
!
dBj dVj
·
dv
dz
To shorten notation, let
Ai,j =
j−1
Y
k=i
dBk
du
!
dBj
dx
Ci,j =
j−1
Y
k=i
dBk
du
!
dBj
dv
such that
n
X
dUj+1
dVj
dUi
=
Ai,j · EQ
|FTj + Ci,j ·
.
dz
dz
dz
(2)
j=i
This last equation would be natural in a forward (automatic) differentiation, since
dV
n+1 dUn dUn−1
we calculate 0 = dUdz
, dz , dz , . . . , together with dzj forwardly. Note that
forward here refers to the order of the operations in the algorithms (which runs
backward over the indices j).
c 2017 Christian Fries
5
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
!
.
Automatic Backward Differentiation for American Monte-Carlo
3.2
Christian Fries
Backward Differentiation
For the application of a backward (adjoint) automatic differentiation Equation (2)
would require a mixture of backward differentiation for that we first calculate
dUj
dz and followed by a forward application of (2). However, we can calculate the
derivative in a single backward differentiation sweep:
d Q
We start with (1) for i = 1. Since we are only interest in dz
E (U1 ) =
d
Q
E dz U1 , we can take expectation and get
d
dB1 dU2 dB1 dV1
Q
Q dB1
Q dU2
E
U1 = E
·E
|FTi +
·
+
·
dz
dx
dz
du
dz
dv
dz
Now we may use
dU2
Q dB1
Q dB1
Q
Q dU2
E
= E
E
·E
|FT1
|FT1 ·
dx
dz
dx
dz
to get
d
dB1
dU2 dB1 dV1
dB1
EQ
U1 = EQ
EQ
|FT1 +
·
+
·
. (3)
dz
dx
du
dz
dv
dz
Plugging (1) into (3) and repeating the previous argument for i = 2, . . . , k − 1 we
get iteratively the forward equation
k
X
dVj
dUk+1
d
∗
U1 = EQ A∗1,k ·
+
C1,j
·
.
EQ
dz
dz
dz
j=1
where
A∗1,i
∗
C1,i
A∗1,0
dBi
dBi
∗
= E
A1,i−1
|FTi + A∗1,i−1
dx
du
dBi
= A∗1,i−1
dv
= 1
Q
Using i = n we have with Un+1 = 0 that
EQ
d
U1
dz
= EQ
n
X
j=1
∗
C1,j
·
dVj
.
dz
∗ have an intuitive interpretation in a
The recursive definitions of A∗1,i , C1,i
backward (automatic) differentiation algorithm: in the algorithm the conditional
expectation operator on Ui+1 is replaced by taking the conditional expectation of
the adjoint differential.
c 2017 Christian Fries
6
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
4
Christian Fries
Numerical Results
As a test case we calculate the delta of a Bermudan digital option. This product pays
if S(Ti ) − Ki > Ũ (Ti ) in Ti if no payout has been occurred,
1
where Ũ (T ) is the time T value of the future payoffs, for T1 , . . . , Tn . Note that
Ũ (Tn ) = 0, such that the last payment is a digital option.
This product is an ideal test-case: the valuation of Ũ (Ti ) is a conditional
expectation. In addition conditional expectation only appears in the indicator
function, such that keeping the exercise boundary (the condition) fixed, would result
in a delta of 0. On the other hand, the delta of a digital option payoff is only driven
by the movement of the indicator function, since
d
df(S)
E(1(f (S(T )) > 0)) = φ(f −1 (0))
,
dS0
dS0
where φ is the probability density of S and 1(·) the indicator function. See [6]
The results are depicted in Figure 1 for T1 = 1, T2 = 2, T3 = 3, T4 = 4, and
K1 = 0.5, K2 = 0.6, K3 = 0.8, K4 = 1.0 for a model with S0 = 1.
The implementation of this test case is available in [5].
Bermudan Digital Option Delta
1
Central Finite Difference Approximation
0.9
Stochastic Automatic Differentiation
0.8
0.7
Delta
0.6
0.5
0.4
0.3
0.2
0.1
0
0.000001 0.00001
0.0001
0.001
0.01
0.1
1
10
Finite Difference Shift Size
Figure 1: Delta of a Bermudan digital option using finite differences (red) and
stochastic AAD. The calculations were repeated with 4 different Monte-Carlo
random number seeds (hence there 4 red data series and 4 data series). The
finite difference approximation is biased for large shifts size The blue dashed
line depicts the value of the delta of the conditional expectation operator is
ignored. This generates a bias du to the correlation of derivative and underlying.
c 2017 Christian Fries
7
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
5
Christian Fries
Implementation Design
Using the framework described in [7] we are able to implement the methodology
with a minimum of code complexity: we require only two additional lines of code:
In [4] all random variable objects of a Monte-Carlo simulation are implementing a
common interface RandomVariableInterface. This interface offers methods
for arithmetic operators like add, sub, mult, exp, but also an operator
RandomVariableInterface getConditionalExpectation(
ConditionalExpectationEstimatorInterface estimator);
The default implementation of this method is just
default RandomVariableInterface getConditionalExpectation(
ConditionalExpectationEstimatorInterface estimator)
{
return estimator.getConditionalExpectation(this);
}
In other words: all conditional expectations of random variables have to be called
through this method - which enables us to track conditional expectation in the
operator tree.
To enable adjoint automatic differentiation we inject a special implementation of RandomVariableInterface which records the operator tree and
can perform the AAD. The implementation of the differential of the operator
getConditionalExpectation now consists of two parts: First, the partial
derivative of the operator with respect to its arguments is set to 1.0:
case CONDITIONAL_EXPECTATION:
return new RandomVariable(1.0);
- this implies that for the arguments the conditional expectation is replaced by the
identity operator. Second, the backward propagation of the differential contains the
additional check for the operator, applying it to the differential if required:
if(operatorType == OperatorType.CONDITIONAL_EXPECTATION) {
ConditionalExpectationEstimatorInterface estimator = (
ConditionalExpectationEstimatorInterface)operator;
derivative = estimator.getConditionalExpectation(derivative);
}
6
Conclusion
In this note we presented a simple modification of the backward automatic differentiation of algorithms which require the calculation (estimation) of a conditional
expectation: the conditional expectation operator has to be applied to the adjoint
differential, while the remaining algorithm remains unchanged.
Our method has important advantages over a direct differentiation of the conditional expectation estimator ([1, 3]):
1. Do not differentiate an approximation: It is often not clear that differentiation
of an approximation given a good approximation of the differential. In
c 2017 Christian Fries
8
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
Christian Fries
our approach it is not necessary to differentiate the approximation of the
conditional expectation operator. A simple example is if the conditional
expectations is approximated by fixed bins, which results in a pice-wise
constant approximation (see [6]) and the (automatic) differentiation of the
conditional expectation would be zero.
2. The approach allows to use a different approximations of the conditional
expectation operator in the valuation and in the algorithmic differentiation.
This can improve the accuracy of the differentiation. For example, it is
possible to choose different regression basis functions.
Most importantly, the approach greatly simplifies the implementation of the
algorithm. In [5] the implementation basically consists of adding two lines of code.
c 2017 Christian Fries
9
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
Christian Fries
References
[1] A NTONOV, A LEXANDRE: Algorithmic Differentiation for Callable Exotics
(April 4, 2017). http://ssrn.com/abstract=2839362.
[2] C APRIOTTI , L UCA ; G ILES , M IKE: Algorithmic Differentiation: Adjoint Greeks Made Easy (2011). http://ssrn.com/abstract=
1801522.
[3] C APRIOTTI , L UCA ; J IANG , Y UPENG ; M ACRINA , A NDREA: AAD and
Least Squares Monte Carlo: Fast Bermudan-Style Options and XVA Greeks
(September 23, 2016). https://ssrn.com/abstract=2842631
or http://dx.doi.org/10.2139/ssrn.2842631
[4] FINMATH . NET: finmath-lib: Mathematical Finance Library: Algorithms
and methodologies related to mathematical finance. https://github.
com/finmath/finmath-lib
finmath-lib automatic differentiation exten[5] FINMATH . NET:
tions: Enabling finmath lib to utilize automatic differentiation
algorithms (e.g. AAD). https://github.com/finmath/
finmath-lib-automaticdifferentiation-extensions
[6] F RIES , C HRISTIAN P.: Mathematical Finance. Theory, Modeling, Implementation. John Wiley & Sons, 2007. ISBN 0-470-04722-4.
http://www.christian-fries.de/finmath/book.
[7] F RIES , C HRISTIAN P.: Stochastic Automatic Differentiation: Efficient
Implementation of Automatic Differentiation for Monte-Carlo Simulations.
(June, 2017). https://ssrn.com/abstract=2995695
[8] G ILES , M ICHAEL ; G LASSERMAN , PAUL: Smoking adjoints: fast Monte
Carlo Greeks (2006). Risk, January 2006. https://www.risk.net
Notes
Suggested Citation
F RIES , C HRISTIAN P.: Automatic Differentiation of the AmericanMonte-Carlo Backward Algorithm.
(June, 2017).
https:
//ssrn.com/abstract=3000822
http://www.christian-fries.de/finmath/
stoachasticautodiff
c 2017 Christian Fries
10
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
Automatic Backward Differentiation for American Monte-Carlo
Christian Fries
Classification
Classification: MSC-class: 65C05 (Primary)
ACM-class: G.1.4; G.3; I.6.8.
JEL-class: C15, G13.
c 2017 Christian Fries
11
Version 0.4.0 (20170627)
http://www.christianfries.com/finmath/stoachasticautodiff
| 8cs.DS
|
Sign properties of Metzler matrices with applications
Corentin Briat
arXiv:1512.07043v3 [math.DS] 7 Nov 2016
Department of Biosystems Science and Engineering, ETH-Zürich, Switzerland
Abstract
Several results about sign properties of Metzler matrices are obtained. It is first established that checking
the sign-stability of a Metzler sign-matrix can be either characterized in terms of the Hurwitz stability of the
unit sign-matrix in the corresponding qualitative class, or in terms the negativity of the diagonal elements of
the Metzler sign-matrix and the acyclicity of the associated directed graph. Similar results are obtained for
the case of Metzler block-matrices and Metzler mixed-matrices, the latter being a class of Metzler matrices
containing both sign- and real-type entries. The problem of assessing the sign-stability of the convex hull of
a finite and summable family of Metzler matrices is also solved, and a necessary and sufficient condition for
the existence of common Lyapunov functions for all the matrices in the convex hull is obtained. The concept
of sign-stability is then generalized to the concept of Ker+ (B)-sign-stability, a problem that arises in the
analysis of certain jump Markov processes. A sufficient condition for the Ker+ (B)-sign-stability of Metzler
sign-matrices is obtained and formulated using inverses of sign-matrices and the concept of L+ -matrices.
Several applications of the results are discussed in the last section.
Keywords: Sign-stability; Metzler matrices; Positive systems
2010 MSC: 15A48, 34D05
1. Introduction
Metzler matrices [1] are often encountered in various fields such as economics [2] and biology [3–6].These
matrices have also been shown to play a fundamental role in the description of linear positive systems, a
class of systems that leave the nonnegative orthant invariant [7]. Due to their particular structure, such
systems have been the topic of a lot of works in the dynamical systems and control communities; see
e.g. also [8–14] and references therein. In this paper, we will be interested in the sign-properties of such
matrices, that is, those general properties that can be deduced from the knowledge of the sign-pattern. This
problem initially emerged from economics [15], but also found applications in ecology [16–19] and chemistry
[20, 21]. The rationale for this approach stems from the fact that, in these fields, the interactions between
different participants in a given system are, in general, qualitatively known but quantitatively unknown. This
incomplete knowledge is very often a direct consequence of the difficulty in identifying and discriminating
models because of their inherent complexity and scarce experimental data. In this context, it seems relevant
to study the properties of the system solely based on the sign pattern structure or, more loosely, from the
pattern of nonzero entries. This is referred to as qualitative analysis [22]. For instance, the sign-stability of
general matrices (the property that all the matrices having the same sign-pattern are all Hurwitz stable) has
been studied in [15, 19, 23]. An algorithm, having worst-case O(n2 ) time- and space-complexity, verifying
the conditions in [23] has been proposed in [24]. Many other problems have also been addressed over the
past decades; see e.g. [25–29] and references therein.
The first problem addressed in this paper is the sign-stability problem for which several general necessary
and sufficient conditions exist [22]. By adapting these results, we show that the sign-stability of a Metzler
Email address: [email protected], [email protected] (Corentin Briat)
URL: www.briat.info (Corentin Briat)
Preprint submitted to Elsevier
November 8, 2016
sign-matrix can be assessed from the Hurwitz stability of a single particular matrix, referred to as the unit
sign-matrix, lying inside the qualitative class. Therefore, for this class of matrices, checking the Hurwitz
stability of an uncountably infinite and unbounded family of matrices is not more difficult than checking the
stability of a given matrix. Lyapunov conditions, taking in this case the form of linear programs [30], can
hence be used for establishing the sign-stability of a given Metzler sign-matrix. An alternative condition is
formulated in terms of the acyclicity of the graph associated with the sign-pattern, a property that can be
easily checked using algorithms such as the Depth-First-Search algorithm [31]. This result is then generalized
to the case of block matrices for which several results, potentially enabling the use of distributed algorithms,
are provided. Sign-stability results are then generalized to the problem of establishing the sign-stability of
all the matrices located in the convex hull of a finite number of ”summable” Metzler sign-matrices. This
problem is highly connected to the analysis of linear positive switched systems, a problem that has been
extensively considered in the literature: see e.g. [32–34]. Necessary and sufficient conditions for the existence
of a common Lyapunov function for all the matrices in the convex-hull of such matrices are also provided.
A novel stability concept, referred to as Ker+ (B)-sign-stability, is then considered in order to formulate a
constrained stability concept that arises in the analysis and the control of positive nonlinear systems and
stochastic reaction networks [4, 6, 35]. It is shown that the obtained sufficient conditions characterizing
Ker+ (B)-sign-stability of a Metzler sign-matrix can be brought back to a combinatorial problem known to
be NP-complete [36]. Finally, mixed-matrices consisting of matrices with both sign- and real-type entries are
considered and their sign-properties clarified, again in terms of algebraic and graph theoretical conditions
analogous to those obtained in the pure sign-matrix case. Several application examples are provided in the
last section.
Outline: Section 2 recalls some important definitions and known results. Section 3 is devoted to the
analysis of the sign-stability of Metzler sign-matrices. The analysis of the sign-stability of block matrices
is carried out in Section 4. Conditions for the sign-stability of the convex-hull of Metzler sign-matrices are
obtained in Section 5. Section 6 is concerned with relative structural stability analysis whereas Section 7
addresses partial sign-stability. Application examples are finally discussed in Section 8.
Notations: The set of real, real positive and real nonnegative numbers are denoted by R, R>0 and
R≥0 , respectively. These notations generalize to vectors where Rn>0 and Rn≥0 are used to denote vectors with
positive and nonnegative entries, respectively. The sets of integers and nonnegative integers are denoted by
Z and Z≥0 , respectively. The n-dimensional vector of ones is denoted by 1n . The block-diagonal matrix
with diagonal elements given by Mi ∈ Rni ×mi is denoted by diagi {Mi }.
2. Preliminaries
We define in this section the terminology that will be used in the paper as well as some elementary
results.
Definition 2.1. Let us define the following matrix types and associated sets:
• A sign-matrix is a matrix taking entries in S := {⊖, 0, ⊕}. The set of all n × m sign-matrices is
denoted by Sn×m . The set S has a full-order structure with the order ⊖ < 0 < ⊕. In this regard, ⊖ is
considered here as a negative entry and ⊕ as a positive one.
• A Metzler matrix with entries in R ∪ S is a square matrix having nonnegative off-diagonal entries.
The set of all n × n Metzler matrices with entries in R ∪ S is denoted by Mn := (R ∪ S)n×n .
n×m
n×m
• The cones of n × m nonnegative matrices with entries in S or R are denoted by S≥0
and R≥0
,
respectively.
• The sets MSn and MRn are used to denote the sets of Metzler sign-matrices and Metzler real matrices,
respectively. The considered orders are the natural one and the one defined above for sign-matrices.
2
Definition 2.2. The qualitative class of a matrix A ∈ (S ∪ R)n×m is the set of matrices given by
∈ R>0 if [A]ij = ⊕
Q(A) := M ∈ Rn×m
[M ]ij ∈ R<0 if [A]ij = ⊖
.
= [A]ij otherwise
Definition 2.3. Let M ∈ Sn×m . Then, the unit sign-matrix U associated with M , denoted by U = sgn(M ),
is defined as
if [S]ij = 0,
0
[U ]ij :=
(1)
1
if [S]ij = ⊕,
−1 if [S]ij = ⊖
where sgn(·) is the extension of the signum function to sign-matrices.
Definition 2.4 ([23]). For any matrix A ∈ (S ∪ R)n×n , we define its associated directed graph as DA =
(V, ED ) where V = {1, . . . , n} and
ED = {(j, i) : aij 6= 0, i, j ∈ V, i 6= j} .
Definition 2.5. A matrix M ∈ Sn×n is sign-stable if all the matrices M ′ ∈ Q(M ) are Hurwitz stable.
Definition 2.6. A matrix M ∈ Sn×n is potentially stable if Q(M ) contains at least one matrix that is
Hurwitz stable.
Example 2.7. Let us consider the matrices
⊖
M1 =
⊕
⊖
⊕
and M2 =
⊕
⊖
⊖
.
⊖
(2)
The matrix M1 is not sign-stable since the determinant is not sign definite; i.e. det(M1 ) = ⊕ + ⊖. Indeed,
the entries (1,1), (1,2) and (2,2) being given, we can always find a sufficiently large entry (2,1) such that
the determinant is negative and, hence, the matrix not Hurwitz stable. On the other hand, the matrix M2
is sign-stable since trace(M2 ) = ⊖ and det(M2 ) = ⊕; i.e. the coefficients of the characteristic polynomial
are always positive and hence all the matrices in the qualitative class are Hurwitz stable. However, the
sign-matrix M1 is potentially sign-stable as some matrices in the qualitative class are Hurwitz stable.
Lemma 2.8 (Hurwitz stability of Metzler matrices, [1]). Let us consider a matrix A ∈ MRn . Then,
the following statements are equivalent:
(a) The matrix A is Hurwitz stable.
(b) A is nonsingular and A−1 ≤ 0.
(c) There exists a vector v ∈ Rn>0 such that v T A < 0.
3. Sign-stability of Metzler sign-matrices
The sign-stability of matrices is a problem that has been extensively studied and for which a complete
solution exists. Such a complete solution is, for instance, given in [23, Theorem 2] where a set of necessary
and sufficient conditions is provided in terms of algebraic and graph theoretical concepts. We propose here to
study the particular case of Metzler sign-matrices and derive a collection of conditions that are more tailored
to that kind of matrices than the general ones stated in [23]. These conditions will be shown to be simpler
than those in [23] from both a theoretical and a computational viewpoint. It will indeed be proved that
the sign-stability of a Metzler sign-matrix can be established by checking the Hurwitz stability of a single
matrix in the associated qualitative class. From a computational viewpoint, we also demonstrate that the
Depth-First Search (DFS) algorithm [31] can be used in order to establish the sign-stability of the matrix.
This algorithm has worst-case time- and space-complexity of O(n2 ) and O(n), respectively.
3
3.1. Preliminary results
The following results will play a key role in the derivation of the main results of the section.
Lemma 3.1 ([7]). Let A ∈ MRn . If A is Hurwitz stable, then A has negative diagonal elements.
Lemma 3.2. Let us consider a matrix M ∈ MSn with negative diagonal elements and assume that its
associated directed graph DM is an elementary cycle. Then, the matrix sgn(M ) is irreducible and its PerronFrobenius eigenvalue, denoted by λP F (sgn(M )), is equal to 0 with 1n as associated right-eigenvector.
Proof : Since DM is an elementary cycle, then sgn(M ) can be decomposed as sgn(M ) = P − In where
P is a cyclic permutation matrix with zero entries on the diagonal. Hence, sgn(M ) is irreducible since P
is irreducible. Furthermore, it is immediate to see that sgn(M )1 = 0 and hence that 0 is an eigenvalue of
sgn(M ) with positive right-eigenvector equal to 1. From the irreducibility property and the Perron-Frobenius
theorem [1], we have that this eigenvalue is the Perron-Frobenius one and that it is unique. The proof is
complete.
♦
Lemma 3.3. Let us consider a matrix M ∈ MRn ∩ {−1, 0, 1}n×n. Then, the following statements are
equivalent:
(a) M is Hurwitz stable (in the sense that M interpreted as a real matrix is Hurwitz stable).
(b) The diagonal elements of M are negative and the directed graph DM is acyclic.
Proof : Proof of (a) ⇒ (b). We use the contrapositive. We have two cases. The first one is when at
least one of the diagonal entries is nonnegative and, in this case, Lemma 3.1 proves the implication. The
second one is when there exists at least a cycle in the graph DM . Let us assume that there is a single cycle
of order k ≥ 2 in DM (the case of multiple cycles can be addressed in the same way). Define the matrix
f ∈ MRn ∩ {−1, 0, 1}n×n with negative diagonal elements and for which the graph D f is obtained from
M
M
f is a cyclic permutation of the
DM by removing all the edges that are not involved in the cycle. Then, M
matrix
P − Ik
0
(3)
0
−In−k
where P is a cyclic permutation matrix with diagonal elements equal to 0. The left-upper block then
corresponds to the nodes involved in the cycle whereas the right-lower block corresponds to those that are
f (in the same permutation basis). Hence, from the Perron-Frobenius
not. Clearly, we have that M ≥ M
f). Again due
theorem and the theory of nonnegative/Metzler matrices [1], we have that λP F (M ) ≥ λP F (M
f
to the Perron-Frobenius theorem and Lemma 3.2, we have that λP F (M ) = λP F (P − Ik ) = 0 and hence
λP F (M ) ≥ 0, proving then that M is not Hurwitz stable.
Proof of (b) ⇒ (a). Assume that statement (b) holds. Since the graph DM is acyclic, then there
exists a permutation matrix P such that the matrix P T M P is upper-triangular with negative elements on
the diagonal and nonnegative elements in the upper-triangular part. Therefore, the matrix M is Hurwitz
stable since its eigenvalues coincide with its diagonal elements, which are negative. The proof is complete.
♦
3.2. Main results
We can now state our main result on the sign-stability of Metzler sign-matrices:
Theorem 3.4. Let A ∈ MSn×n . Then, the following statements are equivalent:
(a) The matrix A is sign-stable.
(b) The matrix sgn(A) is Hurwitz stable.
4
(c) There exists v ∈ Rn>0 such that v T sgn(A) < 0.
(d) The directed graph DA is acyclic and the diagonal elements of A are negative.
(e) There exists a permutation matrix P such that P T AP is upper-triangular with negative diagonal elements.
(f) The diagonal elements of A are negative and the DFS algorithm applied to the graph DA returns a
spanning tree with no back edge1 .
Proof : The proofs that (a) implies (b) and that (e) implies (a) are obvious. The proof that (b) is equivalent
to (c) follows from Lemma 2.8. The proof that (d) is equivalent to (b) follows from Lemma 3.3. The proof
that (d) is equivalent to (e) follows from the definition of an acyclic graph and, finally, the proof that (d) is
equivalent to (f) follows from the property that a graph is acyclic if and only if one of its associated spanning
trees has no back edge; see [31].
♦
Remark 3.5. The condition of statement (b) in Theorem 3.4 states that it is enough to check the Hurwitz
stability of the unique unit sign-matrix inside the qualitative class in order to establish the sign-stability of
the sign-matrix and the Hurwitz stability of the entire qualitative class. It is important to stress that this
condition is peculiar to Metzler matrices as demonstrated by the following example. Indeed, the matrix
⊖ ⊕ ⊕
A = ⊖ ⊖ ⊖
(4)
0 ⊕ ⊖
is not sign-stable because of the presence of a cycle of order 3
unit sign-matrix given by
−1 1
sgn(A) = −1 −1
0
1
is Hurwitz stable since sgn(A) + sgn(A)T is negative definite.
in its directed graph. However, the associated
1
−1
(5)
−1
Remark 3.6. The proof that (a) is equivalent to (d) in Theorem 3.4 could have been based on the general
sign-stability result [23, Theorem 2] that involves both algebraic and graph theoretical conditions. By applying
them to the particular case of Metzler sign-matrices, we obtain that a Metzler sign-matrix is sign-stable if
and only if the diagonal elements are negative and its directed graph does not contain any cycle. For general
matrices, cycles of length 2 are still allowed provided that the products of the entries involved in the same
cycle are negative. This is clearly not possible in the context of Metzler matrices since all the off-diagonal
entries are nonnegative.
Remark 3.7 (Computational complexity). The statement (c) indicates that the Hurwitz stability of
the real Metzler matrix sgn(A) can be easily checked since the problem of finding a vector v > 0 such that
v T sgn(A) < 0 is a linear programming problem [30] that can be solved using modern optimization algorithms
in an efficient way. By virtue of statement (f ), an alternative way for checking the sign-stability of a Metzler
sign-matrix is through the negativity of its diagonal elements and the acyclicity of its associated directed
graph. Indeed, checking the negativity of the diagonal elements and the existence of a cycle in a given
graph are both simple problems, the latter being solvable using algorithms such as the Depth-First-Search
algorithm [31] which has a worst-case time-complexity of O(n2 ) and a worst-case space-complexity of O(n).
Complexity-wise, the general conditions of [23, Theorem 2] can be checked using the algorithm proposed in
[24] which has worst-case time- and space-complexity of O(n2 ).
1
The spanning tree is the tree generated by the DFS algorithm while searching the graph and a back edge is an edge that
points from a node to one of its ancestors in the spanning tree [31].
5
For completeness, we provide several necessary conditions for the sign-stability of Metzler sign-matrices:
Proposition 3.8 (Necessary conditions). The Metzler sign-matrix A ∈ MSn is not sign-stable if one of
the following statements hold:
(a)
(b)
(c)
(d)
[A]ii ≥ 0 for some i = 1, . . . , n.
[A]ij [A]ji > 0 for some i, j = 1, . . . , n, i 6= j.
There is a cycle in the directed graph DA .
The matrix A is irreducible.
Proof : Statement (a) follows from Lemma 3.1 while the others follow from the existence of at least one
cycle in the directed graph associated with the sign-matrix.
♦
We conclude this section with a result on the potential sign-stability of Metzler sign-matrices:
Theorem 3.9 (Potential sign-stability). Let us consider a Metzler sign-matrix A ∈ MSn . Then, the
following statements are equivalent:
(a) The diagonal entries of A are negative.
(b) The matrix A is potentially sign-stable.
Proof : Proof of 2) ⇒ 1) Since the matrix A is Metzler and is partially sign-stable, then there exists a
matrix A′ ∈ Q(A) that is Hurwitz stable. From Lemma 3.1, this matrix necessarily has negative diagonal
elements and, hence, A must have negative diagonal elements.
Proof of 1) ⇒ 2) Assume the matrix A has negative diagonal entries. Let us then consider the matrix
Mε , ε > 0, defined as
if [A]ij = ⊖,
−1
[Mε ]ij =
(6)
ε
if [A]ij = ⊕,
0
if [A]ij = 0
and note that Mε ∈ Q(A). Clearly, when ε = 0, the matrix M0 is Hurwitz stable and all the eigenvalues are
equal to -1. Using the fact that the eigenvalues of Mε are continuous with respect to ε, then we can conclude
that there exists an ε̄ > 0 such that for all 0 ≤ ε < ε̄, the matrix Mε is Hurwitz-stable, showing then that
A is potentially sign-stable. The proof is complete.
♦
It is finally interesting to note that Theorem 3.4 can also be used to obtain analogous sign-stability
conditions for nonnegative sign-matrices. Note, however, that sign-stability needs to be defined here as the
Schur stability of all the matrices in the qualitative class; i.e. all the matrices in the qualitative class have
spectral radius less than unity. This result is stated below:
n×n
Theorem 3.10. Let A ∈ S≥0
. Then, the following statements are equivalent:
(a)
(b)
(c)
(d)
(e)
(f)
All the matrices in Q(A) have spectral radius less than unity.
All the matrices in Q(A) have zero spectral radius.
The matrix sgn(A) − In is Hurwitz stable.
There exists v ∈ Rn>0 such that v T (sgn(A) − In ) < 0.
The directed graph DA is acyclic and the diagonal elements of A are all equal to zero.
There exists a permutation matrix P such that P T AP is upper-triangular and has zero elements on the
diagonal.
(g) There is no back edge in DA and the diagonal elements of A are all equal to zero.
6
Proof : We need to prove here that (a) is equivalent to (b). The equivalence with the other statements
n×n
follows from Theorem 3.4 and the fact that a nonnegative matrix A ∈ R≥0
is Schur stable if and only if the
Metzler matrix A − In is Hurwitz stable. Clearly, (b) implies (a), so let us focus on the reverse implication
and assume that (b) does not hold. Hence, there exists a matrix M ∈ Q(A) such that ρ(M ) = ǫ > 0. If
ǫ ≥ 1, then (a) does not hold and, on the other hand, if ǫ ∈ (0, 1), then using the fact that αM ∈ Q(A)
for any α > 0, we can conclude that, by choosing a large enough α > 0, we will have ρ(αM ) > 1 and the
statement (a) does not hold again. This proves that (a) implies (b).
♦
4. Sign-stability of block matrices
The objective of this section is to extend the results of the previous one to the case of Metzler block
sign-matrices. At first sight, it may seem irrelevant to consider Metzler block sign-matrices as they can be
analyzed using the tools derived in the previous section. The main rationale, however, is that these matrices
are commonplace in fields such as control theory where they can be used to represent, for instance, interconnections of linear (positive) dynamical systems. The objective is then to find conditions that characterize
the sign-stability of a given Metzler block sign-matrix based on the sign-stability of its elementary diagonal
blocks and some additional conditions capturing the interactions between the different diagonal blocks (i.e.
the off-diagonal blocks). This paradigm has led to many important results such as the small-gain theorem
[37, 38] and its linear positive systems variants [9, 12].
First of all, several existing results about the linear S-procedure [39, 40] are recalled and some novel
ones are also obtained. The S-procedure is an essential tool of control theory which allows for checking
the negativity of some conditions under some constraints in a computationally tractable way. These results
are then used to obtain exact linear programming conditions characterizing the sign-stability of Metzler
block real matrices. Finally, the obtained conditions are then adapted to the analysis of the sign-stability
of Metzler block-matrices. It is emphasized that such an approach enables the use of distributed algorithms
for solving this problem for very large systems in a very efficient way.
4.1. Preliminary results
The main theoretical tool considered here is the so-called S-procedure [39] and, more specifically, its
linear version which is known to provide an exact characterization of the nonnegativity of a linear form on
a set defined by linear inequalities (e.g. a compact convex polytope). The linear version of the S-procedure
is recalled below for completeness.
Lemma 4.1 (Linear S-procedure, [39, 40]). Let σi ∈ Rn , i = 0, . . . , N , and assume that the set
S≥0 := y ∈ Rn : σiT y ≥ 0, i = 1, . . . , N
(7)
is nonempty. Then, the following statements are equivalent:
(a) We have that σ0T y ≥ 0 for all y ∈ S≥0 .
(b) There exist some multipliers τi ≥ 0, i = 1, . . . , N , such that the inequality
σ0T y −
N
X
τi σiT y ≥ 0
(8)
i=1
holds for all y ∈ Rn .
(c) There exists a multiplier vector τ ∈ RN
≥0 such that the inequality
σ0T y − τ T σ̄y ≥ 0
T
holds for all y ∈ Rn where σ̄ := colN
i=1 (σi ).
7
(9)
Proof : The equivalence between the first statements follows from the results in [39, 40]. The last one is
simply is an equivalent vector reformulation of the statement (b).
♦
With this result in mind, we can now derive an analogous result on the nonnegative orthant that involves
equality instead of inequality constraints. Based on the linear S-procedure, the following result can be
proved:
Proposition 4.2. Let σi ∈ Rn , i = 0, . . . , N , and assume that the set
S0+ := y ∈ Rn≥0 : σiT y = 0, i = 1, . . . , N
(10)
is nonempty. Then, the following statements are equivalent:
(a) We have that σ0T y ≥ 0 for all y ∈ S0+ .
(b) There exist some multipliers ηi ∈ R, i = 1, . . . , N , such that the inequality
σ0T y
+
N
X
ηi σiT y ≥ 0
(11)
i=1
holds for all y ∈ Rn≥0 .
(c) There exists a multiplier vector η ∈ RN such that σ0T − η T σ̄ y ≥ 0 for all y ∈ Rn≥0 where σ̄ :=
T
colN
i=1 (σi ).
T
(d) There exists a multiplier vector η ∈ RN such that σ0T − η T σ̄ ≥ 0 where σ̄ := colN
i=1 (σi ).
Proof : Defining
S0+
S≤0 := y ∈ Rn : −σiT y ≥ 0, i = 1, . . . , N
(12)
S≥0 ∩ S≤0 ∩ Rn≥0 .
then we can see that
=
Applying then the linear S-procedure, i.e. Lemma 4.1, we get
that the statement (a) is equivalent to the existence of some scalars τi ≥ 0, i = 1, . . . , 2N , such that
σ0T y +
N
X
[τi − τi+N ] σiT y ≥ 0
i=1
for all y ∈ Rn≥0 . Letting then ηi := τi − τi+N ∈ R yields the equivalence between the two first statements.
The equivalence between the three last statements follows from the facts that
σ0T y −
N
X
i=1
ηi σiT y = σ0T − η T σ̄ y
n
T
n
with η = colN
i=1 (ηi ) and that for a given vector c ∈ R , we have that c y ≥ 0 for all y ∈ R≥0 if and only if
T
c ≥ 0.
♦
We can now state the following important result:
n ×nij
i
Theorem 4.3. Let Ai ∈ MRni and Bij ∈ R≥0
statements are equivalent:
(a) The Metzler matrix
is Hurwitz stable.
A1
..
.
BN 1 CN 1
n ×nj
ij
, Cij ∈ R≥0
B12 C12
BN 2 CN 2
8
, i, j = 1, . . . , N , i 6= j. Then, the following
. . . B1N C1N
..
.
...
AN
(13)
i
(b) There exist vectors vi ∈ Rn>0
, ℓij ∈ Rnij , i, j = 1, . . . , N , i 6= j, such that the inequalities
viT Ai +
N
X
ℓTji Cji
j6=i
viT Bij
−
ℓTij
< 0
(14)
≤ 0
hold for all i, j = 1, . . . , N , j 6= i.
Proof : The equivalence between the two statements can be shown using Proposition 4.2. From Lemma
2.8, a Metzler matrix A ∈ MRn is Hurwitz stable if and only if there exists a vector v ∈ Rn>0 such that
v T A < 0, which is equivalent to the existence of a vector v ∈ Rn>0 and a small enough scalar ε > 0 such that
v T Ax ≤ −εv T x for all x ∈ Rn≥0 . By applying this result to the matrix (13) and decomposing v =: colN
i=1 (vi ),
ni
i
vi ∈ Rn>0
, and x =: colN
(x
),
x
∈
R
,
following
the
same
partitioning,
we
get
that
the
matrix
(13) is
i
i
i=1
>0
Hurwitz stable if and only if the following feasibility problem has a solution:
Find
(P)
s.t.
i
vi ∈ Rn>0
, ε > 0
N
X
i=1
for all
viT
Ai xi +
N
X
j=1
j6=i
N
X
viT xi ,
≤
−ε
Bij Cij xj
i=1
i
xi ∈ Rn≥0
, i = 1, . . . , N
Letting now wij := Cij xj , i, j = 1, . . . , N , i 6= j, we get that the inequality constraint in Problem (P) is
equivalent to the fact that the inequality
N
X
i=1
N
X
Bij wij
viT
(A
+
εI)x
+
i
i
≤ 0.
(15)
j=1
j6=i
Hence, we get a problem analogous to that of Proposition 4.2 with the difference that we have vector
equality constraints instead of scalar ones. However, by virtue of the statement (c) or (d) of Proposition
4.2, vector of constraints can be dealt using vector multipliers (this can also be retrieved by scalarizing the
constraints, considering scalar multipliers and vectorizing the result). Invoking then Proposition 4.2, (b),
adapted to vector equality constraints, we get that the above inequality condition holds under the vector
equality constraints if and only if there exist some vector multipliers ℓij ∈ Rnij , i, j = 1, . . . , N , i 6= j, such
that
N
X
i=1
N
N
X
X
+
ℓTij (Cij xj − wij ) ≤ 0
B
w
(A
+
εI)x
+
viT
ij
ij
i
i
(16)
i,j=1
j6=i
j=1
j6=i
n
ij
i
holds for all xi ∈ Rn≥0
, i, j = 1, . . . , N , i 6= j. Reorganizing the terms in the above
and all wij ∈ R≥0
inequality yields
N
N
N
X
X
X
T
T
v (Ai + εI) +
viT Bij − ℓTij wij ≤ 0
ℓji Cji xi +
i
i=1
(17)
i,j=1
j6=i
j=1
j6=i
where the swapping of the indices in the ℓij stems from the fact ℓij acts on terms depending on xj while
ℓji acts on terms depending on xi . Using finally the equivalence between the statements (c) and (d) of
nij
i
Proposition 4.2, we get that the feasibility of (17) for all xi ∈ Rn≥0
is equivalent to the
and all wij ∈ R≥0
feasibility of
N
X
ℓTji Cji ≤ 0 and viT Bij − ℓTij ≤ 0
(18)
viT (Ai + εI) +
j=1
j6=i
9
for all i, j = 1, . . . , N , i 6= j, which is equivalent to the conditions (14). The proof is complete.
♦
We can see from this result that the stability of the overall block-matrix can be broken down to the
stability analysis of N subproblems (one for each of the vi ’s) involving coupled multipliers that explicitly
capture the topology of the interconnection. An important advantage of this formulation lies in its convenient
form allowing for the derivation of efficient parallel algorithms for establishing the Hurwitz stability of Metzler
block-matrices.
4.2. Main results
Before stating the main result of the section, we define the Minkowski product between two qualitative
classes associated with nonnegative sign-matrices as the set
Q(A1 )Q(A2 ) := {M1 M2 : M1 ∈ Q(A1 ), M2 ∈ Q(A2 )}
(19)
n1 ×m
2
where A1 ∈ S≥0
and A2 ∈ Sm×n
. It is also important to introduce the three following product rules for
≥0
dealing with products of nonnegative sign-matrices: 1) ⊕ + ⊕ = ⊕; 2) ⊕ · ⊕ = ⊕; 3) 0 · ⊕ = 0. It seems
that product of sign-matrices has never been introduced before and is defined for the first time. In addition,
the Minkowski products of qualitative classes also seems to be new. A possible reason comes from the fact
that the set of sign-matrices is not closed under addition or multiplication. However, the set of nonnegative
sign-matrices is closed under these operations.
Remark 4.4. Note that because of these product rules, we have Q(A1 )Q(A2 ) ⊂ Q(A1 A2 ) and, in general,
Q(A1 )Q(A2 ) 6= Q(A1 A2 ). Similarly, sgn(A1 A2 ) 6= sgn(A1 ) sgn(A2 ) in general.
Remark 4.5. Note that because of these product rules, we have that Q(A1 )Q(A2 ) ⊂ Q(A1 A2 ) and that, in
general, Q(A1 )Q(A2 ) 6= Q(A1 A2 ). For instance, if
⊕
and A2 = ⊕ ⊕ ,
A1 =
⊕
then Q(A1 )Q(A2 ) is the set of all 2 × 2 rank-1 positive matrices. However, since
⊕ ⊕
,
A1 A2 =
⊕ ⊕
then Q(A1 A2 ) is the set of all 2 × 2 positive matrices. On the other hand, if
⊕
and A2 = ⊕ 0 ,
A1 =
⊕
then we have that Q(A1 )Q(A2 ) = Q(A1 A2 ). A similar statement can be made for unit sign-matrices. We
indeed have, in general, that sgn(A1 A2 ) 6= sgn(A1 ) sgn(A2 ). For instance, this is the case for the matrices
⊕
⊕ ⊕
,
and A2 =
A1 =
⊕
0 ⊕
for which we have that
sgn(A1 A2 ) =
2
1
.
and sgn(A1 ) sgn(A2 ) =
1
1
The above remark emphasizes the difficulty arising from the possible loss of independence between the
entries of a matrix resulting from the multiplication of two sign-matrices. With this warning in mind, we
can state the following result:
10
n ×n
n ×n
ij
j
i
ij
. Assume, moreover, that Q(B̃ij C̃ij ) =
Theorem 4.6. Let Ãi ∈ MSni and B̃ij ∈ S≥0
, C̃ij ∈ S≥0
Q(B̃ij )Q(C̃ij ) for all i, j = 1, . . . , N , i 6= j. Then, the following statements are equivalent:
(a) The Metzler sign-matrix
. . . B̃1N C̃1N
..
.
...
ÃN
B̃12 C̃12
Ã1
..
à =
.
B̃N 1 C̃N 1
B̃N 2 C̃N 2
(20)
is sign-stable.
(b) The matrix sgn(Ã) is Hurwitz stable.
i
(c) There exist vectors vi ∈ Rn>0
, ℓij ∈ Rnj , i, j = 1, . . . , N , j 6= i, such that the inequalities
viT sgn(Ãi ) +
N
X
ℓTji
< 0
ℓTij
< 0
(21)
j6=i
viT
sgn(B̃ij C̃ij ) −
for all i, j = 1, . . . , N , j 6= i.
(d) The diagonal elements of à are negative and the directed graph Dà is acyclic.
Proof : The assumption that Q(B̃ij C̃ij ) = Q(B̃ij )Q(C̃ij ) for all i, j = 1, . . . , N , i 6= j, indicates that
multiplications between nonnegative sign-matrices do not create qualitative classes that strictly include the
initial ones. This indicates, by virtue of Theorem 3.4 and Remark 4.5, that the sign-stability of the matrix
à is equivalent to the Hurwitz stability of sgn(Ã), which proves the equivalence between (a) and (b). The
equivalence with (d) also follows from Theorem 3.4 and Remark 4.5. We prove now the equivalence between
(b) and (c). To this aim, let us consider the matrix
sgn(Ã1 )
..
sgn(Ã) =
.
sgn(B̃N 1 C̃N 1 )
. . . sgn(B̃1N C̃1N )
..
.
.
sgn(ÃN )
sgn(B̃N 2 C̃N 2 ) . . .
sgn(B̃12 C̃12 )
(22)
Note that, by virtue of Remark 4.5, we have that sgn(B̃ij C̃ij ) 6= sgn(B̃ij ) sgn(C̃ij ) in general. We now
observe that the above matrix is of the form of the matrix (13) where Ai = Ãi , Bij = sgn(B̃ij C̃ij ) and
♦
Cij = Inj . Applying then Theorem 4.3 yields the equivalence between the statements (b) and (c).
The above result admits a potentially simpler formulation from a computational viewpoint (in the case
where nij < nj ) whenever the assumption sgn(B̃ij C̃ij ) = sgn(B̃ij ) sgn(C̃ij ) holds for all i, j = 1, . . . , N ,
i 6= j. In this case, we have the following result:
n ×n
n ×n
ij
j
i
ij
. Assume, moreover, that Q(B̃ij C̃ij ) =
Theorem 4.7. Let Ãi ∈ MSni and B̃ij ∈ S≥0
, C̃ij ∈ S≥0
Q(B̃ij )Q(C̃ij ) and sgn(B̃ij C̃ij ) = sgn(B̃ij ) sgn(C̃ij ) for all i, j = 1, . . . , N , i 6= j. Then, the following
statements are equivalent:
(a) The matrix à in (20) is sign-stable.
i
(b) There exist vectors vi ∈ Rn>0
, ℓij ∈ Rnij , i, j = 1, . . . , N , j 6= i, such that the inequalities
viT
sgn(Ãi ) +
N
X
ℓTji sgn(C̃ji ) <
j6=i
viT sgn(B̃ij )
for all i, j = 1, . . . , N , j 6= i.
11
−
ℓTij
<
0
0
(23)
Proof : The proof follows the same lines as the proof of Theorem 4.6 with the difference that we identify
the matrix sgn(Ã) with the matrix (13) together with Ai = Ãi , Bij = sgn(B̃ij ) and Cij = sgn(C̃ij ), and
where we have used the assumption that sgn(B̃ij C̃ij ) = sgn(B̃ij ) sgn(C̃ij ) for all i, j = 1, . . . , N , i 6= j. The
rest of the proof follows from the application of Theorem 4.3.
♦
When the assumption that Q(B̃ij C̃ij ) = Q(B̃ij )Q(C̃ij ) for all i, j = 1, . . . , N , i 6= j, is removed in the
above results, then the conditions stated in the statements (b) and (c) are only sufficient for (a) to hold, but
not necessary.
4.3. Example
Let us consider the following matrices
A1
A2
=
=
⊖
⊕
⊖
0
0
, B12
⊖
0
, B21
⊖
=
=
⊕
, C21
0
0
, C12
⊕
T
0
,
⊕
T
⊕
.
=
0
=
(24)
Clearly, the assumptions of Theorem 4.6 are satisfied and the conditions (23) write
−v11 + v12
−v21 + ℓ12
< 0, −v12 + ℓ21
< 0,
−v22
<
<
0, v11 − ℓ12
0, v22 − ℓ21
≤ 0,
≤ 0
(25)
where vi := col(vi1 , vi2 ). These conditions are equivalent to v12 < v11 , ℓ12 ∈ [v11 , v21 ) and ℓ21 ∈ [v22 , v12 ). The
conditions are readily seen to be feasible. For instance, we may pick v1 = (5, 3), v2 = (7, 1), ℓ12 = 6 and
ℓ21 = 2, proving then that the block matrix Ā is sign-stable. If, however, the matrix A2 is changed to
⊖ ⊕
,
A2 =
0 ⊖
then Ā is not sign-stable anymore and the conditions (23) are infeasible.
5. Sign-stability of a convex-hull of Metzler matrices
We address in this section the problem of establishing whether, for some given family of Metzler signmatrices Ai ∈ MSn , i = 1, . . . , N , all the matrices in the convex-hull co(A1 , . . . , AN ) (when it is well-defined)
are sign-stable. We first derive several intermediate results that will allow us to state under what conditions
on the matrices A1 , . . . , AN , the convex-hull co(A1 , . . . , AN ) exists; i.e. co(A1 , . . . , AN ) ⊂ MSn . We will
show that this problem is equivalent to the sign-summability problem of the matrices A1 , . . . , AN . Then,
we show that establishing whether all the matrices in the convex hull of the matrices in co(A1 , . . . , AN )
are sign-stable is equivalent to checking the sign-stability of N + 1 Metzler sign-matrices, a problem having
solutions that can be formulated in various ways including graph theoretical conditions or linear programs.
The second problem is concerned with the existence of common quadratic/linear Lyapunov functions for a
sign-summable family of Metzler sign-matrices A1 , . . . , AN . It will be shown that a necessary and sufficient
condition for the existence of such Lyapunov functions is that all the matrices in the convex-hull be signstable. This result can be opposed to a well-known result stating that for a family of Metzler matrices,
the stability of the associated convex-hull does not necessarily imply the existence of common Lyapunov
function; e.g. [34].
Definition 5.1. Let A ⊂ MSn be a set of Metzler sign-matrices. The qualitative class of A is defined as
[
Q(A) =
Q(M ).
(26)
M∈A
12
Definition 5.2. Let Ai ∈ MSn , i = 1, . . . , N , be given Metzler sign-matrices. We say that the convex-hull
of the matrices A1 , . . . , AN defined as
)
(N
N
X
X
αi = 1
(27)
αi Ai : α ≥ 0,
co(A1 , . . . , AN ) :=
i=1
i=1
n
is well-defined if co(A1 , . . . , AN ) ⊂ MS .
Definition 5.3. Let us consider a family of Metzler sign-matrices AiP
∈ MSn , i = 1, . . . , N . We say that
n
the family is sign-summable or that the matrices are sign-summable if N
i=1 Ai ∈ MS .
Nonnegative matrices are trivially sign-summable as all the entries are nonnegative and, hence, the signsummability problem did not occur in the previous section. However, the sign of the diagonal elements of
Metzler matrices is indefinite and this may cause their sum to have indefinite sign. A simple necessary and
sufficient condition for the sign-summability of Metzler sign-matrices is given in the result below together
with a full characterization of well-defined convex-hulls of Metzler sign-matrices:
Proposition 5.4. Let us consider a family of Metzler sign-matrices Ai ∈ MSn , i = 1, . . . , N . Then, the
following statements are equivalent:
(a) The matrices A1 , . . . , AN are sign-summable.
(b) For each fixed j = 1, . . . , n, the non-zero elements in the set {[Ai ]jj : i = 1, . . . , N } have the same sign.
(c) The convex-hull co(A1 , . . . , AN ) is well-defined and we have that
)
(N
X
N
βi Ai : β ∈ {0, 1} , β 6= 0
(28)
co(A1 , . . . , AN ) =
i=1
together with
Q(co(A1 , . . . , AN )) =
[
β∈{0,1}N
β6=0
Q
N
X
i=1
βi Ai
!
.
(29)
Proof : Proof that (a) is equivalent to (b). First note that because of the Metzler structure of the
matrices, the nonnegative off-diagonal elements can be summed freely and only the diagonal elements are
problematic. In this regard, the fact that (a) implies (b) is immediate. To prove the converse, we use
contraposition and assume that (a) does not hold. This implies there exists a k ∈ {1, . . . , n} for which the
set {[Ai ]kk : i = 1, . . . , N } contains at least one positive entry and one negative entry. Let i1 , i2 ∈ {1, . . . , N }
be such that [Ai1 ]kk = ⊕ and [Ai2 ]kk = ⊖. This then implies that the sum Ai1 + Ai2 does not exist since
the entry [Ai1 + Ai2 ]kk = ⊕ + ⊖ has indefinite sign.
Proof that (a) is equivalent to (c). Assume that the statement (c) holds and hence that co(A1 , . . . , AN )
is well-defined or, equivalently,
thatPall the matrices in the convex-hull are P
Metzler sign-matrices. Observing
P
N
N
1
now that the matrix N
A
=
A
∈
co(A
,
.
.
.
,
A
)
implies
that
1
N
i=1 i
i=1 i
i=1 Ai is a Metzler sign-matrix,
N
which implies in turn that the matrices A1 , . . . , AN are sign-summable and that (a) holds. Assume now that
(a) holds. To prove that the convex-hull is well-defined, first note that for any αi > 0, we have that αi Ai = Ai
as Q(αi Ai ) = Q(Ai ) and that for αi = 0 we have that αi Ai = 0n×n since Q(αi Ai ) = {0n×n }. Hence, only
the sign of the αi ’s matters (i.e. whether they are zero or positive) and, in this respect, we can substitute
each αi in the definition of the convex-hull by a binary parameter βi ∈ {0, 1} where βi = 1 if αi > 0 and
0 otherwise. This naturally leads to the alternative expression for the convex-hull given in (28). Using now
the sign-summability property of the matrices implies that the convex-hull is well-defined. The expression
(29) for the qualitative-class of the convex-hull then immediately follows from Definition 5.1.
♦
Now that the sign-summability of Metzler sign-matrices has been fully characterized, we can move on to
the main objective of the section, that is, the characterization of the sign-stability of all the matrices in a
well-defined convex-hull of Metzler sign-matrices:
13
Theorem 5.5. Let us consider a family of Metzler sign-matrices Ai ∈ MSn , i = 1, . . . , N . Then, the
following statements are equivalent:
(a)
(b)
(c)
(d)
The convex-hull co(A1 , . . . , AN ) is well-defined and only contains sign-stable matrices.
PN
For all β ∈ {0, 1}N , β 6= 0, the matrix A(β) := i=1 βi Ai exists and is sign-stable.
P
The matrices A1 , . . . , AN have negative diagonal elements and the matrix N
i=1 Ai is sign-stable.
The diagonal elements of the matrices A1 , . . . , AN are negative and the directed graph DĀ is acyclic.
Proof : The equivalence between the statements (a) and (b) follows from Proposition 5.4 while the equivalence between the statements (c) and (d) follows from Theorem 3.4. Let us prove then the equivalence
between (b) and (c). Assume that (b) holds. Hence, for all β ∈ {0, 1}N , β 6= 0, the matrix A(β) exists
N
and is sign-stable. Let {ei }N
i=1 be the natural basis of R , then we get that A(ei ) = Ai . Therefore, for any
i = 1, . . . , N , the matrix Ai is sign-stable
and, hence, it has negative diagonal elements. Considering, finally,
PN
β = 1N , we get that A(1N ) = i=1 Ai is sign-stable and the statement (c) holds.
PK
Assume now that (c) holds and let A(β) := i=1 βi Ai . From the facts that the Metzler matrices Ai have
PN
negative diagonal elements and β 6= 0, then we can state that A(β) ≤ i=1 Ai for all β ∈ {0, 1}N , β 6= 0.
PN
Hence, we have that sgn(A(β)) ≤ sgn( i=1 Ai ) and thus
!!
N
X
for all β ∈ {0, 1}N , β 6= 0
(30)
Ai
λP F (sgn(A(β))) ≤ λP F sgn
i=1
PN
and where λP F (·) denotes the Perron-Frobenius
eigenvalue. By virtue of Theorem 3.4, we know that i=1 Ai
P
N
is sign stable if and only if sgn
i=1 Ai is Hurwitz stable. As a consequence, the inequality (30) implies
PN
that if i=1 Ai is sign stable, then A(β) is sign stable for any β ∈ {0, 1}N , β 6= 0. The proof is complete. ♦
5.1. Existence of common Lyapunov functions
We first recall the following result from [41] that we slightly adapt to our setup:
Lemma 5.6 ([41]). Let us consider N upper-triangular matrices Ai ∈ Rn×n , i = 1, . . . , N . Then, the
following statements are equivalent:
(a) The matrices Ai , i = 1, . . . , N , are Hurwitz stable.
(b) There exists a symmetric positive definite matrix Q ∈ Rn×n such that ATi Q + QAi is negative definite
for all i = 1, . . . , N .
We then have the following result:
Theorem 5.7. The following statements are equivalent:
(a) All the matrices in co(A) are sign-stable.
(b) For all (A′1 , . . . , A′N ) ∈ Q(A1 ) × . . . × Q(AN ), there exists a positive definite matrix Q ∈ Rn×n such that
ÃT Q + Qà is negative definite for all à ∈ co{A′1 , . . . , A′N }.
(c) For all (A′1 , . . . , A′N ) ∈ Q(A1 ) × . . . × Q(AN ), there exists a positive vector v ∈ Rn>0 such that v T Ã < 0
for all à ∈ co(A′ ).
Proof : The implication that (b) implies (a) and (c) implies (a) are immediate. Assume that (a) holds and
note that, by virtue of Theorem 3.4, the sign-stability of the Ai ’s and their sums is equivalent to saying
that they can be simultaneously expressed as Hurwitz stable upper-triangular matrices through appropriate
simultaneous row/column permutations. Assuming now that they are in such a form then the conclusion
directly follows from Lemma 5.6. The proof that (a) implies (c) follows from fact that for all upper-triangular
Hurwitz stable A′i ∈ Q(Ai ), i = 1, . . . , N , we can recursively construct a v ∈ Rn>0 such that v T A′i < 0 for all
i = 1, . . . , N .
♦
14
6. Ker+ (B)-sign-stability of Metzler sign-matrices
6.1. Preliminaries
In this section, we will work with a larger class of sign-matrices, referred to as indefinite sign-matrices,
taking entries in S⊙ := S ∪ {⊙} where ⊙ denotes a sign indefinite entry that can be obtained from the rule
⊕ + ⊖ = ⊙. To avoid confusion, sign-matrices will sometimes be referred to as definite sign-matrices.
n×m
Definition 6.1 (Sign-qualitative class). The sign-qualitative class associated with a matrix A ∈ S⊙
is defined as
∈S
if [A]ij = ⊙
n×m
SQ(A) := M ∈ S
.
[M ]ij
= [A]ij otherwise
It is important to stress that, due to the presence of sign-indefinite entries ⊙, all the matrices in SQ(A) do
not have the same sign-pattern. When there is no sign-indefinite entry in A, then SQ(A) = A. In accordance
with Definition 5.1, we have that
[
Q(A′ )
(31)
Q(SQ(A)) :=
A′ ∈SQ(A)
which defines the set of all real matrices sharing the same sign-pattern of at least one matrix in the signqualitative class of the matrix A.
Let us define now the concept of Ker+ (B)-sign-stability:
Definition 6.2. Let A ∈ MSn×n and B ∈ Rn×ℓ , ℓ < n, full-rank. We say that A is Ker+ (B)-sign-stable
if for all A′ ∈ Q(A), there exists a v ∈ Rn>0 such that v T A′ < 0 and v T B = 0.
The main reason for considering the above concept is that finding such a vector is a problem which arises,
for instance, in the analysis of certain stochastic Markov jump processes that can represent certain biological
processes [4, 6, 42]. When B = 0, the usual notion of sign-stability is retrieved.
We will also need the concept of an inverse of a sign-matrix. This existence problem of inverses of signmatrices has been well-studied in the literature; see e.g. [26, 27, 43, 44]. We consider here the following
definition:
Definition 6.3. We say that the matrix A ∈ MSn is invertible if there exists a matrix à ∈ Sn×n such that
for all M ∈ Q(A), there exists a matrix N ∈ Q(Ã) such that M N = I. When this is the case, we denote the
inverse of A, Ã, by A−1 .
We then have the following result:
Theorem 6.4. Assume that the sign-matrix A ∈ MSn is sign-stable. Then,
(a) the inverse of the sign-matrix A exists, is nonpositive, and is such that [A−1 ]ii = ⊖, i = 1, . . . , n, and
⊖
if there is a path from node j to node i in the directed graph DA .
[A−1 ]ij =
0
otherwise
for all i, j = 1, . . . , n, i 6= j.
(b) we have that Q(A)−1 := M −1 : M ∈ Q(A) ⊂ Q(A−1 ).
Proof : To prove statement (a), we first assume, without loss of generality, that the sign-matrix A
is upper-triangular with negative diagonal elements. The key idea behind the proof is to demonstrate the
existence of an inverse for the matrix A by showing that the inverses of all the matrices in the qualitative
class Q(A) all share the same sign-pattern. To this aim, let us introduce the following matrix
M (m) := −
n
X
mi,i ei eTi +
n
X
j>i
i
15
mi,j ei eTj
(32)
where mi,j ≥ 0 for all i, j = 1, . . . , N , j ≥ i. Note, moreover, that we have
Q(A) = {M (m) : mi,j > 0 if (i, j) ∈ I(A), mi,j = 0 otherwise}
(33)
I(A) := {(i, j) : [A]ij 6= 0, i, j = 1, . . . , N, j ≥ i} .
(34)
where
Since, by assumption, the matrix A is sign-stable, then its diagonal elements are negative (and hence mi,i > 0
for all i = 1, . . . , N ), the directed graph DA is acyclic and all the matrices in Q(A) are invertible. Considering
further the system of linear equations given by y = M (m)x and recursively solving for x yields
xn
=
xn−1
=
=
xn−2
=
..
.
xn−k
=
−1
yn
mn,n
1
(−yn−1 + mn−1,n xn )
mn−1,n−1
−1
mn−1,n
yn−1 +
yn
mn−1,n−1
mn,n
−1
mn−2,n−1
mn−2,n
mn−2,n−1 mn−1,n
yn−2 +
yn
yn−1 +
+
mn−2,n−2
mn−1,n−1
mn−1,n−1 mn,n
mn,n
1
mn−k,n−k
−yn−k +
n
X
mn−k,i xi
i=n−k+1
!
(35)
, k = 3, . . . , n − 1.
As expected the diagonal entries of M (m)−1 are just the inverse of the diagonal entries of M (m). The
off-diagonal entries of M (m)−1 are defined recursively over all possible paths in the graph DA from one
node to another. For instance, the entry [M (m)−1 ]n−2,n is nonzero if there is a path of length one from
the node n to the node n − 2 in the directed graph DA (i.e. mn−2,n 6= 0) or a path of length two passing
through the node n − 1 (i.e. mn−2,n−1 mn−1,n 6= 0). This argument easily generalizes to all the entries
of the inverse of the matrix M (m)−1 by exploiting the acyclicity of the graph DA and the nonnegativity
of the off-diagonal elements of M (m). As a consequence, all the nonzero entries are necessarily negative
and, hence, the matrix M (m)−1 is nonpositive. Uniqueness follows from the fact the sign-pattern of the
inverse only depends on the paths in the graph DA which are, in turn, uniquely defined by the sign-matrix
A. Consequently, an entry of M (m)−1 is 0 if and only if there is no path from j to i in DA , otherwise it
is negative. This proves statement (a). To prove statement (b), first remark that since A is a sign-stable
Metzler sign-matrix, then from statement (a) we have that the matrix A−1 is well-defined and is nonpositive
with negative diagonal elements. Hence, statement (b) holds since all the inverses of the matrices in Q(A)
have the same sign-pattern as A−1 . The proof is complete.
♦
Two important facts are stated in the above result. First of all, the sign-pattern of the inverse signmatrix A−1 is uniquely defined by the sign-pattern of A provided that A is sign-stable. Secondly, that for
any sign-stable matrix A ∈ Sn×n , the nonzero entries of the inverse matrix A−1 can simply be deduced from
the examination of the graph DA . For comparison, it requires at most n2 operations, which is better than
usual real matrix inversion algorithms; see e.g. [45]. A drawback, however, is that by considering the matrix
A−1 , we lose some information since the class of matrices that is considered is larger.
Remark 6.5. It is important to stress that, for any A ∈ MSn , A sign-stable, we have that Q(A)−1 6=
Q(A−1 ). Indeed, the inverse of a Hurwitz stable real nonpositive matrix with negative diagonal is not necessarily Metzler, even in the upper-triangular case. In other words, the matrix inversion operation on the set
of sign-stable upper-triangular nonpositive sign-matrices is not well-defined in the sense that the sign-pattern
16
is non-uniquely defined. For instance, we have that
−1 1
1
−1 −1 −2 −4
0 −1 −1 −5
but M −1 = 0 −1 1
M =
0
0
0 −1
0 −1 −1
0
0
0
0
0
0 −1
−2
4
1
−1
(36)
is not a Metzler matrix.
The following concept will be crucial for stating the main result of this section:
Definition 6.6 (L+ -matrices, [36]). Let A ∈ Sm×n . We say that A is an L+ -matrix if for all A′ ∈ Q(A),
the dual cone of {A′ x : x ≥ 0} defined as {y : y T A′ ≥ 0} is equal to {0}.
Let us also recall the following result:
Theorem 6.7 ([36], Theorem 2.4). Let R ∈ Sm×n . Then, the following statements are equivalent:
(a)
(b)
(c)
(d)
(e)
R is an L+ -matrix.
R has no zero row and for all R′ ∈ Q(R), there exists a v ∈ Rn>0 such that R′ v = 0.
For all nonzero diagonal matrices D ∈ {−1, 0, 1}m×m, some column of DR is nonzero and nonnegative.
For all nonzero diagonal matrices D ∈ {−1, 0, 1}m×m, some column of DR is nonzero and nonpositive.
For all R′ ∈ Q(R), we have that {R′ x : x ∈ Rn≥0 } = Rm ; i.e. the cone generated by R′ is all of Rm .
We are now ready to state the main result of this section which provides sufficient conditions for the
Ker+ (B)-sign-stability of a given matrix A ∈ MSn :
Proposition 6.8. Let A ∈ MSn×n be sign-stable and B ∈ Rn×ℓ , ℓ < n, be full-rank. Assume that one of
the following equivalent statements holds:
(a) All the matrices in SQ(B T A−T ) are L+ -matrices.
(b) All the matrices in SQ(B T A−T ) have no zero row and for all M ∈ Q(SQ(B T A−T ), there exists a
w ∈ Rn>0 such that M w = 0.
(c) For all nonzero diagonal matrices D ∈ {−1, 0, 1}ℓ×ℓ and all M ∈ SQ(B T A−T ), some column of DM is
nonzero and nonnegative.
(d) For all nonzero diagonal matrices D ∈ {−1, 0, 1}ℓ×ℓ and all M ∈ SQ(B T A−T ), some column of DM is
nonzero and nonpositive.
(e) For all M ∈ Q(SQ(B T A−T ), we have {M x : x ∈ Rn≥0 } = Rℓ .
Then, the matrix A is Ker+ (B)-sign-stable.
Proof : The equivalence between the statements of the result directly follows from Theorem 6.7. Hence,
we simply need to prove that (b), for instance, implies that the matrix A is Ker+ (B)-sign-stable. From
statement (b), we have that for all M ∈ Q(SQ(B T A−T ), there exists a w ∈ Rn>0 such that M w = 0 and
observe, moreover, that the following inclusions hold
Q(A)−1 B ⊆ Q(A−1 )B ⊆ Q(SQ(A−1 B)).
(37)
The first inclusion follows from the statement (b) of Theorem 6.4 while the second one is immediate from
the fact that any matrix in Q(A−1 )B necessarily belongs to Q(SQ(A−1 B)) since SQ(A−1 B) contains all
the possible sign patterns that Q(A−1 )B could have. Hence, if all the matrices in SQ(B T A−T ) are L+ matrices, then for all M ∈ B T Q(A)−T , there exist a vector w ∈ Rn>0 such that M w = 0. Decomposing M
as M = B T Ã−T , where à ∈ Q(A) is Hurwitz stable, yields wT Ã−1 B = 0. Letting v T := −wT Ã−1 , we get
that v > 0 since w > 0 and Ã−1 ≤ 0. Hence, we have that −v T B = 0 or, equivalently, v T B = 0 together
17
with v T Ã = −wT < 0, which proves that the matrix A is Ker+ (B)-sign-stable.
♦
From the statements (c) and (d) of Theorem 6.7, we can see that checking whether a matrix is an L+ matrix is a combinatorial problem which turns out to be NP-complete (see [25, 36]). However, the approach
remains applicable for matrices R having a small number of rows. Applied to our problem, the number of
sign-matrices on which to apply the column/row condition (see the statements (c) and (d) of Proposition
6.8) is equal to (3ℓ − 1)p sign-matrices where p is the cardinal of SQ(B T A−T ). Interestingly, the complexity
does not directly depend on n, but indirectly depends on n through the cardinality p since a larger n is likely
to lead to a higher cardinality for the set SQ(B T A−T ). Finally, it is important to stress than the problem
remains tractable as long as the values of ℓ and p are sufficiently small.
7. Sign-stability of mixed matrices
As we have seen in the previous sections, sign-stability requires strong structural properties for the
considered sign-matrices. The idea here is to extend the scope of sign-stability to mixed matrices which
contain both sign and real entries. The idea is to enlarge the structures of the matrices (i.e. not necessarily
triangular) that can be considered using such an approach. In this case, we will shown that cycles can
be allowed for the nodes associated with real entries. We notably prove here that the sign-stability of a
class of mixed matrices can be exactly characterized in terms of Hurwitz stability conditions, sign-stability
conditions and the non-existence of cycles in a particular bipartite graph obtained from the mixed matrix.
7.1. Preliminaries
The following definition states an immediate extension of the sign-stability property of sign-matrices to
the case of mixed matrices:
n×n
Definition 7.1. A mixed matrix M ∈ (R ∪ S)
stable.
is sign-stable if all the matrices M ′ ∈ Q(M ) are Hurwitz
The following lemma states a result regarding the stability of Metzler block matrices which is analogous
to the Schur complement formula:
Lemma 7.2. Let us consider a matrix M ∈ MRn1 +n2 that we decompose as
M11 M12
M =:
M21 M22
(38)
1 ×n2
2 ×n1
where M11 ∈ MRn1 , M22 ∈ MRn2 , M12 ∈ Rn≥0
and M21 ∈ Rn≥0
. Then, the following statements are
equivalent:
(a) The matrix M is Hurwitz stable.
−1
(b) The matrices M11 and M22 − M21 M11
M12 are Hurwitz stable.
−1
(c) The matrices M22 and M11 − M12 M22 M21 are Hurwitz stable.
Proof : We only prove the equivalence between the two first statement. The proof of the equivalence between
the first and third statements is analogous. To this aim, assume that M is Hurwitz stable. Hence, there
exist two positive vectors λ1 , λ2 of appropriate dimensions such that we have
M11 M12 λ1
< 0.
(39)
M21 M22 λ2
Since the matrices M12 and M21 are nonnegative, the above inequality implies that M11 λ1 < 0 and M22 λ2 <
0, i.e. the matrices M11 and M22 are both Hurwitz stable. Moreover, the inequality (39) is equivalent to the
existence of a sufficiently small ǫ > 0 such that
M11 λ1 + M12 λ2 < 0 and M21 λ1 + M22 λ2 ≤ −ǫλ2 .
18
(40)
Solving for λ2 in the second inequality yields
λ2 ≥ −(M22 + ǫIn2 )−1 M21 λ1 ≥ 0
(41)
where the second inequality comes from the fact that M22 + ǫIn2 is Metzler and Hurwitz, which implies that
(M22 + ǫIn2 )−1 ≤ 0. Using now the fact that M12 ≥ 0, we get that
(42)
M11 − M12 (M22 + ǫIn2 )−1 M21 λ1 < 0.
and finally observing now that
implies
−1
− M22
≤ −(M22 + ǫIn2 )−1
−1
M11 − M12 M22
M21 λ1 < 0.
This proves the result. The proof of the converse simply consists of reversing the arguments.
(43)
(44)
♦
The following result, which is essential in proving the main result of this section, provides a graph
interpretation for the spectral radius of the product of two nonnegative matrices to be zero:
n×n
n×n
Lemma 7.3. Let us consider two square nonnegative matrices M1 ∈ R≥0
and M2 ∈ R≥0
. Then, the
following statements are equivalent:
(a) We have that ρ(M1 M2 ) = 0.
(b) There is no cycle in the directed bipartite graph B = (V1 , V2 , E) where Vi = {v1i , . . . , vni }, i = 1, 2,
E := E1 ∪ E2 ,
E1 := {(vi2 , vj1 ) ∈ V1 × V2 : [M1 ]ji 6= 0} and E2 := {(vi1 , vj2 ) ∈ V1 × V2 : [M2 ]ji 6= 0}.
Proof : Assume that (b) holds and let us define the matrix
0 S1
S :=
S2 0
(45)
(46)
n×n
where S1 , S2 ∈ S≥0
and sgn(Si ) = sgn(Mi ), i = 1, 2. Note, moreover, that the directed graph DS = (V, ES )
(see Definition 2.4) coincides with the graph B = (V1 , V2 , E) with V = V1 ∪ V2 and ES = E. Therefore, the
graph B is acyclic if and only if DS is. From Theorem 3.10, this is equivalent to saying that all the matrices
in Q(S) have zero spectral radius and, equivalently, that they can all be expressed in upper-triangular form
with zero diagonal elements modulo some permutation. Noting now that for any X ∈ Q(S) with Xi ∈ Q(Si ),
we have that ρ(X) = 0 if and only if ρ(X1 X2 ) = 0 implies that ρ(M1 M2 ) = 0 since Mi ∈ Q(Si ), i = 1, 2.
This proves the implication (b) ⇒ (a). Assume now that (a) holds and let
0 M1
M=
M2
0
from which we have that ρ (M ) = 0 ⇔ ρ(M1 M2 ) = 0. This then implies that M can be put in uppertriangular form modulo some permutation, which implies that the graph DM is acyclic. Noting, finally, that
the graph DM = (V, EM ) coincides with B yields the result.
♦
7.2. Main result
Let us consider in this section mixed matrices of the form
Aσ Cϕ
Aσϕ =
Bϕ Aϕ
n ×n
n ×n
(47)
σ
ϕ
ϕ
σ
where Aσ ∈ MSnσ , Aϕ ∈ MRnϕ , Bϕ ∈ R≥0
. The following result that can be seen as
and Cϕ ∈ R≥0
the “mixed matrix analogue” of Theorem 3.4 as it contains both algebraic and graph theoretical conditions
characterizing the sign-stability of mixed matrices. While graph theoretical conditions are easier to check,
algebraic ones can be combined with linear optimization problems in the same spirit as in Theorem 3.4:
19
Theorem 7.4. The following statements are equivalent:
(a) The mixed matrix Aσϕ defined in (47) is sign-stable.
(b) The following statements hold:
(b1) the matrix Aϕ is Hurwitz stable, and
−1
(b2) the matrix M − Mϕ where Mϕ := Cϕ Aϕ
Bϕ is Hurwitz stable for all M ∈ Q(Aσ ) or, equivalently,
−1
we have that ρ(M Mϕ ) < 1 for all M ∈ Q(Aσ ).
(c) The following statements hold:
(c1) Aσ is sign-stable,
(c2) Aϕ is Hurwitz stable,
(c3) we have that ρ(Mσ Mϕ ) = 0 where Mσ := sgn(A−1
σ ).
(d) The following statements hold:
(d1) Aσ is sign-stable,
(d2) Aϕ is Hurwitz stable,
(d3) There is no cycle in the directed bipartite graph B = (Vσ , Vϕ , E) where E = Eσ ∪ Eϕ with
Eσ := {viϕ , vjσ ) ∈ Vϕ × Vσ : [Mσ ]ji 6= 0} and Eϕ := {(viσ , vjϕ ) ∈ Vσ × Vϕ : [Mϕ ]ji 6= 0}.
(48)
(e) The following statements hold:
(e1) Aσ is sign-stable,
(e2) Aϕ is Hurwitz stable,
(e3) There is no cycle in the directed graph DAσϕ containing nodes in both Vσ = {v1σ , . . . , vnσσ } and
Vϕ = {v1ϕ , . . . , vnϕϕ } where Vσ contains the first nσ nodes associated with DAσϕ and Vϕ the nϕ last
ones.
(f) The following statements hold:
(f1) sgn(Aσ ) is Hurwitz stable,
(f2) Aϕ is Hurwitz stable,
(f3) The matrix sgn(Mσ Mϕ ) − I is Hurwitz stable.
Proof : Proof of (a) ⇔ (b). The matrix Aσϕ is sign-stable if and only if for all M ∈ Q(Aσ ), we have that
the matrix
M Cϕ
(49)
Bϕ Aϕ
is Hurwitz stable. Since the above matrix is Metzler, then we know from Lemma 7.2 that it is Hurwitz
−1
stable if and only if M and M − Cϕ Aϕ
Bϕ are both Hurwitz stable. Note also that from the fact that Aσ is
necessarily sign-stable, then there exists an M ∗ ∈ Q(Aσ ) such that M ∗ − Mϕ is Hurwitz stable. Hence, this
means that the matrix M − Cϕ A−1
ϕ Bϕ is Hurwitz stable for all M ∈ Q(Aσ ) if and only if det(M − Mϕ ) 6= 0
for all M ∈ Q(Aσ ). From the determinant formula, we get that this condition is equivalent to saying that
det(I − M −1 Mϕ ) 6= 0 for all M ∈ Q(Aσ ) and hence ρ(M −1 Mϕ ) < 1 for all M ∈ Q(Aσ ).
Proof of (b) ⇔ (c). We prove first that (b) implies (c). To this aim, assume that ρ(M −1 Mϕ ) < 1 for all
M ∈ Q(Aσ ) and that Aσ is in upper-triangular form with negative diagonal elements (this last statement
comes from the fact that Aσ is sign-stable and Theorem 3.4). Hence, from Theorem 6.4, its inverse is welldefined and we have that A−1
σ is nonpositive with negative diagonal elements. Since the diagonal elements of
A−1
(which
are
the
inverse
of the diagonal elements of Aσ ) can be arbitrarily large, then the spectral radius
σ
condition of statement (b) can only be satisfied if Mϕ is upper-triangular with zero diagonal elements, which
implies that ρ(M −1 Mϕ ) = 0 for all M ∈ Q(Aσ ) and, hence, that ρ(Mσ Mϕ ) = 0.
To prove the converse, let us assume that ρ(Mσ Mϕ ) = 0. Using the fact that the matrices Mσ and Mϕ
are both nonpositive and that Mσ has negative diagonal elements, then this implies that ρ(M −1 Mϕ ) = 0 for
all M ∈ Q(Aσ ) and, hence, that ρ(M −1 Mϕ ) < 1 for all M ∈ Q(Aσ ).
20
Proof of (c) ⇔ (d). The proof of this statement directly follows from Lemma 7.3.
Proof of (d) ⇔ (e). The equivalence follows from the fact that the cycles in DAσϕ that contain nodes in
both Vσ and Vϕ are exactly those in the graph B.
Proof of (c) ⇔ (f). Clearly the statements (c1) and (c2) are equivalent to (f1) and (f2), respectively. The
condition that ρ(Mσ Mϕ ) = 0 has been shown to be equivalent to saying that Mσ Mϕ is upper-triangular
modulo some cyclic permutation and that its diagonal elements are all equal to zero. Assuming then that
ρ(Mσ Mϕ ) = 0, this implies that the matrix sgn(Mσ Mϕ ) has diagonal elements equal to zero and is uppertriangular, which implies, in turn, that the Metzler matrix sgn(Mσ Mϕ ) − I is Hurwitz stable. To prove the
converse, let us assume that the Metzler matrix sgn(Mσ Mϕ ) − I is Hurwitz stable. This, then implies, from
Theorem 3.4 that it must be upper-triangular with negative elements on the diagonal, which implies then
that sgn(Mσ Mϕ ) is triangular with diagonal elements equal to zero. This proves the result.
♦
Remark 7.5 (Linear program condition). Interestingly, the conditions of statement (4) of Theorem 7.4
nϕ
σ
can be exactly reformulated as the following linear feasibility problem: Find v, z ∈ Rn>0
and w ∈ R>0
such
T
T
T
that v sgn(Aσ ) < 0, w Aϕ < 0 and z (sgn(Mϕ Mσ ) − I) < 0.
7.3. Example
Let us consider the matrix
Aσϕ
⊖ ⊕
0 ⊖
=
Bϕ
1
0
0
0
.
−1 2
1 −5
(50)
0
.
1
(51)
Clearly, Aσ is sign-stable and Aϕ is Hurwitz stable. Assume now that Bϕ is given by
Bϕ =
0
0
In this case, the matrices Mσ and Mϕ are given by
Mσ = −
and we have that
1
0
2 0 1
1
and Mϕ = −
1
3 0 0
2 0 1
Mσ Mϕ =
3 0 0
(52)
(53)
which has clearly zero spectral radius. Therefore, the matrix Aσϕ in (50) is sign-stable. The same conclusion
can be drawn by checking the no-cycle condition in the graphs.
Now, if we let
0 0
Bϕ =
1 0
then we have that
1 0
1 0 0
and Mσ Mϕ =
Mϕ = −
3 1 0
3 0
1
.
1
Clearly, we have that ρ(Mσ Mϕ ) = 1/3 and, hence, the matrix Aσϕ in (50) is not sign-stable.
21
8. Applications
The goal of this section is to apply some of the obtained results to some problems arising in the structural
analysis of linear positive time-delay systems (Section 8.1.1 and Section 8.1.2), linear positive switched
systems (Section 8.1.3) and linear positive impulsive systems (Section 8.1.4). The problem of establishing the
structural attractiveness and forward invariance of a compact set for a class of nonlinear positive dynamical
systems is considered in Section 8.2. Finally, the problem of establishing the structural ergodicity of a class
of continuous-time Markov jump process arising in the analysis of biochemical reaction networks is addressed
in Section 8.3.
8.1. Linear positive dynamical systems
8.1.1. Continuous-time systems with discrete delays
Linear positive systems with discrete-delays arise in problems such as in power-control in networks [46, 47]
and have been theoretically studied in several papers; see e.g. [9, 48]. Let us consider the following linear
system with constant discrete-delays [49–51]:
ẋ(t)
=
A0 x(t) +
N
X
Ai x(t − hi )
i=1
x(s)
=
(54)
φ(s), s ∈ [−h̄, 0]
where A0 , Ai ∈ Rn×n , hi > 0, i = 1, . . . , N , x ∈ Rn , φ ∈ C([−h̄, 0], Rn ) and h̄ := maxi {hi }. It is known [48]
that the above system is positive for all delays hi ≥ 0 if and only if (a) the matrix A0 is Metzler and (b)
the matrices Ai , i = 1, . . . , N , are all nonnegative;
and that it is asymptotically stable for any delays hi ≥ 0,
P
i = 1, . . . , N , if and only if the matrix N
A
is
Hurwitz
stable.
i=0 i
Remark 8.1 (Extension to time-varying delays). By virtue of the results in [9? ? ], the above result
remains valid when time-varying delays are considered instead of constant ones.
Before stating the main result, it seems important to recall that the sign-summability of Metzler and nonnegative matrices has been fully characterized in Proposition 5.4 (note that nonnegative matrices form a
subset of Metzler matrices). With this in mind, we can state the following result:
Proposition 8.2. Let M0 ∈ MSn and Mi ∈ Sn≥0 , i = 1, . . . , N . Then, the following statements are equivalent:
(a) For all A0 ∈ Q(M0 ) and all Ai ∈ Q(Mi ), i = 1, . . . , N , the linear positive system with delays (54) is
asymptotically stable; i.e. the linear positive system with discrete-delays (54) is structurally stable.
(b) For all A0 ∈ Q(M0 ) and all Ai ∈ Q(Mi ), i = 1, . . . , N , there exists a vector v ∈ Rn>0 such that
P
vT ( N
i=0 Ai ) < 0.
PN
(c) The matrices M0 , . . . , MN are sign-summable and the matrix i=0 Mi is sign-stable.
(d) The matrices M0 , . . . , MN are sign-summable and there exists a permutation matrix P such that the
PN
matrix P T ( i=0 Mi )P is upper-triangular.
P
N
M
< 0 holds.
(e) There exists a vector v ∈ Rn>0 such that v T sgn
i
i=0
Proof : The equivalence between the statements (b), (c), (d) and (e) follows from Theorem 3.4 while the
proof that (a) and (b) are equivalent follows from the stability result for linear positive systems with delays. ♦
22
8.1.2. Discrete-time systems with discrete delays
Let us now consider the case of linear discrete positive systems [52]. To this aim, we consider the following
system [53]
N
X
Ai x(k − hi )
x(k + 1) = A0 x(k) +
(55)
i=1
x(s) =
φs , s ∈ {−h̄, . . . , 0}
where A0 , Ai ∈ Rn×n , hi ∈ Z≥0 , i = 1, . . . , N , x ∈ Rn , φs ∈ Rn , s = −h̄, . . . , 0 and h̄ := maxi {hi }. It is
known [52] that the above system is positive for all for all delays hi ∈ Z≥0 if and only if the matrices Ai ,
i = 0, . . . , N , are all nonnegative;
and that it is asymptotically stable for any delays hi ≥ 0, i = 1, . . . , N
PN
if and only if the matrix i=0 Ai is Schur stable. With these results in mind, we can state the following
result:
Proposition 8.3. Let Mi ∈ Sn≥0 , i = 0, . . . , N . Then, the following statements are equivalent:
(a) For all Ai ∈ Q(Mi ), i = 0, . . . , N , the linear positive system with delays (55) is asymptotically stable;
i.e. the linear positive system with discrete-delays (55) is structurally stable.
PN
(b) For all Ai ∈ Q(Mi ), i = 0, . . . , N , there exists a vector v ∈ Rn>0 such that v T ( i=0 Ai − I) < 0.
P
(c) The matrix N
i=0 Mi is sign-stable.
PN
(d) There exists a permutation matrix P such that the matrix P T ( i=0 Mi )P is upper-triangular.
i
h
P
N
(e) There exists a vector v ∈ Rn>0 such that v T sgn
i=0 Mi − I < 0 holds.
Proof : The proof follows the same lines as the proof of Proposition 8.2 with the difference that Theorem
3.10 is used instead of Theorem 3.4.
♦
8.1.3. Continuous-time switched systems
Linear positive switched systems have been widely studied in the literature [32–34] as they can represent
a wide variety of real-world processes; see e.g. [54, 55]. Let us consider, as a starting point, the following
linear switched system [56]:
ẋ(t) = Aσ(t) x(t)
(56)
x(0) = x0
where x ∈ Rn and σ : R≥0 → {1, . . . , N } is a piecewise constant switching signal and Ai ∈ Rn×n . It is known
that [32, 33] the above linear switched system is positive if and only if Ai ∈ MRn for all i = 1, . . . , N .
Proposition 8.4. Let Mi ∈ MSn , i = 1, . . . , N . Then, the following statements are equivalent:
(a) For all Ai ∈ Q(Mi ), i = 1, . . . , N , the linear switched system (56) is asymptotically stable under arbitrary
switching i.e. the system (56) is structurally stable under arbitrary switching.
(b) For all Ai ∈ Q(Mi ), i = 1, . . . , N , the linear time-varying positive system
ẋ(t) = A(t)x(t)
(57)
where A(t) ∈ co(A1 , . . . , AN ), is robustly asymptotically stable.
PN
(c) The diagonal elements of the matrices Mi , i = 1, . . . , N , are negative and the matrix i=1 Mi is signstable.
(d) For all Ai ∈ Q(Mi ), i = 1, . . . , N , there exists a positive definite matrix Q ∈ Rn×n such that ÃT Q + QÃ
is negative definite for all à ∈ co(A1 , . . . , AN ).
(e) For all Ai ∈ Q(Mi ), i = 1, . . . , N , there exists a vector v ∈ Rn>0 such that v T à < 0 for all à ∈
co(A1 , . . . , AN ).
23
Proof : The proof of the equivalence between the statements (a) and (b) follows from [57, Theorem 3] while
the equivalence between the statements (c), (d) and (e) follows from Theorem 5.5 and Theorem 5.7. Clearly,
statement (d) implies statement (b) using standard results on switched systems; see e.g. [56]. Finally, statement (b) implies (c) since if the conditions of statement (c) does not hold, then there will exist some matrices
Ai ∈ Q(Mi ), i = 1, . . . , N , for which the set co(A1 , . . . , AN ) will contain unstable matrices. The proof is
completed.
♦
8.1.4. Continuous-time impulsive systems
Linear positive impulsive systems have been less studied than positive switched system but some results
can be found in [14]. Let us consider, as a starting point, the following linear impulsive system:
ẋ(t)
x(t+
k)
x(t0 )
= Ax(t), t 6= tk
= Jx(tk ), k = 1, . . .
= x0
(58)
where x ∈ Rn is the state of the system. The sequence of jumping instants {tk }∞
k=0 is assumed to be
increasing and to grow unbounded. It is known [14] that this system is positive if and only if A ∈ MRn and
n×n
J ∈ R≥0
.
n×n
Proposition 8.5. Let MA ∈ MSn and MJ ∈ S≥0
. Then, the following statements are equivalent:
(a) For all A ∈ Q(MA ) and all J ∈ Q(JA ), there exists a vector λ ∈ Rn>0 such that
λT A < 0 and λT (J − I) < 0.
(59)
(b) MA has negative diagonal, MJ has zero diagonal and the matrix MA + MJ is sign-stable.
Moreover, when one of the above equivalent statements holds, then for all A ∈ Q(MA ) and all J ∈ Q(JA ),
the linear positive impulsive system is asymptotically stable for any increasing sequence {tk }∞
k=0 such that
tk → ∞ as k → ∞.
Proof : The proof that the first statement implies the stability under arbitrary jump sequence can be found
in [14] and relies on the use of the linear copositive Lyapunov function V (x) = λT x, λ ∈ Rn>0 . Assume that
(b) holds, then MA is sign-stable and MA + MJ is upper-triangular modulo some simultaneous row/column
permutations which both imply that MJ has zero diagonal entries or, equivalently, that for all J ∈ Q(MJ ),
the matrix J − In is Metzler and Hurwitz stable. Since these matrices are upper-triangular, we can finally
use Theorem 5.7 to prove that (a) holds. Assume now that (a) holds. Hence, we have that MA and MJ − In
are sign-stable and, hence, MA and MJ has negative and zero diagonal, respectively. Hence, MA and MJ
are sign-summable and we have that λT (A + J − In ) < 0 for all A ∈ Q(MA ), J ∈ Q(MJ ). This implies
that MA + MJ − In is sign-stable or, equivalently, that it is upper-triangular modulo some simultaneous
row/column permutations, which is equivalent to saying that MA + MJ is sign-stable.
♦
8.2. Nonlinear positive systems
Let us consider the following class of nonlinear positive systems that arises, among others, in reaction
network theory
ẋ(t) = Ax(t) + Bf (x(t)) + b
(60)
x(0) = x0
where x ∈ Rn≥0 , f : Rn≥0 → Rℓ≥0 , b ∈ Rn≥0 , A ∈ Rn×n and B ∈ Zn×ℓ .
24
Proposition 8.6. Let A ∈ MRn and assume that there exist a v ∈ Rn>0 and an ε > 0 such that v T A ≤ −εv T
and v T B = 0. Then, for all x0 ∈ Rn≥0 , the trajectories of the system (60) are bounded and converge to the
compact set
S := {x ∈ Rn≥0 : V (x) ≤ v T b/ε},
(61)
i.e. S is forward-invariant and attractive.
Proof : Let us consider the linear copositive Lyapunov function V (x) = v T x, v ∈ Rn>0 . The derivative
of this function along the solutions of (60) is given by V̇ (x) = v T Ax + v T Bf (x) + v T b. Assuming the
conditions of the theorem hold, then we have that V̇ (x) ≤ −εV (x) + v T b and, therefore, that V (x(t)) ≤
e−εt V (x0 ) + (1 − e−εt )v T b/ε. Hence, for all x0 ∈
/ S , the trajectories of the system (60) converge to the
compact set which proves the result.
♦
We can now derive the following structural version of the above result:
Proposition 8.7. Let M ∈ MSn and suppose further that
(a) M is sign-stable,
(b) Z := (M −1 B)T ∈ Sℓ×n and
(c) Z is an L+ -matrix.
Then, for all A ∈ Q(M ), there exist a v ∈ Rn>0 and an ε > 0 such that v T A ≤ −εv T and v T B = 0.
Therefore, the trajectories of the system (60) are structurally bounded and structurally converge to a compact
set inside the nonnegative orthant.
8.3. Ergodicity of stochastic reaction networks
Reaction networks [42, 58, 59] is a powerful modeling paradigm for representing processes arising, for
instance, in biology [4, 6, 42]. We consider here a stochastic reaction network denoted by (X, R) involving
d molecular species X1 , . . . , Xd that interact through K reaction channels R1 , . . . , RK defined as
Rk :
d
X
i=1
ρk
l
ζk,i
Xi −−−→
d
X
r
ζk,i
Xi , k = 1, . . . , K
(62)
i=1
l
r
where ρk ∈ R>0 is the reaction rate parameter and ζk,i
, ζk,i
∈ Z≥0 . Each reaction is additionally described
by a stoichiometric vector and a propensity function. The stoichiometric vector of reaction Rk is given
r
r
l
l
by ζk := ζkr − ζkl ∈ Zd where ζkr = col(ζk,1
, . . . , ζk,d
) and ζkl = col(ζk,1
, . . . , ζk,d
). In this regard, when
the reaction
R
fires,
the
state
jumps
from
x
to
x
+
ζ
.
We
define
the
stoichiometry
matrix S ∈ Zd×K
k
k
as S := ζ1 . . . ζK . When the kinetics is mass-action, the propensity function of reaction Rk is given by
Qd
d
xi !
/ Zd≥0 . We restrict ourselves here
λk (x) = ρk i=1 (xi −ζ
l )! and is such that λk (x) = 0 if x ∈ Z≥0 and x + ζk ∈
n,i
to reaction networks involving at most bimolecular reactions. That is, the propensities functions are at most
polynomials of degree two and, therefore, we can write the propensity vector as λ(x) = col(w0 , W x, Y (x))
0
u
b
where w0 ∈ Rn≥0
, W x ∈ Rn≥0
and Y (x) ∈ Rn≥0
are the propensity vectors associated the zeroth-, first- and
second-order
reactions,
respectively.
According
to this structure, the stoichiometric matrix is decomposed
as S =: S0 Su Sb . Under the well-mixed assumption, this network can be described by a continuoustime Markov process (X1 (t), . . . , Xd (t))t≥0 with state-space Zd≥0 ; see e.g. [42]. A fundamental result from
[60] states that the ergodicity of an irreducible2 continuous-time Markov process can be checked using a
Lyapunov-like condition, referred to as a Foster-Lyapunov condition. This has led to the following result:
2
The state-space of the Markov process is said to be irreducible if any state can be reached from any other state with positive
probability. When the state-space is finite, then this reduces to the irreducibility of the transition-rate matrix. A method for
proving the irreducibility of the state-space in the infinite case has been proposed in [61].
25
Theorem 8.8 ([4]). Let us consider an irreducible3 bimolecular reaction network (X, R) and define A :=
Su W . Assume that there exists a vector v ∈ Rd>0 such that v T Sb = 0 and v T A < 0. Then, the reaction
network (X, R) is exponentially ergodic and all the moments are bounded and converging.
The above result can be made structural the following way:
Proposition 8.9. Let Z ∈ MSn×n and suppose, further, that
(a) the matrix Z is sign-stable,
(b) the matrix Y := (Z −1 Sb )T is a sign-matrix, and
(c) the matrix Y is an L+ -matrix.
Then, for all A ∈ Q(Z), there exists a v ∈ Rn>0 such that v T A < 0 and v T Sb = 0, which implies, in turn,
that for all Su W = A ∈ Q(Z), the reaction network (X, R) is exponentially ergodic and all the moments are
bounded and converging.
Bibliography
References
[1] A. Berman and R. J. Plemmons. Nonnegative matrices in the mathematical sciences. SIAM, Philadelphia, USA, 1994.
[2] H. Nikaido. Convex structures and economic theory. Academic Press, New York, USA, 1968.
[3] C. Briat and M. Khammash. Computer control of gene expression: Robust setpoint tracking of protein
mean and variance using integral feedback. In 51st IEEE Conference on Decision and Control, pages
3582–3588, Maui, Hawaii, USA, 2012.
[4] A. Gupta, C. Briat, and M. Khammash. A scalable computational framework for establishing long-term
behavior of stochastic reaction networks. PLOS Computational Biology, 10(6):e1003669, 2014.
[5] F. Parise, M. E. Valcher, and J. Lygeros. On the reachable set of the controlled gene expression system.
In 53rd IEEE Conference on Decision and Control, pages 4597–4604, Los Angeles, USA, 2014.
[6] C. Briat, A. Gupta, and M. Khammash. Antithetic integral feedback ensures robust perfect adaptation
in noisy biomolecular networks. Cell Systems, 2:17–28, 2016.
[7] L. Farina and S. Rinaldi. Positive Linear Systems: Theory and Applications. John Wiley & Sons, 2000.
[8] M. Ait Rami and F. Tadeo. Controller synthesis for positive linear systems with bounded controls.
IEEE Transactions on Circuits and Systems – II. Express Briefs, 54(2):151–155, 2007.
[9] C. Briat. Robust stability and stabilization of uncertain linear positive systems via integral linear constraints - L1 - and L∞ -gains characterizations. International Journal of Robust and Nonlinear Control,
23(17):1932–1954, 2013.
[10] A. Rantzer. Distributed control of positive systems. In 50th IEEE Conference on Decision and Control,
pages 6608–6611, 2011.
3
Computationally tractable conditions for checking the irreducibility of reaction networks are provided in [61].
26
[11] T. Tanaka and C. Langbort. Symmetric formulation of the S-Procedure, Kalman-Yakubovich-Popov
Lemma and their exact losslessness conditions. IEEE Transactions on Automatic Control, 58(6):1486–
1496, 2013.
[12] M. Colombino and R. S. Smith. A convex characterization of robust stability for positive and positively
dominated linear systems. IEEE Transactions on Automatic Control, 61(7):1965–1971, 2016.
[13] C. Briat and M. Khammash. Interval peak-to-peak observers for continuous- and discrete-time systems
with persistent inputs and delays. Automatica, 74:206–213, 2016.
[14] C. Briat. Dwell-time stability and stabilization conditions for linear positive impulsive and switched
systems. Submitted, 2016. URL http://arxiv.org/abs/1608.02741.
[15] J. Quirk and R. Ruppert. Qualitative economics and the stability of equilibrium. The review of economic
studies, 32(4):311–326, 1965.
[16] R. M. May. Will a large complex system be stable? Nature, 238:413–414, 1972.
[17] R. M. May. Stability and Complexity in Model Ecosystems. Princeton University Press, Princeton, New
Jersey, 1973.
[18] A. Roberts. The stability of a feasible random ecosystem. Nature, 251:607–608, 1974.
[19] C. Jeffries. Qualitative stability and digraphs in model ecosystems. Ecology, 55:1415–1419, 1974.
[20] J. J. Tyson. Classification of instabilities in chemical reaction systems. The Journal of Chemical Physics,
62:1010–1015, 1974.
[21] B. L. Clarke. Theorems on chemical network stability. The Journal of Chemical Physics, 62:773–775,
1975.
[22] J. Maybee and J. Quirk. Qualitative problems in matrix theory. SIAM Reviews, 11(1):30–51, 1969.
[23] C. Jeffries, V. Klee, and P. van den Driessche. When is a matrix sign stable? Canadian Journal of
Mathematics, 29:315–326, 1977.
[24] V. Klee and P. van den Driessche. Linear algorithms for testing the sign stability of a matrix and for
finding z-maximum matchings in acyclic graphs. Numerische Mathematik, 28(3):273–285, 1977.
[25] V. Klee, R. Ladner, and R. Manber. Signsolvability revisited. Linear Algebra and Its Applications, 59:
131–157, 1984.
[26] M. A. Berger and A. Felzenbaum. Sign patterns of matrices and their inverses. Linear Algebra and its
Applications, 86:161–177, 1987.
[27] C. Thomassen. When the sign pattern of a square matrix determines uniquely the sign patter of its
inverse. Linear Algebra and its Applications, 119:27–34, 1989.
[28] R. A. Brualdi and B. L. Shader. Matrices of sign-solvable linear systems. Cambridge University Press,
Cambridge, UK, 1995.
[29] M. Catral, D. D. Olesky, and P. van den Driessche. Allow problems concerning spectral properties of
sign pattern matrices: A survey. Linear Algebra and its Applications, 430(11-12):3080–3094, 2009.
[30] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, Cambridge, MA,
USA, 2004.
[31] D. Knuth. The Art of Computer Programming, Vol. 1: Fundamental Algorithms, 3rd Edition. AddisonWesley, Boston, USA, 1997.
27
[32] L. Gurvits, R. Shorten, and O. Mason. On the stability of switched positive linear systems. IEEE
Transactions on Automatic Control, 52(6):1099–1103, 2007.
[33] O. Mason and R. N. Shorten. On linear copositive Lyapunov functions and the stability of switched
positive linear systems. IEEE Transactions on Automatic Control, 52(7):1346–1349, 2007.
[34] E. Fornasini and M. E. Valcher. Linear copositive Lyapunov functions for continuous-time positive
switched systems. IEEE Transactions on Automatic Control, 55(8):1933–1937, 2010.
[35] C. Briat and M. Khammash. Robust ergodicity and tracking in antithetic integral control of stochastic
biochemical reaction networks. In 55th IEEE Conference on Decision and Control (Accepted), Las
Vegas, USA, 2016. URL http://arxiv.org/abs/1607.07080.
[36] G.-Y. Lee and B. L. Shader. Sign-consistency and solvability of constrained linear systems. Electronic
Journal of Linear Algebra, 4:1, 1998.
[37] C. A. Desoer and M. Vidyasagar. Feedback Systems : Input-Output Properties. Academic Press, New
York, 1975.
[38] K. Zhou, J. C. Doyle, and K. Glover. Robust and Optimal Control. Prentice Hall, Upper Saddle River,
New Jersey, USA, 1996.
[39] V. A. Yakubovich. S-procedure in nonlinear control theory. Vestnik Leningrad Univ., pages 62–77, 1971.
English translation in Vestnik Leningrad Univ. Math., vol. 4, pp. 73–93, 1977.
[40] U. T. Jönsson.
A lecture on the S-procedure.
Technical report, Division of Optimization and Systems Theory, Royal Institute of Technology, Stockholm, Sweden, 2001.
URL
http://www.math.kth.se/ulfj/5B5746/Lecture.ps.
[41] D. Liberzon, J. P. Hespanha, and A. S. Morse. Stability of switched systems: a Lie-algebraic condition.
System and Control Letters, 37:117–122, 1999.
[42] D. F. Anderson and T. G. Kurtz. Stochastic Analysis of Biochemical Systems, volume 1.2 of Mathematical Biosciences Institute Lecture Series. Springer Verlag, 2015.
[43] C. R. Johnson. Sign patterns of inverse nonnegative matrices. Linear Algebra and its Applications, 55:
69–80, 1983.
[44] C. A. Eschenbach, F. J. Hall, D. L. Harrel, and Z. Li. When does the inverse have the same sign pattern
as the transpose? Czechoslovak Mathematical Journal, 49(124):255–275, 1999.
[45] F. Le Gall. Powers of tensors and fast matrix multiplication. In 39th International Symposium on
Symbolic and Algebraic Computation, pages 296–303, Kobe, Japan, 2014.
[46] G. J. Foschini and Z. Miljianic. A simple distributed autonomous power control algorithm and its
convergence. IEEE Transactions on Vehicular Technology, 42(4):641–646, 2002.
[47] A. Zappavigna, T. Charalambous, and F. Knorn. Unconditional stability of the foschinimiljanic algorithm. Automatica, 48(1):219–224, 2012.
[48] W. M. Haddad and V. Chellaboina. Stability theory for nonnegative and compartmental dynamical
systems with time delay. Systems & Control Letters, 51(5):355–361, 2004.
[49] S. I. Niculescu. Delay effects on stability. A robust control approach, volume 269. Springer-Verlag:
Heidelbeg, 2001.
[50] K. Gu, V. L. Kharitonov, and J. Chen. Stability of Time-Delay Systems. Birkhäuser, Boston, 2003.
28
[51] C. Briat. Linear Parameter-Varying and Time-Delay Systems – Analysis, Observation, Filtering &
Control, volume 3 of Advances on Delays and Dynamics. Springer-Verlag, Heidelberg, Germany, 2015.
[52] M. Buslowicz. Simple stability conditions for linear positive discrete-time systems with delays. Bulletin
of the Polish Academy of Sciences - Technical sciences, 56(4):325–328, 2008.
[53] E. I. Verriest and A. F. Ivanov. Robust stability of delay-difference equations. In 34th IEEE Conference
on Decision and Control, pages 386–391, 1995.
[54] M. Ogura, M. Wakaiki, and V. M. Preciado. Dynamic analysis of bet-hedging strategies as a protection
mechanism against environmental fluctuations. In 55th IEEE Conference on Decision and Control,
2016.
[55] M. Ogura and V. M. Preciado. Stability of spreading processes over time-varying large-scale networks.
IEEE Transactions on Network Science and Engineering, 3(1):44–57, 2016.
[56] D. Liberzon. Switching in Systems and Control. Birkhäuser, New York, 2003.
[57] H. Lin and P. J. Antsaklis. Stability and stabilizability of switched linear systems: A survey of recent
results. IEEE Transactions on Automatic Control, 54(2):308–322, 2009.
[58] F. Horn and R. Jackson. General mass action kinetics. Archive for rational mechanics and analysis,
47(2):81–116, 1972.
[59] M. Feinberg. Complex balancing in general kinetic systems. Archive for rational mechanics and analysis,
49(3):187–194, 1972.
[60] S. P. Meyn and R. L. Tweedie. Stability of Markovian processes III: Foster-Lyapunov criteria for
continuous-time processes. Adv. Appl. Prob. , 25:518–548, 1993.
[61] A. Gupta and M. Khammash. Determining the long-term behavior of cell populations: A new procedure
for detecting ergodicity in large stochastic reaction networks. Technical Report arXiv:1312.2879, ETHZürich, 2013.
29
| 3cs.SY
|
Block-Level Parallelism in Parsing Block
Structured Languages
Abhinav Jangda
arXiv:1509.08068v1 [cs.PL] 27 Sep 2015
Indian Institute of Technology (BHU), Varanasi
[email protected]
Abstract. Softwares source code is becoming large and complex. Compilation of large base code is a time consuming process. Parallel compilation of code will help in reducing the time complexity. Parsing is one
of the phases in compiler in which significant amount of time of compilation is spent. Techniques have already been developed to extract the
parallelism available in parser. Current LR(k) parallel parsing techniques
either face difficulty in creating Abstract Syntax Tree or requires modification in the grammar or are specific to less expressive grammars. Most
of the programming languages like C, ALGOL are block-structured, and
in most languages grammars the grammar of different blocks is independent, allowing different blocks to be parsed in parallel. We are proposing
a block level parallel parser derived from Incremental Jump Shift Reduce Parser by [13]. Block Parallelized Parser (BPP) can even work as
a block parallel incremental parser. We define a set of Incremental Categories and create the partitions of a grammar based on a rule. When
parser reaches the start of the block symbol it will check whether the
current block is related to any incremental category. If block parallel
parser find the incremental category for it, parser will parse the block in
parallel. Block parallel parser is developed for LR(1) grammar. Without
making major changes in Shift Reduce (SR) LR(1) parsing algorithm,
block parallel parser can create an Abstract Syntax tree easily. We believe this parser can be easily extended to LR (k) grammars and also be
converted to an LALR (1) parser. We implemented BPP and SR LR(1)
parsing algorithm for C Programming Language. We evaluated performance of both techniques by parsing 10 random files from Linux Kernel
source. BPP showed 28% and 52% improvement in the case of including
header files and excluding header files respectively.
1
Introduction
Multi core chip architectures are emerging as feasible solution to effectively utilizing the ever growing number of chip. Multi-core chip depends on success in
system software technology (compiler and runtime system), in order to have
thread level parallelism and utilizing on-chip concurrency. With multi-core processors, additional speedups can be achieved by the use of parallelism in dataindependent tasks. There is a gradual shift towards making current algorithms
and design into parallel algorithms. It is rather difficult to achieve lock free and
low cache contention parallel algorithms.
In 70s papers appeared ideas on parallel compilation of programming languages and parallel execution of programs were expected. In those papers discussions on parallel lexical analysis, syntactic analysis and code generation were
done. With VLSI applications, prominent increase in research on parallel compilation is observed.
A compiler contains different phases: lexical analyzer, syntactic analyzer, semantic analyzer and code generator. Parsing or syntax analysis is the phase
of compiler which analyses the program code according to the language. After
analysis, it converts the code into another formal representation which will act
as input for succeeding phases of compiler.
Complexity of software source code is increasing. An effort to compile large
code base is very time consumable. [4] describes two types of parsers: Top Down
and Bottom Up Parsers. Top Down parsers have less power as compared to Bottom Up Parser. LR (k), SLR (k) and LALR (1) are types of Bottom Up Parsers.
With more power Bottom Up Parsers also requires more space and more time
in parsing a string as compared to Top Down parsers. Most of the compiler
compilers like Yacc [17] and Bison [16] creates LR (1) parsers and compilers like
clang [18], Mono C# Compiler [19] etc. uses LR(1) parsers. So, it is evident that
programming languages can be represented easily by LR (1) languages.
Parsing is very time consuming phase of compiler. Parsing different files in
parallel is not enough. As programming languages like C and C++ can includes
different files (using #include) in a single file which results in generation of very
long file. If we can parallel the parsing phase of single file, it will give performance benefits in compiling the source code. Many significant techniques are
already proposed for making parallel parsers ([2], [8], [10], [9], [11]). A parallel
parsing for programming language is given by [14].
A block is a section of code which is grouped together. In a language, a block
may contain class definition, member or method declaration. Another block could
be a block of statements also called compound statement. This block is usually
associated with a function code or if statement or loop. Programming Languages
such as C, C++, Java, Python use the concept of blocks heavily. One of most
important property of parsing blocks is that they all are independent of each
other i.e. each block can be parsed independently of other block. So, we could
parse many blocks in a parallel fashion. In this paper, we will propose a technique to parse various blocks of the code in parallel. It can also work as a block
parallel incremental parser. Our parser is termed as Block Parallelized Parser
(BPP, for short).
Our technique of parallel parsing is based on incremental parsing. An incremental parser is the one that parse only those portions of a program that have
been modified. Whereas an ordinary parser must process the entire program
when it is modified. An incremental parser takes only the known set of changes
done in a source file and updates its internal representation of source file which
may be an Abstract Syntax Tree. By building upon the previously parsed files,
the incremental parser avoids the wasteful re-parsing of entire source file where
most of the cod remains unchanged.
BPP is based on the properties that an incremental parser can parse any
part of a source code without the need of parsing the whole source code and
different blocks in a source code can be parsed independently of other blocks. In
BPP these parts are blocks in a source code. Using the property of incremental
parser, BPP parse each of the blocks independently of other blocks. Each of these
blocks are parsed in their own thread. It can be easily seen that BPP follows a
divide and conquer approach. It divides the source into different blocks, parse
each of them in parallel and at the end conquer these blocks. In our scheme the
conquer step does nothing except waiting for all the BPP Threads to complete
their operations.
There have been many works on incremental parsing [Incremental Parsing
References]. We choose to extend on the works of Incremental Jump Shift Reduce parser of [13]. BPP is derived from Incremental Jump Shift Reduce Parser.
In [13], authors defined Incremental Jump Shift reduce parser for SLR (1) languages only. However, we decided to extend this parser to accept LR(1) language
because LR(1) languages has more power than SLR (1) and nearly all programming languages can be defined in the form of LR(1) grammars. We define the
incremental categories to be a statement containing a block like class definition
or function definition or if statement or for loop statement. Then, we give a
notion of First Non-Terminal symbols of a Non-Terminal symbol. We used this
notion to create partitions of a grammar such that a partition includes an incremental category and its First Non-Terminals. We observed that this scheme
gives us a very interesting property in Incremental Jump Shift Reduce parser.
We used this property to create our Block Parallelized Parser.
Whenever a start of the block symbol is encountered the parser will first
check whether the current block is related to any incremental category and can
it be parsed independently. If BPP is able to find the incremental category for it,
BPP will start parsing the block in parallel. In this paper we developed this BPP
for LR(1) languages but we believe it can be easily extended to LR(k) or can be
easily converted to LALR (1) or SLR (1) grammars. We also see that no major
changes were done to the current LR(1) parsing algorithm and hence, it should
be easy to create an Abstract Syntax Tree. This parser can also work as an incremental parallel parser which can parse different blocks in parallel. Moreover,
it could be seen that there is no requirement of any Thread Synchronization to
communicate between different threads of BPP each of which is parsing a block
in parallel. This is because no two blocks are related in any way for the purpose
of parsing.
We compiled C# implementation using Mono C# Compiler 3.12.1 and executed the implementation using Mono JIT Compiler 3.12.1 on machine running
Fedora 21 with Linux Kernel 3.19.3 with 6 GB RAM and Intel Core i7-3610
CPU with HyperThreading enabled. We found out that our technique showed
28% performance improvement in the case of including header files and 52%
performance improvement in the case of excluding header files.
The following paper is designed as follows. Section 2 shows some previous
work done in parallel parsing. Section 3 and 4 provides the terminology we will
use and the background required to understand our technique. In Section 5 we
will extend Incremental Jump Shift Reduce parser to accept LR(1) grammars. In
Section 6, we introduced the concept of First Non Terminals of a non terminal.
In Section 7 we will use this concept to create partitions of the grammar. We
also showed that by creating partitions using this concept we get a very interesting property. This property would be used by BPP. We have generalized this
property in a theorem and also provided a proof for it. In Section 8 we presents
our Block Parallelized Parser and its parsing algorithm. In Section 9 we will
compare our algorithm with previous work. Section 10 shows our evaluation and
results. In Section 11 and Section 12 we complete our document with conclusion
and related work.
2
Related Work
A lot of previous work has been done in Parallel Parsing of LR (1) and Context
Free Languages. The most recent work done by [2] in parallel parsing of LR(1) is
an extension of an LR substring parser for Bounded Context Languages (developed by Cormack) for Parallel environment. [3] provided a substring parser for
Bounded Context-LR Grammars and Simple Bounded Context-LR Grammars.
[2] distributes the work of parsing the substrings of a language over different
processors. The work was extended to different processors in a Balanced Binary
Tree Fashion and achieved O(log n) time complexity of parsing. But constructing
a Bounded Context LR Grammar for a programming language is also difficult.
C++ is one of the programming languages which cannot be parsed by LR (1)
parsing [12] so creating a Bounded Context Grammar is out of question here.
Parallel and distributed compilation schemes can be divided into two broad
categories, functional decomposition and data decomposition. [1] and [20] talks
about distributed compilation using a scheme based on functional decomposition. Functional decomposition scheme divides different phases of compiler: lexer,
parser, semantic analyzer into functional component and running each of them
on separate processors like an instruction pipeline fashion. The data decomposition scheme divide the input into sections of equal length and parse them in
parallel. BPP is data decomposition scheme which parallel the parser by divide
and conquer approach. The data decomposition scheme was developed by [8],
[7], [9], [11]. These schemes are parsing LR (k) in parallel. They divide the input into sections of equal length and then parse them in parallel. [10], [11], [8]
describes asynchronous algorithms while [9] develops a synchronous algorithm.
[11] develops a parallel LR parser algorithm using the error recovery algorithm
of [21].
[1] has developed an Incremental Parallel Compiler which could be used in
Interactive Programming Environment and he developed an Incremental Parallel Parser also. [22] improves upon the Divide and Conquer Parsing technique
developed by [23]. They show that while the conquer step of algorithm in [23] is
O(n3 ) but under certain conditions it improves to O(log 3 n) .
[14] describes a grammar partitioning scheme which would help in parsing
the language in parallel. In [14] a type of Statement Level Parallelism has been
developed. The grammar is divided into n different sub-grammars corresponding
to n subsets of the language which will be handled by each sub-compiler. For
each n sub grammars required to generate parse tables (using parser generator)
along with driver routine constitute a parser for sub-compiler. For each subcompiler, a requirement of modified scanner is there which recognizes subset of
the language. The technique described in [14] requires a lot of modification to the
Lexical Analyzer. A separate Lexical Analyzer has to be developed for one type
of language. The parser of [14] requires automatic tools for its implementation.
In all the above described techniques constructing Abstract Syntax Tree for
a Block Structured Language is difficult. As Blocks in a Block Structured Language are independent on each other, so they can be parsed independently.
Moreover this scheme would not involve Inter-Thread Communication before a
thread completes its Parsing of Blocks. Hence, no shared memory synchronization methods are required to coordinate between different threads. It could be
easily seen that the creation of an Abstract Syntax Tree is also very easy. With
all these required things in mind we have developed Block Parallelized Parser
for LR (1) languages.
3
Terminology
We assume the notation for Context Free Grammar is represented by G =
(N, T, S, P ) where N is set of non-terminals, T is set of terminals, S is start symbol and P is set of productions of the grammar.
n The language
o generated by G is
∗
∗
given as
L(G) = ω ∈ T |S ⇒ ω
We will use the following conventions.
S, A, B, ... ∈ N
a, b, ... ∈ T
..., w, x ∈ T ∗
X, Y ∈ N ∪ T
α, β, γ, ... ∈ (N ∪ T )∗
Given a grammar G, we represent its augmented grammar as G0 = (N 0 , T 0 , S 0 , P 0 ),
where
N 0 = N ∪ {S 0 }
T 0 = T ∪ {$}
0
P = P ∪ {S 0 → S$}
0
Here S is called the augmented start symbol of G0 and $ is the end of string
marker. We denote a set of End Of String markers as EOS.
In our paper we will represent Block Parallelized Parser as BPP, Jump Shift
Reduce parser as JSR and Incremental Jump Shift Reduce parser as I JSR. An
LR(1) item is represented as [A → α.β, a] , where a is the lookahead symbol.
In a programming language, a block represents a section of code grouped
together. This section of code can be a group of statements, or a group of declaration statements. For example in Java, a block corresponding to class defini-
Fig. 1. Top-Level and Child Blocks of a Java Program.
tion contains declaration statements for fields and methods. Block corresponding to function definition can contain declaration statement for local variables
or expression statements or control flow statements. A Top-Level Block is the
starting block of a program which contains definition for Classes, Functions,
Import/Include statements etc. Child Blocks are contained in either Top-Level
Block or another Child Block. As we proceed further, Block will be in reference
to Child Block. Fig. 1, shows an example of Top-Level and Child Blocks of a
Java Program.
A start block symbol could be ”{” in C style languages, ”begin” in Pascal
style languages is represented as terminal sb . An end block symbol which could
be ”}” in C style languages or ”end” in Pascal style languages is represented as
eb .
4
Background
We now survey LR (1) parsers and their generation algorithm as given by [4].
LR (1) parsers are table driven Shift Reduce parsers. In LR (1), L denotes
left-to-right scanning of input symbols and R denotes constructing right most
derivation in reverse. Some extra information is indicated with each item: the
set of possible terminals which could follow the items LHS. This set of items is
called the lookahead set for the item. Here, 1 signifies that number of lookahead
symbols required are 1.
LR (1) parser consists of an input, an output, a stack, a driver routine. A
driver routine runs its parsing algorithm which interacts with two tables ACTION and GOTO. Any entry in ACTION and GOTO tables are indexed by a
symbol which belongs to N ∪ T 0 and the current state. An entry in both the
tables can be any one of the following:
–
–
–
–
–
If
If
If
If
If
ACTION [j, a] = <S, q> then a Shift Action must be taken.
ACTION [j, a] = <R, A → α > then reduce symbols to a production.
ACTION [j, a] = Accept then grammar is accepted.
ACTION [j, a] = error then Error has occurred.
GOTO [j,A] = q then go to state q.
An LR (1) item is of the form [A → α.β, a], where a is a lookahead symbol.
Construction of LR (1) items requires two procedures CLOSURE and GOTO.
CLOSURE takes a set of items as its argument. GOTO takes a set of items and
a symbol as arguments. Both are defined as follows:
CLOSU RE(I) = I ∪ {[B → .γ, b] | [A → α.Bβ, a] ∈ I
and B → γ ∈ P 0 ∀ b ∈ F IRST (βa)}
GOT O(I, X) = CLOSU RE({[A → αX.β, a] | [A → α.Xβ, a] ∈ I})
Collection of set of LR (1) items is created using CLOSURE and GOTO functions. Items function creates the collection of set of LR (1) items.
items(G0 ) = CLOSU RE({[S 0 → .S, $]}) ∪ {GOT O(I, X) | I ∈ C and X ∈ P 0 }
ACTION and GOTO tables are created using this collection. Following is
the procedure to create these tables:
1. Create collection of set of LR(1) items. Let this collection be C 0 = {I0 , I1 , I2 , ..., In }
2. Let i be the state of a parser constructed from Ii . Entries in ACTION table
are computed as follows:
(a) ACTION [i, a] = shift j, if [A → α.aβ, b] ∈ Ii and GOT O(Ii , a) = Ij
(b) ACTION [i, a] = reduce A → α., if [A → α., a] ∈ Ii and A 6= S 0
(c) ACTION [i, $] = accept, if [S 0 → S., $] ∈ Ii
(d) GOTO [i, A] = j, if GOTO (Ii , A) = Ij
3. All other entries not defined by (b) and (c) are set to error.
4. The Initial State is the one containing the item [S 0 → .S, $].
Most of the programming languages could be constructed from LR (1) grammar. Hence, LR (1) is the most widely used parser. Many parser generators like
YACC and GNU Bison generates an LR (1) parser.
A Jump Shift Reduce [13] (JSR in short) parser is an extension of LR (1)
parser. LR (1) parser generates ACTION and GOTO table for the augmented
grammar G0 . JSR parser first partition the grammar G0 into several sub grammars and creates parsing sub- table of every sub grammar. Hence, the ACTION
and GOTO tables of LR (1) are split into several ACTION and GOTO tables
in JSR parser. JSR parser is equivalent to the LR (1) parser that is it will only
accept languages generated by LR (1) grammar [13].
Let G0 = (N 0 , T 0 , S 0 , P 0 ) be the augmented grammar of grammar G =
(N, T, S, P ). Let us partition the grammar G on the basis of Non Terminals.
Let Gi denotes a partition of the grammar G, such that we have
Gi = (M i , T i , S i , P i )
i
where,
N ⊆ N such that N i ∩ N j = φ | i, j = 1, ..., n, i 6= j
i
Pi = A → α
∈ P | A ∈ N ∀i = 1, ...,i n
i
i
M = N ∪ B ∈ N | ∃A → αBβ ∈ P
T i = a ∈ T i | ∃A → αaβ ∈ P i
Si ∈ N i
Therefore, we have
(∪N i , ∪ T i , S 0 , ∪ P i ) = G0
For every subgrammar Gi , a parsing subtable named Tab(S i ) is built. Each
subtable contains ACTION and GOTO subtables which are represented by
Tab(S i ).ACTION and Tab(S i ).GOTO. In addition to the Shift, Reduce and
Accept action there is an additional Jump action. Jump action is associated
with a sub-table. Whenever a Jump action is encountered then the parsing algorithm Jumps to a sub-table and parse the partition related to that sub-table.
We will now investigate few points about the Incremental Jump Shift Reduce
Parser [13]. Incremental Jump Shift Reduce Parser (I JSR) [13] is based upon the
JSR parser [13]. A set of Incremental Categories will be defined which could be
incrementally parsed by I JSR parser. Given a grammar G = (N, T, S, P ) a set
of Incremental Categories has to be defined IC = {Cj | Cj ∈ N, j = 1, 2, ..., n}
and the Incremental Language of G is
L∗ (G) = ∪L∗ (A),
∗
where A ∈ IC ∪ {S} and L∗ (A) = {α ∈ (T ∪ IC)∗ | A ⇒ α}
For every Incremental Category A, add a production A → A#A , where #A
is an end-of-string for A. For a given grammar G = (N, T, S, P ) with a set of
Incremental Categories IC = {Cj | Cj ∈ N, j = 1, 2, ...n} an Incremental Grammar is defined as
G∗ = (N, T ∪ EOS, S, P ∪ PI C),
where,
EOS is the set of End of String markers = {#j | j = 1, 2, ..n}
PIC = {Cj → Cj #j | Cj ∈ IC, #j ∈ EOS}
A major change by this extension is that now the strings may contain incremental symbols which are also non-terminal symbols. The difference between
ACTION and GOTO tables disappears, as an incremental category can also occur in the input string and can be shifted on the stack. Hence, we would have
only ACTION table and no need for GOTO table. As we have also introduced
EOS set, the ACTION table can now be indexed with symbols belonging to
N ∪ T 0 ∪ EOS. Every Incremental Category will have its own start state and
accept action. We will represent the subtable as Tab(S t ). Entries of table will
be as follows:
If Tab(S t )
If Tab(S t )
If Tab(S t )
If Tab(S t )
accepted.
– If Tab(S t )
– If Tab(S t )
–
–
–
–
[j,
[j,
[j,
[j,
X ] = <S, q> then a Shift Action must be taken.
X ] = <R, A → α. > then a Reduce Action must be taken.
$] = Accept then input is accepted.
#i ] = Accept then input for Incremental Category Ci will be
[j, a] = <J, K> then jump to a subtable Tab(S k ).
[j, a] = error then error occurred.
5
Extending I JSR Parser to accept languages generated
by LR (1) grammars
As JSR Generation Algorithm was already developed for LR (0) items and Incremental JSR Generation Algorithm was developed for SLR (0) items [13]. In this
section, we will first extend the JSR Parser Generation Algorithm to accept LR
(1) Grammar and then we will extend I JSR parser to accept LR(1) Grammar.
Generation of subtables first requires the generation of canonical collection
of sets of augmented items. An augmented item is a triplet represented as <i,
FF, TF>, where i is an LR item, FF called From-Field and TF called To-Field
are the names of parsing sub-tables. From-Field represents the sub-table which
contains the last action performed by parser and To-Field represent the subtable which contains the next action to be performed. Although we focus only
LR (1) items but the procedure for LR (k) items is very similar.
To-Field of an augmented item of a state is determined using TO function.
Let us define the TO function. Function TO calls Function CHOOSE NEXT.
1: procedure TO(Ij )
2:
Ij00 = φ
3:
for all item ik in Ij do
4:
if ik is [A → α.aβ, b] then
5:
add < ik , CHOOSEN EXT (Ij , a) > to Ij00
6:
else if ik is [A → α.Bβ, b] then
7:
add < ik , S i > to Ij00 , where B ∈ N i
8:
else if ik is [A → β., b] then
9:
add < ik , S i > to Ij00 , where A ∈ N i
return Ij00
This function selects a parsing table out of those in which the parsing of the
remaining string could continue. Let < · be a total ordering relation over the
set ST = {S p | p = 1, 2, ..., n} of the names of parsing sub-tables, such that
S i < ·S i+1 , i = 1, 2, ...n − 1
From-Field of an augmented item is enriched using FROM function. FROM
1: procedure CHOOSE NEXT(Ij , a)
2:
let ST 0 be a set of parsing subtables
3:
for all item [H → α.aβ, b] in Ij do
4:
add Sh to ST 0 such that H ∈ Nh
return min < ·ST 0
takes two arguments It00 , which is a set of items enriched with To-Field and Ij ,
whose items we have to enrich with From-Field.
STATES procedure is used to generate the collection of set of JSR items.
1: procedure FROM(It00 , Ij )
2:
Ij0 = φ
3:
for all itemik ∈ Ij do
4:
if ik is [S 0 → .S$, b] then
5:
add < ik , S 1 > toIj
6:
else if ik is [A → αX.β, b] then
7:
add < ik , T F > to Ij0 , where < [A → α.Xβ, b], T F >∈ It00
8:
else if ik is [A → .β, b] then
9:
add < ik , F F > to Ij0 , where < [B → α.Aβ, b], F F >∈ Ij0
return Ij0
STATES algorithm first generates the collection of sets of LR(1) items using
ITEMS procedure which was discussed previously. Afterwards, it calls TO and
FROM functions to generate set of augmented items from the corresponding set
of LR(1) items.
We will now extend I JSR parser to accept LR (1) languages. This extended
1: procedure STATES(G0 )
2:
C 0 = items(G0 )
3:
IA = φ
4:
for all Ij ∈ C 0 do
5:
Ij0 = {< ik , F Fk > | ik ∈ Ij } = F ROM (It00 , Ij ), where Ij =
GOT O(It , X) and It = φ if j = 0
6:
Ij00 = {< ik , T Fk > | ik ∈ Ij } = T O(Ij )
7:
IjA = {< ik , F Fk , T Fk > | < ik , F Fk >∈ Ij0 and < ik , T Fk >∈ Ij00 }
8:
I A = I A ∪ {IjA }
return I A
parser is based on the previous JSR parsing algorithm. The FIRST function
used to compute the set of First symbols related to a non-terminal has to be
modified to include the incremental categories also. The reason being an incremental category can also occur in the input string, can be shifted on the stack
while parsing and can reduce to a production. Moreover, FIRST should now also
include the EOS markers. Hence, the new FIRST becomes
∗
F IRST (A) = {a | A ⇒ aβ, where a ∈ T 0 ∪ EOS ∪ IC}
For an incremental category A, there will be a set of items containing item
[A → .A#A , #A ] and items [B → α.Aβ#A ]. Then the state corresponding to
this set of items will be the start state of the incremental grammar corresponding to A. Correspondingly, there will a single set of items that contains the item
[A → A.#A , #A ]. The state belonging to this set of item is the accepting state
of the incremental grammar corresponding to A. The I JSR parser has initial
and final states for every incremental category.
Now, we can extend the I JSR parser for accepting languages generated by
LR(1) grammars. The procedure I TABS given below is used to construct the
I JSR parsing table.
1: procedure I TABS(G0 )
2:
C 0 = items(G0 )
3:
for all IjA ∈ C 0 do
4:
for all < ih , H, K >∈ IjA do
5:
if ih is of the f orm [A → α.Xβ, b] then
6:
T ab(S K )[j, X] =< S, q > where GOT O(Ij , X) = Iq
7:
if H 6= K then
8:
T ab(S H )[j, X] =< J, K >
9:
else if ih is of the f orm [A → α., X] then
10:
T ab(S K )[j, X] =< R, A → α >
11:
if H 6= K then
12:
T ab(S H )[j, X] =< J, K >
13:
else if ih is of the f orm [S 0 → S.$, $] then
0
14:
T ab(S S )[j, $] = accept
15:
else if ih is of the f orm [A → A.#A , #A ] then
16:
T ab(S K )[j, #A ] = accept
6
First Non Terminals
In this section we will define the concept of First Non Terminals.
We define a set of First Non Terminals for a non terminal A as a set of
non-terminals that appear at the beginning of any sentential form derived from
A i.e. a set of non terminals B such that there exists a derivation of the form
∗
A ⇒ Bβ. F IRST N T (A) represents the set of First Non Terminals for A and
can be represented in set notations as: [ n
o
∗
F IRST N T (A) =
B | A ⇒ Bβ
B
To compute F IRST N T (A) for any non-terminal A, apply the following rules
until no more terminals can be added to the F IRST N T (A) set.
1. If A is a terminal, then F IRST N T (A) = φ
2. If A is a non-terminal and A → B1 B2 ...Bk is a production for some k ≤
1, then place Bi and everything in F IRST N T (Bi ) in F IRST N T (A) if
∗
B1 , B2 , ...Bi−1 ⇒
3. If A → is a production, then F IRST N T (A) = φ
EXAMPLE 1: If we have following productions:
S → DAB
S→C
A → aB
B→b
C→c
D→d
Then find F IRST N T (S)?
SOLUTION 1:
Due to first two productions of S we have, S
S
F IRST N T (S) = F IRST N T (A) F IRST N T (C) {A, C}
F IRST N T (A) = F IRST N T (a) = φ
F IRST N T (C) = F IRST N T (c) = φ
7
Using First Non Terminals to create partitions
In this section we will use the concept of First Non Terminals to create partitions
of grammar. I JSR Parser will use these partitions to develop its tables. We will
see that this kind of partitioning leads to a very interesting property in I JSR
Parsing algorithm. We will generalize this property in a theorem and will also
prove it.
We will partition the grammar in such way that:
– Every Incremental Category will have its own partition.
– The partition of Incremental Category will contain First Non Terminals of
that incremental category also.
– Intersection of set of First Non-Terminals of any two incremental categories
must be empty.
– All other remaining non-terminals including the augmented and start symbol
are included in the first partition.
Given a grammar G∗ = (N, T ∪ EOS, S, P ∪ PIC ) with a set of Incremental Category, IC = {Ct | Ct ∈ N } we define partitions of non-terminals as
N 2 , N 3 , ..., N n such that:
N t = {Ct } ∪ F IRST N T (Ct ) and Ct 6= S
and
F IRST N T (Ct ) ∩ F IRST N T (Cs ) = φ, where t, s = 1, 2, ..., n and t 6= s
And first partition,
[
N1 = N − Nt
t
Example 2: Partition the grammar given in Example 1 using first non terminals as given above and create I JSR parsing table. Then parse the string ”dabb”
and parse string ”ab” incrementally for incremental category A.
Solution 2: First we have to create an augmented grammar of the given
grammar by adding production S 0 → S$. Next we can create two partitions of
grammar as given by the following partitions of non terminals.
N 1 = {S 0 , S, B, C, D}
N 2 = {A}
As the incremental category we want is only A. Hence, there will be two
partitions, N 2 containing A and its first non terminals and N 1 will contain the
remaining non terminals. Moreover, we would also have to add an EOS marker for
A, let us say it is #1 . We can generate the I JSR table using I TABS procedure
described in Section 5.
Let us parse the string ”dabb”. Table 1 shows the series of actions taken
while parsing ”dabb”. In this case the start state will be the start state of table
T ab(S 0 ) i.e. 0. Table 2 shows the series of actions taken when ”ab” is parsed
incrementally with the incremental category A. In this case the start state will
be the start state of table Tab (A) i.e. 2.
Table 1. Parsing of ”dabb”
STACK INPUT ACTION
0
dabb$ Shift 5
0d5
abb$ Reduce D → d
0D2
abb$ Jump A
0D2
abb$ Shift 7
0D2a7
bb$ Jump S’
0D2a7
bb$ Shift 11
0D2a7b11
b$ Reduce B → b
0D2a7B10
b$ Jump A
0D2a7B10
b$ Reduce A → aB
0D2A6
b$ Shift 9
0D2A6b9
$ Reduce B → b
0D2A6B8
$ Reduce S → DAB
0S1
$ Accept
Table 2. Parsing of ”ab”
STACK INPUT ACTION
2
ab#1 Shift 7
2a7
b#1 Jump S’
2a7
b#1 Shift 11
2a7b11
#1 Reduce B → b
2a7B10
#1 Jump A
2a7B10
#1 Reduce A → aB
2A6
#1 Accept
In Table 1, when stack state reaches 0D2 there is a Jump to the Table Tab
(A). From this point until when stack state changes to 0D2A6, the actions taken
are same as the actions of Table 2 and in the same table i.e. Tab (A). Moreover,
in between these Stack states in Table 1 ”ab” is parsed to A.
We can generalize this example in the sense that same series of actions are
taken when parsing a string and when parsing its substring incrementally for
its incremental category. In the current example all the same actions happens
in the same table because we created the partitions in such a way that all the
first non terminals are in that partition. If the partitions were not created in the
way described, it could have happened that these actions would happen in the
different sub tables.
This technique is utilized by our BPP and it is generalized and proved in the
theorem below.
Theorem 1 Given an Incremental Grammar G∗ = (N, T, P 0 , S 0 ) with set of
incremental categories IC = {Ct | Ct ∈ N } such that the non terminal partition,
N t related to incremental category Ct contains only Ct and F IRST N T (Ct ). For
∗
an incremental category Ct and any terminal b ∈ F IRST (Ct ), if Ct ⇒ bγ and
during parsing of the word w = ”µbγδ” the parser reaches at a state q in the
subtable of Ct after performing the shift action on b then during the incremental
parsing of the word wt = ”bγ” for the incremental category Ct the parser will
also reach the state q after performing the shift action on b in the sub table of
Ct .
Proof. Outline: We will divide the proof in 4 cases. For each case, we will first
find two sets of JSR items reached after performing shift on b one during the
parsing of word w and other during the incremental parsing of word wt related
to incremental category Ct . We will then show that both of these sets contains
same JSR items which implies the above theorem.
It is given that Ct is an incremental category and Ct 6= S and let N t be the
partition containing only Ct and F IRST N T (Ct ). Let S t be the sub-table related
to Ct . For any non-terminal A ∈ F IRST N T (Ct ) and A → bβ, we must have
A ∈ N t.
During incremental parsing of the word wt = bγ for the incremental category
Ct , the state before performing actions related to b will be the start state of the
sub table T ab(Ct ). Let that start state be m.
We will have four cases on the basis of whether the grammar has productions of
the form, B → cXCt β and B → Ct β
Case 1: If B → cXCt β ∈ P 0 and B → Ct β ∈ P 0
Let the set of LR(1) items related to the start state of T ab(Ct ) i.e. state m
be Im and X lies in some partition other than N t i.e. X ∈ N h and t 6= h.
As noted by [13] the start state of T ab(Ct ) must contain the item [Ct →
.Ct #t , d], where d is a lookahead symbol. So we have [Ct → .Ct #t , d] ∈ Im . It
is evident that the item [Ct → .Ct #t , d] should be result of a closure of another
item. The only such item we can see is [B → cX.Ct β, j], where j is some lookahead symbol. So, we must have [B → cX.Ct β, j] ∈ Im . As discussed in Section
2, to get a set of LR(1) items we have to apply CLOSURE ([B → cX.Ct β, j]).
Hence, we have
[B → c.XCt β, j] ∈ Im
[Ct → .Ct #t , d] ∈ Im
[Ct → .Aγβ, e] ∈ Im
[A → .bβ, f ] ∈ Im
where e and f are some lookahead symbols.
Let Io be the set of LR(1) items such that Im = GOT O(Io , X). So, [B →
cX.Ct β, j] ∈ Io . Let Io00 be a set of JSR items enriched with TO fields corresponding to all LR(1) items in Io . After applying TO procedure over Io we would get
the TO field for item [B → cX.Ct β, j] as S h because X ∈ N h .
To get TO fields for all JSR items corresponding to LR (1) items of the set
Im we have to apply TO procedure over Im . We could see that the TO fields for
items [B → cX.Ct β, j], [Ct → .Ct #t , d], [Ct → .Aγ, e], [A → .bβ, f ] will be S t .
To enrich JSR items for all the LR(1) items in Im with FROM field, we would
apply FROM procedure as F ROM (Io00 , Im ). Now we could see that the FROM
field of JSR item of LR(1) item [Ct → .Aγ, e] will be equal to the FROM field
of [B → cX.Ct β, j] which in turn is equal to the TO field of [B → c.XCt β, j],
A
which is S h . So, the set of JSR items (Im
) corresponding to Im contains the
following items:
A
[B → cX.Ct β, j, S h , S t ] ∈ Im
h
t
A
[Ct → .Ct #t , d, S , S ] ∈ Im
A
[Ct → .Aγ, e, S h , S t ] ∈ Im
h
t
A
[A → .bβ, f, S , S ] ∈ Im
Let after performing shift operation in the state m over the symbol b, parser
reaches state n. So, we must have
In = GOT O(Im , a) = CLOSU RE({[A → b.β, g] | ∀ g ∈ F IRST (βf )})
Moreover, the TO and FROM fields of all items in the above state will be S t .
We have obtained the set of JSR items reached after performing shift over the
symbol b during incremental parsing of word wt for incremental category Ct .
Also, the subtable at this state is T ab(S t ).
We will now obtain the set of JSR items reached after performing shift over
the symbol b during the parsing of word w for incremental category Ct . Let us
∗
∗
consider the derivation S ⇒ ηB ⇒ µbγδ such that, B doesn’t derive η. In the
above derivation only one production out of B → cXCt β and B → Ct β will be
used. Hence, we have two cases of the basis of which production is used.
Case 1.1 : If production B → cXCt β is used, then the set of items related
to state reached just after performing shift on c in the above derivation say Ix
will contain the LR(1) item [B → c.XCt β, j] besides other items of the form
[X → .δ, k], where k ∈ F IRST (Ct βc). After applying TO procedure on Ix , we
should get the TO field for LR(1) item [B → c.XCt β, j] as S h . Let us represent
these set of items enriched with TO field as Ix00 .
After parsing X next state will be obtained by performing GOTO over Ix
with symbol X. Let that state be Iy . Now we must have,
Iy = GOT O(Ix , X)
.
After performing GOTO we get
[B → cX.Ct β, j] ∈ Iy
[Ct → .Ct #t , d] ∈ Iy
[Ct → .Aγ, e] ∈ Iy
[A → .bβ, f ] ∈ Iy
After applying TO procedure over Iy , we get the TO fields of items [B →
cX.Ct β, j] [Ct → .Ct #t , d] [Ct → .Aγ, e] [A → .bβ, f ] as S t . To enrich JSR items
of LR(1) items in Iy we would apply FROM procedure as F ROM (Ix00 , Iy ). Now,
we could see that the FROM field of JSR item for [Ct → .Aγ, e] will be equal
to the FROM field of JSR item for [B → cX.Ct β, j] which in turn is equal to
the TO field of [B → c.XCt β, j], which is S h . So, the set of JSR items (IyA )
corresponding to Iy contains the following items:
[B → cX.Ct β, j, S h , S t ] ∈ IyA
[Ct → .Ct #t , d, S h , S t ] ∈ IyA
[Ct → .Aγ, e, S h , S t ] ∈ IyA
[A → .bβ, f, S h , S t ] ∈ IyA
Let the state reached after performing shift of b from the state Iy be Iz . Then,
Iz = GOT O(Iy , b) = CLOSU RE({[A → b.β, g] | ∀ g ∈ F IRST (βf )})
Moreover, the TO and FROM fields of all JSR items of Iz will be S t .
So, we have Iz = In . This shows that the states z and n are same. Let us name
these states as q. Also, the TO fields for the items of Iz and In are same i.e.
S t . Hence, in this case Shift on terminal b in states m and x results only in one
state q and in the sub-table for incremental category Ct .
Theorem is proved in this case.
Case 1.2 : If B → cXCt β is used. Let x is the state reached before performing the
shift on b. Then, the set of items, say Ix related to state x will contain following
items:
[B → .Ct β, j] ∈ IyA
[Ct → .Ct #t , d] ∈ IyA
[Ct → .Aγ, e] ∈ IyA
[A → .bβ, f ] ∈ IyA
After applying TO operation to Ix , the TO fields of all JSR items for above
LR(1) items will be S t .
Let Iy be the state reached after performing shift on b in the state Ix . So,
Iy = GOT O(Ix , b) = CLOSU RE({[A → b.β, g] | ∀ g ∈ F IRST (βf ) })
Also, the TO and FROM fields of all the JSR items for the above set of LR
(1) items will be S t .
Hence, Iy = In . This shows that the states y and n are same. Let us name
the state as q. Moreover, the TO fields of LR(1) items of Iy and In are S t . Hence,
in this case Shift on terminal b in the states m and x results only in one state q
in the subtable of Ct .
Theorem is proved in this case.
As Theorem has been proved in Case 1.1 and Case 1.2. So, for Case 1
also the Theorem has been proved.
We have proved for the case containing both productions. Two other cases
are when only one of these productions is present. The proof of both of these
cases are very similar to the Case 1.
Please note that with the given set of conditions in the Theorem, we couldn’t
have the case in which none of the productions belong to this set.
THEOREM 1 is crucial to BPP. In the succeeding sections we will use this
theorem to create our parallel parsing algorithm.
8
Block Parallelized Parser
In this section we will present our Block Parallelized Parser for LR (1) grammars. We will first give the intuition and working of BPP. Then we will present
our algorithm and give a proof that our parser can accept all the LR(k) and
LALR (k) languages which can be accepted by a Shift Reduce LR(k) and LALR
(k) parser.
Let G0 = (N 0 , T 0 , P 0 , S 0 ) be the augmented grammar. The incremental categories are the Non Terminals associated with the blocks to be parsed in parallel. For example, if we want to parse class definitions in parallel then we
can define class-definition as the incremental category. Other examples can be
function-definition, if-statement, for-loop, while-loop if they can be parsed in
parallel. For most of the programming languages including C, C++, Java, C#
above blocks can be parsed in parallel. In general we can define an incremental category to be the non terminals which derive a string containing the start
of the block symbol, sb and ending with the end of the block symbol eb . In
mathematical terms,nfor BPP the set of incremental category IC
o is defined as:
∗
IC = Ct | if Ct → αX ∈ P 0 then Ct ⇒ αsb βeb
In the C Programming Language, a function-definition can be represented by
the following context free grammar productions:
f unction-def inition → type name (arguments) block
block → sb statement∗ eb
statement → type name ;
statement → if -stmt
statement → f or-loop
statement → while-loop;
if -stmt → if (expression) block
while-loop → while (expression) block
f or-loop → for (expression; expression; expression) block
According to the definition of Incremental Categories above, we can see that
function-definition, if-stmt, while-loop, for-loop follows the condition for an incremental category.
In a programming language there could be many types of blocks like in Java,
a block related to the class definition may contain only method definitions and
declaration of member variables while a block related to the method definition
would contain statements including expression statements, variable declarations,
if statement or loop etc. This means in a programming language not all blocks
contains same type of statements. Hence, encountering the start symbol of block
doesn’t give us enough information about what kind of statements the block
would contain. To overcome this problem we will modify the productions of incremental category such that a reduce action will happen when the start symbol
is encountered. Modify the productions of Ct ∈ IC in such a way that every
production Ct → αt sb Xt eb is split into two productions:
Ct → At sb Xt eb
At → αt
If the productions of incremental category Ct is structured as above then
during the parsing of a word related to this incremental category there will be
reduce action to reduce the current symbols to production At → αt when sb
becomes the look ahead symbol. As each At is related to only one Ct and viceverse, we can easily determine which incremental category has to be parsed next.
Now we are in a stage to define Block Parallelized Grammar. Given a grammar G = (N, T, P, S) and a set of incremental categories
n
o
∗
IC = Ct | Ct ∈ αsb Xeb ∈ P 0 and Ct ⇒ αsb βeb
we define Block Parallelized Grammar GP = (N P , T, P P , S) such that
P P = P ∪ {Ct → At sb Xt eb , At → αt | Ct → αt Xt ∈ P } −
{Ct → αt Xt | Ct → αt Xt ∈ P }
N P = {At | Ct → At sb Xt eb , At → αt ∀ Ct → αt Xt ∈ P }
and N P is partitioned using FIRSTNT as given in Section 7.
Now we can use THEOREM 1 to create BPP. Let us have a string w =
∗
∗
”ωaδsb ηeb µ” where ω, δ, η, µ ∈ T ∗ , At ⇒ aδ and Xt ⇒ sb ηeb . During the
parsing w, when sb is encountered we should have a reduce action to At , based
on the production At → αt . Now, we can get Ct associated with At . According
to THEOREM 1, during parsing of the word w if the parser reaches at state q
in the sub table of Ct after performing shift action on a then during the incremental parsing of the word wt = ”aδsb ηeb ” for the incremental category Ct the
parser will also reach the state q in the sub table of Ct after performing the shift
action on a. This means, we can replace the state reached just before performing
shift action on a with the start state of subtable of Ct and wt can now be parsed
incrementally.
It is now evident that why the partitions of Non Terminals should be created
as described in Section 7. If the partitions are not created in this way, then it
may happen after a shift on a during the parsing of w and incremental parsing of wt may reach the same state but not in the same sub-table. By creating
partitions as described in Section 8, we make sure that when sb is encountered
by BPP then the newly created parallel parser knows in which sub-table it has
to continue the parsing in. On the other hand if partitions are not created as
described in Section 8, then newly created parallel parser wouldnt know in which
sub-table it has to continue parsing in.
This property is used by BPP to parse the block incrementally. Algorithm 1
is the BPP parsing algorithm of incremental table S t . If t = 1, then the algorithm is for the Top Level Block. Otherwise it is for other incremental category.
This parsing algorithm is for any block be it top level block or child block. Lines
1-4 initializes different local variables. Lines 5-32 is the main loop of algorithm
which does the whole work. Line 6, gets the top state on the stack. Lines 7-9
pushes the next state on the stack if there is a shift operation. Similarly, lines
10-11 changes current table if there is a jump operation. Line 12-27 are executed
if there is reduce action which reduces according to the production A → β. Line
14 checks if current input symbol is a start of the block symbol and if reduce
action reduces At → αt for an incremental category Ct . If yes then Lines 15-22
gets Ct related to At , pops |αt | states from stack and pushes these states and
start state to a new stack, creates and starts a new BPP for Ct and shifts to
the end of block. In this case next symbol will become Ct . If check of Line 14
fails then it means this is a regular reduce action not associated with any block.
Lines 24-27, pops |β| states from stack and shifts to a new state. Line 28 returns
if there is an accept action. Accept action can be for both Top Level block and
Child Block. Line 30 reports error if none of the above cases are satisfied.
Fig. 2 shows an example of how BPP works. It will start parsing the block
Algorithm 1 Block Parallelized Parsing Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
a = start symbol of input
$ = last symbol of input
h = initial parsing subtable
stack = stack of states
while true do
s = stack.top()
if T ab(S h )[s, a] == shif t t then
stack.push(t)
a = next symbol of input
else if T ab(S h )[s, a] == jump k then
h=k
else if T ab(S h )[s, a] == reduce A → β then
b=a
if a == sb and T ab(S h )[s, a] == reduce At → αt then
get Ct related to At
stackt = new stack of states
stackt .push(start state of Ct )
pop |αt | states f rom stackt and push them to stackt
create new block parser related to Ct with stackt
start new block parser
go to the end of this block
a = Ct
else
pop |β| states f rom stack
t = stack.top()
if T ab(S h )[t, A] == shif t p then
stack.push(p)
else if T ab(S h )[s, a] == accept then return
else
error
Fig. 2. Example of BPP parsing source code
of function f. When it will encounter an if block a new BPP in another thread
will be created which will parse if block. Parent BPP will move ahead to the
end of if block and will also create another thread to parse else block. In this
way input is divided into different threads parsing each block.
In this algorithm we have tried to minimize the amount of serial work to be
done to get to the end of the block for the new block parser. One BPP doesnt
have to do any communication with other BPPs. Also, there are no side effects
of the above algorithm. All the variables which are being modified are local variables. Hence, there is no need of synchronization. This also reduces any amount
of cache contention between different processors. Generation of Abstract Syntax
Tree or Parsing Tree is easy using above algorithm and it requires very little
change in the above algorithm.
It may be argued that the step go to the end of this block is a serial bottleneck for the parallel algorithm. [15] describes an algorithm to perform lexical
analysis of string in O(log n) time using O(n) processors in a parallel fashion.
When performing lexical analysis in parallel as described in [15], lexer could
store the start symbols of a block with its corresponding end symbol for that
block. Now determining the end of block is just a matter of searching through
the data structure. Many ways exist to make this searching as fast as possible
like using a Binary Search Tree or a Hash Table.
9
Comparison with other Parallel Parsing Algorithms
In this section we will show how our technique is better than other techniques.
[14] developed a technique which divides whole grammar into n sub- grammars
which are individually handled by n sub-compilers. Each sub-compiler needs its
own scanner which can scan a sub-grammar. It requires an automatic tool to
generate sub-compiler. This technique requires significant changes in not only in
Parser and Grammar but also in Lexical Analyzer phase. Contrast to this our
Block Parallelized Parser is easy to generate as our technique does not require
any change in the grammar and lexical analyzer and it is pretty easy to modify
current YACC and Bison Parser Generator tools to support the generation of
our parser.
LR substring parsing technique described in [2] is specifically for Bounded
Context (1, 1) grammars. There are no limitations like this to Block Parallelized Parser. Although in this paper we have shown how we can create an LR
(1) Block Parallelized Parser but we believe it can be extended to LR (k) class
of languages and also could be used by LALR (1) parser. Hence, our technique
accepts a larger class of grammars.
[8], [7], [9], [11] all develops algorithm for parsing LR (k) class of languages
in parallel. These techniques and in all other techniques the creation of Abstract
Syntax Tree is not as easy as itis in our technique. Moreover our technique is
simpler than all others.
Hence, we could see that Block Parallelized Parser is easy to construct, accepts wider class of languages and supports an easy construction of Abstract
Syntax Tree.
10
Implementation and Evaluation
We implemented Lexer, Block Parallelized Parser and Shift Reduce LR (1) parser
for C Programming Language supporting a few GNU C extensions required for
our tests. Implementation was done in C# Programming Language. To simplify our implementation we only included function-definition as the Incremental
Category for BPP. Moreover, function-definition would still give us a sufficient
amount of parallelism as we would see in the evaluation. We modified the Lexer
phase so that it will keep track of the position of s b and its corresponding e b
. This information was stored in the form of a C# Dictionary (which is implemented as a Hash Table) with the position of s b as the key and position of e b as
the value. As, thread creation has significant overhead so we used C# TaskParallel Library which is Thread Pool implementation in C#. Our implementation
doesnt have a C preprocessor implementation. So, we first used gcc to perform
preprocessing and the preprocessed file is used as input to our implementation.
We evaluated the performance of BPP with Shift Reduce LR (1) parser by
parsing 10 random files from the Linux Kernel source code. We compiled C#
implementation using Mono C# Compiler 3.12.1 and executed the implementation using Mono JIT Compiler 3.12.1 on machine running Fedora 21 with Linux
Kernel 3.19.3 with 6 GB RAM and Intel Core i7-3610 CPU with HyperThreading enabled.
In C Programming Language, preprocessing #include could actually generate very long files. Normally, the header files contains declarations not function
definitions. So, this leads to less amount of parallelism being available. Hence we
decided to evaluate with including header files and excluding header files. Fig. 3
shows the performance improvement with respect to Shift Reduce LR(1) parser
of 10 random Linux Kernel files. Fig. 3 shows performance improvement for both
cases including the header files and excluding the header files. As expected we
Fig. 3. Results of parsing 10 Linux Kernel Source Files.
could see that the performance improvement with excluding the header files is
more than the performance improvement including the header files.
The performance improvement in the case of excluding header files matters most for the programming languages like Java, Python, C# where most of
the program is organized into blocks the results because in these programs the
amount of parallelism available is high.
The average performance improvement in the case of excluding header files
is 52% and including header files is 28%.
11
Conclusion
In this document we present our Block Parallelized Parser technique which could
parse the source code in a parallel fashion. Our approach is a divide and conquer approach in which, the divide step divides the source code into different
blocks and parse them in parallel whereas the conquer step only waits for all
the parallel parsers to complete their operation. It is based on the Incremental
Jump Shift Reduce Parser technique developed by [13]. Our technique doesnt
require any communication in between different threads and doesnt modify any
global data. Hence, this technique is free of thread synchronization. We develop
this technique for LR (1) languages and we believe that it can be extended to
accept LR (k) languages and could be converted to an LALR (1) parser easily.
Our technique doesnt do any major changes in the parsing algorithm of a Shift
Reduce Parser hence the Abstract Syntax Tree can be created in the same way
as it has been creating in Shift Reduce Parser. Moreover, our parser can also
work as an Incremental Block Parallelized Parser. We implemented Block Parallelized Parser and Shift Reduce LR (1) Parser for C Programming Language
in C#. The performance evaluation of BPP with Shift Reduce LR (1) parser
was done by parsing 10 random files from the Linux Kernel source code. We
compiled C# implementation using Mono C# Compiler 3.12.1 and executed the
implementation using Mono JIT Compiler 3.12.1 on machine running Fedora 21
with Linux Kernel 3.19.3 with 6 GB RAM and Intel Core i7-3610 CPU with
HyperThreading enabled. We found out that our technique showed 28% performance improvement in the case of including header files and 52% performance
improvement in the case of excluding header files.
12
Future Work
Our parser accepts LR (1) languages we would like to extend it to accept LR (k)
languages. In our technique, the parser determines when to create a new parallel
parser thread. If the responsibility of this decision can be given to the lexical
analysis phase then the lexical analysis can actually start the parsers in parallel.
This will lead to significant performance advantage. Moreover, our technique
has been applied to languages which doesnt have indentation in their syntax like
the way Python has. [24] shows an efficient way to parse the language which has
indentation as a mean to determine blocks. Our parser can be extended to accept
those languages also. We are working towards developing a Block Parallelized
Compiler which could compile different blocks of a language in parallel. Block
Parallelized Parser is one of the components of a Block Parallelized Compiler.
Semantic Analysis phase also share the same properties as the Syntax Analysis
phase. In Programming Languages, an entity like variable or type is declared
before using it. So, in this case also a lot can be done to actually parallelize the
semantic analysis phase.
References
1. Neal M. Gafter and Thomas J. LeBlanc. Parallel Incremental Compilation. Department of Computer Science, University of Rochester: Ph. D. Thesis. 113
2. Gwen Clarke and David T. Barnard. 1996. An LR Substring Parser Applied in a
Parallel Environment, Journal of Parallel and Distributed Computing.
3. G. V. Cormack. 1989. An LR Substring Parser for noncorrecting syntax error recovery, ACM SIGPLAN Notices
4. Alfred V. Aho, Monica S. Lam, Ravi Sethi and Jeffery D. Ullman. 2007. Compilers
Principle Tools and Techniques Second Edition, Prentice Hall
5. Floyd, R. W. Bounded Context Synctactic Analysis. Comm. ACM 7, 21 February
1961
6. J. H. Williams, 1975. Bounded context parsable grammars. Information and Control
28, 314-334
7. R. M. Schell, 1979. Jr. Methods for constructing parallel compilers for use in a
multiprocessor environment. Ph. D. thesis, Univ. of Illinois at Urbana-Champaign
8. J. Cohen and S. Kolodner. 1985. Estimating the speedup in parallel parsing, IEEE
Trans. Software Engrg. SE-11
9. C. N. Fischer, 1975. On parsing context free languages in parallel environments. Ph.
D. thesis, Cornell Univ.
10. M. D. Mickunas and R. M. Schell. 1978. Parallel Compilation in a multi-processor
environment (extended abstract). J. Assoc. Comput
11. D. Ligett, G. McCluskey, and W. M. McKeeman, Parallel LR Parsing Tech. Rep.,
Wang Institute of Graduate Studies, July 1982
12. Edward D. Willink, Meta-Compilation for C++, University of Surrey, June 2001
13. Pierpaolo Degano, Stefano Mannucci and Bruno Mojana, Efficient Incremental LR
Parsing for Syntax-Directed Editors, ACM Transactions on Programming Languages
and Systems, Vol. 10, No. 3, July 1988, Pages 345-373.
14. Sanjay Khanna, Arif Ghafoor and Amrit Goel, A PARALLEL COMPILATION
TECHNIQUE BASED ON GRAMMAR PARTITIONING, 1990 ACM 089791-3485/90/0002/0385
15. W. Daniel Hillis and Guy L. Steele, DATA PARALLEL ALGORITHMS, Communications of the ACM, 1986
16. Bison, http://www.gnu.org/software/bison
17. YACC, http://dinosaur.compilertools.net/yacc
18. Clang, http://clang.llvm.org
19. Mono, http//mono-project.com
20. El-Essouki, W. Huen, and M. Evans, Towards a Partitioning Compiler for Distributed Computing System, IEEE First International Conference on Distributed
Computing Systems, October 1979.
21. Thomas J. Pennello and Frank DeRemer, A Forward Move Algorithm for LR Error
Recovery, Fifth Annual ACM Symposium on Principles of Programming Language.
22. Jean-Philipe Bernardy and Koen Claessen, Efficient Divide-and-Conquer Parsing
of Practical Context-Free Languages.
23. L. Valiant. General context-free recognition in less than cubic time. J. of Computer
and System Sciences, 10(2):308-314, 1975.
24. Micheal D. Adams. Principled Parsing for Indentation-Sensitive Languages, ACM
Symposium on Principles of Programming Languages 2013. January 23-25.
| 6cs.PL
|
arXiv:1412.6144v1 [cs.NE] 14 Dec 2014
The Computational Theory of Intelligence:
Applications to Genetic Programming and Turing
Machines
Daniel Kovach
December 22, 2014
Abstract
In this paper, we continue the efforts of the Computational Theory
of Intelligence (CTI) by extending concepts to include computational
processes in terms of Genetic Algorithms (GA’s) and Turing Machines
(TM’s). Active, Passive, and Hybrid Computational Intelligence processes
are also introduced and discussed. We consider the ramifications of the
assumptions of CTI with regard to the qualities of reproduction and virility. Applications to Biology, Computer Science and Cyber Security are
also discussed.
1
Introduction
In this work, we carry on the ideas proposed in [15, 16]. In [15], we posited some
very basic assumptions as to what intelligence means as a mapping. Using our
analysis, we formulated some axioms based on entropic analysis. In the paper
that followed [16] we talked about how ideas or memes can aggregate together
to form more complex constructs.
Here, we extend the concept of intelligence as a computational process to
computational processes themselves, in the classic Turing sense.
The structure of this paper is as follows. In section 2, we provide deeper
insights as to different types of computational intelligence, specifically in the
light of executable code, section 3 the qualities of reproduction, virility, and
nested programs, section 4 presents some experimental results, and sections 5
and 6 discuss applications. In the conclusion, we recapitulate and discuss future
work.
2
Extending Concepts
Consider the definition of computational intelligence from [15]. Given the two
sets S and O, which represent input and output sets, respectively, we introduce
the mapping
1
I : S → O,
(1)
where I essentially makes a prediction based on the input. The disparity
between a prediction and reality can then be used to update I in such a way
that it can climb the gradient of the learning function L so as to improve future
performance. Although we touched upon the application of this framework to
genetic algorithms [15], it will be insightful to revisit it in more detail.
2.1
Active, Passive, and Hybrid Computational Intelligence
Let us clarify the terminology. By genetic algorithm (GA), we mean a collection or population Π = {πi } of potential solutions to a particular task, which
evolves as time progresses according to some fitness function, F. The function
F is synonymous with learning function L from our previous formulation. It
evaluates the quality of Π as a solution to the task. Specifically, if our task or
program P is a function of Π, then the fitness function F is nothing more than
F = F(P (Π)).
(2)
Note that the population Π is not only our input set S but our output set O
in the representation 1. Furthermore, the mapping I is nothing but a collection
of mutation operations to update the Π at iteration t to iteration t + 1. That is
I : Πt → Πt+1 .
(3)
Some common mutation operations of genetic algorithms are [8]:
1. Reproduction: selecting individuals within a population and copying them
without alteration into the new population.
2. Crossover: produces new offspring from elements of two other parents.
3. Mutation: one of the central concepts of genetic algorithms. Mutation
perturbs the population or its constituents in some way. This includes altering the elements of the population directly, or changing their placement
in the population in some way such as swapping their indices.
4. Editing: a manner in which more complicated functionality can be reduced
to simpler algebraic rules.
5. Adding/Deleting: the ability to add a new element to a population or
remove an existing element from the population.
6. Encapsulating: the process of creating a new element which represents an
aggregate of functionality from one or more other elements in the population.
2
Some consideration should be given to the statement
It+1 = It + ∇Ft
(4)
from [15]. For this equation to make sense, not only must the algebraic
operation ’+’ be meaningfully defined, but the gradient operator implies that F
is differentiable. The gradient operation makes sense only if the domain of F is
a Banach space . In this context, we cannot make such a claim. Thus equation
4 becomes
It+1 = It + ∆F,
(5)
where ∆F = F(Πt ) − F(Πt−1 ). Observe that although the mathematical
formulation is different, the intent still holds. The ’+’ operator communicates
the number of permutations proportional to ∆F.
Since the information is retained in the population, and only improved
through evolutionary perturbations, we consider this an example of passive computational intelligence. This stands in contrast to active computational intelligence, where we are updating the mapping I itself. Any combination of the two
methods can be considered a form of hybrid computational intelligence.
2.2
Genetic Algorithms
Recall the entropic considerations from [15], where we specified that
• Computational intelligence is a process that locally minimizes and globally
maximizes entropy.
Since the population Π is the output set, it is the entropy of the population that
must be minimized by the intelligence process. Equivalently, we can say that
as the amount of iterations becomes sufficiently large, the difference between
Πt and Πt−1 becomes increasingly negligible. In other words, assuming the
subtraction operation is meaningfully defined, we have
lim Πt+1 − Πt = 0.
t→∞
(6)
This makes sense, for we expect that as time increases, the fitness of our
solution reaches an optimum, so that eventually there is no need to update. In
other words, our population becomes more consistent over time.
Of course, the typical considerations regarding genetic algorithms such the
necessity of avoiding local optima, the proper choice of a fitness function, the
complexity of the algorithm itself, among others, still hold [8].
As for global entropy considerations, we must consider the role of the population in its realization under its program P . As mentioned before, the total
entropy of a system can be written as:
S=
N
X
i=1
3
si
(7)
where {si } represent each respective source of entropy. In particular, under
the considerations of this section, we have
S = sΠ + sP ,
(8)
where sΠ and sP represent the entropy due to the population itself and the
entropy from the program P , respectively.
To calculate the entropy, we used the Renyi entropy of order α, where α ≥ 0
and α 6= 1 defined as
"N
#
X
1
α
log
P [xi ] .
(9)
Hα (X) =
1−α
i=1
where P[xi ] is the probability associated with state xi ∈ X.
2.3
Turing Machines
Let us consider the discussions of the previous sections in terms of code that
is executable in the Turing sense. We will apply the formulations we have
considered thus far and make improvements to refine the theory in this light.
Recall that a Turing machine M is a hypothetical device that manipulates
symbols on a tape, called a Turing tape, according to a set of rules. Despite its
simplicity, the Turing machine analogy can be used to describe certain aspects
of modern computers [21], and more [13].
The Turing tape has some important characteristics. The tape contents,
called the alphabet, encode a given table of rules. Using these rules the machine
manipulates the tape. It can write and read to and from the tape, change the
configuration, or state Σ of the machine, and move the tape back and forth [11].
Typically, the instructions on the Turing tape consist of at least one instruction that tells the machine to begin computation, or a START instruction, and
one to cease computation, called a STOP or HALT instruction [11]. Other instructions might include MOVE, COPY, etc, or even instructions that encode
no operation at all. Such instructions are commonly called NOOP’s (NO OPeration). The total number of instructions depends on the situation and the
application at hand, but for the purposes of this paper, we will consider only
tape of finite length for which the instruction set is also finite and contains at
least the START and STOP instructions. In fact, for our purposes, the minimum amount of opcodes a program needs to run, or to be executable, are the
START and STOP instructions.
Now, consider the set of all instructions available to a particular M . To be
consistent with our notation thus far, we will denote this set as π
e, where
π
e = {π0 = START, ..., πn = STOP}
(10)
From this set, a particular sequence of instructions Π can be formed. Although many synonymous terms abound for this sequence of instructions, including a code segment, executable, or tape, we will refer to it simply as the
4
code. We consider the execution or implementation Ψ to be the execution of
the code on the Turing machine M , which produces could produce another code
segment, and change the machine state. Expressed in notation:
e ×Σ
e→Π
e × Σ.
e
Ψ:Π
(11)
e and Σ
e represent the set of all code and machine states, respectively.
where Π
For the majority of this paper, it is the code that is of primary concern, and
thus we can omit the machine state to simplify the notation when this context
is understood:
e →Π
e
Ψ:Π
(12)
Our fitness function F is a measure the efficaciousness of the implementation
of our program by some predefined standard. Later, we will present some specific
examples used in application.
We are still free to think of the intelligence mapping I as a perturbation
mapping as we did in section 2.1. The only difference is the physical interpretation of the term ’population’. This case concerns machine code instructions.
As with genetic algorithms, perturbing code may advance or impede the fitness
of the program. In fact, it may even destroy the ability to execute at all, if one
either the START or STOP operations are removed.
A final note should be made as to the entropic considerations of the Turing
Machine. In this light, we must take into account all the workings of the machine
and thus 8 becomes
S = sΠ + sM
(13)
where sM is the entropy incurred by the Turing Machine, which includes its
state Σ.
3
Reproduction, Viruses, and Nested Programs
In this section, we will continue the above analysis, and discussed some qualities
that emerge based on our assumption.
3.1
Formalisms of Code and Other Qualities
We can apply set theoretic formalisms in the usual sense. For example, consider
a code Π and an instruction, π ∈ π
e. Then the notation π ∈ Π communicates
the fact that π is an instruction in Π. Further, set theoretic operations follow
e In the case of the union operation, we have
naturally. Let Π1 , Π2 ∈ Π.
Π1 ∪ Π2 = {Π1 , Π2 }
(14)
which is a set containing the two programs Π1 and Π2 . Similarly, with the
intersection operation, we have
5
c ∈ Π1 ∩ Π2 ⇔ c ∈ Π1 ∧ c ∈ Π2 ,
(15)
or the set of all contiguous code segments c = {π} that are contained by
both Π1 and Π2 . Note that this code segment need not be executable. Also,
if it is of particular utility in the program it may be degenerate in that it may
appear multiple times throughout the code.
For set inclusion, we say that
Π1 ⊂ Π2 ⇔ π ∈ Π2 ∀π ∈ Π1
(16)
Proper containment evolves naturally if Π1 = Π2 .
3.2
Metrics and Neighborhoods
The differences between code and real numbers or sets in conventional mathematics make it at first difficult to form a distance metric d on them. Still, some
have been explored such as the Levenshtein distance [9], Damerau-Levenshtein
distance [2], Hamming distance [10], and the Jaro-Winkler distance [12].
Given a metric d, we can form an open ball [7], or neighborhood B centered
at some code Π0 consisting of all programs Π such that for some real number
,d(Π0 , Π) < . In other words
B(Π0 , ) = {Π : d(Π0 , Π) < }.
(17)
We can utilize the developments above to relax the definition of set inclusion
with respect to the metric, d. Specifically, define
Π1 ⊂ Π2 ⇔ d(Π1 , Π) < , Π ⊂ Π2
Of course, similar considerations can be extended to proper containment.
Such code segments that lie suitably close to each other with respect to d will
be referred to as polymorphic with respect to d.
3.3
Reproduction
Consider the mapping
Ψi (Π) = Π0
(18)
where Π0 is in the neighborhood of Π. Of particular interest is when Π0 not
only lies within the neighborhood of Π but produces a machine state Σ0 suitably
close to that of Π, as valued by some meaningful metric. We note in particular
the condition when Π0 also encodes for reproduction. Unless in the presence of
polymorphic programs [20], we will assume the code copies itself exactly, and
replace Π0 with Π in notation.
Sequential program execution will yield the same result
Ψt = Ψ(Ψt−1 ), Ψ0 = Ψ(Π)
6
(19)
where
Ψi (Π) = Π, ∀i.
(20)
Suppose that a particular tape is able to produce, say N copies of itself.
Then the entropy of the system is simply
S = sM +
N
X
sΠi ,
(21)
i=0
where s0 refers to the original copy of the tape. By comparison with equation
13, the entropy is proportional to the number of progeny in addition to the
machine state.
3.4
Viruses
Thus far, we have been considering only predefined programs. But a well known
concept from nature concerns when external code is injected into a host program, or otherwise hijacks the computational resources of the system. One key
characteristic of viruses is their ability to proliferate. Consider ν such that
ν ∈ Ψ(Π), ν ⊆ Π
(22)
If the viral code, ν is executable, we will call the virus ν-executable. If the
viral code also can reproduce, that is ν ∈ Ψ(ν), we will call this quality νreproductive. Finally, it is often the case that viral code is not reproductive or
executable. Its task is to merely copy or install some payload [18], here denoted
by ρ. In this case,
ρ ∈ Ψ(Π), ρ ⊆ ν ⊆ Π
(23)
Viral code most often is pernicious, though this is not necessarily the case.
For example, up to 8% of the human genome derives from retroviruses [3]. To
determine the nature of the effect of the virus, we look at the change in fitness
before and after its influence.
The influence
T of viral code on
T program fitness falls into three categories.
Consider Π, ν Π = ∅, and Π0 , ν Π0 = ν, Π = Π0 − ν. In other words, Π0 code
in the presence of a virus, from the otherwise pristine Π. The three cases are
F(Π0 ) − F(Π) > 0, F(Π0 ) − F(Π) = 0, and F(Π0 ) − F(Π) < 0. These cases are
of enough significance that they deserve their own nomenclature. We will refer
to these cases as commensalistic, symbiotic, and parasitic, respectively, out of
consistency with their counterparts from biological sciences.
3.5
Nested Programs
One final case to be considered is when a program or multiple programs can
contribute to produce aggregate programs of greater complexity. In such a case,
the program, when executed, would not only alter machine state and produce
7
code as in 11 but will also contain a product code segment, and product code
machine state,
e ×Σ
e→Π
e ×Σ
e ×Π
e1 × Σ
e1
Ψ:Π
(24)
Introduce the program
e1 × Σ
e1 → Π
e1 × Σ
e1
Ψ1 = Π
(25)
Of course, we could nest this process and define
e1 × Σ
e1 → Π
e1 × Σ
e1 × Π
e2 × Σ
e 2,
Ψ1 = Π
(26)
with Ψ from 24 denoted as Ψ0 , we could write
en × Σ
en
Ψn [Ψn−1 [...Ψ0 [Π0 × Σ0 ]...]] → Π
(27)
where the superscript indicates the level of nesting of the program.
Just as we discussed programs producing multiple progeny, we can also let
them produce multiple product code. We will denote each copy with a subscript
im = {0, 1, ..., Im }. Thus the above becomes
0
0
0
en
en
Ψn [Ψn−1
in−1 [...Ψi0 [Πi0 × Σi0 ]...]] → Πin × Σin .
(28)
To determine the total entropy, simply sum over all product code.
S = sM +
Ik X
n
X
e kj × Σ
e kj
Π
(29)
j=0 k=0
This time we see that the entropy of the code varies exponentially with
product and copy code.
4
Experimentation
Consider the following scenario. We have a set of instructions, and a construct
for a tape that encodes them. In the following experiments, the code is created
randomly. We apply a GA to the code to acheive interesting results. The code
involved had no write back capabilities, and thus the experiment was purely
passive.
4.1
Genetics and Codons
A major focus for future work is the application of this research to computational
molecular biology. Hence, we chose to implement the instruction set that nature
has chosen for genetics in the following experiments.
For the base instruction alphabet, we will choose that of RNA [19]. This
instruction set is composed of the ’bits’ Adenine (A), Uracil (U), Guanine (G),
8
and Thymine (T). Each instruction can be represented as three of these quaternary bits, for a total of 64 possible outcomes. We have chosen this quaternary
codon formulation to be the basis of our tape.
In both experiments, the choice of operational codes, or opcodes, are taken
directly from these 64 quaternary bit codons. The choice to represent a opcode
by a particular codon is completely arbitrary. All codons appear with the same
probability in experiment, with the exception of the STOP opcode, as we will
see in section 4.2.
We present two instructions sets, and two different experiments. In the first
experiment, we determine how many iterations our genetic algorithm takes to
produce code that is executable, and code that reproduces itself. In the second
experiment, we compare the amount of reproductions of code with the total
entropy incurred by the simulation.
4.2
First Instruction Set
For the first experiment we constructed a robust instruction set consisting of
the following:
1. START (AAA): Commences program execution.
2. STOP (AUA, ATC, ATG): Halts program execution. Observe that there
are three STOP codons, but only one START codon. This mirrors the
amount of codons in RNA and DNA observed in nature [4].
3. BUILD FR (CUC): Copies to product code starting from the next respective instruction to that which comes before the BUILD TO (GCG)
instruction.
4. COND (UUC, UUA, GAA): Sets an internal variable in the Turing machine state called a flag. A flag can be thought of as an internal Boolean
variable. Every time one of the COND opcodes is encountered the sign of
the flag is switched, that is COND = ¬COND.
5. IF (AAU): Only executes the following instruction if the COND flag has
been set.
6. COPY ALL (AAG): Copies the entire tape to progeny.
7. COPY FR (CCC): Copies to progeny code starting from the next respective instruction to that which comes before the COPY TO (GGG)
instruction.
8. JUMP FAR FR, (CUU) and JUMP NEAR FR, (AGA): both require a
JUMP TO, (CAC, GUG) instruction. The JUMP family of instructions
is designed to continue program execution at the index of the JUMP TO
instruction. The former will find the JUMP TO address that is farthest
away, while the JUMP NEAR finds the closest JUMP TO instructions. If
there is no multiplicity in the JUMP TO instruction, then JUMP FAR FR,
(CUU) and JUMP NEAR FR will produce the same result.
9
9. REM FR, (GCU) and REM TO, (UAA): Removes all code in between
these instructions.
4.3
Second Instruction Set
Although the code of the previous section was robust, it was redundant. For
example, the functionality of the code to copy itself to progeny should be a
desireable trait of the code itself, not just an opcode that accomplishes the
task in one step. Further, the remove operations can be handled via mutation
operations (albeit far less efficiently) in the genetic algorithm itself.
Note that in Experiment 1, most operations have a partner or conjugate
opcode. The necessity of conjugate pairs of instructions follows from the fact
that two pieces of information are required to complete the instruction, as shown
below:
[ i n s t r u c t i o n ] . . . < code > . . . [ i n s t r u c t i o n d u a l ]
We will call such instructions dual. The necessity for a predefined conjugate
opcode can be mitigated by using the concept of addressing. Instead of conjugate
opcodes, the instruction immediately following the dual can be thought of as
an argument. The machine uses this address (unless the address is represents
opcode) as a manner of locating the conjugate to the dual and thus reducing
the instruction set. Specifically:
[ i n s t r u c t i o n ] [ a d d r e s s ] . . . < code > . . . [ a d d r e s s ]
Under these considerations (and omitting the ability to build to a product),
we have need for only six instructions: START, STOP, IF, COND, COPY, and
JUMP. The first four instructions are exactly the same as in Experiment 1.
The COPY instruction concatenates section of the tape to progeny, and the
JUMP instruction continues program execution at a specified address. These
final two are conjugates, the duals of which are handled via the addressing
method entailed above. Notice although the lexicon is different, they have the
same functionality as their respective counterparts in the first instruction set.
4.4
Experiment 1: Execution and Reproduction
The goal of the first instruction set was to produce code that was executable, and
code that could reproduce. The code was perturbed by a GA randomly. That
is, no fitness function was used in the perturbation of the code strings. Each
epoch of the simulation terminated once an executable or reproducable code
was discovered. For each experiment, a maximum of 1,000,000 computational
iterations were allowed.
The results of the first experiment using the first instruction set are shown
in table 1. The results of the first experiment using the second instruction set
are summarized in table 2. Observe that the number iterations in the first
instruction set was much less than that of the second, which is likely due to the
disparity in robustness between the two. Also note how wildly the results vary
10
Table 1: Results: Experiment 1, Instruction Set 1
Experiment
Executable Code
Reproductive Code
Average
184.083
330.964
Std. Dev
126.202
243.924
Experiments
100,000
96,964
Table 2: Results: Experiment 1, Instruction Set 2
Experiment
Executable Code
Reproductive Code
Average
283.387
2003.24
Std. Dev
210.359
96504.8
Experiments
1,000,000
1,000,000
especially in terms of the reproductive capabilities of the second instruction set
as demonstrated by the large standard deviation in results.
4.5
Experiment 2: Reproduction and Entropy
Upon answering the basic questions like the amount of computational iterations
necessary to produce executable and reproductive code, the next experiment
focused on the total entropy incurred by a given code segment and its offspring.
Here, recall that we are considering the entropy produced by not only a given
tape, but its progeny as well. We will continue with the alphabets of Experiment
1 and 2, respectively.
We bounded the total computational iterations per experimental run at
1,000,000. The maximum progeny allowed was 50. Mutations were applied
by the GA according to a fitness function which evaluated the quality of the
code with respect to its Renyi Entropy (of order α = 2).
The results are summarized in table 3. Observe that the entropy and the
amount of reproductions appear to be correlated as demonstrated by the r-values
in table 3 and in the correlation maps in Figure 1. Note that some programs
appear to have acheived reproductive qualities without substantially increasing
entropy but the converse is not true.
Although the actual data summarized in table 3 was one of the primary goals
of this experiment, the operational characteristics of the code itself were also of
interest. The graphics in figure 2 represent some interesting code behaviors. In
the graphs below, each opcode was represented by a number in the following
manner: START = 0, COPY = 1, JUMP = 2, IF = 3, COND = 4, STOP = 5.
With these numeric values we can visualize program behavior graphically.
Although a myriad of behaviors were observed, in Figure 1 we have selected some graphs of the different behaviors. Of specific interest were the
non-terminating programs. Each of these selections also exhibited reproductive
capabilities. Observe that although towards the beginning of program execution, a variety of instructions were executed, eventually the programs under
consideration converged to some sort of periodic behavior, a tendency we will
call asymptotic periodicity.
11
Table 3: Results: Reproduction vs. Entropy
Total Simulations
Average in Reproductions
Std. Dev. in Reproductions
Average in Entropy
Std. Dev. in Entropy
r-value
(a) Reproduction vs. Entropy (Alphabet 1)
Alphabet 1
1,025,043
17.1433
23.7111
12.4556
18.5525
0.870675
Alphabet 2
577,876
39.3921
20.1516
34.8463
17.6445
0.983243
(b) Reproduction vs. Entropy (Alphabet 2)
Figure 1: Graphs of reproductive iterations plotted against the total entropy of
the simulation, for the first and second instruction sets, respectively.
The graphs selected not only demonstrated asymptotic periodicity but reproduced continuously in the sense that the progeny produced reached the maximum allowable threshold.
5
Applications to Biology
Readers familiar with genetics will notice a strong correlation between our computational formulations and genetics. In fact, genetics has been an inspiration
for this research, specifically with regard to passive and hybrid forms of CTI.
Recent research has been devoted to the crossover between genetics and computer science. In fact, it has been demonstrated in [13] that DNA computing
is Turing complete. This has grand implications for the very nature of the
definition of life itself.
The definition of life has always been somewhat controversial. In [1], life is
defined as a characteristic that distinguishes objects with self sustaining processes from those that do not. From the vantage point of Biology, life is defined
as a characteristic of organisms that exhibit all or most of the following characteristics: homeostasis, organization, metabolism, growth, adaptation, response
12
(a) Program Behavior from Sample 457
(b) Program Behavior from Sample 506
(c) Program Behavior from Sample 1156
(d) Program Behavior from Sample 1519
Figure 2: Four examples of asymptotically periodic code behaviors taken from
samples of code that reproduced continuously in experiment 2.
13
to stimuli, and reproduction [14]. Many sources stress the chemical nature of
life, and mandate that living things contain Carbon, Hydrogen, Nitrogen, Oxygen, Phosphorous and Sulfur [17].
In this paper, we introduce a definition for life as a self sustaining passive or
hybrid CTI process. Embedded in this definition are well known assumptions
regarding qualities one commonly associates with living things. It is known
that the purpose of life is to survive, a fact that is redundant in the light of
CTI. The fact that life is fitness maximizing, and increases entropy follows
directly from the framework of CTI as well. Homeostasis follows naturally
from the entropy minimizing portion of the fundamental axiom of CTI. If an
organism is in homeostasis, this implies that the internal state space is ordered
and transitions can be predicted with high probability which implies a lower
entropy. Further, note the similarities with asymptotic periodicity from the
samples shown in 2. Of course, no discussion about life would be complete
without mentioning reproduction, which we showed in this paper follows as a
natural consequence of CTI.
As far as we know, Earth is the only planet in the universe known to sustain life. However, recent advances in planetary sciences have discovered the
possibility of habitable zones in other solar systems. This coupled with the observed robustness of life on Earth in extreme environments, has raised interest
the application of biological concepts to the search for life beyond the confines
of Earth. The vastness of these environments forces us to rethink and generalize
what we know as life itself. One fantastic example of such an esoteric formulation of life is the discovery of arsenic based life right here on Earth. In December
of 2010, NASA unveiled a discovery of a bacterium that integrated Arsenic into
its cellular structure and even its DNA. They found that the bacterium was
able to substitude arsenic for phosphorous, one of the six essential ingredients
of life. This challenges conventional notions of life and seems to encourage a
more general definition. In this paper, we advocate a computational approach
to the definition of life over a chemical one.
6
Applications to Computer Science and Cyber
Security
Applications to cyber security abound as well. In the past, computer viruses
typically served little more than the entertainment of the author. Now, malicious softaware, or malware, is a commoditized industry. Patrons will hire
teams of programmers to create specialized malware for end goals such as espionage, ransome, or identity theft. As such, mitigating these threats has become
a priority in most industries and even an interest in national security. For example, the US Defense Advanced Research Projects Agency (DARPA) has issued
a ’Cyber Grand Challenge’ offering up to $2 million for designs of a fully automated software defense system, capable of mitigating software exploitation in
real time [5]. This is just one example of increased efforts in this area. In fact,
14
the US president is proposing to increase the cyber defense budget to $14 billion
in 2014, more than a billion dollar increase from that of the previous year [6].
How might this research be of merit in this field? Currently, most antimalware software (AM) stores hashes of actual malware samples against which
to compare new potential threats. This has multiple disadvantages in that it ties
users to a provider which may be an issue if internet connectivity is a problem.
Frequently such hashes fail with polymorphic code.
Recently, much research has been conducted looking at a behavioral approach to malware classification. While this is an exciting area of resarch,
approaches of this nature are typically burdened with false positives rendering
the solutions impractical.
This reserach offers the potential to view executable code in a new light
offering insights into two key areas of interests: the nature of viruses, and the
executable code itself, offering a new paradigm to the software industry. Software serves a purpose. It has solid attainable goals to suit the end user. Concurrently, it must function within a specified range of parameters to preserve
privacy, security, and even safety. The effectiveness of a program under these
guidelines can be viewed as its fitness.
Regardless of whether we are analysing virus behavior or developing code
that optimizes its fitness, we need a reliable indicator as to its behavior. One
of the first successful defenses of malware was to attack the very functions the
AM product would call to detect it! This is akin to asking the burgler if he is
in your house. Thus behavior detecting components must function at a ’lower’
level than those to which the malware may have access. Insuring this can be
quite challenging. Nevertheless, if such a framework were in place, the advantages have enormous potential. In terms of viruses and payload detection, we
could apply analysis from section 3.4. Equivalently, we could monitor program
performance to scan for points of inadequacy. Moreover, in the event of a breach
or infection, we could employ the methods of CTI as a means by which the code
could heal itself. Again, this necessitates a proper fitness function to evaluate the ’health status’ of the executable, and trusted behavioral data from the
system.
Of course, there is a wide rift between the musings of theory and the result
of implementation, but this framework seems to provide a promising direction
for future efforts in this arena.
7
Conclusion
To recapitulate, in this paper, we extended the CTI framework presented in
[15, 16] to include genetic algorithms and Turing machines. We defined classes
of CTI including active, passive and hybrid. Finally, we demonstrated that
reproduction emerges as a consequence of the axioms of CTI, both theoretically,
and experimentally.
The concepts we presented have great potential for development in future
work. As far as the theoretical concepts of this paper are concerned, we have
15
a lot to do in terms of exploring the potential of passive, active, and hybrid
processes. It remains the stance of this paper that hybrid CTI processes are
present optimal solutions but this remains to be shown. We would also like
to further explore simulation and programming using reproductive nested code.
Further, we would like to repeat the experiments of section 4 with different
instruction sets, and observe not only the behaviors, but the phenotypes of
the solutions in general, to observe overarching patterns among diverse coding
alphabets that acheive the same result.
In terms of multidisciplinary studies we have only scratched the surface of
the relationship between these concepts and computational molecular biology,
computer science, and cyber security.
Further, there is much to do in the way of advancing the theoretical framework of CTI itself. In the next paper, we will visit the ramifications of adding
feedback into the intelligence process. Another paper will focus on the global
properties of intelligent agent.
References
[1] The American Heritage Dictionary. New York: Dell, 2012. Print.
[2] Bard, Gregory V. ”Spelling-error Tolerant, Order-independent Passphrases via the Damerau-Levenshtein String-edit Distance Metric” Proc.
of Proceedings of the Fifth Australasian Symposium on ACSW Frontiers.
Vol. 69. N.p.: n.p., n.d. 117-24. Print.
[3] Belshaw, R. ”Long-term Reinfection of the Human Genome by Endogenous
Retroviruses.” Proceedings of the National Academy of Sciences 101.14
(2004): 4894-899. Print.
[4] Brenner, S., A. O. W. Stretton, and S. Kaplan. ”Genetic Code: The
‘Nonsense’ Triplets for Chain Termination and Their Suppression.” Nature
206.4988 (1965): 994-98. Print.
[5] ”Cyber Grand Challenge (CGC).” DARPA RSS. N.p., n.d. Web. 13 Dec.
2013.
[6] ”Defense News & Career Advice ClearanceJobs Defense News & Career
Advice.” Defense News. N.p., n.d. Web. 13 Dec. 2013.
[7] Dshalalow, Jewgeni H. Foundations of Real Analysis. [S.l.]: Chapman &
Hall (Us), n.d. Print.
[8] Floreano, Dario, and Claudio Mattiussi. Bio-inspired Artificial Intelligence:
Theories, Methods, and Technologies. Cambridge, MA: MIT, 2008. Print.
[9] Gusfield, Dan. Algorithms on Strings, Trees, and Sequences: Computer
Science and Computational Biology. Cambridge [England: Cambridge UP,
1997. Print.
16
[10] Hamming, Richard W. ”Error Detecting and Error Correcting Codes.” Bell
System Technical Journal 29.2 (n.d.): 147-60. Print.
[11] Hopcroft, John E., and Jeffrey D. Ullman. Introduction to Automata Theory, Languages, and Computation. Reading, MA: Addison-Wesley, 1979.
Print.
[12] Jaro, Matthew A. ”Probabilistic Linkage of Large Public Health Data
Files.” Statistics in Medicine 14.5-7 (1995): 491-98. Print.
[13] Jonoska, Nataša, and Nadrian C. Seeman. DNA Computing :. Berlin:
Springer, 2002. Print.
[14] Koshland, Daniel E., Jr. ”The Seven Pillars of Life.” Science 295.5563
(2009): 2215-216. Print.
[15] Kovach, Daniel J., Jr. The Computational Theory of Intelligence: Information Entropy. Proc. of Internation Conference of Nonlinear Analysis and
Applied Mathematics 2013, Greece, Rodos. N.p.: n.p., n.d. Print.
[16] Kovach, Daniel J., Jr. The Computational Theory of Intelligence: Data
Aggreagation. Proc. of Internation Conference of Nonlinear Analysis and
Applied Mathematics 2013, Greece, Rodos. N.p.: n.p., n.d. Print.
[17] Neuhauss, Scott. Handbook for the Deep Ecologist. N.p.: IUniverse, 2005.
Print.
[18] ”Payload Definition from PC Magazine Encyclopedia.” Payload Definition
from PC Magazine Encyclopedia. N.p., n.d. Web. 13 Dec. 2013.
[19] Snustad, D. Peter., and Michael J. Simmons. Principles of Genetics. Hoboken, NJ: Wiley, 2012. Print.
[20] Srinivasan, Raghunathan. Protecting Anti-virus Software under Viral Attacks. Thesis. Arizona State University, 2007. N.p.: n.p., n.d. Print.
[21] Strathern, Paul. Turing and the Computer. New York: Anchor, 1999. Print.
17
| 9cs.NE
|
The boundary of the outer space of a free product
Camille Horbez
arXiv:1408.0543v3 [math.GR] 24 Apr 2016
Abstract
Let G be a countable group that splits as a free product of groups of the form
G = G1 ∗ · · · ∗ Gk ∗ FN , where FN is a finitely generated free group. We identify
the closure of the outer space PO(G, {G1 , . . . , Gk }) for the axes topology with the
space of projective minimal, very small (G, {G1 , . . . , Gk })-trees, i.e. trees whose arc
stabilizers are either trivial, or cyclic, closed under taking roots, and not conjugate
into any of the Gi ’s, and whose tripod stabilizers are trivial. Its topological dimension is equal to 3N + 2k − 4, and the boundary has dimension 3N + 2k − 5. We also
prove that any very small (G, {G1 , . . . , Gk })-tree has at most 2N + 2k − 2 orbits of
branch points.
Introduction
Let G be a countable group that splits as a free product
G = G1 ∗ · · · ∗ Gk ∗ FN ,
where FN denotes a finitely generated free group of rank N . We assume that N +
k ≥ 2. A natural group of automorphisms associated to such a splitting is the group
Out(G, {G1 , . . . , Gk }) consisting of those outer automorphisms of G that preserve the
conjugacy classes of each of the groups Gi .
The study of the group Out(FN ) of outer automorphisms of a finitely generated free
group has greatly benefited from the study of its action on Culler and Vogtmann’s outer
space [4], as well as some hyperbolic complexes. The present paper is a starting point
of a work in which we extend results about the geometry of these Out(FN )-spaces to
analogues for free products, with a view to establishing a Tits alternative for the group
of outer automorphisms of a free product [14]. The second main step towards this will
be to define hyperbolic complexes equipped with Out(G, {G1 , . . . , Gk })-actions, and give
a description of the Gromov boundary of the graph of cyclic splittings of G relative to
the Gi ’s [13].
The group Out(G, {G1 , . . . , Gk }) acts on a space PO(G, {G1 , . . . , Gk }) called outer
space. This was introduced by Guirardel and Levitt in [12], who generalized Culler
and Vogtmann’s construction [4] of an outer space CVN associated to a finitely generated free group of rank N , with a view to proving finiteness properties of the group
Out(G, {G1 , . . . , Gk }). The outer space PO(G, {G1 , . . . , Gk }) is defined as the space of all
G-equivariant homothety classes of minimal Grushko (G, {G1 , . . . , Gk })-trees, i.e. metric
1
simplicial G-trees in which nontrivial point stabilizers coincide with the conjugates of
the Gi ’s, and edge stabilizers are trivial.
Outer space can be embedded into the projective space PRG by mapping any tree in
PO(G, {G1 , . . . , Gk }) to the collection of all translation lengths of elements in G. The
goal of the present paper is to describe the closure of the image of this embedding.
The closure of Culler and Vogtmann’s classical outer space has been identified by
Bestvina and Feighn [1] and Cohen and Lustig [2], with the space of projective length
functions of minimal, very small actions of FN on R-trees. An FN -tree is very small
if arc stabilizers are cyclic (possibly trivial) and closed under taking roots, and tripod
stabilizers are trivial.
More precisely, Cohen and Lustig have first proved [2, Theorem I] that CVN is
contained in the space of projective length functions of very small FN -actions on Rtrees. In addition, they have shown that every simplicial, very small FN -tree is a limit
of free and simplicial actions [2, Theorem II]. Bestvina and Feighn have shown that
this remains true of every very small (possibly nonsimplicial) FN -action on an R-tree.
However, their proof does not seem to handle the case of geometric actions that are dual
to foliated band complexes, one of whose minimal components is a measured foliation
on a nonorientable surface, and in which some arc stabilizer is nontrivial. Indeed, in this
case, it is not clear how to approximate the foliation by rational ones without creating
any one-sided compact leaf, and one-sided compact leaves are an obstruction for the
dual action to be very small (arc stabilizers are not closed under taking roots). Building
on Cohen and Lustig’s and Bestvina and Feighn’s arguments, and using approximation
techniques due to Levitt and Paulin [19] and Guirardel [9], we reprove the fact that
CVN identifies with the space of minimal, very small projective FN -trees. Our proof
tackles both cases of simplicial and nonsimplicial trees at the same time (it gives a new
interpretation of Cohen and Lustig’s argument in the simplicial case). We work in the
more general framework of free products of groups. A (G, {G1 , . . . , Gk })-tree is an Rtree, equipped with a G-action, in which all Gi ’s fix a point. A (G, {G1 , . . . , Gk })-tree
will be termed very small if arc stabilizers are either trivial, or cyclic, closed under taking
roots, and not conjugate into any of the Gi ’s, and tripod stabilizers are trivial. We prove
the following theorem.
Theorem 1. The closure PO(G, {G1 , . . . , Gk }) in PRG is compact, and it is the space
of projective length functions of minimal, very small (G, {G1 , . . . , Gk })-trees.
When T is a (G, {G1 , . . . , Gk })-tree with trivial arc stabilizers, we can be a bit more
precise about the approximations we get, and show that T is an unprojectivized limit
of Grushko (G, {G1 , . . . , Gk })-trees Tn , that come with G-equivariant 1-Lipschitz maps
from Tn to T , see Theorem 5.3.
We then compute the topological dimension of the closure and the boundary of the
outer space of a free product of groups. In the case of free groups, Bestvina and Feighn
have shown in [1] that CVN has dimension 3N −4, their result was extended by Gaboriau
and Levitt who proved in addition that ∂CVN has dimension 3N − 5 in [6]. Following
Gaboriau and Levitt’s arguments, we show the following.
Theorem 2. The space PO(G, {G1 , . . . , Gk }) has topological dimension 3N + 2k − 4,
and ∂PO(G, {G1 , . . . , Gk }) has dimension 3N + 2k − 5.
Along the proof, we provide a bound on the number of orbits of branch points and
centers of inversion in a very small (G, {G1 , . . . , Gk })-tree, and on the possible Kurosh
ranks of point stabilizers.
2
We also introduce the slightly larger class of tame (G, {G1 , . . . , Gk })-trees, defined
as those trees whose arc stabilizers are either trivial, or cyclic and non-conjugate into
any Gi , and with a finite number of orbits of directions at branch points and inversion
points. We study some properties of this class, and provide some conditions under which
a limit of tame (G, F)-trees is tame. This class will turn out to be the right class of trees
for carrying out our arguments to describe the Gromov boundary of the graph of cyclic
splittings of G relative to the Gi ’s in [13].
The paper is organized as follows. In Section 1, we review basic facts about free
products of groups, and the associated outer spaces. We prove in Section 2 that the
space V SL(G, {G1 , . . . , Gk }) of projective, minimal, very small (G, {G1 , . . . , Gk })-trees
is compact. In Section 3, we introduce a notion of geometric (G, {G1 , . . . , Gk })-trees, and
explain in particular how to approximate every very small (G, {G1 , . . . , Gk })-tree by a
sequence of geometric (G, {G1 , . . . , Gk })-trees. We then compute the topological dimension of V SL(G, {G1 , . . . , Gk }) in Section 4, and we then identify it with the closure of
outer space in Section 5. We finally introduce the class of tame (G, {G1 , . . . , Gk })-trees,
and discuss some of its properties, in Section 6.
Acknowledgments
I warmly thank my advisor Vincent Guirardel for his help, his rigour and his patience
in reading through first drafts of this work. I am also highly grateful to the referee
for their very careful reading of the paper and their numerous and valuable remarks
and suggestions. These suggestions led in particular to significant clarifications in the
approach to the notion of geometric trees. Finally, I acknowledge support from ANR11-BS01-013 and from the Lebesgue Center of Mathematics.
Contents
1 Background
3
2 Closedness of the space of projective very small (G, F)-trees
7
3 Geometric (G, F)-trees
9
4 Dimension of the space of very small (G, F)-trees
14
5 Very small actions are in the closure of outer space.
25
6 Tame (G, F)-trees
36
1
1.1
Background
Free products of groups and free factors
Let G be a countable group which splits as a free product of groups of the form
G = G1 ∗ · · · ∗ Gk ∗ F,
3
gN
G1
Gk
g1
Figure 1: A standard (G, F)-free splitting.
where F is a finitely generated free group. We let F := {[G1 ], . . . , [Gk ]} be the finite
collection of the G-conjugacy classes of the Gi ’s, which we call a free factor system of
G. The rank of the free group F arising in such a splitting only depends on F. We
call it the free rank of (G, F) and denote it by rkf (G, F). The Kurosh rank of (G, F) is
defined as rkK (G, F) := rkf (G, F) + |F|. Subgroups of G that are conjugate into one of
the subgroups in F will be called peripheral.
A splitting of G is a simplicial tree S, equipped with a minimal and simplicial action
of G (here minimality of the action means that S contains no proper nonempty Ginvariant subtree). We say that a subgroup H ⊆ G is elliptic in S if there exists a point
in S that is fixed by all elements of H. A (G, F)-free splitting is a splitting of G in which
all subgroups in F are elliptic, and all edge stabilizers are trivial. A (G, F)-free factor
is a subgroup of G which is a vertex stabilizer in some (G, F)-free splitting.
Subgroups of free products have been studied by Kurosh in [17]. Let H be a subgroup
of G. Let T be the Bass–Serre tree of the graph of groups decomposition of G represented
in Figure 1, which we call a standard (G, F)-free splitting. By considering the H-minimal
subtree in T , we get the existence of a (possibly infinite) set J, together with an integer
ij ∈ {1, . . . , k}, a nontrivial subgroup Hj ⊆ Gij and an element gj ∈ G for each j ∈ J,
and a (not necessarily finitely generated) free subgroup F ′ ⊆ G, so that
H = ∗j∈J gj Hj gj−1 ∗ F ′ .
This decomposition is called a Kurosh decomposition of H. The Kurosh rank of H (which
can be infinite in general) is defined as rkK (H) := rk(F ′ ) + |J|, it does not depend on
a Kurosh decomposition of H. We let F|H be the set of all H-conjugacy classes of the
subgroups gj Hj gj−1 , for j ∈ J, which does not depend on a Kurosh decomposition of H
either.
When H is a (G, F)-free factor, we have Hj = Gij for all j ∈ J, and all integers ij
are distinct (in particular J is finite). In this case, the free group F ′ is finitely generated.
Hence the Kurosh rank of H is finite.
1.2
Outer space and its closure
An R-tree is a metric space (T, dT ) in which any two points x, y ∈ T are joined by a
unique embedded topological arc, which is isometric to a segment of length dT (x, y).
4
Let G be a countable group, and let F be a free factor system of G. In the present paper, a (G, F)-tree is an R-tree T equipped with an isometric action of G, in which every
peripheral subgroup fixes a unique point. A Grushko (G, F)-tree is a minimal, simplicial metric (G, F)-tree with trivial edge stabilizers, whose collection of point stabilizers
coincides with the conjugates of the subgroups in F. Two (G, F)-trees are equivalent if
there exists a G-equivariant isometry between them.
The unprojectivized outer space O(G, F) is defined to be the space of all equivalence
classes of Grushko (G, F)-trees. Outer space PO(G, F) is defined as the space of homothety classes of trees in O(G, F). We note that in the case where F = {G}, outer space
is reduced to a single point, corresponding to the trivial action of G on a point.
For all (G, F)-trees T and all g ∈ G, the translation length of g in T is defined to be
||g||T := inf dT (x, gx).
x∈T
Theorem 1.1. (Culler–Morgan [3]) The map
i : O(G, F) →
RG
T
7→ (||g||T )g∈G
is injective.
We equip O(G, F) with the topology induced by this embedding, which is called the
axes topology, and we denote by O(G, F) the closure of the image of this embedding.
Culler and Morgan have shown in [3, Theorem 4.5] that if G is finitely generated, then
the subspace of PRG made of projective classes of translation length functions of minimal
G-trees is compact. This can fail to be true in general if G is not finitely generated. For
example, if G = A ∗ B with A not finitely generated, and if (Ai )i∈N is an increasing
sequence of finitely generated subgroups of A whose union equals A, then the projective
translation length functions of the Bass–Serre trees of the splittings G = A ∗Ai (B ∗ Ai )
do not admit any converging subsequence in PRG . However, the key point in Culler–
Morgan’s argument [3, Proposition 4.1] still holds in our context, in the form the following
proposition.
Proposition 1.2. Let R be a standard Grushko (G, F)-tree, in which all edges are
assigned length 1. Then there exists a finite set Y ⊆ G such that for all g ∈ G and all
trees T ∈ O(G, F), we have ||g||T ≤ M ||g||R , where M := maxh∈Y ||h||T .
Proof. The existence of a finite set Y ⊆ G that works for all trees T ∈ O(G, F) is
established in [13, Theorem 4.7]. Proposition 1.2 then follows by taking limits.
Arguing as in Culler–Morgan’s proof of [3, Theorem 4.2], we then obtain the following
result.
Proposition 1.3. The closure O(G, F) is projectively compact.
The goal of the present paper is to identify the closure O(G, F) with the space of
very small (G, F)-trees, which are defined in the following way.
Definition 1.4. A (G, F)-tree T is small if arc stabilizers in T are either trivial, or
cyclic (and non-peripheral). A (G, F)-tree T is very small if it is small, and in addition
nontrivial arc stabilizers in T are closed under taking roots, and tripod stabilizers in T
are trivial.
We note that the trivial action of G on a point is very small in the above sense.
We denote by V SL(G, F) the subspace of PRG made of projective classes of minimal,
nontrivial, very small (G, F)-trees, which we equip with the axes topology.
5
1.3
The equivariant Gromov–Hausdorff topology
The equivariant Gromov–Hausdorff topology on the space of (G, F)-trees.
The space O(G, F) can also be equipped with the equivariant Gromov–Hausdorff topology
[21], which is equivalent to the axes topology [22]. We now recall the definition of the
equivariant Gromov–Hausdorff topology on the space of (G, F)-trees. Let T and T ′ be
two (G, F)-trees, let K ⊂ T and K ′ ⊂ T ′ be finite subsets, let P ⊂ G be a finite subset of
G, and let ǫ > 0. A P -equivariant ǫ-relation between K and K ′ is a subset R ⊆ K × K ′
whose projection to each factor is surjective, such that for all (x, x′ ), (y, y ′ ) ∈ R and all
g, h ∈ P , we have |dT (gx, hy)−dT ′ (gx′ , hy ′ )| < ǫ. A basis of open sets for the equivariant
Gromov–Hausdorff topology is given by the sets O(T, K, P, ǫ) of all (G, F)-trees T ′ for
which there exist a finite subset K ′ ⊂ T ′ and a P -equivariant ǫ-relation R ⊆ K × K ′
[21].
The equivariant Gromov–Hausdorff topology on the space of pointed (G, F)trees. The equivariant Gromov–Hausdorff topology can also be defined on the space of
pointed (G, F)-trees. Let T be a (G, F)-tree, and let (x1 , . . . , xl ) ∈ T l . Let K ⊂ T and
P ⊂ G be finite subsets, and let ǫ > 0. A basis of open sets for the equivariant Gromov–
Hausdorff topology is given by the sets O ′ ((T, (x1 , . . . , xl )), K, P, ǫ) of all pointed (G, F)trees (T ′ , (x′1 , . . . , x′l )) for which there exist a finite subset K ′ ⊂ T ′ and a P -equivariant ǫrelation R ⊆ (K ∪{x1 , . . . , xl })×(K ′ ∪{x′1 , . . . , x′l }) with (xi , x′i ) ∈ R for all i ∈ {1, . . . , l}.
Let T be a (G, F)-tree, let x ∈ T , and let (Tn )n∈N be a sequence of (G, F)-trees that
converges to T in the equivariant Gromov–Hausdorff topology. A sequence (xn )n∈N ∈
Q
n∈N Tn is an approximation of x if the sequence ((Tn , xn ))n∈N of pointed (G, F)-trees
converges to (T, x).
Proposition 1.5. (Horbez [15, Theorem 4.3]) Let (T, u) (resp. (T ′ , u′ )) be a pointed
very small (G, F)-tree, and let ((Tn , un ))n∈N (resp. ((Tn′ , u′n ))n∈N ) be a sequence of
pointed very small (G, F)-trees that converges to (T, u) (resp. (T ′ , u′ )) in the equivariant Gromov–Hausdorff topology. Assume that for all n ∈ N, there exists a 1-Lipschitz
G-equivariant map fn : Tn → Tn′ , such that fn (un ) = u′n . Then there exists a 1-Lipschitz
G-equivariant map f : T → T ′ , such that f (u) = u′ , where T ′ denotes the metric completion of T ′ .
1.4
Graphs of actions and transverse coverings
Let G be a countable group, and F be a free factor system of G. A (G, F)-graph of
actions consists of
• a marked graph of groups G, whose fundamental group is isomorphic to G, such
that all subgroups in F are conjugate into vertex groups of G, together with an
assignment of length (possibly equal to 0) to each edge of G, in such a way that
every edge with peripheral stabilizer has length 0, and
• an isometric action of every vertex group Gv on a Gv -tree Tv (possibly reduced
to a point), in which all intersections of Gv with peripheral subgroups of G are
elliptic and fix a unique point, and
• a point pe ∈ Tt(e) fixed by ie (Ge ) ⊆ Gt(e) for every oriented edge e.
6
A (G, F)-graph of actions is nontrivial if the associated graph of groups is not reduced
to a point. Associated to any (G, F)-graph of actions G is a G-tree T (G). Informally, the
tree T (G) is obtained from the Bass–Serre tree S of the underlying graph of groups by
equivariantly attaching each vertex tree Tv at the corresponding vertex v, an incoming
edge being attached to Tv at the prescribed attaching point, and making each edge from
S whose assigned length is equal to d isometric to the segment [0, d] (edges of length 0
are collapsed). The reader is referred to [9, Proposition 3.1] for a precise description of
the tree T (G). We say that a (G, F)-tree T splits as a (G, F)-graph of actions if there
exists a (G, F)-graph of actions G such that T = T (G).
A transverse covering of an R-tree T is a family Y of nondegenerate closed subtrees
of T such that every arc in T is covered by finitely many subtrees in Y, and for all
Y 6= Y ′ ∈ Y, the intersection Y ∩ Y ′ contains at most one point. It is trivial if Y = {T },
and nontrivial otherwise. The skeleton of Y is the simplicial tree S defined as follows.
The vertex set of S is the set Y ∪ V0 (S), where V0 (S) is the set of all intersection points
between distinct subtrees in Y. There is an edge between Y ∈ Y and y ∈ V0 (S) whenever
y ∈Y.
Proposition 1.6. (Guirardel [11, Lemma 1.5]) A (G, F)-tree splits as a nontrivial
(G, F)-graph of actions if and only if it admits a nontrivial G-invariant transverse covering.
2
Closedness of the space of projective very small (G, F )trees
We denote by V SL(G, F) the space of projective very small, minimal, nontrivial (G, F)trees, as defined in Section 1.2. The goal of the present section is to establish the
following fact.
Proposition 2.1. The space V SL(G, F) is closed in PRG .
In other words, every limit point of a sequence of very small (G, F)-trees is very small.
This was proved by Cohen–Lustig for actions of finitely generated groups on R-trees [2,
Theorem I]: by working in the axes topology, they proved closedness of the conditions
that nontrivial arc stabilizers are cyclic and root-closed, and that tripod stabilizers are
trivial. We will provide a shorter proof of these facts by working in the equivariant
Gromov–Hausdorff topology. A (G, F)-tree T is irreducible if no end of T is fixed by all
elements of G. The Gromov–Hausdorff topology is equivalent to the axes topology on
the space of minimal, irreducible (G, F)-trees [22]. Since limits of trees in O(G, F) are
irreducible, we can carry our arguments in the equivariant Gromov–Hausdorff topology.
A proof of the fact that being small is a closed condition (in the equivariant Gromov–
Hausdorff topology) also appears in [21, Lemme 5.7]. In our setting, we also need to
check closedness of the condition that arc stabilizers are not peripheral. We will make
use of classical theory of group actions on R-trees, and refer the reader to [3] for an
introduction to this theory.
Lemma 2.2. Let T be a minimal (G, F)-tree, and let (Tn )n∈N be a sequence of minimal (G, F)-trees that converges (non-projectively) to T . Let g ∈ G be an element that
fixes a nondegenerate arc in T . Then for all n ∈ N sufficiently large, either g fixes a
nondegenerate arc in Tn , or g is hyperbolic in Tn .
7
Proof. Otherwise, up to passing to a subsequence, we can assume that for all n ∈ N, the
element g fixes a single point xn in Tn . Let [a, b] be a nondegenerate arc fixed by g in
T . Let an (resp. bn ) be an approximation of a (resp. b) in the tree Tn . As dTn (an , gan )
and dTn (bn , gbn ) both tend to 0, the points an and bn are both arbitrarily close to xn .
Therefore, the distance dTn (an , bn ) converges to 0, and a = b, a contradiction.
Lemma 2.3. Let T be a minimal (G, F)-tree, and let (Tn )n∈N be a sequence of minimal
(G, F)-trees that converges to T . Let g ∈ G. Assume that some power gk of g fixes a
nondegenerate arc I in T . If for all sufficiently large n ∈ N, the element g is hyperbolic
in Tn , then g fixes I.
Proof. Let I := [a, b]. Let an (resp. bn ) be an approximation of a (resp. b) in Tn . Since
dTn (an , gk an ) and dTn (bn , gk bn ) both converge to 0, the points an and bn are arbitrarily
close to the axis of g in Tn , and ||g||Tn converges to 0. Hence both dTn (an , gan ) and
dTn (bn , gbn ) converge to 0, so g fixes [a, b].
Proof of Proposition 2.1. Let (Tn )n∈N be a sequence of very small (G, F)-trees that converges to a (G, F)-tree T . Let g ∈ G be a peripheral element. Then for all n ∈ N, the
element g fixes a single point in Tn . Lemma 2.2 implies that g fixes a single point in T .
Let now g, h ∈ G be two elements that fix a common nondegenerate arc [a, b] ⊆ T .
We will show that the group hg, hi is abelian, and hence cyclic because g and h are
nonperipheral. Let an (resp. bn ) be an approximation of a (resp. b) in Tn . Let ǫ > 0,
chosen to be small compared to dT (a, b). Since dTn (an , gan ), dTn (bn , gbn ) ≤ ǫ for n large
enough, while dTn (an , bn ) ≥ dT (a, b) − ǫ, the elements g and h are hyperbolic in Tn , and
their translation axes have an overlap of length greater than 3ǫ. On the other hand, we
have ||g||Tn , ||h||Tn ≤ ǫ. This implies that the elements [g, h], g[g, h]g −1 and h[g, h]h−1
all fix a common nondegenerate arc in Tn . As Tn is very small, the group generated by
these elements is (at most) cyclic, and in addition [g, h] is nonperipheral. This implies
that [g, h] is hyperbolic in any Grushko (G, F)-tree. Both g and h preserve the axis of
[g, h] in a Grushko (G, F)-tree, and hence g and h commute.
Let now g ∈ G be an element, one of whose proper powers gk fixes a nondegenerate arc [a, b] ⊆ T .
We first assume that g fixes a nondegenerate arc in Tn for all n ∈ N, and let In denote
the fixed point set of g in Tn . Since Tn is very small, the element g also fixes In for all
n ∈ N. Let an (resp. bn ) be an approximation of a (resp. b) in Tn . Since dTn (gk an , an )
and dTn (gk bn , bn ) both converge to 0, the arc In comes arbitrarily close to both an and
bn . This implies that both dTn (gan , an ) and dTn (gbn , bn ) converge to 0, and therefore g
fixes [a, b] in T .
Otherwise, up to passing to a subsequence, we can assume that gk , and hence g, is
hyperbolic in Tn for all n ∈ N. It then follows from Lemma 2.3 that g fixes [a, b].
We finally assume that g fixes a nondegenerate tripod in T , whose extremities we
denote by a, b and c. Let m be the center of this tripod, and L > 0 be the shortest
distance in T between m and one of the points a, b or c. Let an (resp. bn , cn , mn ) be
an approximation of a (resp. b, c, m) in Tn , and let ǫ > 0 be such that ǫ < L2 . For
n sufficiently large, the point mn lies at distance at most ǫ from the center m′n of the
tripod formed by an , bn and cn in Tn . In addition, as an , bn and cn all lie at distance at
8
most ǫ from CTn (g), the distance from m′n to one of the points an , bn or cn is at most ǫ.
This leads to a contradiction.
3
Geometric (G, F )-trees
In the present section, we introduce the class of geometric (G, F)-trees, that will be an
important tool in our analysis of the closure of O(G, F). In particular, we will explain
how to approximate any very small (G, F)-tree by geometric ones. Our presentation was
largely inspired by Gaboriau–Levitt’s work in the context of free groups [6]. The idea
of using band complexes and systems of isometries for studying trees dates back to the
work of Rips.
3.1
Approximations by geometric (G, F )-trees
Given a tree T , we define a finite tree in T as the convex hull of a finite collection of points
in T . A finite forest of T is a subset of T having finitely many connected components,
each of which is a finite tree.
Let B := {g1 , . . . , gN } be a fixed free basis of FN , and let S be the standard (G, F)free splitting represented on Figure 1. Let T be a (G, F)-tree, and let K ⊆ T be a
finite forest. Let v0 ∈ S be a vertex with trivial stabilizer, and let v1 , . . . , vk be the
adjacent vertices with stabilizers G1 , . . . , Gk . Notice that the set {v0 , . . . , vk } is a set of
representatives of all G-orbits of vertices in S.
Given a vertex v ∈ V (S) in the orbit of vi (for some i ∈ {0, . . . , k}), we let Kv be the
union, taken over all elements g ∈ G such that gvi = v, of the translates gK. Notice in
particular that for all i ∈ {1, . . . , k}, we have Kvi = Gi .K, and in general the forest Kv
is Stab(v)-invariant. Let
a
Kv .
X :=
v∈V (S)
Let Σ be the foliated complex obtained from X by adding, for each pair {v, v ′ } of adjacent
vertices in S, a band Kv ∩ Kv′ × [0, 1], joining the two copies of Kv ∩ Kv′ sitting in Kv
and in Kv′ . These bands are foliated by the vertical sets of the form {x} × [0, 1].
Notice that Σ can also be viewed as the subset of T × S made of all couples of the
form (x, y), where either y is a vertex of S and x ∈ Ky , or else y belong to the interior of
an edge joining two vertices v and v ′ in S, and x ∈ Kv ∩ Kv′ . The diagonal action of G
on T × S then restricts to a G-action on Σ. The projections to T and S define natural
G-equivariant maps πT : Σ → T and πS : Σ → S. We will let Σ0 := πS−1 (G.v0 )
We now assume that the set K ⊆ T is a finite tree, i.e. K is connected, that
K ∩ gK 6= ∅ for all g ∈ B, and that for all i ∈ {1, . . . , k}, the tree K contains the unique
point xi which is fixed by Gi . In this way, all subsets Kv with v ∈ V (S) are connected.
Then Σ projects onto the tree S, and all fibers of the projection map πS are trees, which
implies that Σ is contractible. It follows from [19, Proposition 1.7] that the leaf space
of Σ made Hausdorff is an R-tree TK . Notice by construction that for all vertices vi
with nontrivial stabilizer in S, the tree Kvi is equal to the union of all trees Kv with v
adjacent to vi . Therefore, every point in TK has a representative in Σ0 .
We will now establish a few properties of TK , thus extending a theorem of Gaboriau
and Levitt [6, Theorem I.1] to the context of (G, F)-trees. Given two R-trees T and T ′ ,
a morphism f : T → T ′ is a map such that every segment in T can be subdivided into
finitely many subsegments, in restriction to which f is an isometry.
9
Theorem 3.1. Let T be a (G, F)-tree, and let K ⊆ T be a finite tree that contains the
fixed point of Gi for all i ∈ {1, . . . , k}, and such that K ∩ gK 6= ∅ for all g ∈ B. Then
TK is the unique (G, F)-tree such that
1. the finite tree K embeds isometrically into TK , and
2. for all g ∈ G and all x, y ∈ K, if g.x = y ∈ T and the points (y, v0 ) and (y, gv0 )
belong to the same leaf of Σ, then g.x = y in TK , and
3. every segment of TK is contained in a finite union of translates g.K with g ∈ G,
and
4. if T ′ is any (G, F)-tree satisfying the first two above properties, then there exists a
unique G-equivariant morphism j : TK → T ′ such that j(x) = x for all x ∈ K.
Proof. The proof of Theorem 3.1 is the same as the proof of [6, Theorem I.1], we will
sketch a proof for completeness. Uniqueness follows from the fourth property: if T ′ is
another (G, F)-tree with the same properties, then there are G-equivariant morphisms
from TK to T ′ and from T ′ to TK . Since morphisms can only decrease translation length
functions, this implies that TK and T ′ are G-equivariantly isometric.
The tree TK is obtained in the following way. Given any two points x, y ∈ Σ0 , we let
δ(x, y) := inf dT (x, u1 ) + dT (u′1 , u2 ) · · · + dT (u′k−1 , y),
σ
where the infimum is taken over the set of all sequences
σ = (x = u′0 , u1 , u′1 , u2 , u′2 , . . . , uk−1 , u′k−1 , uk = y)
of points in Σ, such that for all i ∈ {0, . . . , k − 1}, the segment [u′i , ui+1 ] is horizontal
(i.e. its projection to S is constant), and for all i ∈ {1, . . . , k}, the segment [ui , u′i ] is
vertical (i.e. its projection to T is constant), and hence [u′i , ui+1 ] is contained in a leaf
of Σ. Then δ defines a pseudo-metric on Σ0 , and TK is the metric space (Σ0 , δ) made
Hausdorff.
Property 2 in Theorem 3.1 easily follows from the construction, and Property 3
follows from Properties 1 and 2. Property 4 is obtained by noticing that if T ′ is a tree
satisfying the first two conclusions of the theorem, then if we let δ′ ((x, gv0 ), (x′ , g′ v0 )) :=
dT ′ (gx, g′ x′ ) for all x, x′ ∈ K and all g, g ′ ∈ G, then we have δ′ ≤ δ. Then using the fact
that every segment in T ′ is covered by finitely many G-translates of K, we get that the
identity map on K × Σ0 induces the required morphism from T ′ to T . We refer to the
proof of [6, Theorem I.1] for details of the arguments.
The key point for establishing Property 1 will be to prove that if δ(x, y) = 0, then
x and y belong to the same leaf of the foliation. We will prove more generally that the
infimum in the definition of δ is achieved, and that the πS -images of the points in the
minimizing sequence are aligned in the right order in S. Property 1 in Theorem 3.1 will
follow from this fact: if x, y ∈ K, then any minimizing sequence has to be contained in
K, and therefore δ(x, y) = dT (x, y).
We will now prove that the infimum in the definition of δ is achieved. Given a
sequence σ as above, we let
δ(σ) := dT (x, u1 ) + dT (u′1 , u2 ) + · · · + dT (u′k−1 , y).
In order to prove the above fact, we first notice that if the πS -image of the sequence
σ backtracks, then one can find another sequence σ ′ with δ(σ ′ ) ≤ δ(σ): indeed, if σ
10
contains a subsequence of the form ui , u′i , ui+1 , u′i+1 where πS (ui ) = πS (u′i+1 ) := v,
then this subsequence can be replaced by ui , u′i+1 (indeed, the segment [ui , u′i+1 ] is then
contained in Kv ), and this yields a new sequence σ ′ satisfying δ(σ ′ ) ≤ δ(σ).
In addition, let u′i , ui+1 , u′i+1 be a finite subsequence arising in a sequence σ as above.
Let γ := πS ([ui+1 , u′i+1 ]). Let γK ⊆ K be the subset of K made of those points x ∈ K
such that there exists a leaf segment in Σ containing x and projecting to γ. Then one
does not increase the value of δ(σ) if one replaces ui+1 by the projection y of ui to γK ,
and u′i+1 by the unique point z in Σ in the same leaf as y, such that πS ([y, z]) = γ. This
implies that the supremum can be taken over the finite set made of those points in K
that are extremal in some band of Σ. In particular, this supremum is achieved.
If T ′ is a tree satisfying the first two conclusions of Theorem 3.1, the morphism j
provided by Theorem 3.1 is called a resolution of T ′ .
Definition 3.2. A (G, F)-tree T is geometric if there exists a finite subtree K ⊆ T such
that T = TK .
Let T be a (G, F)-tree, and let (Tn )n∈N be a sequence of (G, F)-trees. The sequence
(Tn )n∈N strongly converges towards T (in the sense of Gillet and Shalen [8]) if for all
integers n ≤ n′ , there exist morphisms jn,n′ : Tn → Tn′ and jn : Tn → T such that for
all n ≤ n′ and all segments I ⊆ Tn , the morphism jn′ is an isometry in restriction to
jn,n′ (I). Strong convergence implies in particular that for all g ∈ G, there exists n0 ∈ N
such that for all n ≥ n0 , we have ||g||Tn = ||g||T . The following theorem essentially
follows from work by Levitt and Paulin [19, Theorem 2.2] and Gaboriau and Levitt [6,
Proposition II.1].
Theorem 3.3. Let G be a countable group, and let F be a free factor system of G. Let
T be a minimal (G, F)-tree. Then there exists a sequence (Tn )n∈N of minimal geometric
(G, F)-trees that strongly converges towards T . If in addition H is a subgroup of G with
finite Kurosh rank that is elliptic in T , then the approximation can be chosen so that H
is elliptic in Tn .
Proof. By carefully choosing an increasing sequence of finite subtrees Kn ⊆ T whose
union is the whole tree T , we will establish that the trees TKn strongly converge towards
T . One point in the proof is to carefully choose Kn so that all trees TKn are minimal:
the idea for this is to choose a subtree Kn all of whose endpoints belong to the G-orbit
of some fixed basepoint x0 ∈ T . Here are the details of the proof.
Let T be a minimal (G, F)-tree. Let (gn )n∈N be an enumeration of G. Assume that
this enumeration is chosen so that if v0 , gn v0 and gm v0 are aligned in this order in S,
then m ≥ n. Let x0 ∈ T , and for all n ≥ 1, let Kn be the convex hull of {gk x0 |k ≤ n}
in T , so that by construction, all extreme points of Kn belong to the G-orbit of x0 . By
minimality, the tree T is the increasing union of the trees Kn , and there exists g ∈ G
such that x0 belongs to the translation axis of g in T . Let n0 ∈ N be large enough so
that Kn0 contains gx0 and satisfies the hypotheses from Theorem 3.1.
For all n ≥ n0 , we let Tn := TKn . The first property in Theorem 3.1 implies that
the distance between x0 and gx0 is the same in Tn and in T , so x0 belongs to the axis
of g in Tn . In addition, the second property in Theorem 3.1, together with our choice
of enumeration of G, implies that Tn is the convex hull of the orbit of x0 . Hence Tn is
minimal.
We claim that the sequence (Tn )n∈N strongly converges towards T . Given any two
integers n ≤ n′ , we have Kn ⊆ Kn′ . Theorem 3.1 applied to Kn′ implies that Tn′
11
contains an isometrically embedded copy of Kn′ , and hence of Kn . In addition, if two
points belong to the same leaf of Σn , then they also belong to the same leaf of Σn′
(where Σn and Σn′ denote the band complexes associated to Kn and Kn′ , respectively).
Therefore Tn′ satisfies the first two properties of Theorem 3.1 with respect to the finite
tree Kn . The last property in Theorem 3.1 (applied to Kn ) then provides morphisms
jn,n′ : Tn → Tn′ for all n ≤ n′ , as well as morphisms jn : Tn → T for all n ∈ N. Let
n ∈ N, and I ⊆ Tn . The third property in Theorem 3.1 then enables us to choose a
finite set Y of elements of G so that I is covered by the translates of Kn in Tn under
elements in Y . We then let n′ be large enough so that Kn′ contains the translates of
Kn in T under the elements in Y . Then jn′ is an isometry from jn,n′ (I) to jn (I). This
shows strong convergence of the trees Tn towards T .
By choosing n large enough so that all elements in a basis of the free part of the
Kurosh decomposition of H, as well as all conjugators arising in this decomposition, are
of the form gk for some k ≤ n, and Kn contains a fixed point of H, we can ensure that
the last property of Theorem 3.3 is satisfied.
Remark 3.4. Notice that if branch points are dense in T , then the tree Kn has edges of
arbitrarily small length as n tends to +∞, from which it follows that simplicial edges in
the geometric approximation of T have lengths going to 0.
3.2
Properties of geometric (G, F )-trees
We now list a few other useful properties of the tree TK , which were proved by Gaboriau
and Levitt in [6] in the case of FN -trees. Any element of G can be written in a unique
way as a reduced word of the form s1 . . . sk , where
• for all i ∈ {1, . . . , k}, either si ∈ B ∪ B −1 , or else si belongs to one of the peripheral
groups Gj , and
• no two consecutive letters are of the form ss−1 or s−1 s with s ∈ B, and
• no two consecutive letters belong to the same peripheral group Gj .
It is cyclically reduced if in addition, it does not start and end with letters that are
inverses to each other, and does not either start and end with letters belonging to a
common peripheral group.
Proposition 3.5. Let T be a (G, F)-tree, let K ⊆ T be a finite subtree that contains the
fixed points of all groups Gj , and such that K ∩ sK 6= ∅ for all s ∈ B, and let g ∈ G be
nonperipheral and cyclically reduced. Then the fixed point set of g in TK is contained in
K.
Proof. The proof goes as in [6, Proposition I.5]. Let a ∈ TK be a fixed point of g.
Choose a representative (x, v) ∈ Σ0 of a, with v = hv0 for some h ∈ G, such that the
distance from v0 to v is minimal. We let x0 := h−1 x. Assume towards a contradiction
that v 6= v0 , i.e. h 6= e. We have (x, v) = (gx, gv) = (x, gv) in TK : the first equality
comes from the fact that g(x, v) = (x, v) in TK , and the second from the fact that
gx = x in T . By applying h−1 , we get that (x0 , v0 ) = (x0 , h−1 ghv0 ). It follows from
Theorem 3.1 that (x0 , v0 ) and (x0 , h−1 ghv0 ) lie in a common leaf of Σ. Let g1 (resp. gk )
be the first (resp. last) letter in the reduced word representing h−1 gh. Then there is a
leaf in Σ joining (x0 , v0 ) to (x0 , g1 v0 ), and there is a leaf in Σ joining (x0 , h−1 ghgk−1 v0 )
to (x0 , h−1 ghv0 ). By translating those by h and g−1 h, respectively, we find leaves in
12
Σ joining (x, v) to (x, hg1 v0 ) and to (x, hgk−1 v0 ). Our choice of v then implies that the
reduced word representing h cannot end with gk , and the reduced word representing h−1
cannot begin with g1 . But since h−1 gh begins with g1 and ends with gk , this forces g
not to be cyclically reduced, a contradiction.
Corollary 3.6. Let T be a small (G, F)-tree, and let K be a finite subtree of T that
contains all fixed points of the Gi ’s, and such that K ∩ sK 6= ∅ for all s ∈ B. For all
g 6= 1 ∈ G, the restriction of the resolution map j : TK → T to the fixed point set of
g is an isometry. In particular, arc stabilizers in TK are either trivial, or cyclic and
non-peripheral. If T is very small, then tripod stabilizers in TK are trivial.
Let k ∈ N. A (G, F)-tree T is k-tame if T is small, and in addition, we have
Fix(gkl ) = Fix(g k ) for all l ≥ 1. We refer to Section 6 for details and equivalent
definitions.
Corollary 3.7. Let T be a small (G, F)-tree, and let K be a finite subtree of T that
contains all fixed points of the Gi ’s, and such that K ∩ sK 6= ∅ for all s ∈ B. For all
k ∈ N, if T is k-tame, then TK is k-tame.
Proof. Let g ∈ G be nonperipheral, and let k ∈ N. Up to passing to a conjugate, we
can assume that g is cyclically reduced, and in this case the fixed point set of g in TK is
contained in K by Proposition 3.5. Let x ∈ K be such that there exists l ∈ N such that
gkl x = x in TK . This implies in particular that gkl x = x in T . We will show that gk x = x
in TK . By Theorem 3.1, there is a leaf segment in Σ joining (x, v0 ) to (x, g kl v0 ). But
since g is cyclically reduced, this implies that there is also a leaf segment in Σ (which
is a subsegment of the previous one) joining (x, v0 ) to (x, gk v0 ). Since gk x = x in T
(because T is k-tame), we have found a leaf segment in Σ joining (x, v0 ) to (g k x, g k v0 ).
This implies that gk x = x in TK , as required.
As a consequence of Theorem 3.3 and the above two corollaries, we get the following
approximation result.
Theorem 3.8. Let G be a countable group, and let F be a free factor system of G. Let
T be a minimal, small (G, F)-tree. Then there exists a sequence (Tn )n∈N of minimal,
small, geometric (G, F)-trees, that strongly converges towards T . If T is very small (resp.
k-tame for some k ∈ N), the approximation can be chosen very small (resp. k-tame).
3.3
Interpretation in terms of systems of isometries
Let K be a finite forest. Given two nonempty closed subtrees A and B of K, a morphism
from A to B is a surjective map φ : A → B, such that A can be subdivided into finitely
many subtrees in restriction to which φ is an isometry. The subtrees A and B are called
the bases of φ. A finite system of morphisms is a pair K = (K, Φ), where K is a finite
forest, and Φ is a finite collection of morphisms between nonempty closed subtrees of K.
In the case where all morphisms φ ∈ Φ are isometries between their bases, we call
K a finite system of isometries. Notice that any finite system of morphisms can be
turned into a finite system of isometries by appropriately subdividing the bases of the
morphisms in Φ.
We now assume that Φ is a finite system of isometries. Given an isometry φ from
Aφ to Bφ , we denote by φ−1 its inverse, which is a partial isometry from Bφ to Aφ .
Given partial isometries φ1 , . . . , φn , we denote by φ1 ◦ · · · ◦ φn the composition of the
13
φi ’s, which is a partial isometry whose domain is the set of all x ∈ Aφn such that for
all i ∈ {2, . . . , n}, we have φi ◦ · · · ◦ φn (x) ∈ Aφi−1 . A word in the partial isometries
in the family Φ and their inverses is reduced if it does not contain any subword of the
form φ ◦ φ−1 or φ−1 ◦ φ. A finite system of isometries K has independent generators if
no reduced word in the isometries in Φ and their inverses represents a partial isometry
of K that fixes some nondegenerate arc.
Let T be a (G, F)-tree, let K ⊆ T be a finite subtree, and let Σ be the corresponding
foliated complex. We recall from Section 3.1 that X is the disjoint union of all trees
Kv . The quotient space F := X/G is a finite forest: one connected component of
F is isometric to Kv0 , and the other components are isometric to the quotient spaces
Kvi /Gi . There is a naturally defined finite system of morphisms on F : every map has
one base contained in Kv0 , and there are N morphisms with image in Kv0 , and for each
i ∈ {1, . . . , k}, there is a morphism from Kv0 to Kvi /G. The morphisms of the second
kind may however fail to be isometries because two segments in a common base of Kv0
may have images in Kvi that belong to the same Gvi -orbit, in which case they have the
same image in the quotient. However, as mentioned above, by subdividing the bases of
these morphisms at the points xi ∈ Kv0 corresponding to the fixed points of the groups
Gi , we get a finite system of isometries associated to K.
Lemma 3.9. Let T be a (G, F)-tree in which no element of G fixes a nondegenerate
arc, and let K ⊆ T be a finite subtree. Then the finite system of isometries associated
to K has independent generators.
Proof. We denote by K this finite system of isometries. Assume towards a contradiction
that a reduced word in the partial isometries in K represents a partial isometry fixing
a nondegenerate arc of X/G. Then one could find a nondegenerate arc I ⊆ K, and an
element g ∈ G, such that for all x ∈ I, there is a leaf joining (x, v0 ) to (gx, gv0 ) in Σ.
But this implies that I = gI in TK . By the fourth property in Theorem 3.1, there is a
morphism from TK to T , so g also fixes a nondegenerate arc in T , a contradiction.
4
Dimension of the space of very small (G, F )-trees
Bestvina and Feighn have shown in [1, Corollary 7.12] that the space of very small FN trees has dimension 3N − 4. Their result was improved by Gaboriau and Levitt in [6,
Theorem V.2], who showed in addition that V SL(FN ) r CVN has dimension 3N − 5.
Following Gaboriau and Levitt’s proof, we extend their computation to the general case
of (G, F)-trees. We recall the notion of the free rank of (G, F) from Section 1.1.
Theorem 4.1. Let G be a countable group, and let F be a free factor system of G, such
that rkK (G, F) ≥ 2. Then V SL(G, F) has topological dimension 3rkf (G, F) + 2|F| − 4.
Theorem 4.2. Let G be a countable group, and let F be a free factor system of G,
such that rkK (G, F) ≥ 2. Then V SL(G, F) r PO(G, F) has topological dimension
3rkf (G, F) + 2|F| − 5.
The proof of Theorem 4.1 will be carried in the next two sections. Additional arguments for getting the dimension of the boundary (Theorem 4.2) will be given in Section 4.4.
14
4.1
The index of a small (G, F )-tree
Let T be a small (G, F)-tree, and let x ∈ T . The Kurosh decomposition of the stabilizer
Stab(x) reads as
Stab(x) = g1 Gi1 g1−1 ∗ · · · ∗ gr Gir gr−1 ∗ F.
We claim that the groups Gi1 , . . . , Gir are pairwise non conjugate in G, which implies in
particular that there are only finitely many free factors arising in the Kurosh decomposition of Stab(x). Indeed, otherwise, we could find i ∈ {1, . . . , k}, and g ∈ G r Stab(x),
such that both Gi and gGi g−1 fix x. This would imply that Gi fixes both x and g−1 x,
and therefore g−1 x = x because no arc of T is fixed by a peripheral element. Hence Gi
and gGi g−1 are conjugate in Stab(x), a contradiction.
Notice that the free group F might a priori not be finitely generated (though it will
actually follow from Corollary 4.5 that it is). We define the index of x as
i(x) = 2 rkK (Stab(x)) + v1 (x) − 2,
where v1 (x) denotes the number of Stab(x)-orbits of directions from x in T with trivial
stabilizer. A point x ∈ T is a branch point if T r{x} has at least 3 connected components.
It is an inversion point if T r {x} has 2 connected components, and some element g ∈ G
fixes x and permutes the two directions at x. The following proposition is a generalization
of [6, Proposition III.1].
Proposition 4.3. For all small minimal (G, F)-trees T and all x ∈ T , we have i(x) ≥ 0.
If T is very small, then i(x) > 0 if and only if x is a branch point or an inversion point.
Proof. If rkK (Stab(x)) ≥ 2, then we have i(x) ≥ 2, and in this case x is a branch point.
If Stab(x) is trivial, then i(x) = v1 (x) − 2, where v1 (x) is the number of connected
components of T r {x}, which is nonnegative because T is minimal, and i(x) > 0 if
and only if x is a branch point. Finally, if rkK (Stab(x)) = 1, then i(x) = v1 (x) ≥ 0.
If i(x) > 0, then either x is a branch point as in the first case, or x has valence 2 and
is therefore an inversion point. If i(x) = 0, and T is very small, the stabilizer of any
direction from x is isomorphic to Stab(x). As tripod stabilizers are trivial in T , this
implies that x is not a branch point.
Let T be a small (G, F)-tree, and let x, x′ ∈ T . If x and x′ belong to the same
G-orbit, then i(x) = i(x′ ). Given a G-orbit O of points in T , we can thus define i(O) to
be equal to i(x) for any x ∈ O. We then let
X
i(T ) :=
i(O).
O∈T /G
We now extend [6, Theorem III.2] and its corollaries [6, Corollaries III.3 and III.4]
to the context of (G, F)-trees. The converse of the second statement of Proposition 4.4
will be proved in Proposition 4.18 below.
Proposition 4.4. For all small (G, F)-trees T , we have i(T ) ≤ 2rkK (G, F) − 2. If T is
geometric, then i(T ) = 2rkK (G, F) − 2.
Using Proposition 4.3, we get the following result as a corollary of Proposition 4.4.
Corollary 4.5. Any very small (G, F)-tree has at most 2 rkK (G, F) − 2 orbits of branch
or inversion points, and the Kurosh rank of the stabilizer of any x ∈ T is at most equal
to rkK (G, F).
15
Proof of Proposition 4.4. First assume that T is geometric. For all i ∈ {1, . . . , k}, we let
xi ∈ T be the fixed point of Gi . Let K ⊆ T be a finite subtree containing all points xi ,
and such that T = TK . Let K = (F, Φ) be the associated system of morphisms on the
compact forest F := X/G (see Section 3.3). We fix a G-orbit of points O ⊂ T . We will
associate to O two graphs S and S ′ in the following way.
Vertices of S are the points in O ∩ F (with a slight abuse of notations, as we are
actually considering the image in F of the set of points in X that correspond to points in
O). There is an edge e from z to z ′ whenever there is a morphism in Φ that sends either
z to z ′ or z ′ to z. We note that Theorem 3.1 implies that S is connected. The graph
S is actually equal to the quotient space under the G-action of the leaf in the foliated
2-complex Σ passing through any point of O ∩ F .
The copies of the points xi in the trees Kvi /Gi ⊆ F are called the special vertices of
F . We let n(S) be the number of special vertices in S. We define the Kurosh rank of S
to be
rkK (S) = 1 + |E(S)| − |V (S)| + n(S).
We similarly define the Kurosh rank of any subgraph G ⊆ S.
We will now assign labels in G to the edges in S. When a morphism in Φ has
both its bases in Kv0 , the corrsponding edge between z and z ′ comes equipped with a
natural (oriented) label, given by the element in the base B of FN corresponding to the
partial isometry. If z ′ is a nonspecial point in Kvi /Gi , its preimage by the corresponding
morphism in Φ consists of finitely many points z1 , . . . , zm which all differ by elements in
Gi . We arbitrarily choose the label of one of the corresponding edges (say the one from
z1 to z ′ ) to be equal to 1, and if gj ∈ Gi satisfies zj = gj z1 , we label the edge from z ′ to
zj by gj . We finally assign label 1 to every edge that contains the special vertex.
Fix x ∈ O ∩ F .
Lemma 4.6. We have rkK (S) = rkK (Stab(x)).
Proof. By adding a vertex with vertex group Gi at each special vertex contained in S,
the graph S has a natural structure of graph of groups, and there is a natural morphism
ρ : π1 (S, x) → G, sending any edge of a path in π1 (S, x) to the product of the labels
it crosses. The morphism ρ is injective and takes its values in Stab(x), and surjectivity
follows from the second assertion of Theorem 3.1. By definition, the Kurosh rank of S
is also equal to the Kurosh rank of its fundamental group (as a graph of groups), so
rkK (S) = rkK (Stab(x)).
We now define a graph S ′ by considering orbits of directions instead of orbits of
points. Vertices of S ′ are the directions in F from points in O ∩ F , and two directions d
and d′ are joined by an edge whenever there exists a morphism in Φ that either sends d
to d′ or d′ to d. In this way, every vertex of S is replaced by vF (x) vertices in S ′ , where
vF (x) denotes the valence of x in F . Every edge e in S joining two points z and z ′ ,
and corresponding to a morphism φ is replaced by vφ (e) edges in S ′ , where vφ (e) is the
valence of z in the domain of φ. There is a natural map π : S ′ → S that sends vertices
to vertices and edges to edges.
Lemma 4.7. The set of components of S ′ is in one-to-one correspondence with the set
of Stab(x)-orbits of directions at x in T . For all directions d ∈ V (S ′ ), the fundamental
group of the component of S ′ that contains d is isomorphic to the stabilizer of d, hence
to {1} or Z.
16
Proof. Let S1′ be a component of S ′ . Then S1′ contains a vertex d0 ∈ V (S1′ ) corresponding
to a direction based at a point y ∈ O ∩ Kv0 . Applying any g ∈ G taking y to x, we get a
direction d := gd0 from x in T . The Stab(x)-orbit of d only depends on the component
S1′ (and not on the choices of d0 and g). This defines a map Ψ from the set of connected
components of S ′ to the set of Stab(x)-orbits of directions at x in T .
We now prove injectivity of the map Ψ. Let d0 and d′0 be two directions in K = Kv0
having the same Ψ-image, then there exists g ∈ G mapping d0 to d′0 . Theorem 3.1,
applied to two nearby points defining the direction d0 , implies that d0 and d′0 belong to
the same component of S ′ , showing injectivity of Ψ.
We now show surjectivity of Ψ. Let d be a Stab(x)-orbit of directions at x in T . There
exists a segment [x, x1 ], such that d is contained in [x, x1 ]. Then [x, x1 ] is contained in
some translate wK with w ∈ G, and w−1 d ⊆ K. This shows that d belongs to the image
of Ψ.
The proof of the second statement of the lemma is similar to the proof of Lemma 4.6.
For all d ∈ V (S1′ ), there is an injective morphism ρ′ : π1 (S1′ ) → Stab(d), where this time
S1′ is an actual graph and not a graph of groups because no element in Gi fixes a
nondegenerate arc in TK . Surjectivity of ρ′ follows from Theorem 3.1, applied to two
nearby points defining d.
We say that a (G, F)-tree T has finitely many orbits of directions if there are finitely
many orbits of directions based at branch or inversion points in T . We define the singular
set Sing ⊆ F as the subset of F made of all branch points in F , all endpoints of the
bases of the morphisms in Φ, and all points xi .
Corollary 4.8. For all x ∈ K, there are only finitely many Stab(x)-orbits of directions
at x in T . In addition, there are finitely many orbits of branch or inversion points in T
(and hence finitely many orbits of directions in T ).
Proof. Let d be a direction at x in T . Then the Stab(x)-orbit of d contains a direction
in K based at a point y ∈ O ∩ K. Since S is connected, there exists g ∈ G such that
(x, v0 ) and (gy, gv0 ) are joined by a leaf of Σ. By dragging d along this leaf, we get that
either d is a direction at x that is contained in K, or d is in the G-orbit of a direction
in K at a point in the singular set Sing. The claim follows because Sing is a finite set.
Using the fact that the orbit of any point of T meets K, the above argument also shows
that the orbit of any branch or inversion point in T meets Sing.
Let G be a finite connected subgraph of S containing all vertices in O ∩ Sing (where
Sing denotes the singular set) and all edges e ∈ E(S) with vφ (e) 6= 2 (where φ is the
morphism corresponding to e). Let G ′ ⊆ S ′ be the π-preimage of G in S ′ . Lemma 4.7,
together with the fact that there are only finitely many Stab(x)-orbits of directions at
any point x ∈ T (Corollary 4.8), shows that up to enlarging G if necessary, we may
assume that π1 (G ′ ) generates the fundamental group of every component of S ′ . Denote
by Gj′ the components of G ′ . As the fundamental group of any finite connected graph X
satisfies
1 − rk(π1 (X)) = |V (X)| − |E(X)|,
we have
X
X
X
(1 − rk(π1 (Gj′ ))) =
vF (x) −
vφ (e).
j
x∈V (G)
17
e∈E(G)
Moreover, we have
2rkK (G) − 2 = −2|V (G)| + 2|E(G)| + 2n(G).
Summing the above two equalities, we get
X
X
X
2rkK (G) − 2 +
(1 − rk(π1 (Gj′ ))) =
(vF (x) − 2) +
(2 − vφ (e)) + 2n(G).
j
x∈V (G)
e∈E(G)
We claim that rkK (G) is bounded independently of the choice of the finite graph G, which
implies that rkK (S) is finite. Indeed, Lemma 4.7 implies that 1 − rk(π1 (Gj′ )) cannot be
negative. In addition, the right-hand side of the equality does not depend on G, because
vF (x) = 2 as soon as x ∈
/ Sing, and vφ (e) = 2 for all edges of S that do not belong to
G. Up to enlarging G if necessary, we can thus assume that π1 (G) = π1 (S) (as graphs of
groups). This implies that G contains any embedded path in S with endpoints in G, and
therefore each component of S ′ contains only one component of G ′ . Lemmas 4.6 and 4.7
then imply that the left-hand side of the above equality is equal to i(O), so
X
X
i(O) =
(vF (x) − 2) +
(2 − vφ (e)) + 2|FO |,
(1)
x∈V (S)
e∈E(S)
where |FO | is the common value of |F|Stab(x) | for all x ∈ O (we recall that F|Stab(x)
denotes the set of conjugacy classes of peripheral subgroups in the Kurosh decomposition
of Stab(x)). We will now sum up the above equality over all orbits of points in F to get
an expression of the index of T . Remark that given a finite tree K, we have
X
(vK (x) − 2) = −2,
x∈K
where vK (x) denotes the valence of x in K. Therefore, as F has 1 + |F| connected
components, we obtain that
X X
(vF (x) − 2) = −2 − 2|F|.
O∈T /G x∈V (S)
We also have
X
2|FO | = 2|F|.
O∈T /G
Finally, since the number of morphisms in Φ is equal to rkK (G, F), we have
X X
(2 − vφ (e)) = 2rkK (G, F).
O∈T /G e∈S
By summing the above three contributions, we obtain
i(T ) = 2rkK (G, F) − 2,
and we are done in the geometric case.
We now turn to the general case, where T need no longer be geometric. Let (Kn )n∈N
be a sequence of finite subtrees of T such that the corresponding geometric (G, F)-trees
18
Tn strongly converge to T (we recall the definition of strong convergence from the paragraph preceding Theorem 3.3). Let x ∈ T be a branch or inversion point, and let s ≤ i(x)
be an integer. As T is a (G, F)-tree, the Kurosh decomposition of Stab(x) reads as
Stab(x) = g1 Gi1 g1−1 ∗ · · · ∗ gr Gir gr−1 ∗ F,
where Gi1 , . . . , Gir are pairwise non conjugate in G, and F is a free group (which might
a priori not be finitely generated). Let Y be a finite subset of Stab(x) made of elements
from a free basis of F and one nontrivial element in each of the subgroups gj Gij gj−1 . Let
d1 , . . . , dq be directions at x in T with trivial stabilizers, in distinct Stab(x)-orbits. We
make these choices in such a way that 2|Y | + q − 2 = s. Because of strong convergence,
it is possible for n large enough to lift x to an element xn ∈ Tn in such a way that
all elements in Y fix xn , and we can similarly lift all directions di to a direction from
xn in Tn . We have v1 (xn ) ≥ q, and the resolution morphism from Tn to T provided
by Theorem 3.1 induces an injective morphism from Stab(xn ) to Stab(x), whose image
contains all elements in Y , and all subgroups gj Gij gj−1 because Tn is a (G, F)-tree. As
the subgroup generated by Y and the collection of subgroups of the form gj Gij gj−1 is
a free factor, we get that rkK (Stab(x)) ≥ |Y |. Hence i(xn ) ≥ s. As this is true for all
s ≤ i(x), we get that i(x) ≤ i(xn ). Since lifts to Tn of branch or inversion points in
distinct G-orbits in T belong to distinct G-orbits of Tn , it follows from the first part of
the argument that i(T ) ≤ 2 rkK (G, F) − 2.
4.2
Bounding Q-ranks, and the dimension of V SL(G, F )
We now compute the dimension of V SL(G, F), following the arguments in [6, Sections
IV and V]. Let T be a minimal, small (G, F)-tree. We denote by L the additive subgroup
of R generated by the values of the translation lengths ||g||T , for g varying in G. The
Z-rank rZ (T ) is the rank of the abelian group L, i.e. the minimal number of elements
in a generating set of L (it is infinite if L is not finitely generated). The Q-rank rQ (T )
is defined to be the dimension of the Q-vector space L ⊗Z Q. Notice that we always
have rQ (T ) ≤ rZ (T ). Let Y be the set of points in T which are either branch points
or inversion points. We define Λ as the subgroup of R generated by distances between
points in Y . We have 2Λ ⊆ L ⊆ Λ, see [6, Section IV]. The following two propositions
were stated by Gaboriau and Levitt in the case of nonabelian actions of finitely generated
groups on R-trees without inversions. Their proofs adapt to our framework.
Proposition 4.9. (Gaboriau–Levitt [6, Proposition IV.1]) Let T be a small (G, F)-tree,
and let {g1 , . . . , gN } be a free basis of FN . Then the set {||gi ||T }i∈{1,...,N } generates L/2Λ.
Proposition 4.10. (Gaboriau–Levitt [6, Proposition IV.1]) Let T be a small (G, F)tree, and let {pj }j∈J be a set of representatives of the G-orbits of branch and inversion
points in T . Then for all j0 ∈ J, the set {dT (pj0 , pj )}j∈Jr{j0 } generates Λ/L.
We refer the reader to [6, Proposition IV.1] for a proof of the above two facts. We
mention that these proofs are based on the following lemma, which follows from standard
theory of group actions on R-trees.
Lemma 4.11. (Gaboriau–Levitt [6, Proposition IV.1]) Let T be a small (G, F)-tree.
• For all branch or inversion points p, q, r ∈ T , we have d(p, r) = d(p, q)+d(q, r) mod 2Λ.
• For all branch or inversion points p ∈ T and all g ∈ G, we have d(p, gp) =
||g||T mod 2Λ.
19
G1
Gk
Figure 2: A (3rkf (G, F) + 2|F| − 4)-simplex in PO(G, F).
• For all g, h ∈ G, we have ||gh||T = ||g||T + ||h||T mod 2Λ.
Proposition 4.12. Let T ∈ V SL(G, F) be a geometric tree, and let b be the number of
orbits of branch or inversion points in T . Then rZ (T ) ≤ rkf (G, F) + b − 1.
Proof of Proposition 4.12. It follows from the proof of Theorem 3.1 that Λ is generated
by distances between points in the finite singular set Sing. So Λ is finitely generated, and
therefore L is finitely generated (recall that 2Λ ⊆ L ⊆ Λ). Hence Λ/2Λ is isomorphic to
(Z/2Z)rZ (T ) , and the upper bound on rZ (T ) follows from Propositions 4.9 and 4.10.
We also recall the following result from [6, Proposition IV.2].
Proposition 4.13. (Gaboriau–Levitt [6, Proposition IV.2]) Let T ∈ V SL(G, F) be a
nongeometric tree obtained as the strong limit of a continuous system TK(t) of geometric
trees. Then
rQ (T ) ≤ lim inf rZ (TK(t) ),
t→+∞
and
rQ (T ) < lim sup rZ (TK(t) ).
t→+∞
Proposition 4.14. For all T ∈ V SL(G, F), we have rQ (T ) ≤ 3rkf (G, F) + 2|F| − 3.
Proof. When T is geometric, Proposition 4.12 implies that rQ (T ) ≤ rZ (T ) ≤ rkf (G, F)+
b−1. Corollary 4.5 shows that b ≤ 2rkf (G, F)+2|F|−2, and the claim follows. When T is
nongeometric, it is a strong limit of a system of geometric trees. Indeed, the construction
of the approximation in Theorem 3.3 can be done in a continuous way by choosing a
continuous increasing family of finite trees K(t) ⊆ T , though some of the trees TK(t)
may fail to be minimal in general. The claim then follows from Proposition 4.13.
Proposition 4.15. (Gaboriau–Levitt [6, Proposition V.1]) Let G be a countable group,
let F be a free factor system of G, and let k ≥ 1 be an integer. The space of projectivized
length functions of (G, F)-trees with Q-rank smaller than or equal to k has dimension
smaller than or equal to k − 1.
Proof of Theorem 4.1. Theorem 4.1 follows from Propositions 4.14 and 4.15, since outer
space PO(G, F) contains (3rkf (G, F) + 2|F| − 4)-simplices, obtained for instance by
varying the edge lengths of a graph of groups that has the shape displayed on Figure 2.
20
4.3
Very small graphs of actions
In this section, we mention a decomposition result which was proved by Levitt for actions
of finitely generated groups on R-trees having finitely many orbits of branch points [18,
Theorem 1]. The proof uses the fact that every such action on a tree T is finitely
supported, i.e. there exists a finite tree K ⊂ T such that every arc I ⊂ T is covered
by finitely many translates of K. The fact that minimal (G, F)-actions on R-trees are
finitely supported was noticed by Guirardel in [11, Lemma 1.14]. Using finiteness of
the number of orbits of branch and inversion points in a very small (G, F)-tree, Levitt’s
theorem adapts to our more general framework.
Theorem 4.16. (Levitt [18, Theorem 1]) Let G be a countable group, and let F be a
free factor system of G. Then every tree T ∈ V SL(G, F) splits uniquely as a graph of
actions, all of whose vertex trees have dense orbits, such that the Bass–Serre tree of the
underlying graph of groups is very small, and all its edges have positive length.
4.4
Additional arguments for computing the dimension of V SL(G, F ) r
PO(G, F )
We start by recalling the following well-known fact. We recall that a (G, F)-tree T has
finitely many orbits of directions if there are finitely many orbits of directions at branch
or inversion points in T .
Proposition 4.17. Let T be a (G, F)-tree with dense orbits. If T is small, and has
finitely many orbits of directions (in particular, if T is very small, or small and geometric), then all stabilizers of nondegenerate arcs in T are trivial.
Proof. Let e ⊆ T be a nondegenerate arc in T , and assume there exists a nontrivial
element g ∈ G such that ge = e. We can find two distinct directions d, d′ in e based at
branch or inversion points of T (oriented in the same way), and an element h ∈ G so
that d′ = hd. Notice in particular that h is hyperbolic in T , so h is nonperipheral, and
hg, hi is not cyclic. The points at which these directions are based can be chosen to be
both arbitrarily close to the midpoint of e, and in this case g and hgh−1 fix a common
nondegenerate subarc of e. As T is small, this implies that g and hgh−1 commute. Hence
h preserves the axis of g in any Grushko (G, F)-tree, which implies that g and h generate
a cyclic subgroup of G, a contradiction.
The following proposition extends [6, Theorem III.2].
Proposition 4.18. Let T be a small (G, F)-tree. If T is nongeometric, then i(T ) <
2rkK (G, F) − 2.
Proof. We know from Proposition 4.4 that i(T ) ≤ 2rkK (G, F) − 2. Assume towards a
contradiction that i(T ) = 2rkK (G, F) − 2.
Let Y ⊂ T be a finite set that contains one point from each G-orbit with positive
index, and let x ∈ Y . The Kurosh decomposition of Stab(x) reads as
Stab(x) = Hi1 ∗ · · · ∗ Hij ∗ F,
where F is a finitely generated free group, and Hil is G-conjugate to Gil for all l ∈
{1, . . . , j}. Let q be the rank of F . Let (Tn )n∈N be an approximation of T constructed
as in the proof of Theorem 3.3. We can assume that Kn has been chosen so that the
21
extremities of Kn are branch points or inversion points in T (this can be achieved by
choosing for x0 a branch point or inversion point of T , with the notations from the proof
of Theorem 3.3). As in the proof of Proposition 4.4 in the nongeometric case, we choose
directions d1 , . . . , dr , so that 2j + 2q + r − 2 = i(x), and n ∈ N so that we can associate
a point xn ∈ Tn to each x ∈ Y .
As i(Tn ) = i(T ), the orbit of every branch or inversion point of Tn with positive index
contains some xn . Furthermore, every direction from xn with trivial stabilizer belongs
to the Stab(xn )-orbit of the lift d′β of one of the directions dβ to Tn .
The morphism jn : Tn → T is not an isometry, otherwise T would be geometric.
Hence there exist y ∈ Tn , and two adjacent arcs e1 and e2 at y whose jn -images have
a common initial segment. If y is a branch point or an inversion point with positive
index, it follows from the above paragraph that both e1 and e2 have nontrivial stabilizer
(otherwise we would have i(T ) < i(Tn )). As T is small, the stabilizers of e1 and e2
generate a cyclic subgroup of G, so there exists g ∈ G that fixes both e1 and e2 in Tn .
This contradicts injectivity of jn in restriction to the fixed point set of g (Proposition 3.6).
If y is a branch or inversion point with index 0, then y has cyclic stabilizer, and there
exists g ∈ G that stabilizes all adjacent edges. Again, this contradicts injectivity of jn
in restriction to the fixed point set of g. If y is neither a branch point nor an inversion
point, then Theorem 3.1 implies that e1 and e2 are contained in the interior of a common
G-translate of Kn , because extremal points of Kn have been chosen to be branch or
inversion points in T . This again leads to a contradiction, since the restriction of jn to
this translate of Kn is an isometry.
The following proposition is an extension of [6, Theorem IV.1].
Proposition 4.19. For all very small (G, F)-trees T ∈
/ O(G, F), we have rQ (T ) <
3rkf (G, F) + 2|F| − 3.
Proof. When T is nongeometric, the claim follows from Propositions 4.13 and 4.14. We
will assume that T is geometric and show that rZ (T ) < 3rkf (G, F) + 2|F| − 3. We have
rZ (T ) < +∞ (Proposition 4.12), and Λ/2Λ is isomorphic to (Z/2Z)rZ (T ) .
If the number of distinct orbits of branch or inversion points in T is strictly smaller
than 2rkK (G, F) − 2, then rZ (T ) < 3rkf (G, F) + 2|F| − 3 by Proposition 4.12, and
we are done. Otherwise, let p1 , . . . , p2rkK (G,F )−2 be a set of representatives in K of
the orbits of branch or inversion points in T . Proposition 4.4 implies that for all j ∈
{1, . . . , 2rkK (G, F) − 2}, we have i(pj ) ≤ 1, and hence i(pj ) = 1 by Proposition 4.3.
If T is a simplicial tree, then Λ is generated by the lengths of the edges of the quotient
graph of groups. In particular, Proposition 4.14 implies that the maximal number of
edges of a simplicial tree in V SL(G, F) is 3rkf (G, F) + 2|F| − 3. All vertices of T have
index 1. Therefore, if x ∈ T is a vertex, we either have Stab(x) = {e} and v1 (x) = 3, or
rkK (Stab(x)) = 1 and v1 (x) = 1. Using the fact that T is very small, we get that every
vertex v of T satisfies one of the following possibilities, displayed on Figure 3: either v
1. has valence 3, and trivial stabilizer, or
2. projects in the quotient graph of groups to a valence 1 vertex whose stabilizer is
peripheral, or
3. projects in the quotient graph of groups to a valence 1 vertex whose stabilizer is
isomorphic to Z, and not peripheral, or
22
Gi
Z
Z
Z
1
2
3
4
Z
Z
Z
5
Figure 3: Vertices of index 1 in a very small simplicial (G, F)-tree.
Z
Z
Z
Figure 4: Simplicial trees in V SL(G, F) r PO(G, F) do not have maximal Z-rank.
4. projects in the quotient graph of groups to a valence 2 vertex with stabilizer isomorphic to Z and not peripheral, adjacent to both an edge with trivial stabilizer
and an edge with Z stabilizer, or
5. projects in the quotient graph of groups to a valence 3 vertex with stabilizer isomorphic to Z and not peripheral, adjacent to one edge with trivial stabilizer, and
two edges with Z stabilizers.
As T ∈
/ O(G, F), some vertex in T satisfies one of the last three possibilities. If some
vertex in T satisfies the third possibility, then one can replace the point with Z-stabilizer
in the quotient graph of groups by a loop-edge. This operation yields a new minimal,
very small simplicial tree T ′ for which rZ (T ′ ) > rZ (T ), so rZ (T ) < 3rkf (G, F) + 2|F| − 3.
Otherwise, the graph of groups T /G contains a concatenation of edges that all have the
same Z stabilizer, whose two extremal vertices have valence 2, and are adjacent to an
edge with trivial stabilizer, and whose interior vertices have valence 3, and are adjacent
to a single edge with trivial stabilizer, see Figure 4. Figure 4 illustrates how to construct
a tree T ′ with strictly more orbits of edges than T , so that rZ (T ′ ) > rZ (T ). Again, we
have rZ (T ) < 3rkf (G, F) + 2|F| − 3.
Assume now that T has dense orbits. Notice that Λ/2Λ = Λ/L + L/2Λ, so by
Propositions 4.9 and 4.10, it suffices to prove that the rank of Λ/L is strictly less than
b−1, where b denotes the number of orbits of branch or inversion points in T . Let K ⊆ T
be a finite subtree such that T = TK , chosen in such a way that every terminal vertex of
K is either a branch point or an inversion point in T . The (G, F)-tree T has trivial arc
stabilizers by Proposition 4.17, so the associated system of isometries K = (F, Φ) has
independent generators by Lemma 3.9 (notice here that we are considering the system
of isometries and not the system of morphisms). Using [7, Proposition 6.1], we get that
X
|F | =
|Aφ |,
(2)
φ∈Φ
where |F | (resp. |Aφ |) denotes the total length of F (resp. of Aφ ). We have
X
|F | =
dT (q, r),
e=[q,r]∈E(K)
23
where the sum is taken over all edges in F , after subdividing F at all points in the
singular set Sing (notice that in view of our choice of K, all points in Sing correspond to
branch points or inversion points in T ). We denote by V (F ) the set of vertices in F . Our
hypothesis on the extremal vertices of K, together with Lemma 4.11, implies that for all
points q, r ∈ V (F ), the length of [q, r] is equal modulo L to the sum dT (p1 , pi )+dT (p1 , pj ),
where pi (resp. pj ) belongs to the G-orbit of q (resp. r). Denoting by O(pi ) the orbit of
pi for all i ∈ {1, . . . , 2rkK (G, F) − 2}, we have
2rkK (G,F )−2
X
X
vF (x) mod L,
dT (p1 , pi ) ×
|F | =
i=1
x∈V (F )∩O(pi )
and similarly, for all φ ∈ Φ, we have
2rkK (G,F )−2
|Aφ | =
X
i=1
dT (p1 , pi ) ×
X
x∈V (Aφ )∩O(pi )
vφ (x) mod L,
where V (Aφ ) denotes the set of vertices of Aφ , which is contained in V (F ), and vφ (x)
denotes the valence of x in Aφ . Using the above two equalities, Equation (2) gives a
linear relation in Λ/L between the numbers dT (p1 , pi ), where the coefficient of dT (p1 , pi )
is equal to
X
X
vF (x) −
vφ (x) .
φ∈Φ
x∈V (F )∩O(pi )
For all i ∈ {1, . . . , 2rkK (G, F) − 2}, the index of pi is equal to 1. Therefore, Equation (1)
from the proof of Proposition 4.4 implies that
X
X
vF (x) −
vφ (x)
φ∈Φ
x∈V (F )∩O(pi )
is odd. Equation (2) thus leads to the nontrivial relation
2rkK (G,F )−2
X
dT (p1 , pj ) = 0 mod L
j=2
between the generators of Λ/L, so rZ (T ) < 3rkf (G, F) + 2|F| − 3.
In general, let G be the decomposition of T as a graph of actions provided by Theorem 4.16. We assume that T is not simplicial, and let Tv be a nontrivial vertex tree
of this decomposition. Then Tv is a very small (Gv , FGv )-tree with dense Gv -orbits. Let
T ′ be the very small (G, F)-tree obtained from T by collapsing all vertex trees in the
G-orbit of Tv to points. By definition of the index, we have
i(T ) − i(T ′ ) = i(Tv ) − (2rkK (Gv ) − 2).
As T is geometric, Proposition 4.4 implies that the left-hand side of the above equality
is nonnegative, while the right-hand side is nonpositive. This implies that i(Tv ) =
2rkK (Gv ) − 2. Using Proposition 4.18, this shows that the tree Tv is geometric. Assume
that the number of distinct Gv -orbits of branch or inversion points in the minimal subtree
of Tv is strictly smaller than 2rkK (Gv ) − 2. Then one of these orbits has index at least
24
2 in Tv , and hence in T . This implies that the number of distinct G-orbits of branch
or inversion points in T is strictly smaller than 2rkK (G, F) − 2, and we are done by
Proposition 4.12. We are thus left with the case where the number of distinct Gv -orbits
of branch or inversion points in Tv is equal to 2rkK (Gv ) − 2. As distinct G-translates
of Tv are disjoint in T , all these Gv -orbits of branch or inversion points are distinct
when viewed as G-orbits of points in T . We denote by p1 , . . . , p2rkK (Gv )−2 a set of
representatives of the Gv -orbits of branch or inversion points of Tv . In this case, as Tv
has dense orbits, the analysis from the above paragraph provides a nontrivial relation
between the generators dTv (p1 , pi ) of Λ(Tv )/L(Tv ). The numbers dTv (p1 , pi ) may also
be viewed as part of a generating set of Λ(T )/L(T ), and we have a nontrivial relation
between these generators. Again, this implies that rZ (T ) < 3rkK (G, F) + 2|F| − 3.
Proof of Theorem 4.2. Theorem 4.2 follows from Propositions 4.15 and 4.19, because
V SL(G, F) r PO(G, F) contains a (3rkf (G, F) + 2|F| − 5)-simplex made of simplicial
(G, F)-trees (except in the case where G = G1 ∗ G2 and F = {G1 , G2 }, for which
PO(G, F) is reduced to a point and V SL(G, F) r PO(G, F) is empty). An example of
such a simplex is given by varying edge lengths in a graph of groups obtained from the
graph of groups displayed on Figure 2 by collapsing a loop, or merging two points corresponding to subgroups G1 and G2 , and adding an edge with nontrivial cyclic stabilizer
generated by a nonperipheral element in G1 ∗ G2 , for instance.
5
Very small actions are in the closure of outer space.
In the classical case where G = FN is a finitely generated free group of rank N , and
F = ∅, Cohen and Lustig have shown that a minimal, simplicial FN -tree lies in the
closure cvN if and only if it is very small [2]. Bestvina and Feighn [1] have extended
their result to all minimal FN -actions on R-trees. However, it seems that their proof does
not handle the case of actions that contain both nontrivial arc stabilizers, and minimal
components dual to measured foliations on compact, nonorientable surfaces. Indeed, for
such actions, it is not clear how to approximate the foliation by rational ones without
creating any one-sided leaf (in which case the action we get is not very small). If the
action has trivial arc stabilizers (i.e. if the dual band complex contains no annulus),
then the argument in [1, Lemma 4.1] still enables to get an approximation by very small,
simplicial FN -trees, by using the narrowing process described in [9, Section 7]. However,
this argument does not seem to handle the case of trees having nontrivial arc stabilizers.
We will give a proof of the fact that cvN is the space of very small, minimal, isometric
actions of FN on R-trees that does not rely on train-track arguments for approximating
measured foliations on surfaces by rational ones. Our proof also gives an interpretation
of Cohen and Lustig’s for simplicial trees. We again work in our more general framework
of (G, F)-trees, and show the following result.
Theorem 5.1. Let G be a countable group, and let F be a free factor system of G. The
closure O(G, F) (resp. PO(G, F)) is the space of (projective) length functions of very
small (G, F)-trees.
In particular, Theorem 5.1 states that every minimal, very small (G, F)-tree T can be
approximated by a sequence of Grushko (G, F)-trees. When T has trivial arc stabilizers,
we can be a bit more precise about the nature of the approximation we get.
25
Definition 5.2. Let T be a (G, F)-tree. A Lipschitz approximation of T is a sequence
(Tn )n∈N of (G, F)-trees that converges to T , and such that for all n ∈ N, there exists a
1-Lipschitz G-equivariant map from Tn to T .
Lipschitz approximations seem to be useful: they were a crucial ingredient in [15] for
tackling the question of spectral rigidity of the set of primitive elements of FN in cvN .
They also turn out to be a useful ingredient for describing the Gromov boundary of the
(hyperbolic) graph of (G, F)-cyclic splittings in [13]. The quotient volume of a very small
(G, F)-tree T is defined as the sum of the edge lengths of the graph of actions underlying
the Levitt decomposition of T (where we recall the definition of the Levitt decomposition
from Theorem 4.16). In other words, if S is the simplicial tree obtained after collapsing
all subtrees with dense orbits in the decomposition of T , then the quotient volume of S
is the volume of the quotient graph S/G. In particular, if T has dense orbits, then its
quotient volume is equal to 0.
Theorem 5.3. Let G be a countable group, and let F be a free factor system of G. Then
every minimal (G, F)-tree T with trivial arc stabilizers admits a Lipschitz approximation by (unprojectivized) Grushko (G, F)-trees, whose quotient volumes converge to the
quotient volume of T .
5.1
Reduction lemmas
To prove Theorem 5.1, we are left showing that every very small minimal (G, F)-tree
T can be approximated by a sequence of Grushko (G, F)-trees. By Theorem 3.8, we
can approximate every minimal, very small (G, F)-tree T by a sequence of minimal,
very small, geometric (G, F)-trees. This approximation is a Lipschitz approximation.
If T has dense orbits, then Remark 3.4 implies that the quotient volumes of the trees
in the approximation converge to the quotient volume of T , which is equal to 0. More
generally, in the case where the Levitt decomposition of T has trivial edge stabilizers,
then Lemmas 5.4 and 5.5 below enable us to find a Lipschitz approximation of T by trees
Tn whose quotient volumes converge to the quotient volume of T , such that all trees with
dense orbits in the Levitt decomposition of the trees Tn are geometric. In this case, the
argument from the last paragraph of the proof of Proposition 4.19 shows that the trees
Tn themselves are geometric. To complete the proof of Theorems 5.1 and 5.3, we are
left understanding how to approximate minimal, very small, geometric (G, F)-trees by
minimal Grushko (G, F)-trees.
Our proof of Theorems 5.1 and 5.3 will make use of the following lemmas, which
enable us to approximate very small (G, F)-trees that split as graphs of actions, as
soon as we are able to approximate the vertex actions. Lemma 5.5 is a version of
Guirardel’s Reduction Lemma in [9, Section 4], where we keep track of the fact that
the approximations of the trees are Lipschitz approximations. In Lemma 5.6, we tackle
the problem of approximating trees with nontrivial arc stabilizers by Grushko (G, F)trees. Our argument may be seen as an interpretation of Cohen and Lustig’s twisting
argument for approximating such trees [2]. We consider graphs of actions, instead of
restricting ourselves to simplicial trees. We first recall Guirardel’s Reduction Lemma
from [9, Section 4]. In the statements below, all limits are nonprojective.
Lemma 5.4. (Guirardel [9, Section 4]) Let T be a very small (G, F)-tree that splits as
a graph of actions G. Assume that all pointed vertex actions (T v , (uv1 , . . . , uvk )) admit an
approximation by a sequence of pointed (Gv , F|Gv )-actions ((Tnv , (uv1,n , . . . , uvk,n )))n∈N , in
26
which the approximation points are fixed by the adjacent edge stabilizers. For all n ∈ N,
let Tn be the (G, F)-tree obtained by replacing all vertex actions (T v , (uv1 , . . . , uvk )) by
their approximation (Tnv , (uv1,n , . . . , uvk,n )) in G. Then (Tn )n∈N converges to T .
We say that a sequence ((Tn , (u1n , . . . , ukn )))n∈N of pointed (G, F)-trees is a Lipschitz
approximation of a pointed (G, F)-tree (T, (u1 , . . . , uk )) if ((Tn , (u1n , . . . , ukn )))n∈N converges to (T, (u1 , . . . , uk )), and for all n ∈ N, there exists a 1-Lipschitz G-equivariant
map fn : Tn → T such that for all i ∈ {1, . . . , k}, we have fn (uin ) = ui . Guirardel’s
Reduction Lemma can be refined in the following way.
Lemma 5.5. Let T be a very small (G, F)-tree with trivial arc stabilizers, that splits as
a (G, F)-graph of actions G. If all pointed vertex trees (T v , (uv1 , . . . , uvk )) of G admit Lipschitz approximations by pointed Grushko (Gv , F|Gv )-trees, in which the approximation
points uv1,n , . . . , uvk,n are fixed by the adjacent edge stabilizers, then T admits a Lipschitz
approximation by Grushko (G, F)-trees.
Lemma 5.6. Let T be a minimal, very small (G, F)-tree, that splits as a (G, F)-graph
of actions over a one-edge (G, F)-free splitting (where the vertex actions need not be
minimal). If the minimal subtrees of all vertex trees of G (with respect to the action of
their stabilizer Gv ) admit approximations by minimal Grushko (Gv , F|Gv )-trees, then T
admits an approximation by minimal Grushko (G, F)-trees.
Figures 5, 7, and 8 provide examples of trees for which the vertex actions of the
splitting are not minimal (but they are minimal in the sense of pointed trees when we keep
track of the attaching points). These are the crucial cases of Lemma 5.6, in which we deal
with the problem of approximating trees with nontrivial arc stabilizers. Considering nonminimal vertex actions is crucial to deal with the simplicial case in Theorem 5.1 (when
there are edges with nontrivial stabilizers), and Lemma 5.6 provides a new interpretation
of Cohen and Lustig’s argument for dealing with this case. Lemma 5.6 will also be crucial
for dealing with the case of geometric actions of surface type containing nontrivial arc
stabilizers.
Proof. We will provide a detailed argument in the case where the (G, F)-free splitting S
is a free product, and explain how to adapt the argument to the case of an HNN extension.
Case 1 : The splitting S is of the form G = A ∗ B.
The following description of G is illustrated in Figure 5. We denote by L the length
of the edge of G, which might be equal to 0. Denote by T A and T B the vertex trees of
G, and by uA ∈ T A and uB ∈ T B the corresponding attaching points. The trees T A and
A
B
T B may fail to be minimal, we denote by Tmin
and Tmin
their minimal subtrees. Up
A
to enlarging L if necessary, we can assume that the set T A r Tmin
is either empty (in
A
the case where T is minimal), or consists of the orbit of a single point in the closure of
A , or consists of the orbit of a nondegenerate half-open arc with nontrivial stabilizer.
Tmin
We will explain how to approximate the tree (T A , uA ) by a sequence of pointed
Grushko (A, F|A )-trees. By approximating the pointed tree (T B , uB ) in the same way,
our claim then follows from Guirardel’s Reduction Lemma (Lemma 5.4).
If T A is minimal, we can approximate (T A , uA ) by a sequence of pointed Grushko
A
A
(A, F|A )-trees (TnA , uA
n ) by assumption (by choosing un to be an approximation of u in
the tree TnA , provided by the definition of the equivariant Gromov–Hausdorff topology).
27
A y TA
vA
A
A y Tmin
B y TB
lA
L
uA
uB
hcA i hcA i
lB
hcB i hcB i
vB
B
B y Tmin
Figure 5: The splitting of T as a graph of actions in Case 1 of the proof of Lemma 5.6.
A consists of the orbit of a single point
This also remains true in the case where T A r Tmin
A
A
u in the closure of Tmin . Indeed, in this case, we can first approximate uA by a sequence
A )N , and then choose for each n ∈ N an approximation uA of u′
of points (u′n )n∈N ∈ (Tmin
n
n
A .
in an approximation of Tmin
A consists of the orbit of a nondegenerate arc [uA , v A ]
We now assume that T A r Tmin
with nontrivial stabilizer, whose length we denote by lA . We choose the notations so
A . We will also first assume that T A
that v A ∈ Tmin
min is not reduced to a point. As T
is very small, the stabilizer hcA i of the arc [uA , v A ] is cyclic, closed under taking roots,
and non-peripheral. As tripod stabilizers are trivial in T , the point v A is an endpoint of
A fixed by cA . If this arc is nondegenerate, then we let w A be its other
the subarc of Tmin
endpoint. Otherwise, we let wA be any point that is not equal to v A .
A
A
Let (Tmin,n
)n∈N be an approximation of Tmin
by minimal Grushko (A, F|A )-trees.
A
A
A
Denote by vn (respectively wn ) an approximation of v A (resp. wA ) in the tree Tmin,n
,
provided by the definition of convergence in the equivariant Gromov–Hausdorff topology.
A
We can assume that for all n ∈ N, the point vnA belongs to the axis of cA in Tmin,n
.
We refer to Figure 6 for an illustration of the following construction. For all n ∈ N,
A A
let (TnA , uA
n ) be the pointed tree obtained from (T , u ) in the following way. We start
A
A
by equivariantly unfolding the arc [u , v ] to obtain a tree TeA that contains an edge
e0 of length lA with trivial stabilizer. We then equivariantly replace the pointed tree
A , v A ) in the graph of actions defining T
eA by its approximation (T A , v A ), to get
(Tmin
min,n n
A
A
A
f
a tree Tn . Finally, we define the tree (T , u ) in the following way: the tree (T A , uA )
n
n
n
n
fn A , uA ) by fully folding the edge e0 along the axis of cA in T A ,
is obtained from (T
min,n
A A
A
A
A
A
f
in a direction that does not contain wn . We denote by fn : (Tn , u ) → (Tn , un ) the
folding map.
A A
We now prove that the pointed trees (TnA , uA
n ) converge to (T , u ). Lemma 5.4
fn A , uA ) converge to (TeA , uA ). For all n ∈ N, there is a 1implies that the trees (T
A
fn , uA ) → (TnA , uA
Lipschitz G-equivariant map fnA : (T
n ). This implies that for all
A
A
g ∈ G, we have dTnA (un , gun ) ≤ d f A (uA , guA ). Therefore, up to possibly passing to
Tn
A
A
a subsequence, the pointed trees (TnA , uA
n ) converge to a pointed tree (T∞ , u∞ ) in the
Gromov–Hausdorff equivariant topology, that is minimal in the sense of pointed G-trees.
A , uA ),
Proposition 1.5 shows that there exists a 1-Lipschitz map f A : (TeA , uA ) → (T∞
∞
A denotes the metric completion of T A . We will show that f A factors through
where T∞
∞
A , uA ), and that g A is an isometry between (T A , uA ) and
a map gA : (T A , uA ) → (T∞
∞
A , uA ). This will imply that the pointed trees (T A , uA ) converge to (T A , uA ).
(T∞
n
n
∞
A
We first notice that for all n ∈ N, the map fnA is an isometry in restriction to Tmin,n
.
A
A
By taking limits, this implies that the A-minimal subtree Tmin of T isometrically emA . In addition, for all n ∈ N, the point uA belongs to the axis of cA in
beds into T∞
n
28
fn A
T
A
A y Tmin,n
cA
wnA
vnA
lA
uA
TeA
convergence
A
A v
A y Tmin
lA
{1}
uA
fold
uA
n
TA
A
A v
A y Tmin
fnA
lA
hcA i
uA
hcA i
gA = id
convergence
(TnA , uA
n)
A , uA )
(T∞
∞
Figure 6: The situation in Case 1 of the proof of Lemma 5.6.
Tn . By definition of the equivariant Gromov–Hausdorff topology on the set of pointed
A
A
(G, F)-trees, this implies that cA fixes uA
∞ in T∞ . Similarly, the element c fixes all
A
A
A
A
points of the image of [u , v ] in T . Therefore, the map f factors through a map
A
A can
A , uA ). As T A
gA : (T A , uA ) → (T∞
∞
min isometrically embeds into T∞ , the map g
only decrease the length of the segment [uA , v A ], and fold this segment over a subarc of
[v A , wA ].
A
Let g ∈ A be an element that is hyperbolic in Tmin
(we recall that we have asA
sumed Tmin not to be reduced to a point), such that dT A (v A , gv A ) = dT A (wA , gwA ) +
2dT A (v A , wA ). In particular, we have dT A (uA , guA ) = dT A (wA , gwA )+2lA +2dT A (v A , wA ).
Using the fact that we folded in a direction that did not contain wn , together with
the definition of the equivariant Gromov–Hausdorff topology, we get that the distance
A
A
A
A
A
A
dTnA (uA
n , gun ) gets arbitrarily close to dTnA (wn , gwn ) + 2l + 2dT A (v , w ) as n tends to
A
A
A
A
A
A + 2d
A
A
+∞, so dT∞
A (u∞ , gu∞ ) = dT A (w , gw ) + 2l
T A (v , w ). This implies that g is
A , uA ), and we are done.
an isometry from (T A , uA ) to (T∞
∞
A
A
If Tmin is reduced to a point, then it can be approximated by a sequence (Tmin,n
)n∈N
1
of Grushko (A, F|A )-trees, where all edge lengths are equal to n , which we choose to be
all homothetic to each other. We also choose two distinct constant sequences vnA and wnA
A
in the trees Tmin,n
, and construct the trees TnA as above. Let g ∈ A be any element such
that dTnA (vnA , gvnA ) = dTnA (wnA , gwnA ) + 2dTnA (vnA , wnA ) for all n ∈ N. Arguing similarly as
A
A
A
A defined as
above, we get that dT∞
A (u∞ , gu∞ ) = 2l . This again implies that the map g
above is an isometry.
Case 2 : The splitting S is of the form G = C∗.
29
fA
hc2 i
C
C y Tmin
l2
v2
v1
hc1 i
hc2 i
u2
l3
t
l1 u
1
hc1 i
C y TC
Figure 7: The splitting of T as a graph of actions in Cases 2.1, 2.2 and 2.5 of the proof
of Lemma 5.6.
The vertex tree T C of G may fail to be minimal. We denote by u1 and u2 two points in
T C in the orbits of the attaching points (the points u1 and u2 may belong to the same
C of the C-minimal
G-orbit). We denote by v1 and v2 their projections to the closure Tmin
subtree of T . One of the following cases occurs.
Case 2.1 : The segments [u1 , v1 ] and [u2 , v2 ] are nondegenerate, and their stabilizers
are nontrivial and nonconjugate in C.
In other words, the tree T splits as a graph of actions that has the shape displayed on
Figure 7, where l1 , l2 > 0, and the stabilizers hc1 i and hc2 i are nonconjugate. We allow
the case where v1 and v2 belong to the same G-orbit. For all i ∈ {1, 2}, we let wi be
C , if this arc is nondegenerate,
such that [vi , wi ] is the maximal arc fixed by ci in Tmin
and we let wi be any point distinct from vi otherwise (as in Case 1, one has to slightly
C is reduced to a point). Let T
eC be the tree obtained from
adapt the argument when Tmin
C
T by replacing the edges [u1 , v1 ] and [u2 , v2 ] by edges of the same length with trivial
stabilizer. For all i ∈ {1, 2}, let vn,i (resp. wn,i ) be an approximation of vi (resp. wi ) in
C . We can assume v
an approximation of Tmin
n,i to belong to the translation axis of ci .
C
C
fn , u1 , u2 ) be the approximation of (Te , u1 , u2 ) obtained from an approximation
Let (T
C
of Tmin by adding an edge of length l1 (resp. l2 ) with trivial stabilizer at vn,1 (resp.
fn C by G-equivariantly fully folding the edge
vn,2 ). Let TnC be the tree obtained from T
[ui , vn,i ] along the axis of ci , in a direction that does not contain wn,i , for all i ∈ {1, 2}.
fn C → TnC the corresponding morphism. Arguing as in Case 1, one
We denote by fnC : T
shows that the trees (TnC , fnC (u1 ), fnC (u2 )) converge to (T C , u1 , u2 ). Let now Tn be the
tree obtained by replacing (T C , u1 , u2 ) by its approximation (TnC , fnC (u1 ), fnC (u2 )) in the
graph of actions G. Lemma 5.4 implies that the trees Tn converge to T .
Case 2.2 : The segments [u1 , v1 ] and [u2 , v2 ] are nondegenerate and have nontrivial stabilizers that are conjugate in C, and no two nondegenerate subsegments of [u1 , v1 ] and
[u2 , v2 ] belong to the same G-orbit.
Again, the tree T splits as a graph of actions that has the shape displayed on Figure 7,
where this time the groups hc1 i and hc2 i are conjugate. Up to a good choice of the
stable letter t, we can assume that hc1 i = hc2 i. As tripod stabilizers are trivial in T , the
30
t
C
C y Tmin
hci
v := v1 = v2 l1
hci
l3
u1
u2 hci
hci
l2
C y TC
Figure 8: The splitting of T as a graph of actions in Case 2.3 of the proof of Lemma 5.6.
C . Again, we let T
eC be the tree
segment [v1 , v2 ] is the maximal arc fixed by c1 in Tmin
C
obtained from T by replacing the edges [u1 , v1 ] and [u2 , v2 ] by edges of the same length
with trivial stabilizer. For all i ∈ {1, 2}, let vn,i be an approximation of vi in an approxiC by minimal Grushko (C, F )-trees, which we can assume to belong to the
mation of Tmin
|C
C
fn , u1 , u2 ) be the approximation of (TeC , u1 , u2 ) obtained
translation axis of c1 . Let (T
C
from an approximation of Tmin
by adding an edge of length l1 (resp. l2 ) with trivial
fn C by G-equivariantly
stabilizer at vn,1 (resp. vn,2 ). The tree TnC is then obtained from T
fully folding the edges [u1 , vn,1 ] and [u2 , vn,2 ] along the axis of c1 in opposite directions.
The folding directions should not contain the segment [vn,1 , vn,2 ], in case this segment
fn C → T C the corresponding morphism, the
is nondegenerate. Again, denoting by fnC : T
n
trees (TnC , fnC (u1 ), fnC (u2 )) converge to (T C , u1 , u2 ). The trees Tn obtained by replacing
(T C , u1 , u2 ) by (TnC , fnC (u1 ), fnC (u2 )) in G converge to T .
Case 2.3 : Some nondegenerate subsegments of [u1 , v1 ] and [u2 , v2 ] belong to the same
G-orbit, and their common stabilizer is nontrivial.
Using the fact that tripod stabilizers in T are trivial, we can assume that v1 = v2 (and
we let v := v1 = v2 ), and that [u1 , v] ⊆ [u2 , v]. The tree T splits as a graph of actions
that has the form displayed on Figure 8. We let w be such that [v, w] is the maximal
C
arc fixed by c in Tmin
if this arc is nondegenerate, and choose any w 6= v otherwise (as
C
in Case 1, one has to slightly adapt the argument if Tmin
is reduced to a point). Let
C
e
T be the tree obtained from T by replacing the segment [u2 , v] by a segment of same
length l1 + l2 with trivial stabilizer. Let vn (resp. wn ) be an approximation of v (resp.
C . We can assume v to belong to the translation axis
w) in an approximation Yn of Tmin
n
C
f
e
of c in Yn . Let (Tn , u2 ) be the approximation of (T C , u) obtained from Yn by adding
an edge of length l1 + l2 with trivial stabilizer at vn . The tree TnC is then obtained from
fn C by G-equivariantly fully folding the edge [u2 , vn ] along the axis of c, in a direction
T
fn C → TnC the corresponding morphism,
that does not contain wn . Denoting by fnC : T
the trees (TnC , fnC (u1 ), fnC (u2 )) converge to (T C , u1 , u2 ). Again, the trees Tn obtained by
replacing (T C , u1 , u2 ) by (TnC , fnC (u1 ), fnC (u2 )) in G converge to T .
Case 2.4 : Some nondegenerate subsegments of [u1 , v1 ] and [u2 , v2 ] belong to the same
G-orbit, and they have trivial stabilizer.
Then T splits as a graph of actions of the form displayed on Figure 9. This case may be
viewed as a particular case of Case 1.
31
C
C y Tmin
t
C y TC
Figure 9: The splitting of T as a graph of actions in Case 2.4 of the proof of Lemma 5.6.
Case 2.5 : One of the subsegments [u1 , v1 ] or [u2 , v2 ] is degenerate.
This case is treated in a similar way as Case 2.1, and left to the reader.
5.2
Dynamical decomposition of a geometric very small (G, F )-tree
Every geometric very small (G, F)-tree splits as a graph of actions, which has the following description.
Proposition 5.7. Any very small geometric (G, F)-tree T splits as a graph of actions
G, where for each nondegenerate vertex action Y , with vertex group GY , and attaching
points v 1 , . . . , v k fixed by subgroups H 1 , . . . , H k , either
• the tree Y is an arc containing no branch point of T except at its endpoints, or
• the group GY is the fundamental group of a 2-orbifold with boundary Σ holding an
e or
arational measured foliation, and Y is dual to Σ,
• there exists a Lipschitz approximation of Y by pointed Grushko (GY , F|GY )-trees
(Yn , (vn1 , . . . , vnk )), whose quotient volumes converge to 0, such that for all n ∈ N,
there exists a morphism fn : Yn → Y , with fn (vni ) = v i for all i ∈ {1, . . . , k}, and
vni is fixed by H i for all i ∈ {1, . . . , k} and all n ∈ N.
We call G the dynamical decomposition of T , it is determined by T . Vertex actions
of the third type are called exotic. In case no vertex action is exotic, we say that T is of
surface type.
The proof of Proposition 5.7 goes as follows. Let T be a geometric (G, F)-tree, and
let K ⊆ T be a finite subtree such that T = TK . Let A := Σ/G, where Σ is the band
complex defined in Section 3.1. Let A∗ be the complement of the singular set in F ⊆ A
(where F is the compact forest defined in Section 3.3), endowed with the restriction of
the foliation of Σ. Let C ∗ ⊆ A∗ be the union of the leaves of A∗ which are closed but
not compact. The cut locus of A is defined as C := C ∗ ∪ Sing. The set A r C is a union
of finitely many open sets U1 , . . . , Up , which are unions of leaves of A∗ . By a classical
result of Imanishi [16], see also [7, Section 3], for all i ∈ {1, . . . , p}, either every leaf of
Ui is compact, or else every leaf of Ui is dense in Ui . Notice that each component Ui is
again dual to a finite system of isometries.
As noticed in [11, Propositions 1.25 and 1.31], Imanishi’s theorem provides a transe be the lift of the cut locus to Σ. Given a
verse covering of T in the following way. Let C
e we let TU be the tree dual to the foliated 2-complex U , i.e. the
component U of Σ r C,
leaf space made Hausdorff of U . Then the family {TU }U is a transverse covering of T .
Each TU is either an arc (in the case where every leaf of the image of U in A is compact)
or has dense orbits (in the case where all leaves of the image of U in A are dense). Associated to this transverse covering of T is a graph of actions, whose vertex groups are
32
finitely generated, and whose vertex actions are dual to foliated 2-complexes. In addition, arc stabilizers in the vertex actions with dense orbits are trivial (Proposition 4.17).
Therefore, we can apply [11, Proposition A.6] to each of the vertex actions with dense
orbits. This provides a classification of vertex actions with dense orbits into three types
(axial, surface and exotic). As our system of partial isometries on F has independent
generators (Lemma 3.9), all vertex actions with dense orbits of the decomposition are
dual to finite systems of isometries with independent generators. This excludes the axial
case, see [5, Proposition 3.4]. The existence of the Lipschitz approximation with the
required properties in the exotic case was proved by Guirardel in [9, Proposition 7.2],
using a pruning and narrowing argument.
We finish this section by mentioning a consequence of Proposition 5.7, that will
turn out to be useful in [14].
Lemma 5.8. Let T be a small, minimal (G, F)-tree. If there exists a subgroup H ⊆
G that is elliptic in T , and not contained in any proper (G, F)-free factor, then T is
geometric of surface type.
Proof. If T is geometric, this follows from Proposition 5.7, so we assume that T is
nongeometric. Up to replacing H by the point stabilizer of T in which it is contained,
we can assume that rkK (H) < +∞. Theorem 3.8 lets us approximate T by a sequence
(Tn )n∈N of small, minimal geometric (G, F)-trees, in which H is elliptic. The trees Tn
come with morphisms onto T . As T is nongeometric, Corollary 3.6 ensures that the trees
Tn contain an edge with trivial stabilizer. This implies that H is contained in a proper
(G, F)-free factor, a contradiction.
5.3
Trees of surface type
Let T be a very small geometric (G, F)-tree of surface type (where we recall the definition
from the paragraph below Proposition 5.7). Let G be the dynamical decomposition of
T , and let S be the skeleton of the corresponding transverse covering. It follows from
Proposition 5.7 that there are three types of vertices in S, namely: vertices of surface
type, of arc type, and vertices associated to nontrivial intersections between the trees of
the transverse covering, which we call vertices of trivial type. All edge stabilizers in S
are cyclic (possibly finite or peripheral). Indeed, stabilizers of edges adjacent to vertices
of surface type are either trivial, or they are cyclic, and represent boundary curves or
conical points of the associated orbifold. Both edges adjacent to vertices of arc type have
the same stabilizer, equal to the stabilizer of the corresponding arc in T , which is cyclic
because T is very small.
Definition 5.9. Let T be a very small geometric (G, F)-tree of surface type, and let σ
be a compact 2-orbifold arising in the dynamical decomposition G of T . Let g ∈ G be an
element represented by a boundary curve of σ. We say that g is used in T if either g is
peripheral, or g is conjugate into some adjacent edge group of G. Otherwise g is unused
in T .
Proposition 5.10. Let T be a minimal, very small, geometric (G, F)-tree of surface
type. Then either there exists an unused element in T , or T splits as a (G, F)-graph of
actions over a one-edge (G, F)-free splitting.
Our proof of Proposition 5.10 is based on the following lemma. Given a group G,
and a family Y of subgroups of G, we say that G splits as a free product relatively to
33
Y if there exists a splitting of the form G = A ∗ B, such that every subgroup in Y is
conjugate into either A or B.
Lemma 5.11. Let G be a countable group, and let F be a free factor system of G. Let T
be a minimal, simplicial (G, F)-tree, whose edge stabilizers are all cyclic and nontrivial
(they may be finite or peripheral). Then there exists a vertex v in G, such that Gv splits
as a free product relative to incident edge groups and subgroups in F|Gv .
Let T, T ′ be two simplicial (G, F)-trees. A map f : T → T ′ is a G-equivariant edge
fold (or simply a fold) if there exist two edges e1 , e2 ⊆ T , incident to a common vertex
in T , such that T ′ is obtained from T by G-equivariantly identifying e1 and e2 , and
f : T → T ′ is the quotient map. A fold f : T → T ′ is determined by the orbit of the
pair of edges (e1 , e2 ) identified by f . We say that f is
• of type 1 if e1 and e2 belong to distinct G-orbits of oriented edges in T , and both
e1 and e2 have nontrivial stabilizer, and
• of type 2 if e1 and e2 belong to distinct G-orbits of oriented edges in T , and either
e1 or e2 (or both) has trivial stabilizer in T , and
• of type 3 if e1 and e2 belong to the same G-orbit of oriented edges in T .
Assume that f : T → T ′ is a fold. We note that if H ⊆ G is a subgroup of G that
fixes an edge e1 ⊆ T ′ , and ee1 is an edge in the f -preimage of e1 in T , then H fixes an
extremity of ee1 . We start by making the following observation.
Lemma 5.12. Let T and T ′ be two simplicial (G, F)-trees with cyclic edge stabilizers.
Assume that T ′ is obtained from T by performing a fold f of type 2 or 3. If e′1 and
e′2 are two edges of T that are identified by f , then either e′1 or e′2 (or both) has trivial
stabilizer.
Proof of Lemma 5.11. Let T0 be any Grushko (G, F)-tree. All point stabilizers in T0 are
elliptic in T , so up to possibly collapsing some edges in T0 , and subdividing edges of T0 ,
there exists a simplicial map f : T0 → T (i.e. sending vertices to vertices and edges to
edges). By [23, 3.3], the map f can be decomposed as a sequence of G-equivariant edge
folds fi : Ti → Ti+1 .
We can assume that along the folding sequence, we always perform edge folds of type
1 before performing edge folds of type 2 and 3, and we always perform edge folds of type
2 before edge folds of type 3, for as long as possible. This is possible because the number
of orbits of edges decreases when performing a fold of type 1 or 2.
We claim that we can also assume that folds are maximal in the following sense: if
g fixes an edge e in T , then we never identify a preimage e′ of e with a translate of the
form gk e′ without identifying it with ge′ . Assume otherwise, and let i be the first time
at which a non-maximal fold occurs. Let e′ be a preimage of e in Ti , having a vertex y
stabilized by gk , so that we fold e′ and gk e′ when passing from Ti to Ti+1 . By our choice
of i and the fact that edge stabilizers in T are cyclic, the edge e′ has trivial stabilizer in
Ti (in particular, the fold performed from Ti to Ti+1 is not of type 1). The element g is
also elliptic in Ti . We let x be the point closest to y that is fixed by g in Ti . If x = y,
then we could choose to identify e′ and ge′ when passing from Ti to y. Otherwise, all
edges in the segment [x, y] ⊆ Ti are stabilized by g k . Our choice of i implies that the
stabilizer of their images in T is equal to gk (and not to any proper root of g k ). Since the
34
image of e′ in T is stabilized by g, this implies that we can find two consecutive edges
on the segment [x, y] that are identified in T . This shows that one could perform a fold
of type 1 on the tree Ti , contradicting our choice of folding path.
Let Tk be the last tree along the folding sequence that contains an edge with trivial
stabilizer. The fold fk is either of type 2 or of type 3. It identifies an edge ek of Tk
with trivial stabilizer with some translate gek with g ∈ G (although the pair (ek , gek )
might not be the defining pair of the edge fold). We can assume hgi to be maximal in
the following sense: if g is of the form hl with h ∈ G and l > 1, then ek is not identified
with hek . We claim that Tk+1 = T .
We postpone the proof of the claim to the next paragraph, and first explain how to
derive the lemma from this claim. Let vk be the vertex of ek such that gvk = vk , and
vk′ be the other vertex of ek . Notice that Gvk′ is nontrivial: otherwise, all edges in Tk
adjacent to vk′ would have trivial stabilizer, and there would be at least three distinct
G-orbits of such edges. Since the fold fk involves at most two orbits of edges, there
would be an edge with trivial stabilizer in Tk+1 , contradicting the definition of Tk+1 . If
fk is a fold of type 3, defined by the pair (ek , gek ), then the vertex fk (vk′ ) satisfies the
conclusion of the lemma. Indeed, we have Gfk (vk′ ) = Gvk′ ∗ hgi. If fk is a fold of type 2,
then fk identifies ek with an edge e′k with nontrivial stabilizer, because otherwise Tk+1
would contain an edge with trivial stabilizer. Denote by vk′′ the vertex of e′k distinct
from vk . If vk′ and vk′′ do not belong to the same G-orbit, then we have a nontrivial
splitting Gfk (vk′ ) = Gvk′ ∗ Gvk′′ . If vk′ and vk′′ belong to the same G-orbit, then we have a
nontrivial splitting Gfk (vk′ ) = Gvk′′ ∗ Z. In both cases, this splitting is relative to incident
edge groups and to F|Gv , because all trees along the folding sequence are (G, F)-trees.
We now prove the above claim that Tk+1 = T . Assume towards a contradiction that
Tk+1 6= T . It follows from our choice of folding path, and the fact that T has cyclic edge
stabilizers, that all possible folds in Tk+1 identify two edges e1 and e2 at fk (vk′ ) in distinct
G-orbits, and e1 and e2 have the same nontrivial stabilizer H in Tk+1 . Let ee1 (resp. ee2 )
be an edge in Tk in the fk -preimage of e1 (resp. e2 ). The group H fixes an extremity of
both ee1 and ee2 . If ee1 and ee2 were disjoint, then H would fix the segment between them,
a contradiction (Lemma 5.12). Therefore, the edges ee1 and ee2 are adjacent in Tk . By
our choice of folding path, at least one of them, say ee1 , has trivial stabilizer (otherwise
we could perform a fold of type 1 in Tk identifying ee1 and ee2 ), and hence belongs to the
G-orbit of ek . Since it is possible to fold ee1 and ee2 in Tk , the fold fk is of type 2. Hence
fk identifies ee1 with an edge e3 whose stabilizer is equal to H. Then e3 is adjacent to
e2 , and is identified with e2 in T , so we could have performed a fold of type 1 in Tk , a
contradiction.
Proof of Proposition 5.10. Let S be the skeleton of the dynamical decomposition of T .
If S is reduced to a point, then T is dual to a minimal measured foliation on a compact
2-orbifold σ. Some boundary component c of σ represents a nonperipheral conjugacy
class. Indeed, all boundary components of a compact 2-orbifold cannot be elliptic in a
common free splitting of the fundamental group of the orbifold. Then c is an unused
element in T .
Now assume that S is a nontrivial minimal (G, F)-tree. If S contains an edge with
trivial stabilizer, then this edge defines a (G, F)-free splitting, and T splits as a graph of
actions over this splitting by [10, Lemma 4.7]. Otherwise, let v be a vertex of S provided
by Lemma 5.11.
We note that the vertex v cannot be of arc type, because a vertex of arc type has
a stabilizer equal to the stabilizer of the incident edges. If v is of surface type, it is
35
associated to a compact 2-orbifold σ, equipped with a minimal measured foliation. The
fundamental group of σ splits as a free product relatively to incident edge groups and
to subgroups in F|π1 (σ) . This implies the existence of an unused element in T , otherwise
π1 (σ) would split as a free product in which all boundary components of σ are elliptic.
If v is of trivial type, then we get a one-edge (G, F)-free splitting S0 that is compatible
with S, by splitting the vertex group Gv . Associated to each vertex of S0 with vertex
group Gv′ is a geometric (possibly trivial) Gv′ -action Tv′ . The tree T splits as a graph
of actions over S0 , with the trees Tv′ as vertex actions.
5.4
Approximating very small geometric (G, F )-trees by Grushko (G, F )trees
Proof of Theorems 5.1 and 5.3. Let T be a very small, minimal (G, F)-tree. We want
to show that we can approximate T by a sequence of minimal Grushko (G, F)-trees, and
that the approximation can be chosen to be a Lipschitz approximation, by trees whose
quotient volumes converge to the quotient volume of T , if T has trivial arc stabilizers.
We will argue by induction on rkK (G, F). The claim holds true when rkK (G, F) = 1,
so we assume that rkK (G, F) ≥ 2. The claim also holds true if T is reduced to a point.
Thanks to Theorem 3.3, we can assume T to be geometric. (For the statement about
quotient volumes, see the argument at the beginning of Section 5.1). By Proposition 5.7,
the tree T decomposes as a graph of actions whose vertex actions are either arcs, or of
surface or exotic type. Proposition 5.7 also enables us to approximate all exotic vertex
actions. Using Lemmas 5.4 and 5.5, we can therefore reduce to the case where T is a tree
of surface type (notice that all edges of the decomposition as a graph of actions whose
stabilizer is noncyclic, or nontrivial and peripheral, have length 0).
First assume that there exists an unused element c in T , corresponding to a boundary
curve in a minimal orbifold σ of the dynamical decomposition. One can then narrow the
band complex by width δ > 0 from c to get a Lipschitz approximation of T : this is done
by cutting a segment on Σ of length δ (arbitrarily small) transverse to the boundary curve
c and to the foliation. In this way, all leaves of the foliations become segments (halfleaves of the original foliation on σ are dense), so the tree dual to the foliated complex
by which the minimal foliation on σ has been replaced is simplicial. By choosing δ > 0
arbitrarily close to 0, we can ensure the volume of this tree to be arbitrarily small. In
the new band complex obtained in this way, the orbifold σ has therefore been replaced
by a simplicial component, with trivial edge stabilizers.
We thus reduce to the case where no element of G is unused in T . Lemma 5.10 thus
ensures that T splits as a (G, F)-graph of actions over a one-edge (G, F)-free splitting,
and we can conclude by induction, using Lemmas 5.5 and 5.6.
6
Tame (G, F )-trees
We finish this paper by introducing another class of (G, F)-trees, larger than the class
of very small (G, F)-trees, which we call tame (G, F)-trees. This class will turn out to
provide the right setting for carrying out our arguments in [13] to describe the Gromov
boundary of the graph of cyclic splittings of (G, F).
Definition 6.1. A minimal (G, F)-tree is tame if it is small, and has finitely many
orbits of directions at branch points.
36
There exist small (G, F)-actions that are not tame. A typical example is the following: a sequence of splittings of F2 = ha, bi of the form F2 = (hai ∗ha2 i ha2 i ∗ha4 i · · · ∗ha2n i
k
n
ha2 i) ∗ hbi, in which the edge with stabilizer generated by a2 has length 21k , converges
to a small F2 -tree with infinitely many orbits of branch points.
By the discussion in Section 4.3, the class of tame (G, F)-trees is the right class of
trees in which Levitt’s decomposition makes sense.
Theorem 6.2. (Levitt [18, Theorem 1]) Let G be a countable group, and let F be a free
factor system of G. Then every tame (G, F)-tree splits uniquely as a graph of actions,
all of whose vertex trees have dense orbits for the action of their stabilizer, such that the
Bass–Serre tree of the underlying graph of groups is small, and all its edges have positive
length.
However, the above example of a small (G, F)-tree that is not tame shows that the
space of tame (G, F)-trees is not closed. We will describe conditions under which a limit
of tame trees is tame. It will be of interest to introduce yet another class of (G, F)-trees.
The equivalences in the definition below are straightforward.
Definition 6.3. Let k ∈ N∗ . A small minimal (G, F)-tree is k-tame if one of the
following equivalent conditions occurs.
• For all nonperipheral g ∈ G and all arcs I ⊆ T , if hgi ∩ Stab(I) is nontrivial, then
its index in hgi divides k.
• For all nonperipheral g ∈ G, and all l ≥ 1, we have Fix(g l ) ⊆ Fix(gk ).
• For all nonperipheral g ∈ G, and all l ≥ 1, we have Fix(g kl ) = Fix(gk ).
Notice in particular that 1-tame (G, F)-trees are those (G, F)-trees in which all arc
stabilizers are either trivial, or maximally-cyclic and nonperipheral.
Proposition 6.4. For all k ∈ N∗ , the space of k-tame (G, F)-trees is closed in the space
of small minimal (G, F)-trees.
Proof. Let T be a small, minimal (G, F)-tree, and let (Tn )n∈N be a sequence of k-tame
(G, F)-trees that converges to T . Let g ∈ G, and assume that there exists l ≥ 1 such
that gl fixes a nondegenerate arc [a, b] ⊆ T . If g is hyperbolic in Tn for infinitely many
n ∈ N, then Lemma 2.3 implies that g fixes [a, b]. We can therefore assume that for all
n ∈ N, the fixed point set In of gl is nonempty. Let an (resp. bn ) be an approximation of
a (resp. b) in Tn . Since dTn (gl an , an ) and dTn (gl bn , bn ) both converge to 0, the distances
dTn (an , In ) and dTn (bn , In ) both converge to 0. As Tn is k-tame, this implies that the
distances of both an and bn to the fixed point set of g k in Tn converge to 0. So both
dTn (gk an , an ) and dTn (gk bn , bn ) converge to 0, and therefore gk fixes [a, b] in T . This
shows that T is k-tame.
Proposition 6.5. A minimal (G, F)-tree is tame if and only if there exists k ∈ N∗ such
that T is k-tame.
Proof. Let T be a tame minimal (G, F)-tree. By Theorem 6.2, the tree T splits as a graph
of actions, all of whose vertex actions have dense orbits for the action of their stabilizer.
As tame (G, F)-trees with dense orbits have trivial arc stabilizers by Lemma 4.17, we do
not modify the collection of arc stabilizers of T if we collapse all vertex trees to points.
We can therefore reduce to the case where T is simplicial. In this case, minimality
37
implies that the G-action on T has finitely many orbits of edges, from which it follows
that T is k-tame for some k ∈ N. The converse statement will follow from the following
proposition.
Proposition 6.6. For all k ∈ N∗ , there exists γ(k) ∈ N such that any k-tame minimal
(G, F)-tree has at most γ(k) orbits of directions at branch points.
Proof. Let T be a k-tame minimal (G, F)-tree. We first assume that T is geometric. Let
G be the dynamical decomposition of T . We recall that all vertex trees of G are either
arcs or have dense orbits. As T is geometric, there are finitely many orbits of directions
in T (see Corollary 4.8). This implies that arc stabilizers are trivial in the vertex trees of
G with dense orbits (Proposition 4.17). If x is a branch or inversion point of T contained
in one of the vertex trees with dense orbits Tv of G, then all directions at x contained
in Tv have a positive contribution to the index i(T ). It follows from Proposition 4.4
that there is a bound on the number of such orbits of directions. Therefore, we can
collapse all the vertex trees with dense orbits to points, and reduce to the case where T
is simplicial.
In this case, we argue by induction on rkK (G, F), and show that there is a bound
γ(k, l) on the number of orbits of directions in any k-tame minimal simplicial (G, F)-tree
with rkK (G, F) ≤ l. By splitting one of the vertex stabilizers of the splitting relatively
to incident edge stabilizers if needed (which is made possible by Lemma 5.11, and can
only increase the number of orbits of directions), we can assume that T contains an edge
e with trivial stabilizer. By removing from T the interior of the edges in the orbit of e
in T , we get one or two orbits of trees, whose stabilizers have a strictly smaller Kurosh
rank. Let T ′ be one of the trees obtained in this way, whose stabilizer we denote by G′ .
Then T ′ is k-tame. If T ′ is minimal, then we are done by induction. However, the tree T ′
may fail to be minimal. The quotient graph T ′ /G′ consists of a minimal graph of groups
′
′
Tmin
/G′ , with a segment I = e1 ∪ · · · ∪ en attached to Tmin
/G′ at one of its extremities
(where we denote by ei the edges in I). All edge groups are cyclic, and they satisfy
Gei = Go(ei ) ⊆ Gei+1 for all i ∈ {1, . . . , n − 1}. Since T is k-tame, we have n ≤ k. By
′
induction, there are at most γ(k, l − 1) orbits of directions in Tmin
, so we get a uniform
bound on the possible number of orbits of directions in T .
We have thus shown that there is a uniform bound γ(k) on the possible number
of orbits of directions in a minimal k-tame geometric (G, F)-tree. Let now T be a
minimal nongeometric (G, F)-tree, and assume that T has strictly more than γ(k) orbits
of directions. Arguing as in the proof of Proposition 4.4 in the nongeometric case, we can
find a geometric (G, F)-tree T ′ in which we can lift at least γ(k) + 1 orbits of directions.
Lemma 3.7 shows that the approximation T ′ can be chosen to be k-tame. This is a
contradiction.
We finally establish one more condition under which a limit of tame trees is tame,
which will be used in [13].
Proposition 6.7. Let (Tn )n∈N be a sequence of simplicial metric small (G, F)-trees that
converges to a minimal (G, F)-tree T . If all trees Tn contain a single orbit of edges, then
T is tame.
We will make use of the following lemma.
Lemma 6.8. Let T be a simplicial small (G, F)-tree, with one orbit of edges. Then the
fixed point set of any element of G is a star of diameter at most 2 for the simplicial
metric on T .
38
Proof. Let e1 and e2 be two edges of T stabilized by a common element g ∈ G. As T
has only one orbit of edges, there exists h ∈ G such that he1 = e2 . By choosing an
orientation on e1 , this relation defines an orientation on e2 . Then hge1 = ghe1 , which
implies that h commutes with g because T is small. Hence h is elliptic in T . This implies
that e1 and e2 point in opposite directions in T . As this is true of any pair of edges
stabilized by g, the fixed point set of g has the desired description.
Proof of Proposition 6.7. Up to possibly passing to a subsequence, one of the following
situations occurs.
Case 1 : The length of the unique orbit of edges in Tn converges to 0.
In this case, we will show that T is very small. This implies that T is 1-tame, and
hence tame by Proposition 6.5. We have seen in the proof of Proposition 2.1 that limits
of small (G, F)-trees are small, and limits of trees with trivial tripod stabilizers have
trivial tripod stabilizers. Let g ∈ G be a nonperipheral element, and assume that there
exists l ≥ 1 such that gl fixes a nondegenerate arc [a, b] ⊆ T . Let an (resp. bn ) be
an approximation of a (resp. b) in Tn . If g were elliptic in Tn , then both an and bn
would be arbitrarily close to the fixed point set Xn of gl in Tn , as n goes to +∞. It
follows from Lemma 6.8 that the diameter of Xn in Tn converges to 0 as n tends to +∞.
This contradicts the fact that dTn (an , bn ) is bounded below (because a 6= b). Therefore,
for n large enough, the element g is hyperbolic in Tn . The distances dTn (an , g l an ) and
dTn (bn , gl bn ) converge to 0, so the points an and bn are arbitrarily close to the axis of g
in Tn , and ||g||Tn converges to 0. This implies that dTn (an , gan ) and dTn (bn , gbn ) both
tend to 0, so g fixes [a, b].
Case 2 : There is a positive lower bound on the length of the unique orbit of edges
in Tn .
Up to passing to a subsequence and rescaling Tn by a factor λn > 0 converging to
some λ > 0, we can assume that all trees Tn have edge lengths equal to 1. This implies
that all translation lengths in Tn belong to Z, so all translation lengths in T belong to Z.
It follows that T is a simplicial metric tree (see [20, Theorem 10]), so it has finitely many
orbits of directions. Since a limit of small (G, F)-trees is small, the tree T is tame.
References
[1] M. Bestvina and M. Feighn, Outer limits, preprint,
http://andromeda.rutgers.edu/ feighn/papers/outer.pdf, 1994.
available
at
[2] M.M. Cohen and M. Lustig, Very small group actions on R-trees and Dehn twist
automorphisms, Topology 34 (1995), no. 3, 575–617.
[3] M. Culler and J.W. Morgan, Group actions on R-trees, Proc. London Math. Soc.
55 (1987), no. 3, 571–604.
[4] M. Culler and K. Vogtmann, Moduli of graphs and automorphisms of free groups,
Invent. math. 84 (1986), no. 1, 91–119.
[5] D. Gaboriau, Générateurs indépendants pour les systèmes d’isométries de dimension
un, Ann. Inst. Fourier 47 (1997), no. 1, 101–122.
39
[6] D. Gaboriau and G. Levitt, The rank of actions on R-trees, Ann. Scient. Éc. Norm.
Sup. 28 (1995), no. 5, 549–570.
[7] D. Gaboriau, G. Levitt, and F. Paulin, Pseudogroups of isometries of R and Rips’
theorem on free actions on R-trees, Israel J. Math. 87 (1994), no. 3, 403–428.
[8] H. Gillet and P.B. Shalen, Dendrology of groups in low Q-ranks, J. Differential
Geom. 32 (1990), no. 3, 605–712.
[9] V. Guirardel, Approximations of stable actions on R-trees, Comment. Math. Helv.
73 (1998), no. 1, 89–121.
[10]
, Limit groups and groups acting freely on Rn -trees, Geom. Topol. 8 (2004),
1427–1470.
[11]
, Actions of finitely generated groups on R-trees, Ann. Inst. Fourier 58 (2008),
no. 1, 159–211.
[12] V. Guirardel and G. Levitt, The outer space of a free product, Proc. London Math.
Soc. 94 (2007), no. 3, 695–714.
[13] C. Horbez, Hyperbolic graphs for free products, and the Gromov boundary of the
graph of cyclic splittings, arXiv:1408.0544v2 (2014).
[14]
, The Tits alternative for the automorphism group of a free product,
arXiv:1408.0546v2 (2014).
[15]
, Spectral rigidity for primitive elements of FN , J. Group Theory 19 (2016),
no. 1, 55–123.
[16] H. Imanishi, On codimension one foliations defined by closed one forms with singularities, J. Math. Kyoto Univ. 19 (1979), no. 2, 285–291.
[17] A.G. Kurosh, Die Untergruppen der freien Produkte von beliebigen Gruppen, Math.
Ann. 109 (1934), no. 1, 647–660.
[18] G. Levitt, Graphs of actions on R-trees, Comment. Math. Helv. 69 (1994), no. 1,
28–38.
[19] G. Levitt and F. Paulin, Geometric group actions on trees, Amer. J. Math. 119
(1997), no. 1, 83–102.
[20] J.W. Morgan, Λ-trees and their applications, Bull. Amer. Math. Soc. 26 (1992),
no. 1, 87–112.
[21] F. Paulin, Topologie de Gromov équivariante, structures hyperboliques et arbres
réels, Invent. math. 94 (1988), no. 1, 53–80.
[22]
, The Gromov topology on R-trees, Topol. Appl. 32 (1989), no. 3, 197–221.
[23] J.R. Stallings, Topology of Finite Graphs, Invent. math. 71 (1983), no. 3, 551–565.
40
| 4math.GR
|
SETS OF LENGTHS
arXiv:1509.07462v2 [math.GR] 10 Aug 2016
ALFRED GEROLDINGER
Abstract. Oftentimes the elements of a ring or semigroup H can be written as finite products of
irreducible elements, say a = u1 · . . . · uk = v1 · . . . · vℓ , where the number of irreducible factors is
distinct. The set L(a) ⊂ N of all possible factorization lengths of a is called the set of lengths of a,
and the full system L(H) = {L(a) | a ∈ H} is a well-studied means of describing the non-uniqueness
of factorizations of H. We provide a friendly introduction, which is largely self-contained, to what is
known about systems of sets of lengths for rings of integers of algebraic number fields and for transfer
Krull monoids of finite type as their generalization.
1. Introduction
We all know that every positive integer can be written as a finite product of irreducibles (primes) and
that such a factorization is unique up to the order of appearance. Similar to factorizations in the positive
integers, in many rings and semigroups, elements can also be written as finite products of irreducible
elements, but unlike the case of the integers, such factorizations need not always be unique. It is the
main objective of factorization theory to describe the various aspects of non-uniqueness and to classify
them in terms of invariants of the underlying algebraic structure. Before it was extended to commutative
ring and semigroup theory, factorization theory had its origin in algebraic number theory, and only in
recent years has been extended to non-commutative settings [4, 50]. For further background, we refer the
reader to several monographs and conference proceedings [1, 23, 28, 16, 11]. It is no surprise that this
development has been chronicled over the years by a series of Monthly articles (from [38] to [6, 5]). While
the focus of this interest has been on commutative domains and their semigroups of ideals, such studies
range from abstract semigroup theory to the factorization theory of motion polynomials with application
in mechanism science [37].
Sets of lengths are the most investigated invariants in factorization theory. To fix notation, if an
element a in a semigroup can be written as a product of irreducible elements, say a = u1 · . . . · uk , then
k is called the length of the factorization, and the set L(a) ⊂ N of all possible factorization lengths of
a is called the set of lengths of a. Under a mild condition on the semigroup, sets of lengths are finite
nonempty subsets of N0 , and if there is an element a in the semigroup with |L(a)| > 1 (meaning that a
has factorizations of distinct lengths), then sets of lengths can get arbitrarily long (a precise statement
of this is in Lemmas 2.1 and 2.2).
The goal of this paper is to give a friendly introduction to factorization theory. Indeed, we take
the reader on a tour through sets of lengths which is highlighted by two structure theorems (Theorems
2.6 and 5.3), two open problems (Problem 4.7 and the Characterization Problem at the beginning of
Section 6), and a conjecture (Conjecture 6.7). In Section 2 we introduce sets of distances and unions
of sets of lengths. We provide a full and self-contained proof for the Structure Theorem for Unions of
2010 Mathematics Subject Classification. 11B30, 11R27, 13A05, 13F05, 16H10, 16U30, 20M13.
Key words and phrases. Krull monoids, maximal orders, zero-sum sequences; sets of lengths, unions of sets of lengths,
sets of distances, elasticities.
I would like to thank Daniel Smertnig, Salvatore Tringali, Qinghai Zhong, and all the anonymous referees for their
careful reading. Their comments helped me to eliminate a couple of flaws, to improve the presentation of the paper, and to
increase its readability. Furthermore, I would like to thank the Scott Chapman for his encouragement to write this survey
and all his patience, and Moshe Roitman for showing to me that transfer Krull monoids need not be Mori.
This work was supported by the Austrian Science Fund FWF, Project Number P 28864-N35.
1
2
ALFRED GEROLDINGER
Sets of Lengths (Theorem 2.6), and outline an argument that finitely generated commutative monoids
satisfy all assumptions of that Structure Theorem. In Section 3 we discuss commutative Krull monoids.
This class includes Dedekind domains and hence rings of integers of algebraic number fields, and we
will provide an extended list of examples stemming from a variety of mathematical areas. The central
strategy for studying sets of lengths in a given class of semigroups is to construct homomorphisms
(called transfer homomorphisms) which can be used to transfer analogous results in a simpler class of
semigroups directly back to the more complex class. In Section 4 we discuss transfer homomorphisms,
show that they preserve sets of lengths, and provide a self-contained proof for the fact that there is a
transfer homomorphism from commutative Krull monoids to monoids of zero-sum sequences (which are
monoids having a combinatorial flavor that will often be the simpler class of semigroups to which the
more complex semigroup is reduced). We provide an extended list of transfer Krull monoids (these are
the monoids allowing a transfer homomorphism to a monoid of zero-sum sequences), and then we restrict
our discussion to this class of monoids. In Section 5 we discuss the Structure Theorem for Sets of Lengths
and provide examples showing that all aspects addressed in the Structure Theorem occur naturally. In
Section 6 we discuss sets of lengths of the monoid of zero-sum sequences over a finite abelian group (the
transfer machinery of Section 4 guarantees that these sets of lengths coincide with the sets of lengths of a
ring of integers). They can be studied with methods from Additive Combinatorics and their structure is
by far the best understood (among all classes of monoids). In this setting, unions of sets of lengths and
the set of distances are intervals and they have natural upper bounds (Proposition 6.1). In spite of the
fact that almost all (in a certain sense) sets of lengths are intervals (Theorem 6.3) we conjecture that the
system of sets of lengths is a characteristic for the group (Conjecture 6.7). In order to keep this article
as self-contained as possible, we do not mention arithmetical concepts beyond sets of lengths (such as
catenary and tame degrees), or factorization theory in rings with zero-divisors, or divisibility theory in
non-atomic rings.
2. Basic Notation and Unions of Sets of Lengths
We denote by N the set of positive integers and set N0 = N ∪ {0}. For integers a, b ∈ Z, we denote
by [a, b] = {x ∈ Z | a ≤ x ≤ b} the (discrete) interval between a and b, and by an interval we always
mean a set of this form. Let L, L′ ⊂ Z be subsets of the integers. Then L + L′ = {a + b | a ∈ L, b ∈ L′ }
is the sumset of L and L′ . Thus we have L + ∅ = ∅, and we set −L = {−a | a ∈ L}. For an integer
m ∈ Z, m + L = {m} + L is the shift of L by m. For k ∈ N, we denote by kL = L + . . . + L the k-fold
sumset of L and by k · L = {ka | a ∈ L} the dilation of L by k. A positive integer d ∈ N is called a
distance of L if there are k, ℓ ∈ L with ℓ − k = d and the interval [k, ℓ] contains no further elements
of L. We denote by ∆(L) ⊂ N the set of distances of L. By definition, we have ∆(L) = ∅ if and only
if |L| ≤ 1, and L is an arithmetical progression if and only if |∆(L)| ≤ 1 . For L ⊂ N, we denote by
ρ(L) = sup L/ min L ∈ Q≥1 ∪ {∞} the elasticity of L and we set ρ({0}) = 1.
By a semigroup, we always mean an associative semigroup, and if not stated otherwise, we use multiplicative notation. Let S be a semigroup. We say that S is cancelative if for all elements a, b, c ∈ S,
the equation ab = ac implies b = c and the equation ba = ca implies b = c. All rings and semigroups
are supposed to have an identity, and all ring and semigroup homomorphisms preserve the identity. By
a monoid, we mean a cancelative semigroup. Clearly, subsemigroups of groups are monoids, and finite
monoids are groups. If aS ∩ bS 6= ∅ and Sa ∩ Sb 6= ∅ for all a, b ∈ S, then S has a (unique left and right)
quotient group which will be denoted by q(S). If R is a ring, then the set of cancelative elements R• is
a monoid. A domain D is a ring in which zero is the only zero-divisor (i.e., D• = D \ {0}). We use the
abbreviation ACC for the ascending chain condition on ideals.
Let P be a set. We denote by F ∗ (P ) the free monoid with basis P , the elements of which may be
viewed as words on the alphabet P . We denote by | · | : F ∗ (P ) → N0 the function which maps each word
onto its length. The free abelian monoid with basis P will be denoted by F (P ). Every a ∈ F (P ) has a
SETS OF LENGTHS
3
unique representation of the form
Y
a=
pνp , where νp ∈ N0 and νp = 0 for almost all p ∈ P .
p∈P
Therefore, for every p ∈ P , there is a homomorphism (called the p-adic exponent) vp : F (P ) → N0 defined
by vp (a) = νp . Similar to the
Pcase of free monoids, we denote by |·| : F (P ) → N0 the usual length function,
and we observe that |a| = p∈P vp (a) for all a ∈ F (P ).
Let H be a monoid and let a, b ∈ H. The element a is said to be invertible if there exists an element
a′ ∈ H such that aa′ = a′ a = 1. The set of invertible elements of H will be denoted by H × , and we say
that H is reduced if H × = {1}.
The element a ∈ H is called irreducible (or an atom) if a ∈
/ H × and, for all u, v ∈ H, a = uv implies
×
×
that u ∈ H or v ∈ H . The monoid H is said to be atomic if every a ∈ H \ H × is a product of finitely
many atoms of H. While most integral domains introduced in elementary courses are atomic, not all
such algebraic objects are. An elementary example of a non-atomic monoid can be found in [10, p. 166].
If a ∈ H and a = u1 · . . . · uk , where k ∈ N and u1 , . . . , uk ∈ A(H), then we say that k is the length of the
factorization. For a ∈ H \ H × , we call
LH (a) = L(a) = {k ∈ N | a has a factorization of length k} ⊂ N
the set of lengths of a. For convenience, we set L(a) = {0} for all a ∈ H × . By definition, H is atomic if
and only if L(a) 6= ∅ for all a ∈ H. Furthermore, it is clear that the following conditions are equivalent:
(1) L(a) = {1}; (2) a ∈ A(H); (3) 1 ∈ L(a). If a, b ∈ H, then L(a) + L(b) ⊂ L(ab). If H is commutative,
then Hred = H/H × = {aH × | a ∈ H} is the associated reduced monoid, and H is called factorial if Hred
is free abelian. We say that H is a BF-monoid (or a bounded factorization monoid) if L(a) is finite and
nonempty for all a ∈ H. We call
L(H) = {L(a) | a ∈ H}
the system of sets of lengths of H. So if H is a BF-monoid, then L(H) is a set of finite nonempty subsets
of the non-negative integers. We say that H is half-factorial if |L| = 1 for every L ∈ L(H). Clearly,
factorial monoids are half-factorial, and in 1960 Carlitz showed that the ring of integers of an algebraic
number field is half-factorial if and only if the class group has at most two elements (see Propositions 3.3
and 4.3). Since then, half-factoriality has been a central topic in factorization theory (see, for example,
[13, 25, 41]). We focus in this paper on the structure of sets of lengths in non-half-factorial BF-monoids.
We start with two simple lemmas.
Lemma 2.1. Let H be a monoid.
1. If H satisfies the ACC on principal left ideals and the ACC on principal right ideals, then H is
atomic.
2. Suppose that H is atomic. Then H is either half-factorial or for every N ∈ N there is an element
aN ∈ H such that |L(aN )| > N .
Proof. For ease of discussion, suppose that aH = Ha for all a ∈ H (see [49, Proposition 3.1] for details
in the general case). Assume to the contrary that the set Ω of all nonunits a ∈ H that are not products
of atoms is nonempty. If a ∈ Ω, then a = bc with nonunits b, c ∈ H, and either b ∈ Ω or c ∈ Ω. Thus for
any a ∈ H, there is some a′ ∈ Ω with aH ( a′ H. Starting from an arbitrary a ∈ H, this gives rise to a
properly ascending chain of principal ideals, a contradiction.
To verify the second statement, suppose that H is atomic but not half-factorial. Then there exist
an element a ∈ H, integers k, ℓ ∈ N with k < ℓ, and atoms u1 , . . . , uk , v1 . . . , vℓ ∈ A(H) such that
a = u1 · . . . · uk = v1 · . . . · vℓ . Then for every N ∈ N we have
aN = (u1 · . . . · uk )ν (v1 · . . . · vℓ )N −ν
and hence {ℓN − ν(ℓ − k) | ν ∈ [0, N ]} ⊂ L(aN ).
for all ν ∈ [0, N ] ,
4
ALFRED GEROLDINGER
Let H be a monoid. A function λ : H → N0 is called a right length function (resp. a left length
function) if λ(a) < λ(b) for all a ∈ H and all b ∈ aH \ aH × (resp. all b ∈ Ha \ H × a).
Lemma 2.2. Let H be a monoid.
1. H is a BF-monoid if and only if there is a left length function (or a right length function) λ : H →
N0 .
2. If H is a BF-monoid, then H satisfies the ACC on principal left ideals and on principal right
ideals.
3. Submonoids of free monoids and of free abelian monoids are BF-monoids.
Proof. 1. If H is a BF-monoid, then the function λ : H → N0 , defined by a 7→ max L(a) for every a ∈ H,
is a right length function and a left length function. Now, if there is a right length function λ : H → N0 ,
then we have to show that H is a BF-monoid (the proof is completely analogous for left length functions).
First we observe that, if b ∈ H × and c ∈ H \ H × = bH \ bH × , then λ(c) > λ(b) ≥ 0. We claim that every
a ∈ H \ H × can be written as a product of finitely many atoms, and that sup L(a) ≤ λ(a). If a ∈ A(H),
then L(a) = {1}, and the assertion holds. Suppose that a ∈ H is neither an atom nor a unit. Then a has
a product decomposition of the form
a = u1 · . . . · uk
where k ≥ 2 and u1 , . . . , uk ∈ H \ H × .
For i ∈ [0, k], we set ai = u1 · . . . · ui , with a0 = 1, and hence ai+1 ∈ ai H \ ai H × for all i ∈ [0, k − 1]. This
implies that λ(a) = λ(ak ) > λ(ak−1 ) > . . . > λ(a1 ) > 0 and thus λ(a) ≥ k. Therefore there is a maximal
k ∈ N such that a = u1 · . . . · uk where u1 , . . . , uk ∈ H \ H × , and this implies that u1 , . . . , uk ∈ A(H) and
k = max L(a) ≤ λ(a).
2. Suppose that H is a BF-monoid. Let λ : H → N0 be a right length function and assume to the
contrary that there is a properly ascending chain of principal right ideals a0 H ( a1 H ( a2 H ( . . ..
Then λ(a0 ) > λ(a1 ) > λ(a2 ) > . . ., a contradiction. Similarly, we can show that H satisfies the ACC on
principal left ideals.
3. Use 1., and note that the restriction of a length function is a length function.
Next we introduce a main parameter describing the structure of sets of lengths, namely
[
∆(L)
∆(H) =
L∈L(H)
which is the set of distances of H (also called the delta set of H). We open by showing that ∆(H) satisfies
a fundamental property.
Proposition 2.3. Let H be an atomic monoid with ∆(H) 6= ∅. Then min ∆(H) = gcd ∆(H).
Proof. We set d = gcd ∆(H). Clearly, it suffices to show that d ∈ ∆(H). There are t ∈ N, d1 , . . . , dt ∈
∆(H) and m1 , . . . , mt ∈ Z \ {0} such that d = m1 d1 + . . . + mt dt . After renumbering if necessary,
there is some s ∈ [1, t] such that m1 , . . . , ms , −ms+1 , . . . , −mt are positive. For every ν ∈ [1, t], there
are xν ∈ N and aν ∈ H such that {xν , xν + dν } ⊂ L(aν ) for every ν ∈ [1, s] and {xν − dν , xν } ⊂
L(aν ) for every ν ∈ [s + 1, t]. This implies that
ν
{mν xν , mν xν + mν dν } ⊂ L(am
ν ) for every ν ∈ [1, s]
{−mν xν + mν dν , −mν xν } ⊂
ν
)
L(a−m
ν
and
for every ν ∈ [s + 1, t] .
−m
s+1
ms
t
1
· . . . · a−m
We set a = am
, and observe that L(a) ⊃
t
1 · . . . · as as+1
s
t
s
t
n
o
X
X
X
X
k :=
mν xν −
mν xν , ℓ :=
mν (xν + dν ) −
mν (xν − dν ) .
ν=1
ν=s+1
ν=1
Since ℓ − k = d ≤ min ∆(H), it follows that d ∈ ∆(L(a)) ⊂ ∆(H).
ν=s+1
SETS OF LENGTHS
5
We now introduce unions of sets of lengths. Let H be an atomic monoid, and k, ℓ ∈ N. In the extremal
case where H = H × it is convenient to set Uk (H) = {k}. Now suppose that H 6= H × . We define Uk (H)
to be the set of all ℓ ∈ N such that we have an equation of the form
u1 · . . . · uk = v1 · . . . · vℓ
where u1 , . . . , uk , v1 , . . . , vℓ ∈ A(H) .
In other words, Uk (H) is the union of sets of lengths containing k. Clearly, we have k ∈ Uk (H), and
1 ∈ Uk (H) if and only if k = 1, and U1 (H) = {1}. Furthermore, we have k ∈ Uℓ (H) if and only if
ℓ ∈ Uk (H). Now we define
ρk (H) = sup Uk (H) and λk (H) = min Uk (H) ,
and we call ρk (H) the kth elasticity of H. Since Uk (H) + Uℓ (H) ⊂ Uk+ℓ (H), we infer that
(2.1)
λk+ℓ (H) ≤ λk (H) + λℓ (H) ≤ k + ℓ ≤ ρk (H) + ρℓ (H) ≤ ρk+ℓ (H) ,
and we will use these inequalities without further mention. The invariant
ρ(H) = sup{ρ(L) | L ∈ L(H)} ∈ R≥1 ∪ {∞}
is called the elasticity of H. If k ∈ N and L ∈ L(H) with min L ≤ k, then sup L ≤ ρ(H) min L ≤ kρ(H)
and hence ρk (H) ≤ kρ(H). Thus if the elasticity ρ(H) is finite, then H is a BF-monoid and ρk (H) < ∞
for all k ∈ N.
The kth elasticities ρk (H) were first studied for rings of integers of algebraic number fields and then
in the setting of abstract semigroups [26]. Unions of sets of lengths have been introduced by Chapman
and Smith [12] in the setting of commutative Dedekind domains, and the Structure Theorem for Unions
of Sets of Lengths (as given in Theorem 2.6) has first been proved in [19] (in a commutative setting).
Proposition 2.4. Let H be an atomic monoid with H 6= H × .
1. For every k ∈ N we have ρk (H) = sup{sup L | L ∈ L(H), min L ≤ k} ≥ sup{sup L | L ∈
L(H), k = min L}, and equality holds if ρk (H) < ∞.
2.
n ρ (H)
o
ρk (H)
k
and
ρ(H) = sup
k ∈ N = lim
k→∞
k
k
n λ (H)
o
λk (H)
1
k
= inf
.
k ∈ N = lim
k→∞
ρ(H)
k
k
3. Suppose that ρ(H) < ∞. Then the following statements are equivalent.
(a) There is an L ∈ L(H) such that ρ(L) = ρ(H).
(b) There is an N ∈ N such that kN ρ(H) = ρkN (H) for all k ∈ N.
(c) There is some k ∈ N such that kρ(H) = ρk (H).
If one of the above statements holds, then there is some M ∈ N such that ρk (H) − ρk−1 (H) ≤ M
for all k ≥ 2.
Proof. 1. Let k ∈ N. We define ρ′k (H) = sup{sup L | L ∈ L(H), min L ≤ k}, ρ′′k (H) = sup{sup L |
L ∈ L(H), k = min L}, and obtain by definition ρ′k (H) ≥ ρk (H) ≥ ρ′′k (H). Hence we must prove that
ρk (H) ≥ ρ′k (H), and if ρk (H) < ∞, then also ρ′′k (H) ≥ ρ′k (H). Let a ∈ H with min L(a) = ℓ ≤ k and
u ∈ A(H). Then k ∈ L(auk−ℓ ) implies ρk (H) ≥ sup L(auk−ℓ ) ≥ sup L(a) + k − ℓ ≥ sup L(a), and therefore
ρk (H) ≥ ρ′k (H). Assume now that ρk (H) < ∞, and let a be as above such that sup L(a) = ρ′k (H). Then
min L(auk−ℓ ) ≤ min L(a) + (k − ℓ) = k and therefore
ρ′k (H) ≥ sup L(auk−ℓ ) ≥ sup L(a) + (k − ℓ) ≥ ρ′k (H) .
Thus k = ℓ = min L(a) and ρ′′k (H) ≥ sup L(a) = ρ′k (H).
2. If there is a k ∈ N such that ρk (H) = ∞, then all three terms of the first equation are equal to
infinity. Suppose that ρk (H) < ∞ for all k ∈ N. Then the first equality follows from 1. To verify the
6
ALFRED GEROLDINGER
first limit assertion, let ρ′ < ρ(H) be arbitrary. We must prove that ρk (H) ≥ kρ′ for all sufficiently large
k. If ρ′ < ρ′′ < ρ(H), then there exists some q0 ∈ N such that
qρ′′ + 1
> ρ′
q+1
for all q ≥ q0 ,
and there exists some N ∈ N such that ρN (H) > N ρ′′ . If k > N q0 , then k = N q + r for some q ≥ q0
and r ∈ [1, N ], and ρk (H) ≥ qρN (H) + ρr (H) ≥ qρN (H) + r by Inequality (2.1). Since ρN (H) ≥ N , it
follows that
ρk (H)
qρN (H) + r
qρN (H) + N
qN ρ′′ + N
≥
≥
>
> ρ′ .
k
qN + r
qN + N
qN + N
n
o
Since ρ(H) = sup m
m, n ∈ L, {0} 6= L ∈ L(H) , we have
n
nm
1
= inf
ρ(H)
n
o
m, n ∈ L, {0} 6= L ∈ L(H)
with 1/ρ(H) = 0 if ρ(H) = ∞. The verification of the second limit assertion runs along the same lines
as the proof of the first one (just replace ρk (H) by λk (H) and reverse all inequality signs).
3. In order to show the implication (a) ⇒ (b), let a ∈ H with L = L(a) ∈ L(H) such that ρ(L) = ρ(H)
and set N = min L. Then kN ∈ L(ak ) for all k ∈ N, and thus
sup L(ak )
k sup L(a)
ρkN (H)
ρkN (H)
≥
≥
= ρ(L) = ρ(H) ≥
.
kN
kN
kN
kN
The implication (b) ⇒ (c) is obvious. Suppose that (c) holds and let k ∈ N such that kρ(H) = ρk (H).
Let L ∈ L(H) such that min L = k and ρk (H) = max L. Then kρ(H) = max L = kρ(L) and hence
ρ(L) = ρ(H).
Suppose that the equivalent statements hold and let N ∈ N such that mN ρ(H) = ρmN (H) for all
m ∈ N. Let k ∈ N be given and set k = iN + j with i ∈ N0 and j ∈ [1, N ]. Then we infer that
iN ρ(H) + j ≤ ρiN (H) + ρj (H) ≤ ρiN +j (H) and
(iN + j)ρ(H) − N (ρ(H) − 1) ≤ (iN + j)ρ(H) − j(ρ(H) − 1) ≤ ρiN +j (H) .
Thus kρ(H) − ρk (H) ≤ M ′ := N (ρ(H) − 1) for all k ∈ N. If k ≥ 2, then ρk (H) ≤ kρ(H), (k − 1)ρ(H) −
ρk−1 (H) ≤ M ′ , and hence ρk (H) − ρk−1 (H) ≤ M ′ + ρ(H).
Unions of sets of lengths do have – in a variety of settings – a structure which is only slightly more
general than that of arithmetical progressions. In order to show this, we introduce the concept of almost
arithmetical progressions.
Definition 2.5. Let d ∈ N and M ∈ N0 . A subset L ⊂ Z is called an almost arithmetical progression
(AAP for short) with difference d and bound M if
L = y + (L′ ∪ L∗ ∪ L′′ ) ⊂ y + dZ
where y ∈ Z and L∗ is a nonempty (finite or infinite) arithmetical progression with difference d such that
min L∗ = 0, L′ ⊂ [−M, −1], L′′ ⊂ sup L∗ + [1, M ] (with the convention that L′′ = ∅ if L∗ is infinite).
Clearly, every single finite set is an AAP with a particularly trivial choice of M and d. Suppose we
have an atomic monoid with nonempty set of distances. Then, by Lemma 2.1, sets of lengths become
arbitrarily large, whence the unions Uk (H) are growing as k is growing. The next theorem states (under
the given assumptions) that all unions Uk (H) are AAPs with the critical point that a single choice of M
works for all sufficiently large k. Once this M is chosen, it says that all unions have a ”middle” piece
that becomes larger and larger as k grows and that ”middle” part has a very rigid structure. The poorly
behaved ”end” pieces are bounded in size by a constant and in their distance from the middle.
SETS OF LENGTHS
7
Theorem 2.6 (Structure Theorem for Unions of Sets of Lengths). Let H be an atomic monoid
with finite nonempty set of distances ∆(H) and d = min ∆(H). Suppose that either, ρk (H) = ∞ for
some k ∈ N, or that there is an M ∈ N such that ρk (H) − ρk−1 (H) ≤ M for all k ≥ 2.
1. There exist constants k ∗ and M ∗ ∈ N such that for all k ≥ k ∗ , Uk (H) is an AAP with
difference d and bound M ∗ . Moreover, if ρk (H) < ∞ for all k ∈ N, then the assertion holds for
k ∗ = 1.
2. We have
1
1
|Uk (H)|
ρ(H) −
.
=
lim
k→∞
k
d
ρ(H)
Proof. 1. For every k ∈ N, we have λk (H), k ∈ Uk (H), and since every d′ ∈ ∆(H) is a multiple of d by
Proposition 2.3, it follows that Uk (H) ⊂ λk (H) + dN0 . Thus it remains to show that there exist constants
k ∗ , M ∗ ∈ N such that, for all k ≥ k ∗ ,
(∗)
Uk (H) ∩ [k, ρk (H) − M ∗ ]
and Uk (H) ∩ [λk (H) + M ∗ , k]
are arithmetical progressions with difference d (note, if ρk (H) = ∞, then [k, ρk (H) − M ∗ ] = N≥k ). We
break the proof into three steps.
1.(a) First, we show that the left set in (∗) is an arithmetical progression. Since d ∈ ∆(H), there is an
element a ∈ H and m ∈ N such that
{m, m + d} ⊂ L(a). Since ∆(H) is finite and min ∆(H) = gcd ∆(H)
by Proposition 2.3, ψ := ρ ∆(H) ∈ N. Then the ψ-fold sumset satisfies the containment
U ∗ := ψ{m, m + d} = {ψm, ψm + d, . . . , ψm + ψd} ⊂ L(aψ ) .
We set k0 = ψm and observe that U ∗ ⊂ Uk0 (H), say Uk0 (H) = U ′ ⊎ U ∗ ⊎ U ′′ , where U ′ = {x ∈ Uk0 (H) |
x < k0 } and U ′′ = {x ∈ Uk0 (H) | x > k0 + ψd}. Let k ∗ ∈ N with k ∗ ≥ 2k0 . If there is some ℓ ∈ N with
ρℓ (H) = ∞, then let ℓ0 denote the smallest such ℓ ∈ N, and we suppose further that k ∗ − k0 ≥ ℓ0 . Now
let k ≥ k ∗ be given. Then Uk0 (H) + Uk−k0 (H) =
U ′ + Uk−k0 (H) ∪ U ∗ + Uk−k0 (H) ∪ U ′′ + Uk−k0 (H) ⊂ Uk (H) .
Clearly, we have k ∈ U ∗ + Uk−k0 (H). Since max ∆ Uk−k0 (H) ≤ max ∆(H) and ∆ Uk−k0 (H) ⊂ dN,
it follows that U ∗ + Uk−k0 (H) is an arithmetical progression with difference d. If there is some ℓ ∈ N
such that ρℓ (H) = ∞, then ρk−k0 (H) = ρk (H) = ∞ and
U ∗ + Uk−k0 (H) ∩ N≥k = k + dN0 = Uk (H) ∩ N≥k .
Suppose that ρℓ (H) < ∞ for all ℓ ∈ N. Then
max Uk (H) − max U ∗ + Uk−k0 (H) = ρk (H) − max U ∗ − ρk−k0 (H) ≤ k0 M ,
and hence
U ∗ + Uk−k0 (H) ∩ [k, ρk (H) − k0 M ] = Uk (H) ∩ [k, ρk (H) − k0 M ]
is an arith. progression with difference d. Thus the assertion follows with M ∗ = k0 M .
1.(b). By 1.(a), there are k ∗ , M ∗ ∈ N such that for all k ≥ k ∗ , the set Uk (H) ∩ [k, ρk (H) − M ∗ ] is an
arithmetical progression with difference d. Without restriction we may suppose that M ∗ ≥ k ∗ .
Let k ≥ k ∗ and ℓ = λk (H). We show that Uk (H) ∩ [ℓ + M ∗ , k] is an arithmetical progression with
difference d. Let m ∈ [ℓ + M ∗, k] be such that k − m is a multiple of d. In order to show that m ∈ Uk (H),
we verify that k ∈ Um (H). Since
k ≤ ρℓ (H) and ℓ + M ∗ ≤ m ,
it follows that k ≤ ρℓ (H) ≤ ρm−M ∗ (H) and hence
k + M ∗ ≤ ρm−M ∗ (H) + M ∗ ≤ ρm−M ∗ (H) + ρM ∗ (H) ≤ ρm (H) .
Since k ∈ m + dN0 with k ≤ ρm (H) − M ∗ and Um (H) ∩ [m, ρm (H) − M ∗ ] is an arithmetical progression
with difference d, it follows that k ∈ Um (H).
8
ALFRED GEROLDINGER
1.(c). Suppose that ρk (H) < ∞ for all k ∈ N. In this step, we show that for all k ∈ N the sets Uk (H)
f. Suppose that the assertion holds with the constants k ∗
are AAPs with difference d and some bound M
∗
∗
and M ∈ N. Since for all k ∈ [1, k − 1],
Uk (H) = [λk (H), k − 1] ∩ Uk (H) ∪ {k} ∪ [k + 1, ρk (H)] ∩ Uk (H)
is an AAP with bound M ′ = max{k − λk (H), ρk (H) − k | k ∈ [1, k ∗ − 1]}, it follows that for all k ∈ N
f = max{M ∗ , M ′ }.
the sets Uk (H) are AAPs with difference d and bound M
2. If there is some k ∈ N such that ρk (H) = ∞, then both the left and the right hand side of the
asserted equation are infinite. Suppose that ρk (H) < ∞ for all k ∈ N. By 1. there are k ∗ ∈ N and
M ∗ ∈ dN such that, for all k ≥ k ∗ , Uk (H) ∩ [λk (H) + M ∗ , ρk (H) − M ∗ ] is an arithmetical progression
with difference d. Thus for all k ≥ k ∗ we obtain that
ρk (H) − M ∗ − λk (H) + M ∗ + d
|Uk (H)|
ρk (H) − λk (H) + d
≤
≤
.
dk
k
dk
Since, by Proposition 2.4.2, limk→∞
ρk (H)
k
= ρ(H)
and
limk→∞
λk (H)
k
=
1
ρ(H) ,
the assertion follows.
We end this section with a discussion of finitely presented monoids and of commutative finitely
generated monoids. Let H be a monoid. For every generating set P of H, there is an epimorphism
ψ : F ∗ (P ) → H and F ∗ (P )/ ker(ψ) ∼
= H, where
ker(ψ) = {(x, y) ∈ F ∗ (P ) × F ∗ (P ) | ψ(x) = ψ(y)}
is a congruence relation on F ∗ (P ). If there is a finite generating set P and a finite set of relations
R ⊂ ker(ψ) which generates ker(ψ) as a congruence relation, then H is said to be finitely presented (by P
and R). If R = {(x1 , y1 ), . . . , (xt , yt )}, then we write (as usual) H = hP | Ri = hP | x1 = y1 , . . . , xt = yt i.
Proposition 2.7. If H = hA(H) | Ri is a reduced atomic monoid with a finite set of relations R, then
the set of distances ∆(H) is finite.
Proof. We set A = A(H), ψ : F ∗ (A) → H, R = {(x1 , y1 ), . . . , (xt , yt )}, M = max{ |x1 | − |y1 | , . . . , |xt | −
|yt | }, and assert that ∆(H) ⊂ [1, M ]. Let a ∈ H. Then L(a) = {|x| | x ∈ F ∗ (A) with ψ(x) = a}. We
choose two words v, w ∈ F ∗ (A) with ψ(v) = ψ(w) = a. Since F ∗ (P )/ ker(ψ) ∼
= H and ker(ψ) is generated
by R (as a congruence), there is a sequence of words v = v0 , . . . , vn = w in ψ −1 (a) ⊂ F ∗ (A) where vν
arises from vν−1 by replacing xi by yi for some i ∈ [1,t] and all ν ∈ [1, n]. We set L = {|v0 |, . . . , |vn |}
and obtain ∆(L) ⊂ [1, M ]. Thus it follows that ∆ L(a) ⊂ [1, M ] and hence ∆(H) ⊂ [1, M ].
There are atomic finitely presented monoids H such that ρ(H) = ρk (H) = ∞ for all k ≥ 2. To
provide an example, consider the monoid H = ha, b | a2 = ba2 bi (note that H is an Adyan semigroup
and hence cancelative, see [4, Section 2] for details). Obviously, H is finitely presented and atomic with
A(H) = {a, b}, and ρ2 (H) = ∞ which implies that ρk (H) = ∞ for all k ≥ 2. Since Ha2 bi ( Ha2 bi+1
for all i ∈ N, H does not satisfy the ACC on principal left ideals. As another example, the monoid
H = ha, b | a = babi is a finitely presented monoid which is not atomic (note that a is not a finite product
of atoms). However, this behavior cannot occur in the case of commutative monoids. The next result
shows in particular that finitely generated commutative monoids satisfy all assumptions of Theorem 2.6,
and hence they satisfy the Structure Theorem for Unions of Sets of Lengths.
Proposition 2.8. Let H be a reduced commutative monoid.
1. For a subset A ⊂ A(H) the following statements are equivalent.
(a) H is atomic and A = A(H).
(b) A is the smallest generating set of H (with respect to set inclusion).
(c) A is a minimal generating set of H.
2. H is finitely generated if and only if H is atomic and A(H) is finite.
SETS OF LENGTHS
9
3. Suppose that H is finitely generated. Then H is a BF-monoid with finite set of distances and finite
elasticity. Moreover, there is an L ∈ L(H) such that ρ(L) = ρ(H), and there is an M ∈ N such
that ρk (H) − ρk−1 (H) ≤ M for all k ≥ 2.
Proof. 1.(a) ⇒ (b) Since H is atomic, A is a generating set of H and it remains to show that it is the
smallest one. Let A′ be any generating set of H. If u ∈ A, then u = v1 · . . . · vk with k ∈ N and
v1 , . . . , vk ∈ A′ \ {1}. Since u is an atom, it follows that k = 1 and u = v1 ∈ A′ . The implication (b)
⇒ (c) is obvious.
(c) ⇒ (a) It suffices to verify that A = A(H). Since A is a minimal generating set, it follows that 1 ∈
/ A.
If u ∈ A(H), then u = v1 ·. . . ·vk with k ∈ N and v1 , . . . , vk ∈ A. This implies k = 1, u = v1 ∈ A, and thus
A(H) ⊂ A. Assume to the contrary that there is some u ∈ A \ A(H), say u = vw with v, w ∈ H \ {1}.
Then v = u1 · . . . · um and w = um+1 · . . . · un where n ∈ N≥2 , m ∈ [1, n − 1], and u1 , . . . , un ∈ A.
Therefore we obtain that u = u1 · . . . · un and u ∈
/ {u1 , . . . , un }. Thus A \ {u} is a generating set of H, a
contradiction.
2. This assertion follows directly from 1.
3. By 1. and 2., H is atomic and A = A(H) is the smallest generating set. By Redei’s Theorem,
every finitely generated commutative monoid is finitely presented. Thus ∆(H) is finite by Proposition
2.7. Next we show that there is an L ∈ L(H) such that ρ(H) = ρ(L) < ∞. This implies that H is a
BF-monoid and by Proposition 2.4.3 it follows that there is an M ∈ N such that ρk (H) − ρk−1 (H) ≤ M
for all k ≥ 2.
Let π : F (A) → H be the canonical epimorphism. We set
S = {(x, y) ∈ F (A) × F (A) | π(x) = π(y)} and S ∗ = S \ {(1, 1)} ,
o
n
∗
. Clearly, it is sufficient to show that this supremum
(x,
y)
∈
S
and we observe that ρ(H) = sup |x|
|y|
∗
A
is attained
for some
f : F (A) × F (A) → (NA
0 × N0 , +), defined
Q pairnu(x, y) ∈ S . There is an isomorphism
Q
mu
by ( u∈A u , u∈A u ) 7→ (mu )u∈A , (nu )u∈A . By Dickson’s Theorem [23, Theorem 1.5.3], the set
f (S ∗ ) has only finitely many minimal points, and let T ⊂ S ∗ denote the inverse image of the set of
minimal points. Therefore it suffices to prove that
o
n |x′ |
|x|
′ ′
(x
,
y
)
∈
T
for all (x, y) ∈ S ∗ .
≤ max
|y|
|y ′ |
We proceed by induction on |x| + |y|. If (x, y) ∈ T , then there is nothing to do. Suppose that (x, y) ∈
/ T.
Then there exist (x1 , y1 ) ∈ T such that (x, y) = (x1 x2 , y1 y2 ) with (x2 , y2 ) ∈ F (A) × F (A). It follows
that (x2 , y2 ) ∈ S ∗ , and clearly we have |xj | + |yj | < |x| + |y| for j ∈ {1, 2}. Then
n |x′ |
o
n |x | |x | o
|x|
|x1 | + |x2 |
2
1
′ ′
≤ max
(x
,
y
)
∈
T
=
< max
,
|y|
|y1 | + |y2 |
|y1 | |y2 |
|y ′ |
by the induction hypothesis.
A most interesting class of finitely generated commutative monoids are numerical monoids. Their
study was initiated by Frobenius in the 19th century and they are still a topic of much research due
to their intrinsic relationship with a wide area of mathematics. A monoid H is said to be numerical
if it is a submonoid of (N0 , +) such that the complement N0 \ H is finite. Clearly, numerical monoids
are reduced. Let H be a numerical monoid with H 6= N0 . Since N0 \ H is finite, H has a finite
generating set and hence a smallest generating set. Thus Proposition 2.8 implies that the smallest
generating set is the set of atoms, and that the elasticity and the set of distances are both finite. Suppose
that A(H) = {n1 , . . . , , nt } with t ∈ N and n1 < . . . < nt . We encourage the reader to check that
ρ(H) = nt /n1 and that min ∆(H) = gcd{n2 − n1 , . . . , nt − nt−1 } (compare with Proposition 6.1.2).
These results were the starting points of detailed investigations of the arithmetic of numerical monoids
initiated by Chapman and Garcı́a-Sánchez.
10
ALFRED GEROLDINGER
Clearly, there are natural connections between the arithmetical invariants of factorization theory and
the presentations of a monoid. This point has been emphasized by Garcı́a-Sánchez and it opened the way
to an algorithmic approach towards the computational determination of arithmetical invariants. Many
algorithms have been implemented in GAP (see the GAP Package [14] and a survey by Garcı́a-Sánchez
[20]).
3. Commutative Krull Monoids
It was the observation of the mathematicians of the 19th century that a ring of integers in an algebraic
number field need not be factorial (in other words, it need not satisfy the Fundamental Theorem of Arithmetic). This led to the development of ideals (every nonzero ideal in a ring of integers is a unique product
of prime ideals whence the Fundamental Theorem of Arithmetic holds for ideals) and subsequently to the
development of ”divisor theories” from the elements to the ideals. A divisor theory is a divisibility preserving homomorphism to an object which fulfills the Fundamental Theorem of Arithmetic. Semigroups
allowing a divisor theory are now called Krull monoids.
Commutative Krull monoids can be studied with divisor theoretic and with ideal theoretic tools. We
start with divisor theoretic concepts. Let H and D be commutative monoids. A monoid homomorphism
ϕ : H → D is said to be:
• a divisor homomorphism if a, b ∈ H and ϕ(a) | ϕ(b) (in D) implies that a | b (in H);
• cofinal if for every α ∈ D there is an a ∈ H such that α | ϕ(a) (in D);
• a divisor theory if D is free abelian, ϕ is a divisor homomorphism,
and for every α ∈ D there are
a1 , . . . , am ∈ H such that α = gcd ϕ(a1 ), . . . , ϕ(am ) .
In particular, every divisor theory is a cofinal divisor homomorphism. Let ϕ : H → D be a cofinal divisor
homomorphism. The group
C(ϕ) = q(D)/q ϕ(H)
is called the class group of ϕ. For a ∈ q(D) we denote by [a] = aq ϕ(H) ∈ C(ϕ) the class containing a.
We use additive notation for the class group and observe that [1] is the zero element of the abelian group
C(ϕ). Divisor theories of a given monoid are unique up to isomorphism. If H has a divisor theory, then
there is a free abelian monoid F = F (P ) such that the inclusion ϕ : Hred ֒→ F is a divisor theory, and
the class group
C(ϕ) = C(H) = q(F )/q(Hred )
is called the (divisor) class group of H and G0 = {[p] | p ∈ P } ⊂ C(H) is the set of classes containing
prime divisors. We continue with the most classical example of a cofinal divisor homomorphism and a
divisor theory.
Proposition 3.1. Let R be a commutative domain, I ∗ (R) the monoid of invertible ideals where the
operation is the usual multiplication of ideals, and let ϕ : R• → I ∗ (R) be the homomorphism mapping
each element onto its principal ideal.
1. The map ϕ is a cofinal divisor homomorphism and C(ϕ) is the Picard group Pic(R) of R.
2. If R is a commutative Dedekind domain, then ϕ is a divisor theory and C(ϕ) is the usual ideal
class group of R.
3. If R is the ring of integers of an algebraic number field, then C(ϕ) is finite and every class contains
infinitely many prime ideals.
Proof. 1. A short calculation shows that for two invertible ideals I, J ⊳ R we have J | I in I ∗ (R) if and
only if I ⊂ J. To show that ϕ is a divisor homomorphism, let a, b ∈ R• be given and suppose that bR | aR
in I ∗ (R). Then there is a J ∈ I ∗ (R) such that (aR)J = bR whence a−1 bR = J ⊂ R and a | b in R• . To
show that ϕ is cofinal, let I ∈ I ∗ (R) be given. If a ∈ I, then aR ⊂ I and hence I | aR in I ∗ (R). The
definition of C(ϕ) coincides with the definition of Pic(R).
SETS OF LENGTHS
11
2. Suppose that R is a commutative Dedekind domain. Then every nonzero ideal is invertible and a
product of prime ideals. Thus I ∗ (R) is free abelian. Let I ∈ I ∗ (R). Then I is generated by two elements
a, b ∈ R, whence I = ha, bi = aR + bR = gcd(aR, bR) = gcd(ϕ(a), ϕ(b)). Therefore ϕ is a divisor theory.
3. This can be found in many textbooks on algebraic number theory (see, for example, [23, Theorem
2.10.14] for a summary).
The previous proposition shows that in case of commutative Dedekind domains the embedding in a
monoid of ideals establishes a divisor theory. This holds true in much greater generality and in order to
outline this we mention briefly some key notions on divisorial ideals (see [36] for a thorough treatment of
divisorial ideals).
Let H be a commutative monoid and let A, B ⊂ q(H) be subsets. We denote by (A : B) = {x ∈
q(H) | xB ⊂ A}, by A−1 = (H : A), and by Av = (A−1 )−1 . By an ideal of H we always mean
an s-ideal (thus AH = A holds), and an s-ideal A is a divisorial ideal (or a v-ideal) if Av = A. We
denote by Fv (H) the set of all fractional divisorial ideals and by Iv (H) the set of all divisorial ideals
of H. Furthermore, Iv∗ (H) is the monoid
of v-invertible divisorial ideals (with v-multiplication) and its
quotient group Fv (H)× = q Iv∗ (H) is the group of fractional invertible divisorial ideals. By X(H), we
denote the set of all minimal nonempty prime s-ideals of H and
b = {x ∈ q(H) | there is a c ∈ H such that cxn ∈ H for all n ∈ N} ⊂ q(H)
H
b
is called the complete integral closure of H. We say that H is completely integrally closed if H = H.
Straightforward arguments show that every factorial monoid is completely integrally closed and that a
noetherian commutative domain is completely integrally closed if and only if it is integrally closed.
Theorem 3.2 (Commutative Krull monoids).
Let H be a commutative monoid. Then the following statements are equivalent.
(a) H is completely integrally closed and satisfies the ACC on divisorial ideals.
(b) The map ϕ : H → Iv∗ (H), a 7→ aH for all a ∈ H, is a divisor theory.
(c) H has a divisor theory.
(d) There is a free abelian monoid F such that the inclusion Hred ֒→ F is a divisor homomorphism.
If one of the equivalent statements holds, then H is called a Krull monoid, and Iv∗ (H) is free abelian with
basis X(H).
For a proof of Theorem 3.2 we refer to [23, Section 2.5]. Note, since H is factorial if and only if Hred
is free abelian, it follows that H is factorial if and only if it is Krull with trivial class group. In the
remainder of this section we present a list of examples of commutative Krull monoids stemming from
quite diverse mathematical areas.
Commutative domains. Let R be a commutative domain and H = R• . Then the maps
(
(
Fv (R) → Fv (H)
Fv (H) → Fv (R)
•
◦
(3.1)
ι :
and
ι :
a
7→ a \ {0}
a
7→ a ∪ {0}
are inclusion preserving isomorphisms which are inverse to each other. In particular, if a is a divisorial
semigroup theoretical ideal of H, then a ∪ {0} is a divisorial ring theoretical ideal of R. Thus R satisfies
the ACC on (ring theoretical) divisorial ideal of R if and only if H satisfies the ACC on (semigroup
theoretical) divisorial ideals of H. Since, by definition, R is completely integrally closed if and only H
is completely integrally closed, we obtain that R is a commutative Krull domain if and only if H is a
commutative Krull monoid.
Property (a) in Theorem 3.2 easily implies that noetherian integrally closed commutative domains
are Krull. Furthermore, a commutative Krull domain is Dedekind if and only if it is at most onedimensional. If R is Dedekind, then every ideal is divisorial and Iv∗ (R) = I ∗ (R) (confer Theorem 3.2.(b)
and Proposition 3.1.2).
12
ALFRED GEROLDINGER
Submonoids of commutative domains. Let R be a commutative Krull domain, {0} =
6 f ⊳ R an ideal,
and Γ ⊂ (R/f)× a subgroup. Then the monoid
HΓ = {a ∈ R• | a + f ∈ Γ}
is a Krull monoid, called the (regular) congruence monoid defined in R modulo f by Γ. We refer the
reader to [23, Section 2.11] for more on congruence monoids.
Monadic submonoids of rings of integer-valued polynomials. Let us consider the classical ring
of integer-valued polynomials over the integers. This is the ring
Int(Z) = {f ∈ Q[X] | f (Z) ⊂ Z} ⊂ Q[X] .
We refer the reader to the Monthly article by Cahen and Chabert [9] for a friendly introduction to integervalued polynomials and to their monograph [8] for a deeper study. It is well-known that Int(Z) is an
integrally closed two-dimensional Prüfer domain. It is a BF-domain but it is not Krull. However, every
divisor-closed submonoid of Int(Z), which is generated by one element, is a Krull monoid [44, Theorem
5.2]. We refer to recent work of Frisch and Reinhart [43, 18].
Monoids of regular elements in commutative rings with zero-divisors. By a commutative Krull
ring we mean a completely integrally closed commutative ring which satisfies the ACC on regular divisorial
ideals. The isomorphisms (as given in Equation (3.1)) between monoids of divisorial ideals carry over
from the setting of commutative domains to the setting of commutative rings with zero divisors. Thus,
if a commutative ring R is Krull, then the monoid of cancelative (regular) elements is a Krull monoid,
and the converse holds for v-Marot rings [27, Theorem 3.5].
Monoids of Modules. Let R be a ring and let C be a class of right R-modules which is closed under finite
direct sums, direct summands, and isomorphisms. For a module M in C, let [M ] denote the isomorphism
class of M . Let V(C) denote the set of isomorphism classes of modules in C (we assume here that V(C) is
indeed a set.) Then V(C) is a commutative semigroup with operation defined by [M ]+ [N ] = [M ⊕ N ] and
all information about direct-sum decomposition of modules in C can be studied in terms of factorization
of elements in the semigroup V(C). In particular, the direct-sum decompositions in C are (essentially)
unique (in other words, the Krull-Remak-Schmidt-Azumaya Theorem holds) if and only if V(C) is a
free abelian monoid. This semigroup-theoretical point of view was justified by Facchini [15] who showed
that V(C) is a reduced Krull monoid provided that the endomorphism ring EndR (M ) is semilocal for all
modules M in C. This result allows one to describe the direct-sum decomposition of modules in terms
of factorization of elements in Krull monoids. We refer the reader to the Monthly article by Baeth and
Wiegand [5].
e of a commutative monoid H
Finitely generated monoids and affine monoids. The root closure H
is defined as
e = {x ∈ q(H) | xn ∈ H for some n ∈ N} ⊂ q(H) ,
H
e If
and H is said to be root closed (also the terms normal, full, and integrally closed are used) if H = H.
b
e
H is finitely generated, then H = H. Since finitely generated monoids satisfy the ACC on ideals, they
are Krull if and only if they are root closed (see Theorem 3.2.(a)).
A monoid is called affine if it is a finitely generated submonoid of a finitely generated free abelian
group. It is easy to check that the concepts of normal affine monoids and of reduced finitely generated
commutative Krull monoids coincide (a variety of further characterizations are given in [23, Theorem
2.7.14]). (Normal) affine monoids play an important role in combinatorial commutative algebra.
Monoids of Zero-Sum Sequences. Let G be an additively written abelian group and G0 ⊂ G a
subset. By a sequence over G0 , we mean a finite sequence of terms from G0 where repetition is allowed
and the order is disregarded. Clearly, the set of sequences forms a semigroup, with concatenation as its
operation and with the empty sequence as its identity element. We consider sequences as elements of the
free abelian monoid with basis G0 . This algebraic point of view has turned out to be quite convenient
SETS OF LENGTHS
13
from a notational point of view. But there is much more which we start to outline here and later in
Proposition 4.3. Let
Y
S = g1 · . . . · gℓ =
g vg (S) ∈ F (G0 ) ,
g∈G0
where ℓ ∈ N0 and g1 , . . . , gℓ ∈ G. Then |S| = ℓ is the length of S, supp(S) = {g1 , . . . , gℓ } is the support
of S, −S = (−g1 ) · . . . · (−gℓ ), and σ(S) = g1 + . . . + gℓ is the sum of S. We say that S is a zero-sum
sequence if σ(S) = 0, and clearly the set
B(G0 ) = {S ∈ F (G0 ) | σ(S) = 0} ⊂ F (G0 )
of all zero-sum sequences is a submonoid, called the monoid of zero-sum sequences (also called Block
Monoid) over G0 . Obviously, the inclusion B(G0 ) ֒→ F (G0 ) is a divisor homomorphism and hence B(G0 )
is a reduced commutative Krull monoid by Theorem 3.2.(d). Monoids of zero-sum sequences form a
powerful link between the theory of (general) Krull monoids and additive combinatorics [28, 34]. Thus
all methods from the later area are available for the study of sets of lengths in Krull monoids, and we
will make heavily use of this in Section 6.
Proposition 3.3. Let G be an additive abelian group and G0 ⊂ G a subset.
1. If G0 is finite, then B(G0 ) is finitely generated.
2. The following statements are equivalent.
(a) |G| ≤ 2.
(b) B(G) is factorial.
(c) B(G) is half-factorial.
3. If |G| ≥ 3, then the inclusion B(G) ֒→ F (G) is a divisor theory with class group isomorphic to G
and every class contains precisely one prime divisor.
4. Let G′ be an abelian group. Then the monoids B(G) and B(G′ ) are isomorphic if and only if the
groups G and G′ are isomorphic.
Q
mg
0
, is a monoid isomorphism.
Proof. 1. The map f : NG
0 → F (G0 ), defined by m = (mg )g∈G0 7→
g∈G0 g
G0
−1
The embedding Γ := f (B(G0 )) ֒→ N0 is a divisor homomorphism (i.e, m, n ∈ Γ and m ≤ n implies
that n − m ∈ Γ). By Dickson’s Lemma [23, Theorem 1.5.3], Γ is generated by the finite set of minimal
points Min(Γ), and hence B(G0 ) is generated by f (Min(Γ)).
2. If G = {0}, then B(G) = F (G) ∼
= (N0 , +) is free abelian. If G = {0, g}, then B(G) is free abelian
with basis A(G) = {0, g 2 }. Thus (a) ⇒ (b), and obviously (b) ⇒ (c). In order to verify that (c) ⇒ (a),
it suffices to show that |G| ≥ 3 implies that B(G) is not half-factorial. Suppose that |G| ≥ 3. If there
is some element g ∈ G with ord(g) = n ≥ 3, then U = g n , −U , and V = (−g)g are atoms of B(G) and
(−U )U = V n shows that B(G) is not half-factorial. If there are two distinct elements e1 , e2 ∈ G of order
two, then U = e1 e2 (e1 + e2 ), V0 = (e1 + e2 )2 , V1 = e21 , and V2 = e22 are atoms of B(G) and U 2 = V0 V1 V2
shows that B(G) is not half-factorial.
3. Let |G| ≥ 3. Clearly, the inclusion is a cofinal divisor homomorphism. To show that it is a divisor
theory, let g ∈ G \ {0} be given. If ord(g) = n ≥ 3, then g = gcd g n , g(−g)
. If ord(g) = 2, then there is
an element h ∈ G \ {0, g}, and we obtain that g = gcd g 2 , gh(−g − h) . It is easy to check that the map
Φ : C B(G) = q F (G) /q B(G) = {[S] = Sq B(G) | S ∈ F (G)} → G ,
defined by Φ([S])
= σ(S) is a group isomorphism. Since for every S ∈ F (G), [S] ∩ G = {σ(S)}, every
class of C B(G) contains precisely one prime divisor.
4. This follows from 2.,3., and the fact that a reduced commutative Krull monoid is uniquely determined by its class group and the distribution of prime divisors in its classes ([23, Theorem 2.5.4]).
14
ALFRED GEROLDINGER
4. Transfer Homomorphisms and Transfer Krull Monoids
A central method to study the arithmetic of a given class of monoids H is to construct simpler auxiliary
monoids B (and such constructions are often based on the ideal theory of H) and homomorphisms
θ : H → B (called transfer homomorphisms) which allow us to pull back arithmetical results from B to
H. The concept of transfer homomorphisms was introduced by Halter-Koch in the commutative setting
[35]) and recently generalized to the noncommutative setting ([4, Definition 2.1]).
Definition 4.1. Let H and B be atomic monoids. A monoid homomorphism θ : H → B is called a weak
transfer homomorphism if it has the following two properties.
(T1) B = B × θ(H)B × and θ−1 (B × ) = H × .
(WT2) If a ∈ H, n ∈ N, v1 , . . . , vn ∈ A(B) and θ(a) = v1 · . . . · vn , then there exist u1 , . . . , un ∈ A(H) and
a permutation τ ∈ Sn such that a = u1 · . . . · un and θ(ui ) ∈ B × vτ (i) B × for each i ∈ [1, n].
Property (T1) says that θ is surjective up to units and that only units are mapped onto units. Property
(WT2) says that factorizations can be lifted up to units and up to order. We do not discuss equivalent
formulations or variants of the definition and we do not give the definition of a transfer homomorphism,
but note that the two concepts coincide if H and T are both commutative.
Lemma 4.2. Let H and B be atomic monoids, and let θ : H → B be a weak transfer homomorphism.
1. For every a ∈ H, we have LH (a) = LB θ(a) . In particular, an element a ∈ H is an atom in H if
and only if θ(a) is an atom in B.
2. L(H) = L(B). In particular, ∆(H) = ∆(B), Uk (H) = Uk (B), and ρk (H) = ρk (B) for every
k ∈ N.
Proof. Since 2. follows directly from 1., we prove 1. Let a ∈ H. If n ∈ LB θ(a) , then θ(a) = v1 · . . . · vn
with v1 , . . . , vn ∈ A(B), and thus (WT2) implies that n ∈ LH (a). Conversely, let n ∈ LH (a). Then there
are u1 , . . . , un ∈ A(H) such that a = u1 · . . . · un . Thus θ(a) = θ(u1 ) · . . . · θ(un ), and we have to verify
that θ(u1 ), . . . , θ(un ) ∈ A(B). Let i ∈ [1, n]. Property (T1) implies that θ(ui ) is not a unit. Since B is
atomic, there are m ∈ N and w1 , . . . , wm ∈ A(B) such that θ(ui ) = w1 · . . . · wm . Since this factorization
can be lifted and ui is an atom, it follows that m = 1 and that θ(ui ) = w1 ∈ A(B). Since an element of
an atomic monoid is an atom if and only if its set of lengths equals {1}, the statement follows.
Next we discuss the most classic example of a transfer homomorphism and its application. This is
the homomorphism from a commutative Krull monoid to an associated monoid of zero-sum sequences.
If H is a commutative monoid, then H is Krull if and only if Hred is Krull, and if this holds, then the
canonical epimorphism π : H → Hred is a transfer homomorphism. Thus, in the following proposition we
may restrict to reduced Krull monoids for technical simplicity, but without loss of generality.
Proposition 4.3. Let H be a reduced commutative Krull monoid, F = F (P ) a free abelian monoid such
that the embedding H ֒→ F is a cofinal divisor homomorphism with class group G, and let G0 = {[p] |
p ∈ P } ⊂ G = q(F )/q(H) denote the set of classes containing prime divisors.
Then there is a transfer
homomorphism β : H → B(G0 ). In particular, we have L(H) = L B(G0 ) .
e : F → F (G0 ) be the unique epimorphism defined by β(p)
e
Proof. Let β
= [p] for all p ∈ P . We start with
the following assertion.
e
e
A1. For every a ∈ F , we have β(a)
∈ B(G0 ) if and only if a ∈ H. Thus β(H)
= B(G0 ) and
−1
e
B(G0 ) = H.
β
Proof of A1. Let a = p1 · . . . · pℓ ∈ F where ℓ ∈ N0 and p1 , . . . , pℓ ∈ P . Then
e
β(a)
= [p1 ] · . . . · [pℓ ] ∈ F (G0 ) and σ [p1 ] · . . . · [pℓ ] = [p1 ] + . . . + [pℓ ] = [a] .
Since H ֒→ F is a divisor homomorphism, we have [a] = 0 ∈ G if and only if a ∈ H. Therefore all
assertions follow and we have proved A1.
SETS OF LENGTHS
15
e : H → B(G0 ), and we assert that it is a transfer
Therefore we can define the homomorphism β = β|H
homomorphism. Clearly, H and B(G0 ) are reduced and β is surjective. Thus (T1) reads as
B(G0 ) = β(H)
and β −1 ({1}) = {1} ,
which holds true by A1. We continue with the following assertion.
A2. If a ∈ H, B, C ∈ B(G0 ) and β(a) = BC, then there exist b, c ∈ H such that a = bc, β(b) = B,
and β(c) = C.
Proof of A2. Let a = p1 · . . . · pℓ ∈ H, where ℓ ∈ N0 and p1 , . . . , pℓ ∈ P , and suppose that β(a) = BC, say
B = [p1 ] · . . . · [pk ] and C = [pk+1 ] · . . . · [pℓ ] for some k ∈ [0, ℓ]. By A1, we infer that b = p1 · . . . · pk ∈ H,
c = pk+1 · . . . · pℓ ∈ H and clearly we have a = bc. This completes the proof of A2.
Clearly, (WT2) follows from A2 by a straightforward
induction, and hence β is a transfer homomorphism. Then Lemma 4.2 implies that L(H) = L B(G0 ) .
Definition 4.4. A monoid H is said to be a transfer Krull monoid (over G0 ) if there exists a weak
transfer homomorphism θ : H → B(G0 ) for a subset G0 of an abelian group G. If G0 is finite, then we
say that H is a transfer Krull monoid of finite type.
By Proposition 4.3, every commutative Krull monoid is a transfer Krull monoid. If a monoid H ∗ has
a weak transfer homomorphism to a commutative Krull monoid, say θ : H ∗ → H, then the composition
β ◦ θ : H ∗ → B(G0 ) is a weak transfer homomorphism (with the notation of Proposition 4.3) and hence
H ∗ is a transfer Krull monoid. Thus a monoid is a transfer Krull monoid if and only if it allows a weak
transfer homomorphism to a commutative Krull monoid.
Since monoids of zero-sum sequences are BF-monoids (this can be checked directly or by using Lemma
2.2), Lemma 4.2 shows that transfer Krull monoids are BF-monoids. However, the examples given below
reveal that transfer Krull monoids need neither be commutative nor completely integrally closed nor Mori
(i.e., they do not necessarily satisfy the ACC on divisorial ideals). Before we provide a list of transfer
Krull monoids, we briefly discuss general, not necessarily commutative Krull monoids (for details we
refer to [22]). This concept was introduced by Wauters in 1984 in complete analogy to the ideal theoretic
definition of commutative Krull monoids (compare with Theorem 3.2.(a)).
Suppose that H is a monoid such that aH ∩ bH 6= ∅ and Ha ∩ Hb 6= ∅ for all a, b ∈ H. Then H is
called a Krull monoid (or a Krull order) if it is completely integrally closed and satisfies the ACC on
two-sided divisorial ideals. The isomorphisms in Equation (3.1) between monoids of divisorial ideals carry
over from the setting of commutative domains to the setting of prime Goldie rings. Thus, in analogy to
the commutative setting, we have that a prime Goldie ring is a Krull ring if and only if its monoid of
cancelative elements is a Krull monoid [22, Proposition 5.1]. Moreover, Krull monoids play a central role
in the study of noetherian semigroup algebras. We refer to [11] for recent surveys on non-commutative
Krull rings and monoids.
Examples 4.5. 1. As outlined above, commutative Krull monoids (hence all the examples given is
Section 3) are transfer Krull monoids. But more generally, every normalizing Krull monoid is a transfer
Krull monoid by [22, Theorems 4.13 and 6.5] (a monoid H is said to be normalizing if aH = Ha for all
a ∈ H).
2. Let H be a half-factorial monoid. Since the map θ : H → B({0}), defined by θ(ǫ) = 1 for all ǫ ∈ H ×
and θ(u) = 0 for every u ∈ A(H), is a transfer homomorphism, H is a transfer Krull monoid (over the
trivial group {0}). Only recently M. Roitman showed that commutative half-factorial domains need not
be Mori [45]. Thus, transfer Krull monoids satisfy the ACC on principal left ideals and on principal
right ideals (since they are BF-monoids; see Lemma 2.2) but they do not necessarily satisfy the ACC on
divisorial ideals.
3. Let O be the ring of integers of an algebraic number field K, A a central simple algebra over K,
and R a classical maximal O-order of A. Then R• is a Krull monoid. If every stably free left R-ideal is
free, then R• is a transfer Krull monoid over a ray class group of O (note that this group is finite). If
16
ALFRED GEROLDINGER
there is a stably free left R-ideal that is not free, then R• is not a transfer Krull monoid. This is due to
Smertnig [49, Theorem 1.1 and 1.2], and for related results in a more general setting we refer to [4].
4. Let R be an order in an algebraic number field K, R the integral closure of R (thus R is the ring
of integers of K), and let π : spec(R) → spec(R) be the natural map defined by π(P) = P ∩ R for all
nonzero prime ideals P ⊳ R.
4.(a) If R is seminormal, π is bijective, and there is an isomorphism δ : Pic(R) → Pic(R), then R• is
a transfer Krull monoid over Pic(R) ([25, Theorem 5.8]).
4.(b) Suppose that π is not bijective. Since ρ(R• ) = ∞ by [23, Corollary 3.7.2], R• is not a transfer
Krull monoid of finite type by Theorem 4.6. Moreover, R• is not a transfer Krull monoid over an infinite
abelian group G (compare Theorems 4.6 and 5.5).
5. Let D be a commutative Krull domain, R ⊂ D a subring having the same quotient field such
that D = RD× , D× ∩ R = R× , and (R : D) = m ∈ max(R). Then the inclusion R• ֒→ D• is a transfer
homomorphism and hence R• is a transfer Krull monoid [23, Proposition 3.7.5]. Note that K +M -domains
satisfy the above assumptions. Indeed, let R ( D be commutative domains, m a nonzero maximal ideal
of D, and let K ( L ( D be subfields such that D = L + m and R = K + m. If D is Krull, then the
above assumptions are satisfied.
6. Let R be a bounded HNP (hereditary noetherian prime) ring, and note that a commutative domain
is an HNP ring if and only if it is a Dedekind domain. If every stably free left R-ideal is free, then R• is
a transfer Krull monoid [48, Theorem 4.4].
7. In [3], the authors study monoids of modules over HNP rings and thereby monoids of the following
type occur. Let H0 be a commutative Krull monoid but not a group, D be a commutative monoid with
D 6= {1D }, and define H = (H0 \ H0× ) × D ∪ H0× × {1D }. Then H is a transfer Krull monoid which is
not completely integrally closed [3, Proposition 6.1].
8. In [2], the authors study conditions under which monoids of upper triangular matrices over commutative domains allow weak transfer homomorphisms to the underlying domain. Thus, in case of
commutative Krull domains we obtain transfer Krull monoids. Smertnig established characterizations on
the existence of transfer homomorphisms from full matrix rings over commutative noetherian rings with
no nonzero nilpotent elements to commutative Krull domains [50, Theorem 5.18].
Sets of lengths in transfer Krull monoids (hence in all above examples) can be studied successfully
with the strategy using transfer homomorphisms. Indeed combining Lemma 4.2 and Proposition 3.3 we
are able to apply the structural results for finitely generated monoids (derived in Section 2) to transfer
Krull monoids. This is done in Theorem 4.6 whose proof follows from Propositions 2.8, 3.3, and from
Lemma 4.2.
Theorem 4.6. Let H be a transfer Krull monoid of finite type. Then the set of distances ∆(H) is finite,
the elasticity ρ(H) is finite, the unions Uk (H) of sets of lengths are finite for all k ∈ N, and they satisfy
the Structure Theorem for Unions of Sets of Lengths, as given in Theorem 2.6.
We end this section by posing the following problem (see [30]).
Problem 4.7. Let R be an order in an algebraic number field. Characterize when the monoid of nonzero
elements R• and when the monoid of invertible ideals I ∗ (R) are transfer Krull monoids, resp. transfer
Krull monoids of finite type.
5. The Structure Theorem for Sets of Lengths
In transfer Krull monoids of finite type, not only do unions of sets of lengths do have a well-defined
structure (as given in Theorem 4.6), but the same is true for sets of lengths. We start with a set of
examples which demonstrate that the structure of sets of lengths is richer than that of their unions.
Examples 5.1. Let G be a finite abelian
group and G0 ⊂ G a subset such that B(G0 ) is not half-factorial.
Since min ∆ B(G0 ) = gcd B(G0 ) by Proposition 2.3, it follows that for every B ∈ B(G0 ) and every
SETS OF LENGTHS
y ∈ L(B) we have
L(B) ⊂ y + dZ where
17
d = min ∆ B(G0 ) .
Clearly, every
set of lengths in B(G0 ) is an arithmetical progression with difference d if and only if
∆ B(G0 ) = {d}. We will demonstrate that arithmetical progressions (of arbitrary lengths) actually
occur as sets of lengths, but also several variants of arithmetical progressions do occur naturally.
1. Arithmetical progressions. Let g ∈ G with ord(g) = n ≥ 3. Then U = g n , −U = (−g)n , and
V = (−g)g are atoms, (−U )U = V n , and clearly L (−U )U = {2, n}. For every k ∈ N, we have
L (−U )k U k = 2k + {ν(n − 2) | ν ∈ [0, k]}.
2. Sumsets of arithmetical progressions. Let r, k1 , . . . , kr ∈ N and n1 , . . . , nr ∈ N≥3 . For every
i ∈ [1, r], let gi ∈ G with ord(gi ) = ni and we define Bi = (−gi )ni gini . If hg1 , . . . , gr i = hg1 i ⊕ . . . ⊕ hgr i,
then by 1., L(B1k1 · . . . · Brkr ) =
L(B1k1 )
+ ...+
L(Brkr )
= 2(k1 + . . . + kr ) +
r
X
{ν(ni − 2) | ν ∈ [0, ki ]}
i=1
is the sum of r arithmetical progressions. Clearly, the sum of r long arithmetical progressions with
differences d1 , . . . , dr is an almost arithmetical progression with difference d = gcd(d1 , . . . , dr ).
3. Almost arithmetical progressions (AAPs, see Definition 2.5). We sketch the argument that large
sets of lengths in B(G0 ) are AAPs with difference d = min ∆ B(G0 ) (for a formal statement and proof
we refer to [23, Theorem 4.3.6]).
We proceed as at the beginning of the proof of Theorem 2.6. Clearly,
C0 ∈ B(G0 )
there exist an element
and m ∈ N such that {m, m + d} ⊂ L(C0 ). Since d = gcd ∆ B(G0 , ψ = ρ ∆ B(G0 − 1 ∈ N. Then
L0 = {k0 , k0 + d, . . . , k0 + ψd} ⊂ L(C) where C = C0ψ and k0 = ψm. Now pick any large element
A ∈ B(G0 ), where by large we mean that A is divisible by C. Thus, for some B ∈ B(G0 ), we have
A = BC
and L0 + L(B) ⊂ L(C) + L(B) ⊂ L(A) .
Since L(B) can be viewed as an arithmetical progression with difference d which has gaps (whose number
is controlled by ψ), the sumset L0 + L(B) is an arithmetical progression with difference d. Thus, if A is
large (with respect to the parameters d and ψ depending on G0 ), the set of lengths L(A) contains a long
arithmetical progression with difference d as the central part, whereas the initial and end parts may have
gaps.
4. Almost arithmetical multiprogressions (the definition is given below). Let G1 ⊂ G0 be a subset and
let B ∈ B(G1 ) be such that L(B) is an AAP with difference d, say
L(B) = y + (L′ ∪ L∗ ∪ L′′ ) ⊂ y + dZ
where L∗ is a long arithmetical progression with difference d (the long central part of L(B)) such that
min L∗ = 0. It is not difficult to show that every finite subset of N≥2 can be realized as a set of lengths
(e.g., [23, Proposition 4.8.3]). Thus for any set D ⊂ [0, d] with min D = 0 and max D = d, there is a
zero-sum sequence C with L(C) = x + D for some x ∈ N. Suppose that C ∈ B(G2 ) for a subset G2 ⊂ G0
with hG1 i ∩ hG2 i = {0}. Then
(x + y)+ (L′ + D) ⊎ (L∗ + D) ⊎ (L′′ + D)
⊂(x + y) + L(B) + L(C) ⊂ (x + y) + L(BC) ⊂ (x + y) + D + dZ .
Note that the long central part L∗ + D repeats the set D periodically, whereas the short initial and end
parts L′ + D and L′′ + D may contain gaps. Indeed, if L∗ = {0, d, 2d, . . . , ℓd}, then
L∗ + D = D ∪ (d + D) ∪ . . . ∪ (ℓd + D) ⊂ D + dZ .
Consider transfer Krull monoids of finite type. Then their sets of lengths coincide with sets of lengths
of the monoid of zero-sum sequences. Moreover, the Structure Theorem for Sets of Lengths for these
18
ALFRED GEROLDINGER
monoids states that no other phenomena besides those which we have described in the above examples
can occur. We make this more precise with the following definition.
Definition 5.2. Let d ∈ N, ℓ, M ∈ N0 and {0, d} ⊂ D ⊂ [0, d]. A subset L ⊂ Z is called an almost
arithmetical multiprogression (AAMP for short) with difference d, period D, length ℓ and bound
M , if
L = y + (L′ ∪ L∗ ∪ L′′ ) ⊂ y + D + dZ
where y ∈ Z is a shift parameter,
• the central part L∗ satisfies min L∗ = 0, L∗ = [0, max L∗ ] ∩ (D + dZ), and ℓ ∈ N is maximal such
that ℓd ∈ L∗ ,
• the initial part L′ satisfies L′ ⊂ [−M, −1], and
• the end part L′′ satisfies L′′ ⊂ max L∗ + [1, M ].
Note that AAMPs are finite subsets of the integers, that an AAMP with period D = {0, d} is an AAP,
and that an AAMP with period D = {0, d} and bound M = 0 is a usual arithmetical progression with
difference d. As it was with AAPs (see Definition 2.5), every single finite set is an AAMP with a trivial
choice of parameters (let L∗ be a singleton and set M = max L). To discuss one example of an AAMP
(with natural parameters), let n = pk11 · . . . · pkr r , where r, k1 , . . . , kr ∈ N and p1 , . . . , pr are distinct primes.
We consider the set A = {a ∈ [0, n] | gcd(a, n) > 1} ∪ {0} and observe that A = ∪ri=1 pi N0 ∩ [0, n]. Setting
d = p1 · . . . · pr and D = A ∩ [0, d], we obtain that
A = D + {0, d, 2d, . . . , (n/d − 1)d} ⊂ D + dZ
is an AAMP with difference d, period D, and bound M = 0.
Consider an atomic monoid with nonempty set of distances. Lemma 2.1 shows that sets of lengths
become arbitrarily large. The Structure Theorem for Sets of Lengths (formulated below) states that the
set of distances is finite (whence there are only finitely many periods D with differences in ∆(H)) and
there is one global bound M for all sets of lengths. Thus (with the above notation) long sets of lengths
have a highly structured central part L∗ , and L∗ is the only part of the set of lengths that can become
arbitrarily large whereas the initial and end parts are universally bounded.
Theorem 5.3 (Structure Theorem for Sets of Lengths). Let H be a transfer Krull monoid of finite
type. Then the set of distances is finite and there is some M ∈ N0 such that every L ∈ L(H) is an AAMP
with some difference d ∈ ∆(H) and bound M .
The above theorem was first proved in [21] (in a slightly weaker version), and a detailed proof can be
found in [23, Chapter 4]. To provide an additional example for the validity of the Structure Theorem,
b and with nontrivial conductor
take a commutative Mori domain R with complete integral closure R,
b
b
f = (R : R). If the class group C(R) and the residue class ring R/f are both finite, then the Structure
Theorem holds true [23, Theorems 2.11.9 and 4.6.6] (this setting includes orders in algebraic number
fields). It is an open problem whether the assumption on the finiteness of R/f is necessary for the validity
of the Structure Theorem [24, 39]. On the other hand, for transfer Krull monoids of finite type the
description given by the above Structure Theorem is best possible as the following realization theorem
by Schmid [46] shows.
Theorem 5.4 (A Realization Theorem). Let M ∈ N0 and ∆∗ ⊂ N be a finite nonempty set. Then
there exists a commutative Krull monoid H with finite class group such that the following holds : for
every AAMP L with difference d ∈ ∆∗ and bound M there is some yH,L ∈ N such that y + L ∈
L(H) for all y ≥ yH,L .
We end this section with results which are in sharp contrast to the Structure Theorem. Indeed, they
offer monoids where every finite subset of N≥2 occurs as a set of lengths. Moreover, there is a transfer
Krull monoid H1 and a monoid H2 , which is not a transfer Krull monoid, whose systems of sets of lengths
coincide (the first class is due to a theorem of Kainrath [40] and the second example due to Frisch [17]).
SETS OF LENGTHS
19
Theorem 5.5. For the following classes of monoids we have
L(H) = {L ⊂ N≥2 | L is finite and nonempty} ∪ {0}, {1} .
• H is a transfer Krull monoid over an infinite abelian group G.
• H = Int(Z)• is the monoid of nonzero integer-valued polynomials over Z.
Moreover, Int(Z)• is not a transfer Krull monoid.
6. The Characterization Problem for Systems of Sets of Lengths
Let H be a transfer Krull monoid of finite type. As we have seen in Theorems 4.6 and 5.3, the finite
type property implies the finiteness of the set of distances and the structural results on unions of sets
of lengths and on sets of lengths. In this final section we will always suppose that H is a transfer Krull
monoid over a finite abelian group, and this assumption will imply even stronger results.
Thus let H be a transfer Krull monoid over a finite abelian
group G. Then Lemma 4.2 implies that
L(H) = L B(G) , and as usual we set L(G) := L B(G) . Recall all the examples discussed in Section
3 and in Examples 4.5. In particular, rings of integers of algebraic number fields are the prototypical
examples for transfer Krull monoids over finite abelian groups. Classical philosophy in algebraic number
theory (dating back to the 19th century) states that the class group determines their arithmetic. This
idea can be justified (see [23, Section 7.1]), and concerning lengths of factorizations it holds true by
Proposition 4.3. In the 1970s Narkiewicz posed the inverse question of whether or not arithmetical
behaviour (in other words, behaviour describing the non-uniqueness of factorizations) characterize the
class group. The first affirmative answers ([23, Sections 7.1 and 7.2]) have an artificial flavor because the
given characterizations are based on rather abstract arithmetical properties which are designed to do the
characterization and play only a small role in other parts of factorization theory. Since on the other hand
sets of lengths are of central interest in factorization theory, it has been natural to ask whether their
structure is rich enough to force characterizations, and this question is known as the Characterization
Problem.
The Characterization Problem. Given two finite abelian groups G and G′ with D(G) ≥ 4 such that
L(G) = L(G′ ). Does it follow that G ∼
= G′ ?
Clearly, a necessary condition for an affirmative answer is that G and G′ are isomorphic if and only if
the associated monoids B(G) and B(G′ ) are isomorphic. This necessary condition is guaranteed by Proposition 3.3.4. Answering the Characterization Problem is a long-term goal in the study of sets of lengths
of transfer Krull monoids over finite abelian groups. We start with two elementary results (Propositions
6.1 and 6.2). Then we will be in
Problem in greater detail. As
a position to analyze
the Characterization
usual we set A(G) := A B(G) , ∆(G) := ∆ B(G) , Uk (G) = Uk B(G) , and ρk (G) := ρk B(G) for
every k ∈ N. Since A(G) is finite (see Propositions 2.8 and 3.2), the Davenport constant
D(G) = max{|U | | U ∈ A(G)}
is finite. Clearly, D(G) is the smallest integer ℓ ∈ N such that every sequence S over G of length |S| ≥ ℓ
has a zero-sum subsequence T of length |T | ≥ 1. The significance of D(G) for the study of sets of lengths
will become
clear in our next result. If |G| ≤ 2, then D(G) = |G| and Proposition 3.3.2 implies that
L(G) = {k} | k ∈ N0 , whence ∆(G) = ∅, and Uk (G) = {k} for every k ∈ N. Thus we suppose that
2 < |G| < ∞.
Proposition 6.1. Let G be a finite abelian group with |G| ≥ 3.
1. For every k ∈ N, Uk (G) is an interval, ρ(G) = D(G)/2, ρ2k (G) = kD(G), and
D(G)
⌋.
2
2. ∆(G) is an interval with min ∆(G) = 1 and max ∆(G) ≤ D(G) − 2.
1 + kD(G) ≤ ρ2k+1 (G) ≤ kD(G) + ⌊
20
ALFRED GEROLDINGER
Proof. 1. Let k ∈ N. First, we show that Uk (G) is an interval. Note that it suffices to prove that
[k, ρk (G)] ⊂ Uk (G). Indeed, suppose that this is done, and let ℓ ∈ [min Uk (G), k]. Then ℓ ≤ k ≤ ρℓ (G),
hence k ∈ Uℓ (G) and consequently ℓ ∈ Uk (G).
Thus let ℓ ∈ [k, ρk (G)] be minimal such that [ℓ, ρk (G)] ⊂ Uk (G) and assume to the contrary that ℓ > k.
Let Ω be the set of all A ∈ B(G) such that {k, j} ⊂ L(A) for some j ≥ ℓ, and let B ∈ Ω be such that |B|
is minimal. Then B = U1 · . . . · Uk = V1 · . . . · Vj , where j ≥ ℓ and U1 , . . . , Uk , V1 , . . . , Vj ∈ A(G). Since
j > k, we have B 6= 0|B| , and (after renumbering if necessary) we may assume that Uk = g1 g2 U ′ and
Vj−1 Vj = g1 g2 V ′ , where g1 , g2 ∈ G and U ′ , V ′ ∈ F (G). Then Uk′ = (g1 + g2 )U ′ ∈ A(G), and we suppose
′
that Vj−1
= (g1 +g2 )V ′ = W1 ·. . .·Wt , where t ∈ N and W1 , . . . , Wt ∈ A(G). If B ′ = U1 ·. . .·Uk−1 Uk′ , then
|B ′ | < |B| and B ′ = V1 · . . . · Vj−2 W1 · . . . · Wt . By the minimal choice of |B|, it follows that j − 2 + t < ℓ,
hence t = 1, j = ℓ and ℓ − 1 ∈ Uk (G), a contradiction.
Second, we study ρk (G). We start with the following assertion.
A. If A = 0m B ∈ B(G), with m = v0 (A) ∈ N0 and B ∈ B(G), then
D(G)
.
2
Proof of A. Let A = 0m U1 · . . . · Uℓ where ℓ, m ∈ N0 and U1 , . . . , Uℓ ∈ A(G) \ {0}. Then 2 ≤ |Uν | ≤ D(G)
for all ν ∈ [1, ℓ] and hence
m + 2ℓ ≤ |A| ≤ m + ℓD(G) .
2 max L(A) − m ≤ |A| ≤ D(G) min L(A) − m(D(G) − 1) and ρ(A) ≤
Choosing ℓ = min L(B) and ℓ = max L(B) we obtain the first inequalities, and then we obtain
ρ(A) =
max L(A)
m + max L(B)
max L(B)
D(G)
=
≤
≤
.
min L(A)
m + min L(B)
min L(B)
2
The proof of A and Proposition 2.4.2 imply that ρk (G) ≤ kρ(G) ≤ kD(G)/2. If U = g1 ·. . .·gD(G) ∈ A(G),
k
QD(G)
then (−U )k U k = i=1
(−gi )gi , whence kD(G) ≤ ρ2k (G) and thus ρ2k (G) = kD(G). Furthermore, it
follows that
(2k + 1)D(G)
1 + kD(G) = ρ1 (G) + ρ2k (G) ≤ ρ2k+1 (G) ≤
.
2
Finally, Proposition 2.4.2 implies that ρ(G) = D(G)/2.
2. The proof that ∆(G) is an interval is similar but trickier than that of 1., and we refer to [31]. It is easy
to verify that 1 ∈ ∆(G), and we encourage the reader to do so. Next we prove that max ∆(G) ≤ D(G)−2.
If A′ = 0k A with k ∈ N0 and A ∈ B(G) with 0 ∤ A, then L(A′ ) = k + L(A) and ∆(L(A′ )) = ∆(L(A)).
Thus we have to prove that max ∆(L(A)) ≤ D(G) − 2 for all A ∈ B(G) with 0 ∤ A, and we proceed by
induction on |A|. Suppose that
A = U1 · . . . · Ui = V1 · . . . · Vk , where i < k, U1 , . . . , Ui , V1 , . . . , Vk ∈ A(G) ,
and L(A)∩[i, k] = {i, k}. If |A| ≤ 2D(G), then k ≤ D(G) and k −i ≤ D(G)−2. Suppose that |A| > 2D(G)
and that max ∆(L(A′ )) ≤ D(G) − 2 for all A′ with |A′ | < |A|. If |Vj | ≥ i for all j ∈ [1, k], then
ki ≤ |V1 · . . . · Vk | = |U1 · . . . · Ui | ≤ iD(G)
and hence
k − i ≤ D(G) − 2 .
Suppose that there is a j ∈ [1, k] such that |Vj | < i, say j = 1, V1 | U1 · . . . · Ui−1 , and let U1 · . . . · Ui−1 =
V1 W2 · . . . · Wℓ with ℓ ∈ N and W2 , . . . , Wℓ ∈ A(G). We distinguish two cases.
CASE 1: ℓ ≥ i. Since L(U1 ·. . .·Ui )∩[i, k] = {i, k}, it follows that L(U1 ·. . .·Ui−1 )∩[i, k −2] = ∅ and hence
ℓ ≥ k − 1. We may suppose that ℓ ≥ k − 1 is minimal such that U1 · . . . · Ui−1 satisfies such an equation.
Then the induction hypothesis implies that ℓ − (i − 1) ≤ D(G) − 2 and hence k − i ≤ (ℓ + 1) − i ≤ D(G) − 2.
CASE 2: ℓ ≤ i−1. Since V1 ·. . .·Vk = U1 ·. . .·Ui = V1 W2 ·. . .·Wℓ Ui , it follows that V2 ·. . .·Vk = Ui W2 ·. . .·Wℓ .
Note that L(V2 · . . . · Vk ) ∩ [i, k − 2] = ∅, and suppose that ℓ ≤ i − 1 is maximal such that V2 · . . . · Vk
satisfies such an equation. Then the induction hypothesis implies that (k − 1) − ℓ ≤ D(G) − 2 and hence
k − i = (k − 1) − (i − 1) ≤ k − 1 − ℓ ≤ D(G) − 2.
SETS OF LENGTHS
21
The state of the art on ρ2k+1 (G) is discussed in [47]. For some small groups G the system L(G) can
be written down explicitly.
Proposition 6.2.
1. L(C3 ) = L(C
2 ⊕ C2 ) = y + 2k + [0, k] y, k ∈ N0 .
2. L(C4 ) = y +k+ 1+[0, k] | y, k ∈ N0 ∪ y +2k+2 · [0, k] | y, k ∈ N0 .
3
3. L(C
2 ) = y + (k + 1) + [0, k] y ∈N0 , k ∈ [0, 2] ∪
y + k + [0, k] y ∈ N0 , k ≥ 3 ∪ y + 2k + 2 · [0, k] y, k ∈ N0 .
4. L(C32 ) = {[2k, ℓ] | k ∈ N0 , ℓ ∈ [2k, 5k]}
∪ {[2k + 1, ℓ] | k ∈ N, ℓ ∈ [2k + 1, 5k + 2]} ∪ {{1}}.
Proof. We prove the first statement. The proofs of the remaining statements are similar but more lengthy
(details can be found in [29, Proposition 4.2]). Suppose that G is cyclic of order three, say G = {0, g, −g}.
Then A(G) = {0, U = g 3 , −U, V = (−g)g}, D(G) = 3, and (−U )U = V 3 is the only minimal relation.
Clearly, L(V 3 ) = {2, 3}, L(V 3k ) = {2k, 2k + 1, . . . , 3k} = 2k + [0, k], and L(0y V 3k ) = y + 2k + [0, k] for all
y, k ∈ N0 . Since ∆(G) = {1}, ρ2k (G) = 3k, and ρ2k+1 (G) = 3k + 1 for every k ∈ N by Proposition 6.1,
there are no further sets of lengths.
Suppose that G is an elementary 2-group of rank two, say G = {0, e1 , e2 , e1 + e2 }. Then A(G) =
{0, U = e1 e2 (e1 + e2 ), V1 = e21 , V2 = e22 , V3 = (e1 + e2 )2 }, and hence U 2 = V1 V2 V3 is the only minimal
relation. Now the proof runs along the same lines as above.
One big difficulty in all work on the Characterization Problem stems from the fact that most sets of
lengths over any finite abelian group are intervals. To make this precise we mention two deep results
without proof.
Theorem 6.3 (Sets of lengths which are intervals). Let G be a finite abelian group with |G| ≥ 3.
1. If A is a zero-sum sequence whose support supp(A) ∪ {0} is a subgroup of G, then L(A) is an
interval.
2. If R is the ring of integers of an algebraic number field K with class group G, then
lim
x→∞
#{aR | NK/Q (aR) ≤ x, L(a) is an interval}
= 1.
#{aR | NK/Q (aR) ≤ x}
The first statement is a result in additive combinatorics which can be found in [23, Theorem 7.6.8].
The limit formula is based on the first statement and on the analytic machinery of counting functions
[23, Theorem 9.4.11]. In the 1960s Narkiewicz initiated a systematic study of the asymptotic behavior of
counting functions associated with non-unique factorizations. We refer to the monographs [42, Chapters
7 and 9], [23, Chapters 8 and 9], and to [39] (analytic monoids, introduced in [39], are Krull monoids
which have an abstract norm function satisfying axioms which allow to develop a theory of L-functions).
In spite of Theorem 6.3 and having Propositions 6.1 and 6.2 at our disposal, we start with a more
detailed analysis of the Characterization Problem. We have seen that
L(C1 ) = L(C2 )
and L(C3 ) = L(C2 ⊕ C2 ) .
An abelian group G has Davenport constant D(G) ≤ 3 if and only if it is either cyclic of order |G| ≤ 3 or
isomorphic to C2 ⊕ C2 . Thus we focus on groups whose Davenport constant is at least four. Let G be a
finite abelian group with D(G) ≥ 4, say
G∼
= Cn1 ⊕ . . . ⊕ Cnr with 1 < n1 | . . . | nr
and set D∗ (G) = 1 +
r
X
(ni − 1) .
i=1
Clearly, the system L(G) depends only on G and hence on the group invariants (n1 , . . . , nr ). Thus L(G)
as a whole as well as the invariants controlling L(G) – such as the set of distances ∆(G) and the kth
22
ALFRED GEROLDINGER
elasticities ρk (G) – allow a description in (n1 , . . . , nr ). We demonstrate the complexity of such problems
by considering ρ2 (G).
By Proposition 6.1, we have ρ2 (G) = D(G). The Davenport constant D(G) is one of the most classical
zero-sum invariants which has been studied since the 1960s. If (e1 , . . . , er ) is a basis of G with ord(ei ) = ni
for each i ∈ [1, r], then
r
Y
A = (e1 + . . . + er )
eini −1 ∈ A(G) ,
i=1
and hence D∗ (G) = |A| ≤ D(G). It has been known since the 1960s that equality holds for p-groups
and groups of rank at most two [23, Theorem 5.8.3]. It is an open problem whether equality holds for
groups of rank three, but for every r ≥ 4 there are infinitely many groups G having rank r and for which
D∗ (G) < D(G) holds. We refer to [28, 47] for a survey of what is known on parameters controlling L(G).
We first show a simple finiteness result and then present one result (a proof can be found in [23, Theorem
6.6.3]) revealing characteristic phenomena of L(G) for cyclic groups and elementary 2-groups.
Lemma 6.4. Let G be a finite abelian group with D(G) ≥ 4. Then there are only finitely many abelian
groups G′ (up to isomorphism) such that L(G) = L(G′ ).
Proof. If G′ is an abelian group such that L(G′ ) = L(G), then Proposition 6.1 implies D(G) = ρ2 (G) =
ρ2 (G′ ) = D(G′ ) and hence D∗ (G′ ) ≤ D(G). Since there are only finitely many G′ (up to isomorphism)
such that D∗ (G′ ) is bounded above by a constant, there are only finitely many groups G′ for which
L(G′ ) = L(G) can hold.
Proposition 6.5. Let G be a finite abelian group with D(G) ≥ 4. Then {2, D(G)} ∈ L(G) if and only if
G is either cyclic or an elementary 2-group.
The next theorem gathers what is known on the Characterization Problem.
Theorem 6.6. Let G be a finite abelian group with D(G) ≥ 4, and let G′ be an abelian group with
L(G) = L(G′ ). Then G and G′ are isomorphic in each of the following cases.
1. G ∼
= Cn1 ⊕ Cn2 where n1 , n2 ∈ N with n1 | n2 and n1 + n2 > 4.
2. G is an elementary 2-group.
3. G ∼
= Cnr where r, n ∈ N with n ≥ 2 and 2r < n − 2.
4. D(G) ≤ 11.
Proof of a special case. We give a sketch of the proof for cyclic groups and for elementary 2-groups. Let
G be either cyclic or an elementary 2-group with D(G) ≥ 4, and let G′ be any abelian group with
L(G) = L(G′ ). Since D(G) = ρ2 (G) = ρ2 (G′ ) = D(G′ ) by Proposition 6.1 and since {2, D(G)} ∈ L(G) by
Proposition 6.5, it follows that {2, D(G′ )} ∈ L(G′ ). Again Proposition 6.5 implies that G′ is either cyclic
or an elementary 2-group. There are two proofs showing that the system of sets of lengths of cyclic groups
and that of elementary 2-groups (with the same Davenport constant) are distinct ([28, Corollary 5.3.3,
page 77] or [23, Theorem 7.3.3]), and neither of them is elementary. Both proofs use the Savchev-Chen
Structure Theorem for long zero-sum free sequences over cyclic groups ([28, Theorem 5.1.8, page 61], [34,
Chapter 11]) or related statements. To discuss one approach, let k ∈ N and consider the inequality for
ρ2k+1 (G) given in Proposition 6.1. Elementary examples show in case of elementary 2-groups that we
have equality on the right side, whereas for cyclic groups we have equality on the left side ([28, Theorem
5.3.1, page 75]. Detailed proofs can be found in [28, Corollary 5.3.3, page 77] and [23, Theorem 7.3.3].
Now suppose that G has rank two, say G ∼
= Cn1 ⊕Cn2 where n1 , n2 ∈ N with 1 < n1 | n2 and n1 +n2 > 4,
and let G′ be any abelian group such that L(G) = L(G′ ). The proof that G and G′ are isomorphic has
two main ingredients. First, it is based on the characterization of all minimal zero-sum sequences over G
of length D(G). This has been done in a series of papers by Gao, Geroldinger, Grynkiewicz, Reiher, and
Schmid (see [7] for the characterization and detailed references). Second, it is based on the Structure
SETS OF LENGTHS
23
Theorem for Sets of Lengths (Theorem 5.3), on an associated inverse result ([23, Proposition 9.4.9]), and
on a detailed study of the set of minimal distances ∆∗ (G) ([33]), which is defined as
∆∗ (G) = {min ∆(G0 ) | G0 ⊂ G with ∆(G0 ) 6= ∅} ⊂ ∆(G) .
Detailed proofs of 1., 3., and 4. can be found in [29, 32, 51].
We end this survey with the conjecture stating that the Characterization Problem has a positive answer
for all finite abelian groups G having Davenport constant D(G) ≥ 4.
Conjecture 6.7. Let G be a finite abelian group with D(G) ≥ 4. If G′ is an abelian group with L(G) =
L(G′ ), then G and G′ are isomorphic.
References
[1] D.D. Anderson (ed.), Factorization in Integral Domains, Lect. Notes Pure Appl. Math., vol. 189, Dekker, New York,
1997.
[2] D. Bachman, N. Baeth, and J. Gossell, Factorizations of upper triangular matrices, Linear Algebra Appl. 450 (2014),
138 – 157.
[3] N.R. Baeth, A. Geroldinger, D.J. Grynkiewicz, and D. Smertnig, A semigroup-theoretical view of direct-sum decompositions and associated combinatorial problems, J. Algebra Appl. 14 (2015), 1550016 (60 pages).
[4] N.R. Baeth and D. Smertnig, Factorization theory: From commutative to noncommutative settings, J. Algebra 441
(2015), 475 – 551.
[5] N.R. Baeth and R. Wiegand, Factorization theory and decomposition of modules, Amer. Math. Monthly 120 (2013),
3 – 34.
[6] P. Baginski and S.T. Chapman, Factorizations of algebraic integers, block monoids, and additive number theory, Amer.
Math. Monthly 118 (2011), 901 – 920.
[7] P. Baginski, A. Geroldinger, D.J. Grynkiewicz, and A. Philipp, Products of two atoms in Krull monoids and arithmetical
characterizations of class groups, Eur. J. Comb. 34 (2013), 1244 – 1268.
[8] P.-J. Cahen and J.-L. Chabert, Integer-Valued Polynomials, vol. 48, Amer. Math. Soc. Surveys and Monographs, 1997.
, What you should know about integer-valued polynomials, Amer. Math. Monthly 123 (2016), 311 – 337.
[9]
[10] S.T. Chapman, A tale of two monoids: a friendly introduction to nonunique factorizations, Math. Magazine 87 (2014),
163 – 173.
[11] S.T. Chapman, M. Fontana, A. Geroldinger, and B. Olberding (eds.), Multiplicative Ideal Theory and Factorization
Theory, Proceedings in Mathematics and Statistics, vol. 170, Springer, 2016.
[12] S.T. Chapman and W.W. Smith, Factorization in Dedekind domains with finite class group, Isr. J. Math. 71 (1990),
65 – 95.
[13] J. Coykendall and W.W. Smith, On unique factorization domains, J. Algebra 332 (2011), 62 – 70.
[14] M. Delgado, P.A. Garcı́a-Sánchez, and J. Morais, “numericalsgps”: a gap package on numerical semigroups,
(http://www.gap-system.org/Packages/numericalsgps.html).
[15] A. Facchini, Direct sum decomposition of modules, semilocal endomorphism rings, and Krull monoids, J. Algebra 256
(2002), 280 – 307.
[16] M. Fontana, E. Houston, and T. Lucas, Factoring Ideals in Integral Domains, Lecture Notes of the Unione Matematica
Italiana, vol. 14, Springer, 2013.
[17] S. Frisch, A construction of integer-valued polynomials with prescribed sets of lengths of factorizations, Monatsh. Math.
171 (2013), 341 – 350.
[18]
, Relative polynomial closure and monadically Krull monoids of integer-valued polynomials, in Multiplicative
Ideal Theory and Factorization Theory. Ed. S. T. Chapman, M. Fontana, A. Geroldinger, and B. Olberding. Springer,
2016, pp. 145 – 157.
[19] W. Gao and A. Geroldinger, On products of k atoms, Monatsh. Math. 156 (2009), 141 – 157.
[20] P.A. Garcı́a-Sánchez, An overview of the computational aspects of nonunique factorization invariants, in Multiplicative
Ideal Theory and Factorization Theory. Ed. S.T. Chapman, M. Fontana, A. Geroldinger, and B. Olberding. Springer,
2016, pp. 159 – 181.
[21] A. Geroldinger, Über nicht-eindeutige Zerlegungen in irreduzible Elemente, Math. Z. 197 (1988), 505 – 529.
, Non-commutative Krull monoids: a divisor theoretic approach and their arithmetic, Osaka J. Math. 50 (2013),
[22]
503 – 539.
[23] A. Geroldinger and F. Halter-Koch, Non-Unique Factorizations. Algebraic, Combinatorial and Analytic Theory, Chapman & Hall/CRC, Boca Raton, FL, 2006.
[24] A. Geroldinger and F. Kainrath, On the arithmetic of tame monoids with applications to Krull monoids and Mori
domains, J. Pure Appl. Algebra 214 (2010), 2199 – 2218.
24
ALFRED GEROLDINGER
[25] A. Geroldinger, F. Kainrath, and A. Reinhart, Arithmetic of seminormal weakly Krull monoids and domains, J. Algebra
444 (2015), 201 – 245.
[26] A. Geroldinger and G. Lettl, Factorization problems in semigroups, Semigroup Forum 40 (1990), 23 – 38.
[27] A. Geroldinger, S. Ramacher, and A. Reinhart, On v-Marot Mori rings and C-rings, J. Korean Math. Soc. 52 (2015),
1 – 21.
[28] A. Geroldinger and I. Ruzsa, Combinatorial Number Theory and Additive Group Theory, Advanced Courses in Mathematics - CRM Barcelona, Birkhäuser, 2009.
[29] A. Geroldinger and W. A. Schmid,
A characterization of class groups via sets of lengths,
http://arxiv.org/abs/1503.04679.
[30] A. Geroldinger, W. A. Schmid, and Q. Zhong, Systems of sets of lengths: transfer Krull monoids versus weakly Krull
monoids, http://arxiv.org/abs/1606.05063.
[31] A. Geroldinger and P. Yuan, The set of distances in Krull monoids, Bull. Lond. Math. Soc. 44 (2012), 1203 – 1208.
[32] A. Geroldinger and Q. Zhong, A characterization of class groups via sets of lengths II, J. Théor. Nombres Bordx.
(forthcoming).
[33]
, The set of minimal distances in Krull monoids, Acta Arith. 173 (2016), 97 – 120.
[34] D.J. Grynkiewicz, Structural Additive Theory, Developments in Mathematics, Springer, 2013.
[35] F. Halter-Koch, Finitely generated monoids, finitely primary monoids and factorization properties of integral domains,
in Factorization in Integral Domains. Ed. D.D. Anderson. Lect. Notes Pure Appl. Math., vol. 189, Dekker, New York,
1997, pp. 73 – 112.
, Ideal Systems. An Introduction to Multiplicative Ideal Theory, Marcel Dekker, 1998.
[36]
[37] G. Hegedüs, Z. Li, J. Schicho, and H.P. Schröcker, From the fundamental theorem of algebra to Kempe’s universality
theorem, Internat. Math. Nachrichten 229 (2015), 13–26.
[38] B. Jacobson, Matrix number theory : an example of non-unique factorization, Amer. Math. Monthly 72 (1965), 399 –
402.
[39] J. Kaczorowski, Analytic monoids and factorization problems, Semigroup Forum (forthcoming).
[40] F. Kainrath, Factorization in Krull monoids with infinite class group, Colloq. Math. 80 (1999), 23 – 30.
[41] P. Malcolmson and F. Okoh, Half-factorial subrings of factorial domains, J. Pure Appl. Algebra 220 (2016), 877 – 891.
[42] W. Narkiewicz, Elementary and Analytic Theory of Algebraic Numbers, 3rd ed., Springer, 2004.
[43] A. Reinhart, On the divisor-class group of monadic submonoids of rings of integer-valued polynomials, Commun.
Korean Math. Soc. (forthcoming).
, On monoids and domains whose monadic submonoids are Krull, in Commutative Algebra. Recent Advances in
[44]
Commutative Rings, Integer-Valued Polynomials, and Polynomial Functions. Ed. M. Fontana, S. Frisch, and S. Glaz.
Springer, 2014, pp. 307 – 330.
[45] M. Roitman, Half-factorial domains, manuscript.
[46] W.A. Schmid, A realization theorem for sets of lengths, J. Number Theory 129 (2009), 990 – 999.
, Some recent results and open problems on sets of lengths of Krull monoids with finite class group, in Multi[47]
plicative Ideal Theory and Factorization Theory. Ed. S.T. Chapman, M. Fontana, A. Geroldinger, and B. Olberding.
Springer, 2016, pp. 323 – 352.
[48] D. Smertnig, Factorizations in bounded hereditary noetherian prime rings, http://arxiv.org/abs/1605.09274.
[49]
, Sets of lengths in maximal orders in central simple algebras, J. Algebra 390 (2013), 1 – 43.
, Factorizations of elements in noncommutative rings: A Survey, in Multiplicative Ideal Theory and Factorization
[50]
Theory. Ed. S.T. Chapman, M. Fontana, A. Geroldinger, and B. Olberding. Springer, 2016, pp. 353 – 402.
[51] Q. Zhong,
Sets of minimal distances and characterizations of class groups of Krull monoids,
http://arxiv.org/abs/1606.08039.
University of Graz, NAWI Graz, Institute for Mathematics and Scientific Computing, Heinrichstraße 36,
8010 Graz, Austria
E-mail address: [email protected]
| 4math.GR
|
TORSION FUNCTORS WITH MONOMIAL SUPPORT
arXiv:1207.0610v3 [math.AC] 2 Apr 2014
FRED ROHRER
Abstract. The dependence of torsion functors on their supporting ideals is
investigated, especially in the case of monomial ideals of certain subrings of
polynomial algebras over not necessarily Noetherian rings. As an application
it is shown how flatness of quasicoherent sheaves on toric schemes is related
to graded local cohomology.
Introduction
Over a Noetherian ring (where rings are always understood to be commutative),
torsion functors (and hence local cohomology functors) depend only on the radical of
their supporting ideals. Without Noetherianness this need not hold. Supposing the
supporting ideals to be monomial ideals of finite type of a polynomial algebra one
might hope for it to still hold – after all, monomial ideals behave quite independently
of the base ring and hence might be unaffected by its potential lack of some nice
properties. We will see that this is indeed true. However, the torsion functors with
respect to a monomial ideal of finite type and its radical may be different, unless
the nilradical of the base ring is nilpotent. But using a monomial variant of the
notion of radical it is possible to get a satisfying result that is independent of the
base ring and moreover can be generalised to a class of subalgebras of polynomial
algebras.
There were two main reasons for writing this note. First, the recent preprint
[2] by Botbol and Chardin on Castelnuovo-Mumford regularity over not necessarily Noetherian polynomial algebras contains the claim that local cohomology with
respect to monomial ideals of finite type remains the same when we replace the
supporting ideal by its radical. This aroused suspicions and turned out to be indeed wrong. Luckily, the aforementioned monomial variant of radical can save the
day, and the moral might be that working over not necessarily Noetherian rings
requires some meticulousness even in a monomial setting.
The second reason comes from the study of toric schemes, i.e., “toric varieties
over arbitrary base rings”. An overview of this continuing project including its
motivation can be found in [8]. With this approach to toric geometry one gets
a relation between flatness of quasicoherent sheaves on a toric scheme and graded
Čech cohomology over a corresponding restricted Cox ring with respect to a certain
sequence of monomials. The ideal J generated by these monomials is not necessarily
equal to the restricted irrelevant ideal I, and thus – in case graded Čech cohomology
coincides with graded local cohomology (which it need not do in general) – the
question arises whether or not graded local cohomology distinguishes between J
2010 Mathematics Subject Classification. Primary 13D45; Secondary 14M25.
Key words and phrases. Torsion functor, local cohomology, monomial ideal, Čech cohomology,
flatness, Cox scheme.
The author was supported by the Swiss National Science Foundation.
1
2
FRED ROHRER
and I. By applying our general results we will see that it does not, and thus
we can relate flatness of quasicoherent sheaves on toric schemes with graded local
cohomology with respect to I. (The importance of graded local cohomology with
respect to I stems from its relation to sheaf cohomology as given by the toric
Serre-Grothendieck correspondence ([9, 4.5]).)
1. Equality of torsion functors: The general case
In this section we search for conditions under which torsion functors over arbitrary rings depend only on the radical of the supporting ideals. For the sake of
generality we consider throughout a graded setting, even if the graduation is not
relevant. So, let G be a commutative group and let R be a G-graded ring.
By a projective system of graded ideals we mean a projective system of graded
ideals of R over an ordered set. We denote such a system in the form A = (aj )j∈J ,
where J is understood to be an ordered set. For a projective system of graded ideals
A = (aj )j∈J we denote by ΓA the G-graded A-torsion functor, i.e., the subfunctor
of the identity functor on the category GrModG (R) of G-graded R-modules with
S
ΓA (M ) = j∈J (0 :M aj ) for a G-graded R-module M . Extending our ambient
universe (cf. [1, I.0]) we can consider the (potentially big) set
T(R) := {ΓA | A projective system of graded ideals}
of subfunctors of the identity functor as an ordered set by means of its canonical
ordering. If A = (aj )j∈J and B = (bk )k∈K are two projective systems of graded
ideals, then A is called coarser than B if for every j ∈ J there exists k ∈ K
with bk ⊆ aj ; this defines a preorder relation. Two projective systems of graded
ideals are called equivalent if each of them is coarser than the other; this defines
an equivalence relation. Slightly generalising [3, III.1.2] (and keeping in mind [3,
II.6.9]) we get representatives of equivalence classes under this relation, and thus
obtain an order relation associated with the above preorder relation. This order
relation is readily seen to have a graph, and so we end up with a (small) ordered set
(R) whose elements are representatives of equivalence classes of projective systems
of graded ideals. The relation between (R) and (R) is given by the following
proposition, where by an isomorphism of ordered sets we understand – accordingly
to [3, III.1.3] – an increasing bijection with increasing inverse.
S
S
T
(1.1) Proposition There is an isomorphism of ordered sets
∼
=
S(R) −→
T(R), A 7→ ΓA.
Proof. The map in question is obviously a surjective morphism of ordered sets.
Let A = (aj )j∈J and B = (bk )k∈K be projective systems of graded ideals such
that ΓA is a subfunctor of ΓB . If j ∈ J then R/aj is annihilated by aj , and so
R/aj = ΓA (R/aj ) ⊆ ΓB (R/aj ). Hence, there exists k ∈ K with bk ⊆ aj , and thus
A is coarser than B. The claim follows from this.
T
In particular, it follows that the set (R) is small. For a graded ideal a ⊆ R we
consider the projective system of graded ideals A = (an )n∈N and set Γa := ΓA , and
we denote by Ta the a-adic topology on R. If a, b ⊆ R are two graded ideals, then
(an )n∈N is coarser than (bn )n∈N if and only if Ta is coarser than Tb ([5, III.2.5]).
TORSION FUNCTORS WITH MONOMIAL SUPPORT
3
(1.2) Corollary There is an isomorphism of ordered sets
∼
=
{Ta | a ⊆ R graded ideal } −→ {Γa | a ⊆ R graded ideal }.
(1.3) Corollary a) If a ⊆ R is a graded ideal, then Γa = Γ√a holds if and only
√ n
if there exists n ∈ with a ⊆ a.
√
√
b) If a, b ⊆ R are graded ideals, then Γa = Γb implies
a
=
b, and the converse
√ n
√ m
holds if there exist m, n ∈ with a ⊆ a and b ⊆ b.
N
N
If a ⊆ R is a graded ideal whose radical is of finite type then Γa = Γ√a holds by
1.3 a). Hence, if R is Noetherian (where Noetherianness of a graded ring is always
understood to mean that every graded ideal is of finite type) then we have Γa = Γ√a
for every graded ideal a ⊆ R. Similarly, the hypothesis for the converse to hold in
1.3 b) is fulfilled if the radicals of a and b are of finite type, hence in particular if
R is Noetherian. The seemingly weaker hypothesis that every perfect graded ideal
of R is of finite type yields no improvement – it is equivalent to Noetherianness of
R by a graded version of Cohen’s Theorem [6, 0.6.4.7], proved analogously to the
ungraded one.
The above shows that even when considering only graded ideals of finite type we
do not have a satisfying result. Therefore, we will restrict our attention to a more
specific situation in the next section.
2. Equality of torsion functors: The monomial case
In this section we consider the following situation. Let A be a ring, let I be
a set, and let ψ : ⊕I ։ G be an epimorphism of groups. We denote by R the
polynomial algebra A[(Xi )i∈I ] in indeterminates (Xi )i∈I over A, furnished with
the G-graduation derived from its canonical ⊕I -graduation by means of ψ. By a
Q
monomial (in R) we mean an element of R of the form i∈I Xiµi with µ = (µi )i∈I ∈
⊕I
, and such an element is homogeneous of degree ψ(µ). By a monomial ideal (of
R) we mean an ideal of R generated by a set of monomials in R. Monomial ideals
are graded, but the converse is not necessarily true. For a subset E ⊆ R we write
hEiR or, if no confusion can arise, hEi for the ideal of R generated by E.
The ordered set ⊕I (whose ordering is induced by the product ordering on I )
is Artinian. Therefore, for a monomial ideal a ⊆ R there exists a unique subset
Q
E ⊆ ⊕I such that { i∈I Xiµi | µ ∈ E} is a minimal generating set of a; we denote
this set by E(a) and define a monomial ideal
Q
⌊a⌋ := h i∈I,µi >0 Xi | µ ∈ E(a)iR .
Z
Z
N
N
N
N
This contains but not necessarily equals a.
(2.1) Lemma If a ⊆ R is a monomial ideal, then a is of finite type if and only
if ⌊a⌋ is so.
Proof. If a is of finite type then E(a) is finite, hence ⌊a⌋ is of finite type, too.
Conversely, suppose that ⌊a⌋ is of finite type, and assume that a is not of finite
type. Then, by the pigeonhole principle there exists an infinite subset F ⊆ E(a)
Q
Q
such that for µ, ν ∈ F we have i∈I,µi >0 Xi = i∈I,νi >0 Xi . Thus, there exists a
Q
finite subset J ⊆ I with F ⊆ ⊕J . But then F = E(h i∈J Xiµi | µ ∈ F i) is the
minimal set of generators of a monomial ideal in a polynomial ring in finitely many
indeterminates, hence finite – a contradiction.
N
4
FRED ROHRER
For a ring B we denote by Nil(B) its nilradical. If Nil(B) is finitely generated
then it is nilpotent, but the converse is not necessarily true. In our situation we have
Nil(R) = Nil(A)R, i.e., a polynomial in R is nilpotent if and only if its coefficients
are nilpotent ([4, IV.1.5 Proposition 9]). Hence, Nil(R) is nilpotent or of finite type,
respectively, if and only if Nil(A) is so.
√
(2.2) Lemma a) For a monomial ideal a ⊆ R we√have a = Nil(R) + ⌊a⌋.
√
b) For monomial ideals a, b ⊆ R we have a = b if and only if ⌊a⌋ = ⌊b⌋.
√
√
Proof. a) Clearly, Nil(R) + ⌊a⌋ is contained in a. Conversely, let f ∈ a. Every
term occurring in f being either nilpotent or not, and the sum of the nilpotent
terms being nilpotent, we can write f = g + h with g ∈ Nil(R) and h ∈ R such
that no term occurring in h is nilpotent. So, it suffices to show h ∈ ⌊a⌋. As
√
√
f, g ∈ a it follows h ∈ a, and hence we can suppose that no term occurring
in f is nilpotent. We prove now by induction on the number r of terms occurring
in f that f ∈ ⌊a⌋. If r = 0 this is clear. Suppose that r > 0 and that the claim
holds for strictly smaller values of r. Furnishing ⊕I with a structure of totally
ordered group (e.g., the lexicographic product with respect to some well-ordering
of I) there occurs a term t in f whose ⊕I -degree is the greatest of the ⊕I -degrees
of the terms occurring in f . If n ∈ with f n ∈ a then tn is a term occurring in f n
whose ⊕I -degree is not shared by any other term occurring in f n . Hence, as a is
√
Q
a monomial ideal, i∈I Xiµi divides tn for some µ ∈ E(a), and thus t ∈ ⌊a⌋ ⊆ a.
√
We get f ′ := f − t ∈ a. As the number of terms occurring in f ′ is strictly smaller
than r it follows f ′ ∈ ⌊a⌋,√hence f = f ′ + t ∈ ⌊a⌋ as desired.
√
b) Suppose that a = b and therefore ⌊a⌋ ⊆ Nil(R)+⌊b⌋ by a). For µ ∈ E(a) it
Q
Q
follows i∈I,µi >0 Xi ∈ ⌊a⌋ ⊆ Nil(R) + ⌊b⌋, hence either i∈I,µi >0 Xi ∈ Nil(R) and
Q
thus R = 0, or i∈I,µi >0 Xi ∈ ⌊b⌋. So, by reasons of symmetry we get ⌊a⌋ = ⌊b⌋.
The converse follows from a).
Z
N
Z
Z
Z
(2.3) Proposition a) For a monomial ideal a ⊆ R of finite type we have Γa =
Γ⌊a⌋ .
b) If a, b ⊆ R are monomial ideals of finite type, then
√ the following statements
√
are equivalent: (i) Γa = Γb ; (ii) ⌊a⌋ = ⌊b⌋; (iii) a = b.
N
Proof. a) For µ ∈ E(a) and m ∈
with m ≥ µi for every i ∈ I we have
Q
m
( i∈I,µi >0 Xi ) ∈ a. As ⌊a⌋ is of finite type by 2.1 this implies that there exists m ∈ with ⌊a⌋m ⊆ a. Hence, as a ⊆ ⌊a⌋ the claim follows from 1.2.
b) (i) implies (iii) by 1.3 b), (iii) implies (ii) by 2.2 b), and (ii) implies (i) by
a).
N
This answers our original question in a satisfying way. But since the radical of
a monomial ideal of finite type is not necessarily a monomial ideal we cannot infer
that Γa = Γ√a . Moreover, in case of monomial ideals that are not of finite type
2.3 b) is no longer true. Let us provide some examples of good and bad behaviour.
N
(2.4) Examples Let I = >0 and ψ = IdZ⊕I , let a ⊆ R be a monomial ideal of
finite type, and let b := hXii | i ∈ IiR . If A = [(Yk )k∈N ]/hYk2 | k ∈ Ii then we
have Γa = Γ⌊a⌋ = Γ√a and Γb 6= Γ⌊b⌋ = Γ√b . If A = [(Yk )k∈N ]/hYkk | k ∈ Ii then
we have Γa = Γ⌊a⌋ 6= Γ√a and Γb 6= Γ⌊b⌋ 6= Γ√b 6= Γb .
Q
Q
TORSION FUNCTORS WITH MONOMIAL SUPPORT
5
These examples already hint at the condition on the base ring for Γ⌊a⌋ and Γ√a
(and in case a is of finite type Γa and Γ√a ) to coincide, as presented in the next
result.
(2.5) Proposition The following statements are equivalent: (i) Nil(A) is nilpotent; (ii) For every monomial ideal a ⊆ R we have Γ⌊a⌋ = Γ√a ; (iii) There exists a
monomial ideal a ⊆ R with Γ⌊a⌋ = Γ√a .
N
Proof. “(i)⇒(ii)”: Immediately by 1.2 and 2.2 a). “(iii)⇒(i)”: There exists n ∈
√ n
with a ⊆ ⌊a⌋ (1.2), implying Nil(A)n ⊆ Nil(R)n ∩ A ⊆ ⌊a⌋ ∩ A = 0 (2.2 a)).
Sometimes – for example in toric geometry – it is necessary to consider a slightly
more general situation than the one above, namely monomial ideals in subrings of
R obtained by degree restriction to a subgroup of finite index of G. So, from now
on let H ⊆ G be a subgroup of finite index. We consider the degree restriction
Q
S := R(H) of R to H, i.e., the sub-A-algebra A[ i∈I Xiµi | µ ∈ ⊕I ∧ ψ(µ) ∈ H] of
R generated by the monomials with degree in H, furnished with the H-graduation
Q
such that deg( i∈I Xiµi ) = ψ(µ). By a monomial in S we mean a monomial in
R contained in S, and by a monomial ideal of S we mean an ideal of S generated
by monomials in S. Monomial ideals of S are graded, but the converse is not
necessarily true. For a monomial ideal a of R we consider the degree restriction
a(H) := a ∩ S to H. This is a monomial ideal of S, and if a is of finite type then
so is a(H) (as is seen analogously to [9, 3.1.3 a)]). Conversely, if a is a monomial
ideal (of finite type) of S then haiR is a monomial ideal (of finite type) of R, and
we have (haiR )(H) = a ([9, 2.2.3]). Therefore, every monomial ideal (of finite type)
of S is of the form a(H) for some monomial ideal a (of finite type) of R.
In this situation there are two ways of forming a “monomial radical” of a monomial ideal. We will show now that both yield the same torsion functors, but only
one of them can be used to characterise equality of torsion functors. So, let a ⊆ R
be a monomial ideal. Since H has finite index in G it follows that for µ ∈ E(a)
Q
with the property that i∈I,µi >0 Xim ∈ a(H) , and
there exists a minimal m ∈
we denote this number by mµ . Then,
Q
m
⌊a⌋H := h i∈I,µi >0 Xi µ | µ ∈ E(a)iS ⊆ S
N
N
is a monomial ideal. On the other hand, we can also consider the monomial ideal
⌊a⌋(H) ⊆ S, obtained by degree restriction from ⌊a⌋. Clearly, both these constructions generalise the one above, corresponding to the case H = G.
It is readily checked that a(H) ⊆ ⌊a⌋H ⊆ ⌊a⌋(H) , but none of these inclusions is
necessarily an equality. For the first this is clear by considering the case H = G.
For the second we suppose that G = , deg(X0 ) = 1, and H = 2 . Then, setting
a = hX04 iR we get a(H) = ⌊a⌋H = hX04 iS $ hX02 iS = hai(H) (provided A 6= 0).
Z
Z
(2.6) Lemma If a ⊆ R is a monomial ideal, then a is of finite type if and only
if ⌊a⌋H is so.
Proof. Analogously to 2.1.
(2.7) Proposition For a monomial ideal a ⊆ R of finite type we have
Γa(H) = Γ⌊a⌋H = Γ⌊a⌋(H) .
6
FRED ROHRER
N
Proof. Let t ∈ ⌊a⌋(H) be a monomial. Then, there exist µ ∈ E(a) and ν ∈ ⊕I
Q
Q
with t = i∈I,µi >0 Xi1+νi · i∈I,µi =0 Xiνi , and there exists n ∈
such that for
i ∈ I with µi > 0 we have n(1 + νi ) ≥ µi . This implies that tn is a multiple of
Q
µi
i∈I Xi and therefore belongs to a. Since ⌊a⌋(H) ⊆ S is by 2.1 a monomial ideal
of finite type, the claim follows from 1.2.
N
(2.8) Proposition Let a, b ⊆ R be monomial ideals of finite type.
√
√
a) The following statements are equivalent: (i) Γa(H) = Γb(H) ; (ii) a = b;
(iii) ⌊a⌋ = ⌊b⌋; (iv) ⌊a⌋(H) = ⌊b⌋(H) .
b) If ⌊a⌋H = ⌊b⌋H then Γa(H) = Γb(H) .
p
p
Proof. a) “(i)⇒(iii)”: If Γa(H) = Γb(H) then ⌊a⌋H = ⌊b⌋H holds (2.7, 1.3). Let
Q
nm
µ ∈ E(a). There exists n ∈ >0 with i∈I,µi >0 Xi µ ∈ ⌊b⌋H , so there exists
Q
Q
nm
ν ∈ E(b) such that i∈I,νi >0 Ximν divides i∈I,µi >0 Xi µ . If mν = 0 or mµ = 0
Q
then we get a = R = b and thus the claim. Otherwise, i∈I,νi >0 Xi divides
Q
Q
i∈I,µi >0 Xi ∈ ⌊b⌋ and thus by symmetry the claim. The
i∈I,µi >0 Xi , implying
remaining implications follow trivially
or by
p
p 2.3 b) and 2.7.
b) The hypothesis implies ⌊a⌋H = ⌊b⌋H , and the claim follows with the
same argument as the implication “(i)⇒(iii)” in a).
N
Z
The converse of 2.8 b) is not true. Indeed, supposing again that G = ,
deg(X0 ) = 1, and H = 2 , and setting a = hX04 iR and b = hX02 iR , we get
⌊a⌋ = hX0 iR = ⌊b⌋, but ⌊a⌋H = hX04 iS $ hX02 iS = ⌊b⌋H (provided A 6= 0).
Z
3. Application: Flatness of quasicoherent sheaves on Cox schemes
In this last section we will apply our results to toric geometry. For the necessary
background and unexplained notations we refer the reader to [9].
Before recalling the toric setting we have to make some general remarks on
flatness and Čech cohomology. During these, let G be a commutative group, let
A be a ring, and let R be a G-graded ring such that R0 is an A-algebra. For a
G-graded R-module M we set degsupp(M ) := {g ∈ G | Mg 6= 0}. For a finite
sequence a = (aj )nj=1 in Rhom we denote by C(a, M ) the G-graded Čech cocomplex
of M with respect to a, by C(a, M )i its i-th component for i ∈ , and by H i (a, M )
the i-th G-graded Čech cohomology of M with respect to a for i ∈ . Furthermore,
for a graded ideal a ⊆ R and i ∈ we denote by Hai (M ) the i-th G-graded local
cohomology of M with respect to a. One should note that if R is not Noetherian
then Čech cohomology with respect to a and local cohomology with respect to the
ideal of R generated by a do not necessarily coincide.
Z
Z
Z
(3.1) Lemma Let R′ be a flat G-graded R-algebra, and let M be a G-graded
R-module. If M is flat over A, then so is R′ ⊗R M .
Proof. The functor (R′ ⊗R M ) ⊗A • : Mod(A) → Mod(A) is the composition of
R′ ⊗R (M ⊗A •) : Mod(A) → GrModG (R′ )
with the forgetful functor from GrModG (R′ ) to Mod(R′ ) and the scalar restriction
from Mod(R′ ) to Mod(A), and all these three functors are exact.
(3.2) Lemma Let a = (aj )nj=1 be a finite sequence in Rhom , and let M be a Ggraded R-module. If Maj is flat over A for every j ∈ [1, n], then so is C(a, M )ig for
every i ∈ >0 and every g ∈ G.
N
TORSION FUNCTORS WITH MONOMIAL SUPPORT
7
Proof. Using 3.1 and [5, I.2.3 Proposition 2] this follows immediately from the
definition of the Čech cocomplex.
(3.3) Proposition Let a = (aj )nj=1 be a finite sequence in Rhom , and let M be
a G-graded R-module. If Maj is flat over A for every j ∈ [1, n], then so is Mg for
Sn
every g ∈ G \ i=0 degsupp(H i (a, M )).
Sn
Proof. Taking components of degree g ∈ G \ i=0 degsupp(H i (a, M )) of C(a, M )
yields an exact sequence
0 → C(a, M )0g → C(a, M )1g → · · · → C(a, M )gn−1 → C(a, M )ng → 0
of A-modules. Using 3.2 and [5, I.2.5 Proposition 5] the claim follows then by
induction on n.
R
Now we are ready to consider the toric setting. Let V be an -vector space
of finite dimension, let N be a -structure on V , let Σ be an N -fan in V , let Σ1
denote the set of 1-dimensional cones in Σ, and let R be a ring. The combinatorics
of Σ give rise to a group A and an epimorphism a : Σ1 ։ A, as well as to a
subgroup Pic(Σ) ⊆ A, the so-called Picard group of Σ. Additionally, we choose
a subgroup B ⊆ A of finite index. The Cox ring S associated with Σ1 over R
is defined as the polynomial algebra R[(Zρ )ρ∈Σ1 ] furnished with the A-graduation
derived from its canonical Σ1 -graduation by means of a, and the B-restricted Cox
ring SB is defined as its degree restriction to B. For σ ∈ Σ we define a monomial
Q
bσ :=
with the
Z
ρ∈Σ1 \σ Zρ ∈ S. For σ ∈ Σ there exists a minimal m ∈
m
b
property that Zσ ∈ SB , and we denote this number by mσ . We denote the ring
bmσ by SB,σ . Its component of degree
of fractions obtained from SB by inverting Z
Z
Z
Z
N
σ
0 is independent of the choice of B and is denoted by S(σ) . The facial relation in
Σ allows us to glue the R-schemes Spec(S(σ) ) to obtain an R-scheme Y := YΣ (R),
called the Cox scheme associated with Σ over R. There is a canonical morphism
of R-schemes from Y to the toric scheme associated with Σ over R, which is an
S
isomorphism if and only if Σ is full, i.e., h ΣiR = V . Hence, studying Cox schemes
is essentially the same as studying toric schemes.
There is an essentially surjective functor SB from the category of B-graded SB modules to the category of quasicoherent OY -modules, allowing to translate the
study of quasicoherent sheaves on Cox schemes (and hence on toric schemes) into
the study of B-graded SB -modules. For example, in [9, 4.2.7] it was shown that if
F is a B-graded SB -module such that SB,σ ⊗SB F is flat over R for every σ ∈ Σ,
then SB (F ) is flat over R, and the converse holds if B ⊆ Pic(Σ) (in which case Σ
is simplicial by [9, 1.4.5]). Thus, after choosing a counting Zb of (Zbσmσ )σ∈Σ we can
relate flatness of SB (F ) with Čech cohomology of F .
(3.4) Proposition Suppose that B ⊆ Pic(Σ), and let F be a B-graded SB module. If SB (F ) is flat over R, then Fα is flat over R for every α ∈ B \
S
i b
i∈N degsupp(H (Z, F )).
Proof. Immediately from the above and 3.3.
bσ | σ ∈ ΣiS of S is called the irrelevant ideal, and
The monomial ideal I := hZ
its degree restriction IB ⊆ SB to B is a monomial ideal of finite type. Under the
condition that SB has the ITI-property with respect to IB , i.e., that ΓIB preserves
injectivity of B-graded SB -modules, the toric Serre-Grothendieck correspondence
8
FRED ROHRER
([9, 4.5.4]) states that we can study sheaf cohomology on Y in terms of B-graded
local cohomology over SB with support in IB . Thus, local cohomology with respect
to IB seems to be an appropriate tool in our setting, and the question arises on
b Under the condition that SB
its relation with Čech cohomology with respect to Z.
has the ITI-property with respect to hZbσmσ iSB for every σ ∈ Σ, our results so far
provide a satisfying answer.
(3.5) Corollary Suppose that B ⊆ Pic(Σ) and that SB has the ITI-property with
respect to hZbσmσ iSB for every σ ∈ Σ, and let F be a B-graded SB -module. If SB (F )
S
is flat over R, then Fα is flat over R for every α ∈ B \ i∈N degsupp(HIiB (F )).
bσmσ | σ ∈ ΣiSB = ⌊I⌋B this follows from 2.7, 3.4 and the proof of
Proof. Since hZ
[10, Proposition 7].
It remains the question about a converse of the above result, i.e., Does flatness
of certain components of F imply flatness of SB (F )? Considering the analogous
results in the case of projective schemes one should be able to tackle this problem
using bounds for some notion of Castelnuovo-Mumford regularity appropriate to
our setting (cf. [2], [7]). We hope to address this in some future work.
Acknowledgement: I thank the referee for the very careful reading and for his
comments and suggestions.
References
[1] M. Artin, A. Grothendieck, J. L. Verdier, Séminaire de géométrie algébrique 4. Théorie des
topos et cohomologie étale des schémas I. Lecture Notes in Math. 269. Springer, Berlin, 1972.
[2] N. Botbol, M. Chardin, G-graded Castelnuovo-Mumford regularity. Preprint (2011).
arXiv:1107.2494
[3] N. Bourbaki, Éléments de mathématique: Théorie des ensembles. Chapitres 1 à 4. Hermann,
Paris, 1970.
[4] N. Bourbaki, Éléments de mathématique: Algèbre. Chapitres 4 à 7. Masson, Paris, 1981.
[5] N. Bourbaki, Éléments de mathématique: Algèbre commutative. Chapitres 1 à 4. Masson,
Paris, 1985.
[6] A. Grothendieck, J. A. Dieudonné, Éléments de géométrie algébrique. I: Le langage des schémas (seconde édition). Grundlehren Math. Wiss. 166. Springer, Berlin, 1971.
[7] D. Maclagan, G. G. Smith, Uniform bounds on multigraded regularity. J. Algebraic Geom. 14
(2005) 137–164.
[8] F. Rohrer, On toric schemes. Proceedings of the 32nd Symposium and the 6th Japan-Vietnam
Joint Seminar on Commutative Algebra (Hayama, Japan, 2010). Meiji Institute for Advanced
Study of Mathematical Sciences, 2011. arXiv:1107.2713
[9] F. Rohrer, Quasicoherent sheaves on toric schemes. Expo. Math. 32 (2014) 33–78.
[10] F. Rohrer, Coarsening of graded local cohomology. Comm. Algebra 42 (2014) 593–599.
Institute of Mathematics, Vietnam Academy of Science and Technology, 18 Hoàng
Quốc Việt, 10307 Hà Nội, Việt Nam
E-mail address: [email protected]
| 0math.AC
|
Efficient Constraint Propagation Engines
arXiv:cs/0611009v1 [cs.AI] 2 Nov 2006
CHRISTIAN SCHULTE
School of Information and Communication Technology,
KTH – Royal Institute of Technology,
Isafjordsgatan 39, Electrum 229
SE-164 40 Kista, SWEDEN,
email: [email protected].
PETER J. STUCKEY
NICTA Victoria Laboratory, Dept of Comp. Sci. & Soft. Eng.,
University of Melbourne 3010, AUSTRALIA,
email: [email protected].
Abstract
This paper presents a model and implementation techniques for speeding up constraint propagation. Three fundamental approaches to improving constraint propagation based on propagators as implementations of
constraints are explored: keeping track of which propagators are at fixpoint, choosing which propagator to apply next, and how to combine
several propagators for the same constraint.
We show how idempotence reasoning and events help track fixpoints
more accurately. We improve these methods by using them dynamically
(taking into account current domains to improve accuracy). We define
priority-based approaches to choosing a next propagator and show that
dynamic priorities can improve propagation. We illustrate that the use of
multiple propagators for the same constraint can be advantageous with
priorities, and introduce staged propagators that combine the effects of
multiple propagators with priorities for greater efficiency.
1
Introduction
We consider the problem of solving Constraint Satisfaction Problems (CSPs)
defined in the sense of Mackworth [21], which can be stated briefly as follows:
We are given a set of variables, a domain of possible values for each
variable, and a set (read as conjunction) of constraints. Each constraint is a relation defined over a subset of the variables, limiting
the combination of values that the variables in this subset can take.
The goal is to find a consistent assignment of values to the variables
so that all the constraints are satisfied simultaneously.
1
One widely-adopted approach to solving CSPs combines backtracking tree search
with constraint propagation. This framework is realized in finite domain constraint programming systems, such as SICStus Prolog [18], ILOG Solver [17],
and Gecode [12] that have been successfully applied to many real-life industrial
applications.
At the core of a finite domain constraint programming system is a constraint
propagation engine that repeatedly executes propagators for the constraints of a
problem. Propagators discover and remove values from the domains of variables
that can no longer take part in a solution of the constraints.
Example 1.1 Consider a simple CSP, with variables x1 , x2 , and x3 whose
domain of possible values are respectively x1 ∈ {2, 3, 4}, x2 ∈ {0, 1, 2, 3}, x3 ∈
{−1, 0, 1, 2} and the constraints are x3 = x2 , x1 ≤ x2 + 1, and x1 6= 3.
A propagator for x3 = x2 can determine that x2 6= 3 in any solution of this
constraint since x3 cannot take the value 3. Similarly x3 6= −1. The propagator
then reduces the domains of the variables to x2 ∈ {0, 1, 2} and x3 ∈ {0, 1, 2}. A
propagator for x1 ≤ x2 + 1 can determine that x1 6= 4 since x2 ≥ 3, and x2 6= 0
since x1 ≤ 1 so the domains are reduced to x1 ∈ {2, 3} and x2 ∈ {1, 2}. The
propagator for x1 6= 3 can remove the value 3 from the domain of x1 , leaving
x1 ∈ {2} (or x1 = 2). If we now reconsider the propagator for x3 = x2 we can
reduce domains to x2 ∈ {1, 2} and x3 ∈ {1, 2}. No propagator can remove any
further values.
We have not solved the problem, since we do not know a value for each variable. So after propagation we apply search, usually by splitting the domain of a
variable into two disjoint subsets and considering the resulting two subproblems.
Suppose we split the domain of x2 . One subproblem has x1 ∈ {2}, x2 ∈ {1}
and x3 ∈ {1, 2}. Applying the propagator for x2 = x3 results in x3 ∈ {2}. Since
each variable now takes a fixed value we can check that x1 = 2, x2 = 1, x3 = 1
is a solution to the CSP. The other subproblem has x1 ∈ {2}, x2 ∈ {2} and
x3 ∈ {1, 2}, and leads to another solution.
✷
As can be seen from the example finite domain constraint programming
interleaves propagation with search. In this paper we investigate how to make
a propagation engine as efficient as possible.
There are two important decisions the engine must make: which propagators
should execute, and in which order they should execute. In order to make
constraint propagation efficient, it is clear that the engine needs to take the
following issues into account: avoid unnecessary propagator execution, restrict
propagation to relevant variables, and choose the cheapest possible method for
propagation. In this paper we show how propagation can be speeded up if the
engine takes these issues into account.
The contributions of the paper are as follows:
• We give a formal definition of propagation systems including fixpoint and
event-based optimizations used in current propagation systems.
• We extend event-based propagation systems to use dynamically changing
event sets.
2
• We introduce multiple propagators and staged propagators for a single
constraint for use with propagation queues with priority.
• We give experimental results that clarify the impact of many choices in implementing propagation engines: including idempotence reasoning, static
and dynamic events, basic queuing strategies, priority queues, and staged
propagation.
Plan of the paper The next section introduces propagation-based constraint
solving, followed by a model for constraint propagation systems in Section 3.
Section 4 presents how to optimize propagation by taking idempotence into account, while Section 5 explores the use of event sets. Which propagator should
be executed next is discussed in Section 6, while combination strategies of multiple propagators for the same constraint is discussed in Section 7. Experiments
for each feature are included in the relevant section, and a summary is given in
Section 8. Section 9 concludes.
2
Propagation-based Constraint Solving
This section defines our terminology for the basic components of a constraint
propagation engine. In this paper we restrict ourselves to finite domain integer
constraint solving. Almost all the discussion applies to other forms of finite
domain constraint solving such as for sets and multisets.
Domains A domain D is a complete mapping from a fixed (finite) set of
variables V to finite sets of integers. A false domain D is a domain with D(x) = ∅
for some x ∈ V. A variable x ∈ V is fixed by a domain D, if |D(x)| = 1. The
intersection of domains D1 and D2 , denoted D1 ⊓ D2 , is defined by the domain
D(x) = D1 (x) ∩ D2 (x) for all x ∈ V.
A domain D1 is stronger than a domain D2 , written D1 ⊑ D2 , if D1 (x) ⊆
D2 (x) for all x ∈ V. A domain D1 is stronger than (equal to) a domain D2
w.r.t. variables V , denoted D1 ⊑V D2 (resp. D1 =V D2 ), if D1 (x) ⊆ D2 (x)
(resp. D1 (x) = D2 (x)) for all x ∈ V .
A range is a contiguous set of integers, we use range notation [l .. u] to denote
the range {d ∈ Z | l ≤ d ≤ u} when l and u are integers. A domain is a range
domain if D(x) is a range for all x. Let D′ = range(D) be the smallest range
domain containing D, that is, the unique domain D′ (x) = [inf D(x) .. sup D(x)]
for all x ∈ V.
We shall be interested in the notion of an initial domain, which we denote
Dinit . The initial domain gives the initial values possible for each variable. It
allows us to restrict attention to domains D such that D ⊑ Dinit .
Valuations and constraints An integer valuation θ is a mapping of variables
to integer values, written {x1 7→ d1 , . . . , xn 7→ dn }. We extend the valuation θ
to map expressions and constraints involving the variables in the natural way.
3
Let vars be the function that returns the set of variables appearing in a
valuation. We define a valuation θ to be an element of a domain D, written
θ ∈ D, if θ(xi ) ∈ D(xi ) for all xi ∈ vars(θ).
The infimum and supremum of an expression e with respect to a domain D
are defined as inf D e = inf {θ(e)|θ ∈ D} and supD e = sup {θ(e)|θ ∈ D}.
We can map a valuation θ to a domain Dθ as follows
{θ(x)}
x ∈ vars(θ)
Dθ (x) =
Dinit (x) otherwise
A constraint c over variables x1 , . . . , xn is a set of valuations θ such that
vars(θ) = {x1 , . . . , xn }. We also define vars(c) = {x1 , . . . , xn }.
Propagators We will implement a constraint c by a set of propagators prop(c)
that map domains to domains. A propagator f is a monotonically decreasing
function from domains to domains: f (D) ⊑ D, and f (D1 ) ⊑ f (D2 ) whenever
D1 ⊑ D2 . A propagator f is correct for a constraint c iff for all domains D
{θ | θ ∈ D} ∩ c = {θ | θ ∈ f (D)} ∩ c
This is a very weak restriction, for example the identity propagator is correct
for all constraints c.
A set of propagators F is checking for a constraint c, if for all valuations θ
where vars(θ) = vars(c) the following holds: f (Dθ ) = Dθ for all f ∈ F , iff θ ∈ c.
That is, for any domain Dθ corresponding to a valuation on vars(c), f (Dθ ) is a
fixpoint iff θ is a solution of c. We assume that prop(c) is a set of propagators
that is correct and checking for c.
The output variables output(f ) ⊆ V of a propagator f are the variables
changed by the propagator: x ∈ output(f ) if there exists a domain D such that
f (D)(x) 6= D(x). The input variables input(f ) ⊆ V of a propagator f is the
smallest subset V ⊆ V such that for each domain D: D =V D′ implies that
D′ ⊓f (D) =output(f ) f (D′ )⊓D. Only the input variables are useful in computing
the application of the propagator to the domain.
Example 2.1 [Propagators, input, and output] For the constraint c ≡
x1 ≤ x2 + 1 the function fA defined by fA (D)(x1 ) = {d ∈ D(x1 ) | d ≤ supD x2 +
1} and fA (D)(v) = D(v), v 6= x1 is a correct propagator for c. Its output
variables are {x1 } and its input variables are {x2 }. Let D1 (x1 ) = {1, 5, 8} and
D1 (x2 ) = {1, 5}, then f (D1 ) = D2 where D2 (x1 ) = D2 (x2 ) = {1, 5}.
The propagator fB defined as fB (D)(x2 ) = {d ∈ D(x2 ) | d ≥ inf D x1 − 1}
and fB (D)(v) = D(v), v 6= x2 is another correct propagator for c. Its output
variables are {x2 } and input variables {x1 }.
The set {fA , fB } is checking for c. The domain Dθ1 (x1 ) = Dθ1 (x2 ) = {2}
corresponding to the solution θ1 = {x1 7→ 2, x2 7→ 2} of c is a fixpoint of
both propagators. The non-solution domain Dθ2 (x1 ) = {2}, Dθ2 (x2 ) = {0}
corresponding to the valuation θ2 = {x1 7→ 2, x2 7→ 0} is not a fixpoint (of
either propagator).
✷
4
A propagation solver solv(F, D) for a set of propagators F and an initial
domain D finds the greatest mutual fixpoint of all the propagators f ∈ F . In
other words, solv(F, D) returns a new domain defined by
solv(F, D) = gfp(λd. iter(F, d))(D)
iter(F, D) =
⊓
f ∈F
f (D)
where gfp denotes the greatest fixpoint w.r.t ⊑ lifted to functions.
Note that by inverting the direction of ⊑ we could equally well phrase this
as a least fix point (as in [1]). But the current presentation emphasizes the
reduction of domains as computation progresses.
Domain and bounds propagators A consistency notion C gives a condition
on domains with respect to constraints. A set of propagators F maintains
C-consistency for a constraint c, if solv(F, D) is always C consistent for c.
Many propagators in practice are designed to maintain some form of consistency:
usually domain or bounds. But note that many more do not.
The most successful consistency technique was arc consistency [21], which
ensured that for each binary constraint, every value in the domain of the first
variable, has a supporting value in the domain of the second variable that satisfied the constraint. Arc consistency can be naturally extended to constraints
of more than two variables. This extension has been called generalized arc consistency [24], as well as domain consistency [33, 34] (which is the terminology
we will use), and hyper-arc consistency [22]. A domain D is domain consistent
for a constraint c if D is the least domain containing all solutions θ ∈ D of c,
that is, there does not exist D′ ⊏ D such that θ ∈ D ∧ θ ∈ c → θ ∈ D′ .
Define the domain propagator dom(c), for a constraint c as
dom(c)(D)(x)
dom(c)(D)(x)
=
=
{θ(x) | θ ∈ D ∧ θ ∈ c} where x ∈ vars(c)
D(x)
otherwise
The basis of bounds consistency is to relax the consistency requirement to
apply only to the lower and upper bounds of the domain of each variable x.
There are a number of different notions of bounds consistency [8], we give the
two most common here.
A domain D is bounds(Z) consistent for a constraint c with vars(c) =
{x1 , . . . , xn }, if for each variable xi , 1 ≤ i ≤ n and for each di ∈ {inf D xi , supD xi }
there exist integers dj with inf D xj ≤ dj ≤ supD xj , 1 ≤ j ≤ n, j 6= i such that
θ = {x1 7→ d1 , . . . , xn 7→ dn } is an integer solution of c.
A domain D is bounds(R) consistent for a constraint c with vars(c) =
{x1 , . . . , xn }, if for each variable xi , 1 ≤ i ≤ n and for each di ∈ {inf D xi , supD xi }
there exist real numbers dj with inf D xj ≤ dj ≤ supD xj , 1 ≤ j ≤ n, j 6= i such
that θ = {x1 7→ d1 , . . . , xn 7→ dn } is a real solution of c.
A set of propagators F maintains bounds(α) consistency for a constraint c,
if for all domains D, solv(F, D) is bounds(α) consistent for c.
We can define a bounds(Z) propagator, zbnd(c) for a constraint c as follows:
zbnd(c)(D)
=
D ⊓ range(dom(c)(range(D)))
5
search(Fo , Fn , D)
D := isolv(Fo , Fn , D)
if (D is a false domain)
return false
if (∃x ∈ V.|D(v)| > 1)
choose {c1 , . . . , cm } where C ∧ D |= c1 ∨ · · · ∨ cm
for i ∈ [1 .. m]
if (search(Fo ∪ Fn , prop(ci ), D))
return true
return false
return true
% propagation
% search strategy
Figure 1: Search procedure
It is not straightforward to give a generic description of bounds(R) propagators, rbnd(c), for a constraint c, that just maintains bounds(R) consistency.
Examples 3.2 and 4.3 define three such propagators.
3
Constraint Propagation Systems
A constraint propagation system evaluates the function solv(F, D) during backtracking search. We assume an execution model for solving a constraint problem
with a set of constraints C and an initial domain D0 as follows. We execute
the procedure search(∅, F, D0 ) given in Figure 1 for an initial set of propagators
F = ∪c∈C prop(c). This procedure is used to make precise the optimizations
presented in the remainder of the paper.
Note that the propagators are partitioned into two sets, the old propagators Fo and the new propagators Fn . The incremental propagation solver
isolv(Fo , Fn , D) (to be presented later) takes advantage of the fact that D is
guaranteed to be a fixpoint of the old propagators.
The somewhat unusual definition of search is quite general. The default
search strategy for many problems is to choose a variable x such that |D(x)| > 1
and explore x = inf D x or x ≥ inf D x + 1. This is commonly thought of as
changing the domain D for x to either {inf D x} or {d ∈ D(x) | d > inf D x}.
This framework allows more general strategies, for example x1 ≤ x2 or x1 > x2 .
The basic incremental propagation solver algorithm is given in Figure 2. The
algorithm uses a queue Q of propagators to apply. Initially, Q contains the new
propagators. Each time the while loop is executed, a propagator f is deleted
from the queue, f is applied, and then all propagators that may no longer be
at a fixpoint at the new domain D′ are added to the queue. An invariant of the
algorithm is that at the while statement f (D) = D for all f ∈ F − Q.
The propagation solver isolv leaves two components undefined: choose(Q)
chooses the propagator f ∈ Q to be applied next; new(f, F, D, D′ ) determines
the set of propagators f ′ ∈ F that are not guaranteed to be at their fixpoint at
6
isolv(Fo , Fn , D)
F := Fo ∪ Fn ; Q := Fn
while (Q 6= ∅)
f := choose(Q)
Q := Q − {f }; D′ := f (D)
Q := Q ∪ new(f, F, D, D′ )
D := D′
return D
% select next propagator to apply
% add propagators f ′ ∈ F . . .
% . . . not necessarily at fixpoint at D′
Figure 2: Incremental propagation solver.
the domain D′ . The remainder of the paper investigates how to best implement
these two components.
3.1
Basic Variable Directed Propagation
The core aim of the constraint propagation solver solv(F, D) is to find a domain
that is a mutual fixpoint of all f ∈ F . The incremental solver isolv(Fo , Fn , D)
already takes into account that initially D is a fixpoint of propagators f ∈ Fo .
The role of new is (generally) to return as few propagators f ∈ F as possible.
A basic definition of new is as follows
newinput (f, F, D, D′ ) = {f ′ ∈ F | input(f ′ ) ∩ {x ∈ V | D(x) 6= D′ (x)} =
6 ∅}
Here all propagators f ′ are added whose input variable domains have changed.
By the definition of input variables, if none of them have changed for f ′ , then
f ′ (D′ ) = D′ since f ′ (D) = D if f ′ ∈ F − Q.
Proposition 3.1 newinput maintains the invariant f ′ (D) = D for all f ′ ∈
F − Q at the start of the while loop.
Proof : Consider f ′ ∈ F − Q. Then f ′ (D) = D and if D =input(f ′ ) D′ we have
that D′ ⊓ f ′ (D) =output(f ′ ) f ′ (D′ ) ⊓ D. Then
D′
=
=
=output(f ′ )
=
D′ ⊓ D
D′ ⊓ f ′ (D)
f ′ (D′ ) ⊓ D
f ′ (D′ )
since D′ ⊑ D
since D = f ′ (D)
by definition of input(f ′ )
since f ′ (D′ ) ⊑ D′ ⊑ D
Now D′ =output(f ′ ) f ′ (D′ ) implies D′ = f ′ (D′ ) by the definition of output(f ′ ).
Hence each f ′ in F − Q is at fixpoint at the start of the loop. ✷
The incremental propagation solver isolv with this definition of new (assuming Fo = ∅) is more or less equivalent to the propagation algorithms in [4] and [1,
page 267].
7
Example 3.2 [Incremental propagation] Consider the problem with constraints cC ≡ x1 = 2x2 and cD ≡ x1 = 3x3 represented by the bounds(R)
propagators
fC (D)(x1 ) =
fC (D)(x2 ) =
fC (D)(x) =
D(x1 ) ∩ [2 inf D x2 .. 2 supD x2 ] ,
D(x2 ) ∩ ⌈ 12 inf D x1 ⌉ .. ⌊ 12 supD x1 ⌋ ,
D(x)
x 6∈ {x1 , x2 }
fD (D)(x1 ) =
fD (D)(x3 ) =
fD (D)(x) =
D(x1 ) ∩ [3 inf D x3 .. 3 supD x3 ] ,
D(x3 ) ∩ ⌈ 13 inf D x1 ⌉ .. ⌊ 13 supD x1 ⌋ ,
D(x)
x 6∈ {x1 , x3 },
with initial domains D(x1 ) = [0 .. 17], D(x2 ) = [0 .. 9], and D(x3 ) = [0 .. 6].
Initially no constraint is at fixpoint, so Q = {fC , fD }. fC is selected initially,
and we execute the bounds propagator determining D(x1 ) = [0 .. 16], D(x2 ) =
[0 .. 8]. Since x1 has changed, both fC and fD are added to the queue. Then fD
is executed setting D(x1 ) = [0 .. 15], D(x3 ) = [0 .. 5]. Again both constraints are
re-queued. fC is executed changing the domains to D(x1 ) = [0 .. 14], D(x2 ) =
[0 .. 7]. Then fD changes the domains to D(x1 ) = [0 .. 12], D(x3 ) = [0 .. 4].
Since x1 has changed we have Q = {fC , fD }. Now fC is executed for no change,
and fD is executed for no change. We have reached a fixpoint D(x1 ) = [0 .. 12],
D(x2 ) = [0 .. 6], and D(x3 ) = [0 .. 4].
✷
4
Fixpoint Reasoning
The propagation engine computes a mutual fixpoint of all the propagators.
Clearly, if we can determine that some propagators are at fixpoint without
executing them we can limit the amount of work required by the engine.
4.1
Static Fixpoint Reasoning
A propagator f is idempotent if f (D) = f (f (D)) for all domains D. That is,
applying f to any domain D yields a fixpoint of f .
Example 4.1 [Idempotent propagator] The propagator fE defined by
fE (D)(x1 ) =
fE (D)(x2 ) =
fE (D)(x) =
{d ∈ D(x1 ) | 32 d ∈ D(x2 )}
{d ∈ D(x2 ) | 23 d ∈ D(x1 )}
D(x)
x 6∈ {x1 , x2 }
is the domain propagator for the constraint 3x1 = 2x2 . The propagator fE is
idempotent.
✷
It is not difficult to see that each domain propagator dom(c) is idempotent.
Proposition 4.2 For all constraints c and domains D
dom(c)(D) = dom(c)(dom(c)(D))
8
Proof : Consider θ ∈ c where θ ∈ D. Then by definition θ(x) ∈ dom(c)(D) for all
x ∈ vars(c). Hence θ ∈ dom(c)(D). Since dom(c)(D) ⊑ D we have θ ∈ c∧θ ∈ D
iff θ ∈ c ∧ θ ∈ dom(c)(D). Hence dom(c)(D) = dom(c)(dom(c)(D)). ✷
Example 4.3 [Non-idempotent propagators] While many propagators are
idempotent, some widely used ones are not idempotent. Consider the constraint
3x1 = 2x2 and the propagator fF :
fF (D)(x1 ) = D(x1 ) ∩ ⌈ 23 inf D x2 ⌉ .. ⌊ 32 supD x2 ⌋
fF (D)(x2 ) = D(x2 ) ∩ ⌈ 32 inf D x1 ⌉ .. ⌊ 23 supD x1 ⌋
fF (D)(x) = D(x)
x 6∈ {x1 , x2 }
In almost all constraint programming systems prop(3x1 = 2x2 ) is {fF } where
fF is the bounds(R) propagator for 3x1 = 2x2 . Now fF is not idempotent.
Consider D(x1 ) = [0 .. 3] and D(x2 ) = [0 .. 5]. Then D′ = fF (D) is defined
by D′ (x1 ) = [0 .. 3] ∩ [0 .. ⌊10/3⌋] = [0 .. 3] and D′ (x2 ) = [0 .. 5] ∩ [0 .. ⌊9/2⌋] =
[0 .. 4]. Now D′′ = fF (D′ ) is defined by D′′ (x1 ) = [0 .. 3] ∩ [0 .. ⌊8/3⌋] = [0 .. 2]
and D′′ (x2 ) = [0 .. 4] ∩ [0 .. ⌊9/2⌋] = [0 .. 4]. Hence fF (fF (D)) = D′′ 6= D′ =
fF (D).
✷
We can always create an idempotent propagator f ′ from a propagator f by
defining f ′ (D) = solv({f }, D). Indeed, in some implementations (for example [16]) prop(3x1 = 2x2 ) is defined as the fixpoint of applying fF .
Assume that idem(f ) = {f } if f is an idempotent propagator and idem(f ) =
∅ otherwise. The definition of new is improved by taking idempotence into
account
newsfix (f, F, D, D′ ) = newinput (f, F, D, D′ ) − idem(f )
An idempotent propagator is never put into the queue after application.
Note that without the idempotence optimization each propagator f that
changes the domain is likely to be executed again to check it is at fixpoint. Almost all constraint propagation solvers take into account static fixpoint reasoning (for example ILOG Solver [17], Choco [19], SICStus [18], and Gecode [12]).
Some systems even only allow idempotent propagators (for example Mozart [26]).
4.2
Dynamic Fixpoint Reasoning
Even if a propagator is not idempotent we can often determine that f (D) is
a fixpoint of f for a specific domain D. For simplicity we assume a function
fix(f, D) that returns {f } if it can show that f (D) is a fixpoint for f and ∅
otherwise (of course without calculating f (f (D)), otherwise we gain nothing).
In practice this will be included in the implementation of f .
newdfix (f, F, D, D′ ) = newinput (f, F, D, D′ ) − fix(f, D)
Example 4.4 [Dynamic idempotence] For bounds propagation for linear
equations on range domains we are guaranteed that the propagator is at a fixpoint if there is no rounding required in determining new endpoints [16, Theorem
8].
9
We can define fix(fF , D) for the bounds propagator fF from Example 4.3
for the constraint 3x1 = 2x2 , as returning {fF } if any new bound does not
require rounding, e.g. 2 inf D x2 /3 = ⌈2 inf D x2 /3⌉ or 2 inf D x2 /3 ≤ inf D x1 and
similarly for the other three bounds.
Consider applying fF to the domain D′′ from the same example. Now D′′′ =
fF (D′′ ) is defined by D′′′ (x1 ) = [0 .. 2] ∩ [0 .. ⌊8/3⌋] = [0 .. 2] and D′′′ (x2 ) =
[0 .. 4] ∩ [0 .. ⌊6/2⌋] = [0 .. 3]. Notice that the new bound x2 ≤ 3 is obtained
without rounding ⌊6/2⌋ = ⌊3⌋ = 3. In this case we are guaranteed that the
propagator is at a fixpoint.
✷
Note that the dynamic case extends the static case since for idempotent f it
holds that fix(f, D) = {f } for all domains D. The dynamic fixpoint reasoning
extensions are obviously correct, given Proposition 3.1.
Proposition 4.5 newdfix maintains the invariant f (D) = D for all f ∈ F − Q
at the start of the while loop.
A complexity of either form of fixpoint reasoning is that a great deal of
care has to be taken when we claim a propagator is at fixpoint, particularly for
bounds propagators and for propagators computing with multiple occurrences
of the same variable.
Example 4.6 [Falling into domain holes] Consider a bounds propagator for
x1 = x2 + 1 defined as
fG (D)(x1 )
fG (D)(x2 )
fG (D)(x)
= D(x1 ) ∩ [inf D x2 + 1 .. supD x2 + 1]
= D(x2 ) ∩ [inf D x1 − 1 .. supD x1 − 1]
= D(x)
x 6∈ {x1 , x2 }
We would expect this propagator to be idempotent since there is no rounding
required. Consider the application of fG to the domain D(x1 ) = {0, 4, 5, 6}
and D(x2 ) = {2, 3, 4, 5}. Then fG (D) = D′ where D′ (x1 ) = {4, 5, 6} and
D′ (x2 ) = {2, 3, 4, 5}. This is not a fixpoint for fG because of the hole (that is
1, 2, 4 6∈ D(x1 )) in the original domain of x1 .
✷
Example 4.7 [Multiple variable occurrences] The regular constraint introduced in [27] constrains a sequence of variables to take values described by
a regular expression (or a corresponding finite automaton). A common case for
the regular constraint is to express cyclic patterns by performing propagation
on a sequence of variables where some variables appear multiply.
Assume a propagator fH propagating that the sequence of variables hx1 , x2 , x3 i
conforms to the regular expression (11|00)0 (that is, the values of three variables form either the string 110 or 000).
For a domain D with D(x1 ) = D(x2 ) = D(x3 ) = {0, 1} propagation for
the sequence hx1 , x2 , x3 i is obtained by checking which values are still possible
for each variable by traversing the sequence once. In this particular case, for
D′ = fH (D) we have that D′ (x3 ) = {0} and D′ (x1 ) = D′ (x2 ) = {0, 1} and D′
is a fixpoint for fH .
10
Now assume a sequence hx1 , x2 , x1 i where x1 appears twice. Using the same
strategy as above, a single forward traversal of the sequence yields: D′ = fH (D)
where D′ (x1 ) = {0} and D′ (x2 ) = {0, 1}. However, D′ is not a fixpoint for fH .
✷
The two examples above describe common cases where a propagator is not
idempotent but in many cases still computes a fixpoint. In these cases, dynamic
fixpoint reasoning is beneficial: static fixpoint reasoning would force these propagators to be never considered to be at fixpoint while dynamic fixpoint reasoning
allows the propagator to decide whether the propagator is at fixpoint for a given
domain or not.
In practice dynamic fixpoint reasoning completely subsumes static fixpoint
reasoning and is easy to implement. To implement dynamic fixpoint reasoning
a propagator is extended to not only return the new domain but also a flag
indicating whether it is guaranteed to be at fixpoint.
4.3
Fixpoint Reasoning Experiments
Table 1 shows runtime (walltime) and the number of propagation steps for
the three different variants of fixpoint reasoning considered. The first column
presents absolute runtime values in milliseconds and the number of propagation
steps when no fixpoint reasoning is considered. The two remaining columns
“static” and “dynamic” show the relative change to runtime and propagation
steps when using static and dynamic fixpoint reasoning. For example, a relative
change of +50% means that the system takes 50% more time or propagation
steps, whereas a value of −50% means that the system takes only half the time or
propagation steps. The row “average (all)” gives the average (geometric mean)
of the relative values for all examples. More information on the examples can
be found in Appendix A and on the used platform in Appendix B.
Note that both static and dynamic fixpoint reasoning do not change the
memory requirements for any of the examples.
Static fixpoint reasoning While static fixpoint reasoning reduces the number of propagator executions by 8.9% in average, the reduction in runtime is
modest by 1.6% in average. The reason that the reduction in propagation steps
does not directly translate to a similar reduction in runtime is that the avoided
steps tend to be cheap: after all, these are steps not performing any propagation
as the propagator is already at fixpoint. Examples with significant reduction
in runtime (such as donald-d, minsort-200, picture, and square-5-d) profit
because the execution of costly propagators is avoided (domain-consistent linear equations for donald-d and square-5-d; regular propagators for picture;
minimum propagators involving up to 200 variables for minsort). The behavior
of golomb-10-b and golomb-10-d is explained further below.
Dynamic fixpoint reasoning As expected, both runtime as well as propagation steps are considerably smaller for dynamic fixpoint reasoning compared
11
Table 1: Fixpoint reasoning experiments.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
none
time (ms)
steps
118.31
503 036
106.56
272 398
2 279.36
1 335 657
4.64
15 641
624.25
806 664
0.70
546
30.37
46
0.38
546
1 347.48
2 642 464
2 430.00
2 642 962
35.87
9 344
55.41
2 299
7.46
48 038
342.48
240 991
574.36
387 060
8 571.24
16 785 128
108.87
422 206
1 553.42
150 165
4 433.12
31 424 152
16.15
2 469
517.96
151 609
33 391.24
1 762 492
10 166.24
7 731 557
5 690.00
13 956 982
0.74
2 486
—
—
12
static
time
steps
−3.1%
−24.8%
−2.0%
−5.7%
−0.2%
−6.5%
−0.4%
±0.0%
−0.1%
±0.0%
−2.5%
−8.4%
−6.3%
−6.5%
−5.9%
−16.5%
+4.7%
−17.9%
+4.5%
−18.0%
−0.1%
±0.0%
+2.6%
−3.8%
+0.9%
+2.2%
−11.3%
−8.3%
±0.0%
−0.5%
−3.7%
−18.8%
−0.3%
−0.8%
−4.1%
−13.1%
+0.2%
±0.0%
−1.2%
−16.2%
−0.1%
−0.1%
−6.2%
−16.0%
+2.4%
−14.5%
−4.1%
−16.9%
−2.5%
−1.8%
−1.6%
−8.9%
dynamic
time
steps
−38.9%
−24.9%
−1.5%
−5.7%
−0.2%
−6.5%
+0.1%
±0.0%
−0.2%
±0.0%
−10.4%
−16.5%
−6.4%
−13.0%
−5.1%
−16.5%
+0.1%
−17.5%
+4.3%
−18.0%
−7.2%
−6.2%
+2.6%
−3.8%
+0.9%
+2.2%
−5.1%
−16.9%
−8.7%
−18.2%
−9.2%
−19.7%
−14.6%
−3.6%
−3.4%
−20.5%
±0.0%
±0.0%
−1.2%
−16.2%
−0.1%
−0.1%
−6.1%
−16.1%
−6.8%
−14.7%
−4.3%
−16.9%
−7.9%
−9.1%
−5.6%
−11.5%
Table 2: Fixpoint reasoning experiments with propagator priorities.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
static
time
steps
−2.9% −24.9%
−5.4% −15.6%
−0.7%
−6.5%
−0.2%
±0.0%
−0.7%
±0.0%
−4.8% −22.9%
−6.7%
−5.1%
−5.6% −16.5%
−4.9% −23.5%
−3.3% −23.4%
+0.4%
±0.0%
−4.9% −29.5%
±0.0%
+1.3%
−10.9%
−9.1%
+0.3%
−1.4%
−8.5% −30.4%
−0.7%
−1.2%
+6.5% −13.1%
+0.2%
±0.0%
−1.3% −16.2%
+0.1%
±0.0%
−6.4% −11.1%
−2.5% −25.2%
−6.6% −24.9%
−2.2%
−1.4%
−2.9% −12.7%
dynamic
time
steps
−37.7%
−25.0%
−4.6%
−15.6%
−0.7%
−6.5%
+0.1%
±0.0%
−0.5%
±0.0%
−11.8%
−29.4%
−6.7%
−28.8%
−5.3%
−16.5%
−6.5%
−23.5%
−3.0%
−23.4%
−7.3%
−4.9%
−4.9%
−29.5%
+0.2%
+1.3%
−0.9%
−9.2%
−6.1%
−22.9%
−11.8%
−30.9%
−7.5%
−2.6%
+4.0%
−20.5%
+0.2%
±0.0%
−1.3%
−16.2%
+0.1%
±0.0%
−7.3%
−15.8%
−10.0%
−26.0%
−5.8%
−24.9%
−7.9%
−9.8%
−6.1%
−15.9%
to static reasoning. This is in particular true for examples donald-b and
square-7-b where a bounds-consistent alldifferent constraint can take advantage of reporting whether propagation has computed a fixpoint due to no
domain holes as discussed in Example 4.6 (all-interval-500 shows the same
behavior due to the absolute value propagator used).
Influence of propagation order Some examples show a considerable increase in runtime (in particular, golomb-10-b and golomb-10-d). This is due
to the fact that the order in which propagators are executed changes: costly
propagators are executed often while cheap propagators are executed less often
(witnessed by the decrease in propagation steps).
Section 6 presents priorities that order execution according to propagator
priorities. For now it is sufficient to note that when these priority inversion
problems are avoided through the use of priorities, there is no increase in runtime. Table 2 provides evidence for this. The table shows relative runtime
and propagation steps (relative to no fixpoint reasoning as in Table 1). When
avoiding priority inversion it becomes clear that both static as well as dynamic
13
fixpoint reasoning consistently improve the number of propagation steps and
also runtime. The only exception is picture where even with priorities the
considerable reduction in execution steps does not translate into a reduction in
runtime. This is possibly due to a change in propagation order that affects runtime (that this can have a remarkable effect even with priorities is demonstrated
in Section 6).
5
Event Reasoning
The next improvement for avoiding propagators to be put in the queue is to
consider what changes in domains of input variables can cause the propagator
to no longer be at a fixpoint. To this end we use events: an event is a change
in the domain of a variable.
Assume that the domain D changes to the domain D′ ⊑ D. A typical set of
events defined in a constraint propagation system are:
• fix(x): the variable x becomes fixed, that is |D′ (x)| = 1 and |D(x)| > 1.
• lbc(x): the lower bound of variable x changes, that is inf D′ x > inf D x.
• ubc(x): the upper bound of variable x changes, that is supD′ x < supD x.
• dmc(x): the domain of variable x changes, that is D′ (x) ⊂ D(x).
Clearly the events overlap. Whenever a fix(x) event occurs then a lbc(x) event,
a ubc(x) event, or both events must also occur. If any of the first three events
occur then a dmc(x) event occurs. These events satisfy the following property.
Definition 5.1 [Event] An event φ is a change in domain defined by an event
condition φ(D, D′ ) which states that event φ occurs when the domain changes
from D to D′ ⊑ D. The event condition must satisfy the following property
φ(D, D′′ ) = φ(D, D′ ) ∨ φ(D′ , D′′ )
where D′′ ⊑ D′ ⊑ D. So an event occurs on a change from D to D′′ iff it occurs
in either the change from D to D′ or from D′ to D′′ .
Given a domain D and a stronger domain D′ ⊑ D, then events(D, D′ ) is the
set of events φ where φ(D, D′ ). Suppose D′′ ⊑ D′ ⊑ D, then clearly
events(D, D′′ ) = events(D, D′ ) ∪ events(D′ , D′′ ).
(1)
Most integer propagation solvers use the events defined above, although
many systems collapse ubc(x) and lbc(x) into a single event bc(x) (for example, SICStus [18], ILOG Solver [17], and Gecode [12]). Choco [19] maintains
an event queue and interleaves propagator execution with events causing more
propagators to be added to the queue.
Other kinds of events or variants of the above events are also possible. For
example, (for domains D and D′ with D′ ⊑ D):
14
• bc(x): as discussed above (lbc(x) ∨ ubc(x)).
• two(x): the variable x reduces to a domain of at most two values: |D′ (x)| ≤
2 and |D(x)| > 2.
• ran(x): the variable x reduces to a domain that will always be a range
(two consecutive values or a single value): supD′ x − inf D′ x ≤ 1 and
supD x − inf D > 1.
• pos(x): the variable x reduces to a domain that is strictly positive, that
is inf ′D x > 0 and inf D x ≤ 0 (likewise, an event neg(x) for reduction to a
strictly negative domain).
• nneg(x): the variable x reduces to a domain that is non-negative: inf ′D x ≥
0 and inf D x < 0 (likewise, an event npos(x) for reduction to a non-positive
domain).
• neq(x, d): the variable x can no longer take the value d, that is d ∈ D(x)
and d 6∈ D′ (x)
The events two and ran are useful for tracking endpoint-relevance and rangeequivalence [31]. The neq event has been used in e.g. Choco [19] and BProlog [37] for building AC4 [23] style propagators.
Example 5.2 [Events] Let D(x1 ) = {1, 2, 3}, D(x2 ) = {3, 4, 5, 6}, D(x3 ) =
{0, 1}, and D(x4 ) = {7, 8, 10} while D′ (x1 ) = {1, 2}, D′ (x2 ) = {3, 5, 6}, D′ (x3 ) =
{1} and D′ (x4 ) = {7, 8, 10}. Then events(D, D′ ) is
{ubc(x1 ), dmc(x1 ), dmc(x2 ), fix(x3 ), lbc(x3 ), dmc(x3 )}
Considering the additional events we obtain in addition
{bc(x1 ), two(x1 ), ran(x1 ), bc(x3 ), neq(x1 , 3), neq(x2 , 4), neq(x3 , 0)}
✷
Example 5.3 [Events are monotonic] Events are monotonic: further changes
to a domain do not discard events from previous changes. Consider the property
range(x) capturing that D(x) is a range for a domain D (this property is related
to the event ran(x), lacking the restriction that the domain can have at most
two elements).
The property range(x) is not an event: consider domains D′′ ⊑ D′ ⊑ D
with D(x) = {1, 2, 3, 5}, D′ (x) = {1, 2, 3}, and D′′ (x) = {1, 3}. If range were
an event, then events(D, D′′ ) = events(D, D′ ) ∪ events(D′ , D′′ ). However,
events(D, D′′ ) = {dmc(x), ubc(x)}
whereas
events(D, D′ ) ∪ events(D′ , D′′ ) = {dmc(x), ubc(x), range(x)} ∪ {dmc(x)}
✷
15
5.1
Static Event Sets
Re-execution of certain propagators can be avoided since they require certain
events to generate new information.
Definition 5.4 [Propagator dependence] A propagator f is dependent on
a set of events es(f ) iff
(a) for all domains D if f (D) 6= f (f (D)) then events(D, f (D)) ∩ es(f ) 6= ∅,
(b) for all domains D and D′ where f (D) = D, D′ ⊑ D and f (D′ ) 6= D′ then
events(D, D′ ) ∩ es(f ) 6= ∅.
The definition captures the following. If f is not at a fixpoint then one of
the events in its event set occurs. If f is at a fixpoint D then any change to a
domain that is not a fixpoint D′ involves an occurrence of one of the events in
its set. Note that for idempotent propagators the case (a) never occurs.
For convenience later we will store the event set chosen for a propagator f
in an array evset[f ].
Clearly, if we keep track of the events since the last invocation of a propagator, we do not need to apply a propagator if it is not dependent on any of these
events.
Example 5.5 [Event sets] Event sets for previously discussed propagators are
as follows:
fA
{ubc(x2 )}
fB
{lbc(x1 )}
fE
{dmc(x1 ), dmc(x2 )}
fF
{lbc(x1 ), ubc(x1 ), lbc(x2 ), ubc(x2 )}
This is easy to see from the definitions of these propagators. If they use inf D x
then lbc(x) is in the event set, similarly if they use supD x then ubc(x) is in the
event set. If they use the entire domain D(x) then dmc(x) is in the event set.
✷
Indexical propagation solvers [34, 9, 6] are based on such reasoning. They
define propagators in the form f (D)(x) = D(x) ∩ e(D) where e is an indexical
expression. The event set for such propagators is automatically defined by the
domain access terms that occur in the expression e.
Example 5.6 [Indexical] An example of an indexical to propagate x1 ≥ x2 +1
is
x1 ∈ [inf(x2 ) + 1 .. + ∞]
x2 ∈ [−∞ .. sup(x1 ) − 1]
These range expressions for indexicals define two propagators:
fI (D)(x1 )
fI (D)(x)
fJ (D)(x2 )
fJ (D)(x)
=
=
=
=
D(x1 ) ∩ [inf(x2 ) + 1 .. + ∞]
D(x)
x 6= x1
D(x2 ) ∩ [−∞ .. sup(x1 ) − 1]
D(x)
x=
6 x2
16
The event set for the propagator fI from the definition is {lbc(x2 )} while
the the event set for fJ is {ubc(x1 )}.
✷
Using events we can define a much more accurate version of new that only
adds propagators for which one of the events in its event set has occurred.
newevents (f, F, D, D′ ) = {f ′ ∈ F | evset[f ′ ] ∩ events(D, D′ ) 6= ∅} − fix(f, D)
This version of new (without dynamic fixpoint reasoning) roughly corresponds
with what most constraint propagation systems currently implement.
Proposition 5.7 newevents maintains the invariant f (D) = D for all f ∈ F −Q
at the start of the while loop.
Proof : Consider f ′ ∈ F − Q − {f } different from the selected propagator f .
Then f ′ (D) = D and if f ′ (D′ ) 6= D′ then events(D, D′ ) ∩ es(f ′ ) 6= ∅ by case (b)
of the definition of es(f ′ ), so f ′ ∈ Q at the start of the loop.
Consider selected propagator f . This is removed from Q, but if f (D′ ) 6= D′
then events(D, D′ ) ∩ es(f ) 6= ∅ by case (a) of the definition of es(f ). Clearly
also fix(f, D) 6= ∅. So f ∈ Q at the start of the loop. ✷
5.2
Event Set Experiments
Table 3 shows runtime and number of propagation steps for different event
sets relative to a propagation engine not using events (the engine uses dynamic
fixpoint reasoning but no priorities). The row “average (above)” gives the geometric mean of the relative numbers given in the table whereas “average (all)”
shows the relative numbers for all examples.
General observations A first, quite surprising, observation is that using no
events at all is not so bad. It is the best approach for 10 out of the 25 benchmarks
and for crowded-chess-7 by a considerable margin (between 8.3% and 38.7%).
Another general observation is that a reduction in the number of propagation
steps does not directly translate into a reduction in runtime. This is due to the
fact that all saved propagator executions are cheap: the propagator is already
at fixpoint and does not have to perform propagation.
Note that the ratio between reduction in steps and reduction in runtime
ultimately depends on the underlying system. In Gecode, the system used, the
actual overhead for executing a propagator is rather low. In systems with higher
overhead one can expect that the gain in runtime will be more pronounced.
Event set observations Adding the fix event is particularly beneficial for
benchmarks with many disequalities, particularly queens-400 which only uses
disequalities.
Adding the bc event has significant benefit (up to 5%) when there are linear
equalities as in alpha or bounds(Z) consistent alldifferent as in photo. But
17
Table 3: Event set experiments (runtime and propagation steps).
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
only fix, dmc
time
steps
+0.6%
±0.0%
−2.0%
−4.4%
−2.6%
−16.6%
+1.1%
−0.2%
+8.3%
±0.0%
+0.3%
±0.0%
±0.0%
±0.0%
+0.6%
−10.5%
−0.6%
±0.0%
+0.2%
±0.0%
−0.7%
−47.7%
−0.1%
±0.0%
−12.7%
−47.5%
+0.4%
±0.0%
−1.2%
±0.0%
+0.7%
±0.0%
+1.6%
±0.0%
+2.0%
±0.0%
−87.5%
−99.1%
−10.4%
−38.8%
+2.2%
±0.0%
+0.8%
±0.0%
−0.2%
+0.5%
−1.0%
−8.9%
+2.2%
+0.2%
−8.5%
−24.3%
with
time
+1.0%
−7.3%
−1.4%
+1.7%
+20.2%
−2.1%
+0.9%
−1.4%
±0.0%
+26.4%
+0.1%
±0.0%
−10.8%
+1.1%
+0.2%
+15.1%
−3.3%
+3.5%
−87.5%
−9.6%
+3.5%
+2.7%
+0.6%
−1.7%
+3.3%
−6.7%
18
bc
steps
±0.0%
−19.2%
−16.6%
−0.2%
−8.4%
−9.2%
±0.0%
−19.7%
−3.9%
−3.5%
−49.5%
±0.0%
−48.6%
±0.0%
−1.7%
+3.4%
−27.1%
±0.0%
−99.1%
−38.8%
+36.3%
±0.0%
−8.2%
−16.9%
−7.6%
−26.9%
with lbc, ubc
time
steps
+2.0%
±0.0%
−6.3%
−19.3%
+7.4%
−15.8%
+2.6%
−0.2%
+38.7%
−8.4%
−1.3%
−9.2%
−0.1%
±0.0%
+0.1%
−19.7%
−0.5%
−3.9%
+26.5%
−3.5%
+0.6%
−49.5%
±0.0%
±0.0%
−7.7%
−48.6%
+1.5%
±0.0%
+1.5%
−1.7%
+18.8%
+17.7%
−2.0%
−26.9%
+11.5%
±0.0%
−87.5%
−99.1%
−8.2%
−38.8%
+8.8%
+36.3%
+2.6%
+0.5%
+0.7%
−7.8%
+1.2%
−16.9%
+4.2%
−6.6%
−4.6%
−26.4%
Table 4: Event set experiments with priorities (runtime and propagation steps).
Example
bibd-7-3-60
crowded-chess-7
donald-b
donald-d
golomb-10-b
golomb-10-d
minsort-200
partition-32
picture
square-5-d
square-7-b
average (above)
average (all)
only fix, dmc
time
steps
+6.1%
−3.3%
+8.0%
±0.0%
+0.9%
±0.0%
−0.3%
±0.0%
+0.1%
±0.0%
+0.1%
±0.0%
+0.9%
±0.0%
+1.9%
−1.2%
+2.6%
±0.0%
+0.9%
±0.0%
+0.4%
−0.1%
+1.9%
−0.4%
−7.8% −24.1%
with
time
+0.6%
+20.4%
−1.1%
+0.1%
−0.9%
−0.5%
+1.5%
+1.8%
+3.7%
+3.2%
+0.2%
+2.5%
−7.8%
bc
steps
−3.3%
−0.7%
−9.2%
±0.0%
−5.5%
−7.8%
±0.0%
−1.8%
±0.0%
±0.0%
−10.6%
−3.6%
−27.8%
with lbc, ubc
time
steps
+13.1%
−2.5%
+38.0%
−0.7%
−0.7%
−9.2%
−0.2%
±0.0%
−0.9%
−5.5%
±0.0%
−7.8%
+1.8%
±0.0%
+1.0%
−0.4%
+1.6%
±0.0%
+3.2%
+0.2%
+1.2% −10.6%
+4.8%
−3.4%
−6.3% −27.7%
for other examples, where one would expect some reduction in runtime, the overhead (to be discussed in more detail below) for maintaining a richer event set
exceeds the gains from reducing the number of propagation steps. This is true for
examples such as minsort-200 (minimum propagators), partition-32 (multiplication), and golomb-10-b and square-7-b (linear equations and bounds(Z)
consistent alldifferent).
Splitting the bc event into lbc and ubc events exposes the overhead once
more. There is almost never an improvement in number of propagations, since
only inequalities can actually benefit, and there is substantial overhead.
Influence of propagation order Similar to using fixpoint reasoning, the
use of events also changes the order in which propagators are executed. Table 4
reconsiders all examples that could possibly benefit from bc or lbc, ubc events
and all examples that show a remarkable increase in number of propagation
steps in Table 3.
The numbers confirm that with priorities no considerable increase in runtime
can be observed for all but crowded-chess-7, where the increase in runtime here
is due to a change in propagation order to the introduction of event sets that
does not depend on the relative priorities of the propagators used.
Once priorities are used, lbc and ubc are basically never beneficial.
Memory requirements Table 5 shows the total memory allocated for different event sets relative to a propagation engine not using events (the engine
uses dynamic fixpoint reasoning and priorities). It is important to note that the
memory figures reflect the amount of memory allocated which is bigger than
the amount of memory actually used. In particular, for small examples such
as donald-* the increase in allocated memory just reflects the fact that an
additional memory block gets allocated.
19
Table 5: Event set experiments with priorities (allocated memory).
Example
bibd-7-3-60
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
minsort-200
partition-32
picture
queens-400
queens-400-a
average (above)
average (all)
no events
mem (KB)
6 690.1
198.2
7.8
3.4
5.8
39.7
37.7
32 419.1
160.3
451.6
30 286.6
567.0
—
—
only fix, dmc
mem
+2.9%
+4.0%
±0.0%
±0.0%
+34.3%
+2.6%
±0.0%
+1.1%
+3.7%
+17.7%
+0.2%
+2.8%
+5.4%
+3.9%
with bc
mem
+5.7%
+13.1%
+12.8%
+58.2%
+34.3%
+10.1%
+2.8%
+2.3%
+9.4%
+17.7%
+0.2%
+4.2%
+13.3%
+9.9%
with lbc, ubc
mem
+14.4%
+21.2%
+12.8%
+58.2%
+51.5%
+12.7%
+16.0%
+17.4%
+17.5%
+30.1%
+0.2%
+4.2%
+20.3%
+15.5%
Using just a fix event increases the required memory by less than 4% in
average, while using full event sets increases the required memory by 15.5% in
average.
It must be noted that Gecode (the system used) has a particularly efficient
implementation of event sets: the implementation uses a single pointer for an
entry in an event set; entries for the same variable and the same propagator
but with different events still use a single pointer. The memory overhead is due
to the fact that per variable and supported event, one single pointer for bookkeeping is needed. Hence, the highest overhead can be expected for examples
with many variables but relatively few propagators and small event sets (such
as bibd-7-3-60, crowded-chess-7, and picture). The overhead becomes less
noticeable for examples with many propagators or large event sets and few variables (such as queens-400 and queens-400-a).
Summary In summary, while there is a compelling argument for fix events,
there is only a weak case for bc being supported, and lbc and ubc should not
be used.
5.3
Dynamic Event Sets
Events help to improve the efficiency of a propagation-based solver. Just as we
can improve the use of fixpoint reasoning by examining the dynamic case, we
can also consider dynamically updating event sets as more information is known
about the variables in the propagator.
Monotonic event sets
20
Definition 5.8 [Monotonic propagator dependence] A propagator f is
monotonically dependent on a set of events es(f, D) in the context of domain D
iff
(a) for all domains D0 ⊑ D if f (D0 ) 6= f (f (D0 )) then events(D0 , f (D0 )) ∩
es(f, D) 6= ∅,
(b) for domains D0 and D1 where D0 ⊑ D, f (D0 ) = D0 , D1 ⊑ D0 and
f (D1 ) 6= D1 then events(D0 , D1 ) ∩ es(f, D) 6= ∅.
Clearly given this definition es(f, D) is monotonically decreasing with D.
The simplest kind of event reduction occurs by subsumption.
Definition 5.9 [Subsumption] A propagator f is subsumed for domain D, if
for each domain D′ ⊑ D we have f (D′ ) = D′ .
A subsumed propagator makes no future contribution. If f is subsumed
by D then es(f, D) = ∅ and f is never re-applied. Most current constraint
propagation systems take into account propagator subsumption.
Example 5.10 [Subsumption] Consider the propagator fA and the domain
D with D(x1 ) = [1 .. 3] and D(x2 ) = [3 .. 7]. Then the constraint holds for all
D′ ⊑ D and es(f, D) = ∅.
✷
Changing event sets can occur in cases other than subsumption.
Example 5.11 [Minimum propagator] Consider the propagator fK for x0 =
min(x1 , x2 ) defined by
fK (D)(x0 )
fK (D)(xi )
fK (D)(x)
= D(x0 ) ∩ [min(inf D x1 , inf D x2 ) .. min(supD x1 , supD x2 )]
= D(xi ) ∩ [inf D x0 .. + ∞]
i ∈ {1, 2}
= D(x)
x∈
6 {x0 , x1 , x2 }
The static event set es(fK ) is {lbc(x0 ), lbc(x1 ), ubc(x1 ), lbc(x2 ), ubc(x2 )}. Note
that this propagator is idempotent.
But given domain D where D(x0 ) = [1 .. 3] and D(x2 ) = [5 .. 7] we know that
modifying the value of x2 will never cause propagation. A minimal definition of
es(fK , D) is {lbc(x0 ), lbc(x1 ), ubc(x1 )}.
✷
Example 5.12 [exactly propagator] Another example is a propagator for
the exactly constraint [35]: exactly([x1 , . . . , xn ], m, k) states that exactly m
out of the variables x1 , . . . , xn are equal to a value k. As soon as one of the xi
becomes different from k, all events for xi can be ignored. Originally the events
are dmc(xi ), 1 ≤ i ≤ m, lbc(m), ubc(m) and dmc(k).
Suppose a domain D where D(k) = {1, 3, 8} and D(x3 ) = {2, 5, 6, 10, 11, 12},
then x3 6= k and we know its contribution to the exactly constraint. We can
remove the event dmc(x3 ) from the event set safely.
✷
21
Other examples for monotonic event sets are propagators for the lex constraint and the generalized element constraint. When using a variant of the lex
propagator proposed in [5], events can be removed as soon as the order among
pairs of variables being compared can be decided. For the generalized element
constraint [7] where the array elements are variables, events for a variable from
the array can be safely removed as soon as the variable becomes known to be
different from the result of the element constraint.
Using monotonic dynamic event sets we can refine our definition of new as
follows.
newmevents (f, F, D, D′ )
F ′ := {f ′ ∈ F | evset[f ′ ] ∩ events(D, D′ )} − fix(f, D)
evset[f ] := es(f, D′ )
return F ′
Every time a propagator f is applied its event set is updated to take into account
newly available information.
A related idea is the “type reduction” of [30] where propagators are improved as more knowledge on domains (here called types) becomes available.
For example, the implementation of x0 = x1 × x2 will be replaced by a more
efficient one, when all elements in D(x1 ) and D(x2 ) are non-negative. Here we
concentrate on how the event sets change. The two ideas could be merged as
they are complementary.
Proposition 5.13 newmevents maintains the invariant f (D) = D for all f ∈
F − Q at the start of the while loop.
Proof : The proof is almost identical to that for Proposition 5.7 since we are
working in a context if evset[f ] = es(f, D∗ ) then D ⊑ D∗ .
For propagators using the monotonic event sets, we have the invariant that
f ∈ F − Q iff evset[f ] = es(f, D∗ ) where D∗ is the result of the last time we
executed propagator f .
Suppose we have f ′ ∈ F − Q − {f } where f ′ (D′ ) 6= D′ and f (D) = D then
events(D∗ , D′ ) ∩ es(f, D∗ ) 6= ∅ and since f ′ 6∈ Q we have that events(D∗ , D) ∩
es(f, D∗ ) = ∅ otherwise we would have placed f in the queue already, hence by
the equation (1) events(D, D′ ) ∩ es(f, D∗ ) 6= ∅ so f ′ ∈ Q at the start of the loop.
Consider selected propagator f . Then it is removed from Q but if f (D′ ) 6= D′
then clearly fix(f, D) = ∅ and using case (a) of Definition 5.8 we have that
events(D, D′ ) ∩ es(f, D) 6= ∅. Hence f ∈ Q at the start of the loop. ✷
Fully dynamic event sets Note that for many propagators we can be more
aggressive in our definition of event sets if we allow the event sets to change in
a manner that is not necessarily monotonically decreasing.
Definition 5.14 [General propagator dependence] A propagator f is dependent on a set of events es(f, D) in the context of domain D if for all domains
D1 where D1 ⊏ D and f (D1 ) 6= D1 then events(D, D1 ) ∩ es(f, D) 6= ∅.
22
Using fully dynamic event sets we can refine our definition of new as follows.
newdevents (f, F, D, D′ )
F ′ := {f ′ ∈ F | evset[f ′ ] ∩ events(D, D′ )} − fix(f, D)
if (fix(f, D) = ∅)
F ′ := F ′ ∪ {f }
evset[f ] := es(f, D′ )
return F ′
Every time a propagator f is applied its event set is updated to take into
account newly available information. The only difficult case is that if the definition of dynamic dependency does not capture the events that occur when
moving from D to D′ = f (D). If fixpoint reasoning cannot guarantee a fixpoint
we need to add f to the queue.
Fully dynamic event sets are more powerful than monotonically decreasing
event sets, but in general they require reasoning about the new event sets each
time the propagator is run.
Example 5.15 [Fully dynamic events for minimum] Given the propagator
fK from Example 5.11 and the domain D where D(x0 ) = [0 .. 10], D(x1 ) =
[0 .. 15], and D(x2 ) = [5 .. 10]. D is a fixpoint of fK and a minimal set es(fK , D)
is
{lbc(x0 ), lbc(x1 ), ubc(x1 ), ubc(x2 )}
While at D′ ⊑ D where D′ (x0 ) = [5 .. 9], D′ (x1 ) = [6 .. 9], and D′ (x2 ) =
[5 .. 10], which is also a fixpoint, the minimal set es(fK , D′ ) is
{lbc(x0 ), ubc(x1 ), lbc(x2 ), ubc(x2 )}
For the constraint cK we simply need to maintain a lbc event for some
variable xi in the right hand side with the minimal lbc value.
✷
Proposition 5.16 newdevents maintains the invariant f (D) = D for all f ∈
F − Q at the start of the while loop.
Proof : We have the invariant that f ∈ F − Q iff evset[f ] = es(f, D∗ ) where D∗
is the result of the last time we executed propagator f .
Suppose we have f ′ ∈ F − Q − {f } where f ′ (D′ ) 6= D′ and f (D) = D then
events(D∗ , D′ ) ∩ es(f, D∗ ) 6= ∅ and since f ′ 6∈ Q we have that events(D∗ , D) ∩
es(f, D∗ ) = ∅ otherwise we would have placed f ′ in the queue already, hence
by the equation (1) events(D, D′ ) ∩ es(f, D∗ ) 6= ∅ so f ′ ∈ Q at the start of the
loop.
Consider the selected propagator f . The same reasoning cannot apply since
even though we know that events(D∗ , D′ ) ∩ es(f, D∗ ) 6= ∅, we have no guarantee
that events(D, D′ ) ∩ es(f, D∗ ) is not empty. Now if fix(f, D) = ∅ then possibly
f (D′ ) 6= D′ , but this will force f ∈ Q by the start of the loop. ✷
Effectively the fully dynamic event sets approach relies only on the dynamic
fixpoint reasoning of the propagator f to handle what happens when moving
from D to D′ .
23
Table 6: Dynamic event sets experiments (with priorities).
Example
bibd-7-3-60
crowded-chess-7
sequence-500
o-latin-7-d
average (above)
average (all)
time
+0.4%
−24.0%
−82.3%
−0.7%
−39.5%
−10.8%
monotonic
steps
mem
±0.0%
±0.0%
±0.0%
−15.2%
−78.3%
−49.5%
±0.0%
±0.0%
−31.8%
−19.1%
−8.2%
−4.1%
time
−5.0%
−28.2%
−82.1%
−1.7%
−41.1%
−11.2%
fully dynamic
steps
mem
−5.2%
±0.0%
−34.2%
−17.8%
−78.3%
−49.5%
±0.0%
±0.0%
−39.3%
−19.7%
−9.9%
−4.2%
Fully dynamic event sets are closely related to the watched literals approach
to improving unit propagation in SAT solving [25]. Using watched literals, unit
propagation only considers a clause for propagation if one of two watched literals
in the clause becomes false. Recently, the idea of watched literals has been used
in constraint programming for the Minion solver [13]. Watched literals differ
from the events we concentrate on here since they take into account values
(similar to the neq(x, a) event). Note that dynamic event sets do not usually
have the property of watched literals, in that they do not need to be updated
on backtracking.
5.4
Dynamic Event Sets Experiments
Table 6 shows the comparison of monotonic and fully dynamic event sets to a
propagation solver using static event sets with {fix, bc, dmc} events and priorities to avoid priority inversion as discussed before. The table lists only examples
where dynamic event sets are used.
The propagators using monotonic event sets are as follows: for bibd-7-3-60:
lex; for crowded-chess-7: exactly and element; for sequence-500: exactly;
for o-latin-7-d: lex. Clearly, monotonic event sets lead to a drastic reduction
in both runtime and memory usage, where it is worth noting that the reduction
in time is even more marked than the reduction in propagation steps (each
propagation step becomes cheaper as smaller event sets must be maintained).
Fully dynamic event sets are considered in Boolean-sum propagators used in
bibd-7-3-60 and crowded-chess-7. The difference in improvement between
the two examples can be explained by the fact that crowded-chess-7 uses
Boolean-sums as inequalities while bibd-7-3-60 uses Boolean-sums as equalities
where inequalities offer the potential for considerably smaller event sets [13].
Example bibd-7-3-60 provides another insight: the drastic reduction in
runtime observed in [13] by using watched literals for Boolean-sum propagators
in the Minion solver is most likely not due to small event sets but to other
aspects. A possible aspect is the knowledge about which variables have been
modified when a propagator is executed.
24
Table 7: Queue versus stack experiments.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
6
queue
time (ms)
steps
73.00
377 777
97.31
207 470
2 020.30
1 020 162
4.65
14 860
553.25
660 525
0.62
414
28.53
40
0.36
366
1 342.48
2 095 161
3 201.84
2 091 691
33.54
4 422
56.87
2 211
6.66
25 225
141.84
113 437
532.48
311 135
9 001.24
13 939 101
90.37
296 661
1 583.12
119 406
549.36
268 771
14.40
1 265
98.12
56 048
32 263.12
1 478 403
9 491.84
6 054 656
5 345.00
9 636 675
0.70
2 075
—
—
stack
time
steps
+13626.0%
+32.2%
+47.5%
+99.0%
+7.7%
+102.7%
+1.3%
+2.6%
+4.7%
+68.1%
+36.3%
+31.9%
−5.4%
+7.5%
+7.6%
+23.2%
+870.6%
+171.5%
−11.9%
+98.9%
+299.7%
+118.4%
+44.2%
−6.3%
+7.7%
+53.3%
+1580.1%
+1123.3%
+4.9%
+24.9%
−18.5%
+34.8%
+11.1%
+8.7%
+27.1%
+79.9%
±0.0%
−0.1%
−0.9%
−0.3%
+53.8%
+240.9%
+71.5%
+57.7%
+75.3%
+44.9%
+8.3%
+35.2%
+11.7%
+29.3%
+78.4%
+61.0%
Which Propagator to Execute Next
We now address how to define which propagator f in the queue Q should execute
first, that is how to define the choose function.
The simplest policy to implement is a FIFO (First In First Out) queue
of propagators. Propagators are added to the queue, if they are not already
present, and choose selects the oldest propagator in the queue. The FIFO
policy ensures fairness so that computation is not dominated by a single group
of propagators, while possibly not discovering failure (a false domain) from other
propagators quickly.
The equally simple LIFO (Last In First Out) policy is a stack where propagators not already in the stack are pushed, and choose selects the top of the
stack.
25
6.1
Basic Queuing Strategy Experiments
Table 7 compares using a FIFO queue versus a LIFO stack. The result, according with folklore knowledge, clearly illustrates that a queue must be used. The
few cases where a stack is better are comprehensively outweighed by the worst
cases for a stack.
Later experiments in Section 6.3 using priorities with combinations of FIFO
queues and LIFO stacks reveal that the pathological behavior of all-interval-500
is due to priority inversion. However, the pathological behavior of minsort-200
is due to the use of a LIFO stack.
6.2
Static Priorities
A statically prioritized queue associates with each propagator a fixed priority,
we will assume an integer in the range [0 .. k − 1]. In effect, the queue Q is
split into k queues, Q[0], . . . Q[k − 1] where each Q[i] is a FIFO queue for the
propagators with priority i. Selection always chooses the oldest propagator in
the lowest numbered queue Q[i] that is non-empty. Static prioritization allows
one to ensure that quick propagators are executed before slow propagators.
We give an example of seven static priorities, with names of the integer priorities as follows: unary=0, binary=1, ternary=2, linear=3, quadratic=4,
cubic=5, and veryslow=6. The names are meant to represent the arity of the
constraint, and then the asymptotic runtime of the propagator, once the constraint can handle n variables. So binary is for binary constraints, quadratic
is for constraints that are approximately O(n2 ) for instances with n variables.
Example 6.1 [Propagator priorities] For example, the propagator fL for
even(x1 ) defined by
fL (D)(x1 ) = D(x1 ) ∩ 2⌈ 21 inf D x1 ⌉ .. 2⌊ 21 supD (x1 )⌋
fL (D)(x) = D(x)
x 6= x1
might be given priority unary, while fE and fF might be given priority binary.
The domain propagator defined in [29] for the alldifferent constraint
∧ni=1 ∧nj=i+1 xi 6= xj (with complexity O(n2.5 )) might be given priority quadratic.
The alldifferent bounds(Z) propagator defined in [28] (with complexity
O(n log n)) might be given priority linear.
✷
Priorities in effect force many more fixpoints to be calculated. A fixpoint of
all propagators at priority level i and lower must be reached before a propagator
at priority level i + 1 is run. This means we will often cause more propagators
to run when using priorities, but more cheap propagators!
Example 6.2 [Repeated fixpoints] Consider the execution of a system of
propagators for constraints cC ≡ x1 = 2x2 , cD ≡ x1 = 3x2 , cM ≡ x2 ≤
6 → x1 ≤ x3 + 7 and cN ≡ alldifferent[x1 , x2 , x3 , x4 , x5 ]. We will use the
bounds(R) propagators fC , fD , fM for the first three constraints, and the domain propagator, fN , for alldifferent from [29], for the last constraint. Let
26
the initial domain be D(x1 ) = [0 .. 18], D(x2 ) = [0 .. 9], D(x3 ) = [0 .. 6], and
D(x4 ) = D(x5 ) = [0 .. 3]. The priorities are binary, binary, ternary, and
quadratic respectively. All propagators are at fixpoint.
Suppose the domain of x1 changes to [0 .. 17]. All propagators are enqued.
The priority level binary propagators are run to fixpoint (as in Example 3.2)
giving D(x1 ) = [0 .. 12], D(x2 ) = [0 .. 6], D(x3 ) = [0 .. 4] Then fM is scheduled and causes D(x1 ) = [0 .. 11]. Both fC and fD are enqued, and after
executing fC , fD , fC and fD the next fixpoint of the binary priority propagators is reached: D(x1 ) = [0 .. 6], D(x2 ) = [0 .. 3], D(x3 ) = [0 .. 2]. Then
fM is scheduled again and causes no change. After that, fN is executed, it
reduces the domains of D(x1 ) = [4 .. 6] since all the values [0 .. 3] are required
for the variables x2 , x3 , x4 , x5 . Both fC and fD are enqued, and after executing we reach their fixpoint D(x1 ) = {6}, D(x2 ) = {3}, D(x3 ) = {2}. Once
again fM is executed for no change. Then fN is executed once more obtaining
D(x4 ) = D(x5 ) = {0, 1}. This is the overall fixpoint.
✷
We can adjust the granularity of the priorities: we can have a finer version
of the above priorities with 14 priorities, each priority above with a low and
high version. This allows us to separate, for example, a domain consistent
propagator for the binary absolute value constraint abs(x) = y (binary-low)
from a bounds consistent propagator for the same constraint (binary-high).
This increased granularity will be important in Section 7.
Conversely, we may collapse the priorities into fewer levels, for example into
three levels unary-ternary, linear-quadratic, cubic-veryslow.
Another model for priorities in constraint propagation based on composition
operators is [14]. The model, however, runs all propagators of lower priority
before switching propagation back to propagators of higher priority. This model
does not preempt computing a fixpoint for a low priority. The model always
completes a fixpoint for a given priority level and only then possibly continues
at a higher priority level.
Most systems have some form of static priorities, typically using two priority
levels (for example, SICStus [18], Mozart [26]). The two levels are often not
entirely based on cost: in SICStus all indexicals have high priority and all other
lower priority. While ECLi PSe [36, 15] supports twelve priority levels, its finite
domain solver also uses only two priority levels where another level is used to
support constraint debugging.
A similar, but more powerful approach is used by Choco [19] using seven
priority levels allowing both LIFO and FIFO traversal.
Prioritizing particular operations during constraint propagation is important
in general.
For interval narrowing, prioritizing constraints can avoid slow
convergence, see for example [20].
The prioritizing of propagators by cost is important, inverting the priorities
can lead to significant disadvantages.
Example 6.3 [Inverted priorities] Consider executing Example 6.2 with inverted priorities. We first execute fN then fM for no effect. Then executing fC
27
Table 8: Priority experiments with varying granularities.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
small
time
steps
+1.8%
+0.3%
+1.2%
±0.0%
+0.1%
+16.9%
−1.4%
−10.4%
+8.1%
+91.2%
+1.1%
±0.0%
−0.1%
±0.0%
+1.6%
±0.0%
−33.1%
+33.5%
−49.5%
+29.2%
−1.3%
±0.0%
+10.3%
−9.2%
−3.4%
−19.5%
+0.5%
±0.0%
−10.6%
+4.6%
−38.6%
−18.2%
+0.1%
+3.4%
−1.0%
±0.0%
+0.7%
±0.0%
+0.2%
±0.0%
+0.5%
±0.0%
−0.5%
±0.0%
+1.3%
±0.0%
+0.9%
±0.0%
+5.3%
+13.0%
−5.6%
+3.8%
medium
time
steps
+2.3%
+0.3%
+1.2%
±0.0%
+0.1% +16.9%
−1.5% −10.5%
+9.0% +91.2%
+0.4%
±0.0%
−5.6%
+5.0%
+2.6%
±0.0%
−31.3%
+51.0%
−49.0%
+46.7%
−1.3%
−0.1%
−5.6%
+1.9%
−4.7% −19.9%
+0.5%
±0.0%
−11.0%
+4.6%
−38.7%
−17.8%
+0.2%
+3.4%
−1.5%
±0.0%
+0.5%
±0.0%
+0.2%
±0.0%
+0.3%
±0.0%
+0.9% +23.8%
+0.4%
±0.0%
+0.7%
±0.0%
+5.7% +12.0%
−6.3%
+6.4%
full
time
+2.3%
+1.2%
+0.2%
−1.8%
+8.8%
+0.9%
−5.5%
+2.2%
−31.6%
−48.9%
−0.9%
−5.6%
−3.7%
−0.2%
−10.6%
−38.4%
−7.3%
+0.2%
+0.5%
+0.3%
+0.6%
+1.1%
−23.4%
+0.9%
+5.3%
−7.4%
steps
+0.3%
±0.0%
+16.9%
−10.5%
+91.1%
±0.0%
+5.0%
±0.0%
+51.0%
+46.7%
−0.1%
+1.9%
−19.9%
±0.0%
+4.6%
−17.8%
−1.8%
±0.0%
±0.0%
±0.0%
±0.0%
+23.8%
+22.0%
±0.0%
+12.0%
+7.0%
modifies D(x1 ), so each of fN and fM are enqued and re-executed for no effect.
Then executing fD has the same behavior. Overall we execute the propagators
fM and fN each at least 10 times, as opposed to 3 and 2 times respectively in
Example 6.2. Since they are the most expensive to execute we would expect
this to be slower (this is confirmed immediately below).
✷
6.3
Static Priority Experiments
Priority granularity Table 8 gives runtime and propagation steps of various
priority granularities compared to a propagation engine using a FIFO queue (and
using dynamic fixpoint reasoning, events of types {fix, bc, dmc}, and fully dynamic event sets). The three different experiment capture different priority granularities: “small” uses three priorities (unary-ternary, linear-quadratic,
cubic-veryslow); “medium” uses seven priority levels (from unary to veryslow);
“full” uses 14 priority levels (from unary-high to veryslow-low).
The results illustrate that even when there are substantially more propagations (for example, golomb-10-{b,d}) there can be significant savings. Priori-
28
ties can have a very substantial saving (almost 50% for golomb-10-d) and the
worst case cost on the benchmarks is only 10.3%. Overall while the “medium”
range of priorities is preferable on many benchmarks, the “full” range of priorities gives real speedups on two more benchmarks (photo and square-7-b) and
tends to reduce the worst case behavior of “medium”.
Examples such as queens-400, queens-400-a, and sequence-500 for “small”
only feature propagators with the same priority (the number of propagation
steps remains the same). Hence the increase in runtime by less than 1% describes the overhead of using priorities at all.
Depending on the underlying system, increasing the granularity also requires
more memory. Gecode, as a system based on recomputation and copying, constitutes the worst case in that each copied node in the search tree maintains
queues for all priority levels. However, the average increase in used (not allocated as before) memory is +0.2% for “small”, +0.7% for “medium”, and +1.5%
for “full” and hence can be neglected.
A broad spectrum of priorities will be useful for the optimizations presented
in Section 7. Therefore it is important that while “full” does not offer huge
advantages over “medium”, it neither degrades overall performance nor requires
much memory.
Priorities and stacks Table 9 gives the runtime and propagation steps of
using priorities together with stacks or combinations of stacks and queues for
the different priority levels. All numbers are given relative to a propagation
engine using the “full” priority spectrum with only queues for each priority
level. All propagation engines considered also use the full priority spectrum. The
propagation engine for “for all” uses only stacks for all priority levels, whereas
“for 1, 2, 3-ary” (“for 1, 2-ary”) uses stacks for priority levels unary-high to
ternary-low (unary-high to binary-low) and queues for the other levels.
The numbers for “for all” clarify that the misbehavior of LIFO stacks is
not due to priority inversion. The folklore belief that LIFO stacks are good
for small propagators is refuted by the numbers for “for 1, 2, 3-ary” and “for
1, 2-ary”. Only three examples show improvement in both cases while grocery
“for 1, 2, 3-ary” already exhibits pathological behavior. The measurements show
that queue versus stack does not matter for unary or binary constraints whereas
stacks are wrong for anything else.
Issues to avoid Table 10 shows runtime and propagation steps of using priorities in flawed ways. The experiment “complete fixpoints” refers to the model
proposed in [14] where fixpoints are always completed before possibly switching
to a higher priority level. As to be expected, the number of propagation steps
is reduced, however at the expense of increased runtime. More importantly,
two examples exhibiting substantial slowdown (golomb-10-d and square-5-d)
are particularly relevant as they feature propagators of vastly different priority
levels.
The surprising behavior of bibd-7-3-60 appears to be due to a problem
29
Table 9: Priority experiments with stacks and priorities.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
for all
time
steps
−0.6%
±0.0%
+47.2%
+99.0%
+5.5%
+49.7%
−4.2%
−7.3%
+0.6%
+10.4%
+0.3%
±0.0%
−0.6%
+2.4%
+7.7%
+23.2%
+7.1%
+27.0%
+5.7%
+39.2%
+62.7%
+77.1%
+42.8%
+7.1%
+0.1%
+5.0%
+1545.4%
+1106.4%
−3.3%
+5.2%
−2.2%
−2.9%
−1.3%
−0.8%
+21.1%
+79.9%
−0.1%
−0.1%
−1.0%
−0.3%
+54.1%
+240.9%
+65.2%
+31.6%
+1.7%
+18.6%
+8.9%
+35.2%
−2.5%
+1.7%
+23.9%
+36.0%
30
for 1, 2, 3-ary
time
steps
+1.3%
±0.0%
+2.3%
±0.0%
−4.0%
+1.4%
+1.2%
−2.0%
+2.0%
±0.0%
+1.6%
±0.0%
+1.0%
±0.0%
+2.5%
±0.0%
+9.7%
+27.0%
+7.7%
+39.2%
+2.3%
±0.0%
+45.8%
+7.1%
+2.1%
+5.0%
+2.2%
±0.0%
+1.0%
+0.3%
+2.2%
+2.3%
−1.0%
−0.9%
−1.3%
±0.0%
+1.0%
−0.1%
+1.2%
±0.0%
+1.4%
±0.0%
+1.2%
±0.0%
+1.4%
±0.0%
+2.9%
±0.0%
+3.0%
+0.2%
+3.3%
+2.8%
for 1, 2-ary
time
steps
+0.2% ±0.0%
+1.0% ±0.0%
−5.9% +1.4%
+1.8% −2.0%
+1.7% ±0.0%
+0.9% ±0.0%
+0.3% ±0.0%
+1.7% ±0.0%
+1.2% ±0.0%
+1.0% ±0.0%
+1.5% ±0.0%
+1.6% −0.3%
+1.4% +5.0%
+2.4% ±0.0%
−0.4% +0.3%
+2.2% +0.4%
+0.1% −0.9%
−1.2% ±0.0%
+0.8% −0.1%
+1.2% ±0.0%
+0.8% ±0.0%
+0.8% ±0.0%
+0.8% ±0.0%
+2.6% ±0.0%
+3.0% +0.2%
+0.8% +0.2%
Table 10: Priority experiments: issues to avoid.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
complete
time
−0.8%
+0.5%
−11.7%
−1.2%
−4.2%
+0.9%
+0.9%
+1.6%
−2.5%
+14.6%
+1.9%
+29.3%
−0.9%
+5.5%
+7.3%
+0.4%
+1.2%
−3.6%
−0.7%
+0.3%
+0.2%
+32.5%
+0.5%
+2.6%
−3.7%
+2.5%
fixpoints
steps
±0.0%
±0.0%
−45.8%
−0.9%
−41.1%
±0.0%
±0.0%
±0.0%
−14.5%
−14.2%
+0.1%
−23.4%
±0.0%
+8.6%
−3.8%
−1.2%
±0.0%
±0.0%
±0.0%
±0.0%
±0.0%
+7.9%
±0.0%
±0.0%
−9.0%
−6.6%
31
inverse priorities
time
steps
+13296.5%
+32.1%
+0.3%
±0.0%
−7.7%
−42.8%
+14.4%
+40.1%
−7.4%
−52.0%
+36.8%
+31.9%
+6.2%
−7.1%
+0.2%
±0.0%
+892.0%
−5.7%
+945.4%
−2.2%
+336.8%
+51.1%
+35.3%
+9.1%
+8.7%
+43.5%
+2732.2%
+2246.7%
+105.1%
+18.7%
+108.3%
+66.8%
+18.6%
−1.9%
−3.7%
±0.0%
+0.2%
±0.0%
−0.3%
±0.0%
−0.2%
±0.0%
+50.1%
+7.5%
+177.4%
−7.6%
±0.0%
±0.0%
+14.7%
+11.3%
+107.5%
+18.4%
with how the priorities for the two different kinds of propagators used in this
example (lex and Boolean-sum) are classified. This clarifies that even with a
rich spectrum of priority levels at disposal it remains difficult to assign priority
levels to propagators.
The experiment “inverse priorities” shows numbers for a propagation engine
where a propagator with lowest priority is executed first. The experiment shows
that there is a point to the priority levels, high priority = fast propagator. While
the number of propagations is often reduced the approach is rarely better than
no priorities and sometimes catastrophically worse.
The experiment “inverse priorities” clarifies a very important aspect of priorities: they not only serve as a means to improve performance, they also serve
as a safeguard against pathological propagation order.
6.4
Dynamic Priorities
As evaluation proceeds, variables become fixed and propagators can be replaced
by more specialized versions. If a propagator is replaced by a more specialized
version, also its priority should change.
Example 6.4 [Updating a propagator] Consider the propagator fO for updating x1 in the constraint x1 = x2 + x3 defined by
fO (D)(x1 ) =
fO (D)(x) =
D(x1 ) ∩ [inf D (x2 ) + inf D (x3 ) .. supD (x2 ) + supD (x3 )]
D(x)
x 6= x1
might have initial priority ternary. When the variable x2 becomes fixed to d2
say, then the implementation for x1 can change to
fO (D)(x1 ) = D(x1 ) ∩ [d2 + inf D (x3 ) .. d2 + supD (x3 )]
and the priority can change to binary.
✷
Changing priorities is also relevant when a propagator with n > 3 variables
with priority linear (or worse) reduces to a binary or ternary propagator.
6.5
Dynamic Priority Experiments
Table 11 shows runtime and propagation steps for an engine using dynamic
priorities compared to an engine using all optimizations introduced earlier and
the full priority spectrum. Dynamically changing the priority of propagators as
they become smaller due to fixed variables can lead to significant improvements.
In effect, constraints that become smaller (and thus run at higher priority) are
run first causing the still large constraints to be run less often.
This is in particular true for alpha with initially only propagators for linear equalities with priority linear. When fixing variables during search many
of these propagators are then run at priority levels binary and ternary. It
is worth noting that using dynamic priorities can disturb the FIFO queue behavior: for sequence-500 it appears to be more important to run all exactly
32
Table 11: Dynamic priority experiments.
Example
alpha
cars
crowded-chess-7
o-latin-7-d
picture
sequence-500
average (above)
average (all)
dynamic
time
steps
−25.5% −41.7%
−5.9% −13.4%
−5.0% −26.5%
−7.3%
−7.4%
−1.9%
±0.0%
+6.0% +34.8%
−7.1% −12.1%
−1.7%
−3.0%
propagators at the same priority level. Running some of the exactly propagators at priorities binary and ternary is not beneficial and disturbs the queue
behavior.
Dynamic priorities incur the overhead to compute the priority based on the
number of not yet fixed variables. However, the overhead is still small enough
to make dynamic priorities worthwhile overall.
7
Combining Propagation
There are many ways to define a correct propagator f for a single constraint
c: the art of building propagators is to find good tradeoffs in terms of speed
of execution versus strength of propagation. Typically a single constraint may
have a number of different propagator implementations: the cheapest simple
propagator, a more complex bounds propagator, and a more complex domain
propagator, for example.
Example 7.1 [alldifferent propagators] Consider the propagator fP (D)
for the alldifferent constraint.
E := ∅
for i ∈ [1 .. n]
if (∃d.D(xi ) = {d})
if (d ∈ E) return D⊥ else E := E ∪ {d}
for i ∈ [1 .. n]
if (|D(xi )| > 1) D(xi ) := D(xi ) − E
return D
The propagator does a linear number of set operations in each invocation and is
checking. It can be made idempotent by testing that no variable becomes fixed.
Another propagator for the same constraint is the domain propagator fN
introduced in [29] with complexity O(n2.5 ).
✷
33
Given two propagators, say f1 and f2 in prop(c), where f1 is strictly stronger
than f2 (f1 (D) ⊑ f2 (D) for all domains D), we could choose to implement c by
just f1 or just f2 trading off pruning versus execution time.
Without priorities there is no point in implementing the constraint c using
both propagators, since f1 will always be run and always compute stronger
domains than f2 .
We could possibly merge the implementation of the propagators to create
a new propagator f12 (D) = f1 (f2 (D)). By running the cheaper propagator
immediately first we hope that we can (a) quickly determine failure in some cases
and (b) simplify the domains before applying the more complicated propagator
f1 . While this immediate combination of two propagators in essence is simply
building a new propagator, once we have priorities in our propagation engine
we can use two or more propagators for the same constraint in different ways.
7.1
Multiple Propagators
Once we have priorities it makes sense to use multiple propagators to implement
the same constraint. We can run the weaker (and presumably faster) propagator
f2 with a higher priority than f1 . This makes information available earlier to
other propagators. When the stronger propagator f1 is eventually run, it is able
to take advantage from propagation provided by other cheaper propagators.
Note that this is essentially different from having a single propagator f12
that always first runs the algorithm of f2 and then the algorithm of f1 .
Example 7.2 [Multiple alldifferent] Consider the two propagators fP and
fN defined in Example 7.1 above. We can use both propagators: fP with priority
linear, and fN with priority quadratic. This means that we will not invoke
fN until we have reached a fixpoint of fP and all linear and higher priority
propagators.
Consider the additional propagator fE for the constraint 3x1 = 2x2 , which
has priority binary. Consider the domain D where D(x1 ) = {4, 6}, D(x2 ) =
{6, 9}, D(x3 ) = {6, 7} and D(x4 ) = · · · = D(xn ) = [1 .. n], which is a fixpoint
for fE , fP and fN . Now assume the domain of x3 is reduced to 6. Propagator
fP is placed in queue linear and fN is placed in queue quadratic. Applying
fP removes 6 from the domain of all the domains of x1 , x2 , x4 , . . . , xn , and this
causes fE to be placed in queue binary. This is the next propagator considered
and it causes failure. Propagator fN is never executed.
If we just use fN then we need to invoke the more expensive fN to obtain
the same domain changes as fP , and then fail.
✷
7.2
Staged Propagators
Once we are willing to use multiple propagators for a single constraint it becomes
worth considering how to more efficiently manage them. Instead of using two
(or more) distinct propagators we can combine the several propagators into a
single propagator with more effective behavior.
34
We assume that a propagator has an internal state variable, called its stage.
When it is invoked, the stage determines what form of propagation applies.
Example 7.3 [Staged alldifferent] Consider the alldifferent constraint
with implementations fP and fN discussed in Example 7.1. We combine them
into a staged propagator as follows:
• On a fix(x) event, the propagator is moved to stage A, and placed in the
queue with priority linear.
• On a dmc(x) event, unless the propagator is in stage A already, the propagator is put in stage B, and placed in the queue with priority quadratic.
• Execution in stage A uses fP , the propagator is put in stage B, and placed
in the queue with priority quadratic, unless it is subsumed.
• Execution in stage B uses fN , afterwards the propagator is removed from
all queues (stage NONE).
The behavior of the staged propagator is identical to the multiple propagators for the sample execution of Example 7.1. In addition to the obvious
advantage of having a single staged propagator, another advantage comes from
avoiding the execution of fN when the constraint is subsumed.
✷
In addition to giving other propagators with higher priority the opportunity
to run before the expensive part of a staged propagator, the first stage of a
propagator can already determine that the next second does not need to be
run. This is illustrated by the following example.
Example 7.4 [Staged linear equations] Consider the unit coefficient linear
equation Σni=1 ai xi = d constraint where |ai | = 1, 1 ≤ i ≤ n. We have two implementations, fQ , which implements bounds(R) consistency (considering real
solutions, with linear complexity) for the constraint, and fR , which implements
domain consistency (with exponential complexity).
We combine them into a staged propagator as follows:
• On a bc(x) (or depending on the event types available: lbc(x) or ubc(x))
event, the propagator is moved to stage A, and placed in the queue with
priority linear.
• On a dmc(x) event, unless the propagator is in stage A already, the propagator is put in stage B, and is placed in the queue with priority veryslow.
• Execution in stage A uses fQ , afterwards the propagator is put in stage B,
and placed in the queue with priority veryslow, unless each xi has a range
domain in which case it is removed from all queues (stage NONE).
• Execution in stage B uses fR , afterwards the propagator is removed from
all queues (stage NONE).
35
Table 12: Combination experiments.
Example
donald-b
donald-d
golomb-10-b
golomb-10-d
graph-color
o-latin-7-d
partition-32
photo
picture
square-5-d
square-7-b
average (above)
average (all)
immediate
time
steps
+0.7% ±0.0%
−1.6% −2.4%
+0.3% ±0.0%
−2.7% +0.2%
−0.6% −0.6%
−1.2% +1.0%
−0.6% ±0.0%
±0.0% ±0.0%
+0.3% ±0.0%
−29.6% −7.4%
−0.1% ±0.0%
−3.6% −0.9%
−1.6% −0.4%
multiple
time
steps
−15.4%
+10.4%
±0.0%
+111.9%
−10.1%
+9.2%
−14.5%
+8.6%
+2.0%
+23.0%
−9.6%
+14.3%
−4.3%
+7.4%
−0.9%
+7.8%
−0.1%
±0.0%
−38.3%
+28.8%
−17.5%
−11.0%
−10.7%
+16.1%
−4.8%
+6.8%
staged
time
steps
−16.9% +10.1%
−1.6% +88.1%
−9.8%
+9.2%
−16.1%
+8.6%
−14.4% +12.3%
−14.0% +10.9%
−6.2%
+4.2%
−1.5%
+7.8%
−3.4%
±0.0%
−42.8% +56.0%
−18.6% −10.9%
−14.0% +15.3%
−6.5%
+6.5%
The staged propagator is advantageous since the “fast” propagator fQ can
more often determine that its result D′ = fQ (D) is also a fixpoint for fR .
✷
Staged propagators are widely applicable. They can be used similarly for the
bounds(Z) version of the alldifferent constraint. Another area where staged
propagators can be used is constraint-based scheduling, where typically different
propagation methods with different strength and efficiency are available [2].
Staging is not limited to expensive propagators, it is already useful for binary
(for example, combining bounds and domain propagation for the absolute value
constraint abs(x) = y) and ternary constraints (for example, combining bounds
and domain propagation for the multiplication constraint x × y = z).
It is important to note that staging requires a sufficiently rich spectrum of
priorities. For example, to use staging for binary or ternary propagators as mentioned above, at least two different priority levels must be available for staging.
This explains why the full priority spectrum is useful: here, for binary propagators two priorities binary-high and binary-low are available. Likewise,
ternary-high and ternary-low are available for ternary propagators.
7.3
Combining Propagation Experiments
Table 12 presents runtime and propagation steps of different propagator combination schemes compared to a propagation engine using the full priority spectrum and all optimizations presented so far. The experiment “immediate” uses
a single propagator that always runs the first stage immediately followed by
the second stage. For experiment “multiple”, multiple propagators for different
stages (as discussed in Section 7.1) are used, whereas for experiment “staged”
full staging is used (as described in Section 7.2).
A quite surprising result is that “immediate” offers only modest or even
36
no speedup. The only exception is square-5-d using bounds(R) propagation
immediately before domain propagation for several linear equation propagators.
Using multiple propagators leads to an average reduction in runtime by 10%
with a slowdown for just a single example (graph-color). As to be expected,
the number of propagation steps rises sharply. This is due to the fact that more
propagators need to be run and that, similar to the introduction of priorities,
propagators with high priority are run more often. The exceptional case of
square-7-b where the number of propagation steps decreases appears to be a
fortunate change in propagation order.
Staged propagation offers another level of improvement over using multiple
propagators: all examples now achieve speedup. As fewer propagators must
be executed compared to “multiple”, also the number of propagation steps decreases (apart from square-5-d being another case of changing propagation
order). Due to the reduced overhead compared to “multiple”, even small examples such as graph-color are able to benefit from staged execution.
The exact improvement in runtime of donald-d for “immediate” and “staged”
is due to early fixpoint detection for a big linear equation propagator as discussed in Example 7.4.
The memory requirements for “immediate” and “staged” are unchanged.
The use of multiple propagators for “multiple” leads to an average increase of
6.4% in allocated memory for the examples shown in Table 12.
In summary, staged propagation is very effective for all examples, so it should
clearly be used.
8
Experiment Summary
In Table 13 we summarize the effect of all improvements suggested in this paper.
The naive propagation engine is compared to an engine featuring all techniques
introduced in this paper: dynamic fixpoint reasoning, {dmc, fix, bc} fully dynamic events, dynamic priority based LIFO queuing with the full priority spectrum, and staged propagators.
It is interesting to note that all examples but warehouse show an improvement in runtime and that almost 75% of the examples show an improvement
of at least 10%. The improvement in runtime does not incur a large increase
in memory: the largest increases are for the three donald-* problems, where
the increase is actually negligible in absolute terms and due to the underlying
memory allocation strategy (as discussed in Section 5.2).
The effects of the individual optimizations discussed in this paper could be
summarized as follows. Dynamic fixpoint reasoning subsumes static reasoning
and is easy to implement, it provides a modest improvement in execution times.
Events, while used in all finite domain propagation engines, have less benefit
than perhaps was assumed by developers. Using dynamic events again leads to a
modest improvement in execution times. The fairness of a FIFO queue strategy
is essential for scheduling propagators. While priorities by themselves are not
that important they provide a protection against worst case behavior and enable
37
Table 13: Experiment summary.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
average (all)
no optimizations
time (ms) mem (KB)
118.31
385.4
106.56
22.2
2 279.36
6 688.6
4.64
41.7
624.25
196.9
0.70
7.4
30.37
3.2
0.38
5.4
1 347.48
40.0
2 430.00
37.0
35.87
832.4
55.41
7.7
7.46
770.3
342.48
32 454.5
574.36
242.9
8 571.24
160.4
108.87
37.0
1 553.42
450.5
4 433.12
30 286.1
16.15
566.3
517.96
6 081.5
33 391.24
43.5
10 166.24
160.3
5 690.00
144.2
0.74
29.4
—
—
38
all optimizations
time
memory
−36.1%
+25.0%
−31.2%
+9.8%
−11.0%
+5.8%
−7.7%
+1.2%
−8.1%
−6.4%
−25.7%
+19.0%
−13.0%
+68.2%
−3.3%
+44.4%
−38.5%
+9.3%
−43.7%
+4.8%
−19.9%
+9.0%
−5.9%
+7.9%
−14.7%
+4.2%
−58.8%
−5.1%
−32.5%
+5.5%
−40.4%
+9.2%
−23.8%
+10.1%
−3.0%
+18.0%
−87.6%
+0.2%
−10.3%
+4.4%
−79.8%
−49.4%
−44.2%
+11.9%
−41.2%
+8.7%
−3.9%
+11.7%
+1.2%
+1.0%
−33.3%
+7.2%
the use of multiple propagators. Staging is an important optimization that can
significantly improve performance.
9
Conclusion and Future Work
We have given a formal definition of propagation systems including idempotence,
events, and priorities used in current propagation systems and have evaluated
their impact. We have introduced dynamically changing event sets which are
shown to improve efficiency considerably. The paper has introduced multiple
and staged propagators which are shown to be an important optimization in
particular for improving the efficiency of costly global constraints.
While the improvements to an engine of a propagation based constraint
solver have been discussed for integer constraints, the techniques readily carry
over to arbitrary constraint domains such as finite sets and multisets.
A rather obvious way to further speed up constraint propagation is to consider not only cost but also estimated impact for a propagator. However, while
computing cost is straightforward it is currently not clear to us how to accurately
predict propagation impact.
A
Examples Used in Experiments
All variants of constraint propagation discussed in the paper are experimentally
evaluated. The characteristics of the examples used in evaluation are summarized in Table 14. The column “variables” gives the number of variables in the
example, whereas the column “propagators” shows the number of propagators
as implementations of constraints in the example. The column “search” shows
which search strategy is used to search for a solution (“first” is simple backtracking search for the first solution, “all” is search for all solutions, “best” is
branch-and-bound search for a best solution). The two last columns describe
how many failed nodes are explored during search (column “failures”) and how
many solutions are found (column “solutions”).
A -d at the end of the example name means that domain propagation is used
for all occurring alldifferent and linear equation constraints. Likewise, -b
means that bounds(Z) propagation is used for all alldifferent and bounds(R)
for all linear equation constraints. In contrast, for -v bounds(R) propagation is
used for all linear constraints, whereas naive propagation (eliminating assigned
values as in Example 7.1) is used for alldifferent.
If not otherwise mentioned, bounds consistency is used for arithmetic constraints (including linear constraints) and naive propagation for alldifferent.
• all-interval-500 computes a series of numbers where the distances between adjacent numbers are pairwise distinct (prob007 in [10]). The model
uses a single bounds(Z) consistent alldifferent propagator and many
binary absolute value (abs(x) = y) and ternary minus propagators.
39
Table 14: Example characteristics.
Example
all-interval-500
alpha
bibd-7-3-60
cars
crowded-chess-7
donald-b
donald-d
donald-v
golomb-10-b
golomb-10-d
graph-color
grocery
knights-10
minsort-200
o-latin-7-d
partition-32
photo
picture
queens-400
queens-400-a
sequence-500
square-5-d
square-7-b
square-7-v
warehouse
variables
1 498
26
11 760
60
163
10
10
10
46
46
201
7
2 028
399
147
128
61
625
400
400
500
25
49
49
81
propagators
1 002
21
9 693
93
275
2
2
2
46
46
566
7
2 981
398
133
134
54
50
239 400
3
502
15
19
19
76
40
search
first
all
first
all
first
first
first
first
best
best
first
first
first
first
first
first
best
first
first
first
all
first
first
first
best
failures
0
7 435
1 306
107
30 396
79
5
79
19 929
19 929
37
37
2
0
2 188
160 258
6 995
3 242
10
10
250
41 272
245 208
481 301
20
solutions
1
1
1
6
1
1
1
1
10
10
1
1
1
1
1
1
7
1
1
1
1
1
1
1
4
• alpha and donald are crypto-arithmetic puzzles involving linear equation
propagators and a single alldifferent propagator.
• bibd-7-3-60 is an instance of a balanced incomplete block design problem
with parameters (v, k, l) = (7, 3, 60) (prob028 in [10]). The model involves
Boolean-sum propagators and lex propagators for symmetry breaking.
• cars models the well known car sequencing problem from [35] using element,
exactly, and linear equation propagators (prob001 in [10]).
• crowded-chess-7 places several different chess pieces on a 7 × 7 chessboard [11]. It uses exactly, element, domain consistent alldifferent,
and bounds(R) consistent linear equation propagators.
• golomb-10 finds an optimal Golomb ruler of size 10 (prob006 in [10]) with
the usual model.
• graph-color performs clique-based graph coloring for a graph with 200
nodes. Coloring each clique uses a domain consistent alldifferent propagator.
• grocery is a small crypto-arithmetic puzzle using in particular bounds(R)
consistent multiplication propagators.
• knights-10 finds a sequence of knight moves on a 10×10 chess board such
that each field is visited exactly once and that the moves return the knight
to the starting field. The model uses a naive alldifferent propagator
and a large number of reified binary propagators.
• minsort-200 sorts 200 variables using 200 minimum propagators.
• o-latin-7 finds an orthogonal latin square of size 7 and mostly uses
domain consistent alldifferent propagators.
• partition-32 partitions two 32 number blocks such that their products
match. Uses several bounds(R) multiplication propagators, a single domain consistent alldifferent propagator, and few linear equation propagators.
• photo places 9 persons on a picture such that as many preferences as
possible are satisfied. Uses a large bounds(Z) consistent alldifferent
propagator, a large bounds(R) consistent linear propagator, and many
reified binary propagators.
• picture models a 25×25 picture-puzzle (prob012 in [10]) using 50 regular
propagators.
• queens-400 and queens-400-a places 400 queens on a 400 × 400 chess
board such that the queens do not attack each other. queens-400 uses
quadratically many binary disequality propagators, while queens-400-a
uses three naive alldifferent-propagators.
41
• sequence-500 computes a magic sequence with 500 elements using 500
exactly propagators (prob019 in [10]).
• square-5 (square-7) computes a magic square of size 5 × 5 (7 × 7) using linear equation propagators and a single alldifferent propagator
(prob019 in [10]).
• warehouse solves a warehouse location problem following [32].
B
Evaluation Platform
All experiments use Gecode, a C++-based constraint programming library [12].
Gecode is one of the fastest constraint programming systems currently available,
benchmarks comparing Gecode to other systems are available from Gecode’s
webpage. The version used in this paper corresponds to Gecode 1.3.0 (albeit
slightly modified to ease the numerous experiments in this paper). Gecode has
been compiled with Microsoft Visual Studio Express Edition 2005.
All examples have been run on a Laptop with a 2 GHz Pentium M CPU
and 1024 MB main memory running Windows XP. Runtimes are the average of
25 runs with a coefficient of deviation less than 4% for all benchmarks.
Acknowledgments
Christian Schulte is partially funded by the Swedish Research Council (VR)
under grant 621-2004-4953. We thank Mikael Lagerkvist and Guido Tack for
many helpful suggestions that improved the paper.
References
[1] Krzysztof Apt. Principles of Constraint Programming. Cambridge University Press, Cambridge, United Kingdom, 2003.
[2] Philippe Baptiste, Claude Le Pape, and Wim Nuijten. Constraint-based
Scheduling. Kluwer Academic Publishers, Dordrecht, The Netherlands,
2001.
[3] Nicolas Beldiceanu, Warwick Harvey, Martin Henz, François Laburthe, Eric
Monfroy, Tobias Müller, Laurent Perron, and Christian Schulte. Proceedings of TRICS: Techniques foR Implementing Constraint programming Systems, a post-conference workshop of CP 2000. Technical Report TRA9/00,
School of Computing, National University of Singapore, September 2000.
[4] Frederic Benhamou. Heterogeneous Constraint Solving. In Proceedings of
the Fifth International Conference on Algebraic and Logic Programming,
volume 1139 of LNCS, pages 62–76, Aachen, Germany, 1996. SpringerVerlag.
42
[5] Mats Carlsson and Nicolas Beldiceanu. Revisiting the lexicographic ordering constraint. Technical Report T2002-17, Swedish Institute of Computer
Science, Stockholm, Sweden, 2002.
[6] Mats Carlsson, Greger Ottosson, and Björn Carlson. An open-ended finite
domain constraint solver. In Programming Languages: Implementations,
Logics, and Programs, 9th International Symposium, PLILP’97, volume
1292 of LNCS, pages 191–206, Southampton, United Kingdom, September
1997. Springer-Verlag.
[7] Andre Chamard, Annie Fischler, Dominique-Benoit Guinaudeau, and Andre Guillard. CHIC lessons on CLP methodology. Technical report, Dassault Aviation, 1995.
[8] Chiu Wo Choi, Warwick Harvey, Jimmy Ho-Man Lee, and Peter J.
Stuckey. Finite domain bounds consistency revisited. Technical report,
http://arxiv.org/abs/cs.AI/0412021, 2004.
[9] Philippe Codognet and Daniel Diaz. Compiling constraints in clp(FD).
Journal of Logic Programming, 27(3):185–226, June 1996.
[10] CSPLib. CSPLib: a problem library for constraints, 2006. Available from
http://www.csplib.org.
[11] Henry E. Dudeney. Amusements in Mathematics. Dover, New York, NY,
USA, 1958.
[12] Gecode Team. Gecode: Generic constraint development environment, 2006.
Available from http://www.gecode.org.
[13] Ian P. Gent, Chris Jefferson, and Ian Miguel. Watched literals for constraint
propagation in Minion. In Frédéric Benhamou, editor, Twelfth International Conference on Principles and Practice of Constraint Programming,
volume 4204 of LNCS, pages 182–197. Springer-Verlag, Nantes, France,
September 2006.
[14] Laurent Granvilliers and Eric Monfroy. Implementing constraint propagation by composition of reductions. In Logic Programming: Proceedings of
the 19th International Conference, volume 2916 of LNCS, pages 300–314,
Mumbai, India, 2003. Springer-Verlag.
[15] Warwick Harvey. Personal communication, April 2004.
[16] Warwick Harvey and Peter J. Stuckey. Improving linear constraint propagation by changing constraint representation. Constraints, 8(2):173–207,
2003.
[17] ILOG S.A. ILOG Solver 5.0: Reference Manual. Gentilly, France, August
2000.
43
[18] Intelligent Systems Laboratory. SICStus Prolog user’s manual, 3.11.1.
Technical report, Swedish Institute of Computer Science, Box 1263, 164
29 Kista, Sweden, 2004.
[19] François Laburthe. CHOCO: implementing a CP kernel. In Beldiceanu
et al. [3], pages 71–85.
[20] Olivier Lhomme, Arnaud Gotlieb, and Michel Rueher. Dynamic optimization of interval narrowing algorithms. Journal of Logic Programming, 37(1–
3):165–183, 1998.
[21] Alan K. Mackworth. Consistency in networks of relations. Artificial Intelligence, 8(1):99–118, 1977.
[22] Kim Marriott and Peter J. Stuckey. Programming with Constraints: an
Introduction. The MIT Press, Cambridge, MA, USA, 1998.
[23] Roger Mohr and Thomas C. Henderson. Arc and path consistency revisited.
Artificial Intelligence, 28:225–233, 1986.
[24] Roger Mohr and Gérald Masini. Good old discrete relaxation. In Yves
Kodratoff, editor, Proceedings of the 8th European Conference on Artificial
Intelligence (ECAI 88), pages 651–656, Munich, Germany, 1988. Pitmann
Publishing.
[25] Matthew W. Moskewicz, Conor F. Madigan, Ying Zhao, Lintao Zhang, and
Sharad Malik. Chaff: Engineering an efficient SAT solver. In Proceedings
of the 38th Design Automation Conference, DAC 2001, pages 530–535, Las
Vegas, NV, USA, 2001. ACM.
[26] Mozart Consortium. The Mozart programming system, 1999. Available
from www.mozart-oz.org.
[27] Gilles Pesant. A regular language membership constraint for finite sequences of variables. In Mark Wallace, editor, Tenth International Conference on Principles and Practice of Constraint Programming, volume 3258
of LNCS, pages 482–495. Springer-Verlag, Toronto, Canada, September
2004.
[28] Jean-Francois. Puget. A fast algorithm for the bound consistency of alldiff
constraints. In Proceedings of the 15th National Conference on Artificial
Intelligence (AAAI-98), pages 359–366, Madison, WI, USA, July 1998.
AAAI Press/The MIT Press.
[29] Jean-Charles Régin. A filtering algorithm for constraints of difference in
CSPs. In Proceedings of the Twelfth National Conference on Artificial Intelligence, volume 1, pages 362–367, Seattle, WA, USA, 1994. AAAI Press.
[30] Pierre Savéant. Constraint reduction at the type level. In Beldiceanu et al.
[3], pages 16–29.
44
[31] Christian Schulte and Peter J. Stuckey. When do bounds and domain
propagation lead to the same search space? Transactions on Programming
Languages and Systems, 27(3):388–425, May 2005.
[32] Pascal Van Hentenryck. The OPL Optimization Programming Language.
The MIT Press, Cambridge, MA, USA, 1999.
[33] Pascal Van Hentenryck, Vijay Saraswat, and Yves Deville. Constraint
processing in cc(FD). Draft, 1991.
[34] Pascal Van Hentenryck, Vijay Saraswat, and Yves Deville. Design, implementation and evaluation of the constraint language cc(FD). Journal of
Logic Programming, 37(1–3):139–164, 1998.
[35] Pascal Van Hentenryck, Helmut Simonis, and Mehmet Dincbas. Constraint
satisfaction using constraint logic programming. Artificial Intelligence,
58:113–159, 1992.
[36] Mark Wallace, Stefano Novello, and Joachim Schimpf. Eclipse: A platform
for constraint logic programming. Technical report, IC-Parc, Imperial College, London, GB, August 1997.
[37] Neng-Fa Zhou. Programming finite-domain constraint propagators in action rules. Theory and Practice of Logic Programming, 6(5):483–508,
September 2006.
45
| 6cs.PL
|
arXiv:1802.05324v1 [cs.IT] 14 Feb 2018
1
Advancing System Performance with Redundancy:
From Biological to Artificial Designs
Anh Tuan Nguyen1∗ , Jian Xu1 , Diu Khue Luu1 , Qi Zhao2 , and Zhi
Yang1
1 Biomedical Engineering, University of Minnesota, MN, USA.
2 Computer Science and Engineering, University of Minnesota, MN, USA.
∗ Email: [email protected].
Keywords: Representational redundancy, entangled redundancy, precision enhancement, redundant sensing, muscle redundancy, deep residual networks, bio-inspired designs
Abstract
Redundancy is a fundamental characteristic of many biological processes such as those
in the genetic, visual, muscular and nervous system; yet its function has not been fully
understood. The conventional interpretation of redundancy is that it serves as a fault-
tolerance mechanism, which leads to redundancy’s de facto application in man-made
systems for reliability enhancement. On the contrary, our previous works have demonstrated an example where redundancy can be engineered solely for enhancing other
aspects of the system, namely accuracy and precision. This design was inspired by
the binocular structure of the human vision which we believe may share a similar operation. In this paper, we present a unified theory describing how such utilization of
redundancy is feasible through two complementary mechanisms: representational redundancy (RPR) and entangled redundancy (ETR). Besides the previous works, we
point out two additional examples where our new understanding of redundancy can be
applied to justify a system’s superior performance. One is the human musculoskeletal system (HMS) - a biological instance, and one is the deep residual neural network
(ResNet) - an artificial counterpart. We envision that our theory would provide a framework for the future development of bio-inspired redundant artificial systems as well
as assist the studies of the fundamental mechanisms governing various biological processes.
1 Introduction
Redundancy is a well-known characteristic of many biological processes from the molecular to the systematic level. For example, the human’s genome is highly redundant: a
particular gene can be duplicated at various regions of DNA while multiple genes can
encode the same or similar biochemical functions and phenotype expressions. These
genetic redundancy and functional redundancy are observed in many crucial pathways
2
of the developmental, signaling, and cell cycle processes (Tautz, 1992; Nowak, 1997;
Kafri, 2009). High level of redundancy can also be found in the nervous system. The
neuronal architecture and synaptic interconnections have been shown to be highly redundant which allows them to facilitate complex processes of information processing,
learning, memorizing, and self-repairing. In fact, it is believed that the human brain
is at least twice the size as necessary for its function as a result of neural redundancy
(Glassman, 1987).
In many scenarios, the redundant structure of a biological system can be seen as a
consequence of the evolutionary process. Under the pressure of natural selection, living
organisms develop multiple different strategies that achieve the same goal: survival. It
is not uncommon for distinct strategies that emerge from entirely different evolutionary
pathways to resolve the same biological problem. These strategies could co-exist in
the same ecosystem or even the same organism’s genome creating observable repeated
evolutional behaviors such as functional redundancy, parallel evolution, and convergent
evolution (York, 2017). Redundancy also serves as a defence mechanism against failures which contribute to a higher survival rate. For example, gene duplication has been
shown to mitigate effects of mutations and reduce the chance of catastrophic phenotype
expression (Kafri, 2009). Redundancy also helps the human brain tolerate significant
damages and loss of mass due to injuries or diseases. Damaged neurons and brain
tissue generally do not regrow, yet their redundant structures allow reorganization of
the neuronal circuits to recover many basic brain functions (Glassman, 1987). Lastly,
redundancy increases the organism’s adaptivity. For example, genetic and functional
duplication has been shown to be the basis of phenotypic plasticity which allows an
3
organism to adapt and survive rapidly-changing endogenous and exogenous environmental conditions (Kafri, 2009).
Many of these principles find their application in designing and engineering of artificial systems. However, almost all intentional utilization of redundancy in man-made
systems focus on enhancing reliability, which importance is often overshadowed by the
system’s performance. Also, existing methods for incorporating redundancy involves
the replication of partial or entire systems which require large resources overhead. As
a result, redundant designs such as dual modular redundancy (DMR) or triple modular redundancy (TMR) are mostly found in specialized systems that perform critical
functions such as aircraft controllers, biomedical implants, and computer servers, etc.
In this paper, we argue two counter-intuitive arguments. First, redundancy can be
engineered solely for enhancing systems’ performance regarding accuracy and precision, instead of reliability and plasticity. Secondly, a practical implementation of redundancy is feasible without replication and excessive resource overhead, thus mitigating
the trade-off encountered by conventional designs. The performance boost in our proposed framework is achieved by employing two complementary mechanisms, namely
RPR and ETR. RPR describes how information is redundantly encoded and processed,
while ETR allows realizing of a RPR scheme in actual applications.
In (Nguyen, 2015, 2016), we have shown a simple but practical application where
redundancy resembling the binocular structure of the human vision is applied to enhance the precision of a man-made sensor without incurring compromises often seen in
conventional architectures. In theory, the RPR and ETR principles utilized in our design
can be generalized to different applications, and also serve as a fundamentally structural
4
characteristic of more complex systems. This argument is further asserted in this paper
by examining empirical evidence in two different systems from two distinct fields of
science and engineering. One is the HMS - a biological system where redundancy contributes to generating complex and precise muscle movements; another is the ResNet
- an artificial deep learning architecture where redundancy helps accomplish superior
predicting accuracy compared to conventional methods. By understanding subtle yet
sophisticated roles of redundancy in these systems, we believe that the findings would
not only enrich our knowledge of biological processes but also inform the derivation of
new methods for advancing the performance of man-made designs.
The remains of the paper are organized as follows. Section 2 consolidates our redundant model comprised of RPR and ETR mechanism. Section 3 examines the evidence
suggesting the implication of our model in biological and artificial systems, which include the proposed sensor design, the HMS, and the ResNet. Finally, section 4 concludes our findings and gives discussions on the future development of the proposed
theory.
2 Advancing Performance with Redundancy
Representational redundancy: The vast majority of artificial systems are designed
upon an orthogonal scheme of information representation where each entry of information is encoded by a unique configuration of the system. An entry of information can
be an input value, a desirable output, an intermediate instance or an operation of the
information processing pathway. Such orthogonal systems excel in efficiency because
5
they allow rapidly and unambiguously acquiring, processing and storing of information. However, any encoding/decoding scheme in practice suffers from an inevitable
level of error resulting in the limitation of its accuracy. In many computational models,
this limitation is described by Shannon’s theorem. Because of the uniqueness of the
representation scheme, any error acquired during the sampling, processing and storing
of information cannot be easily corrected without an overhead in term of resources such
as power, bandwidth, and memory, etc.
The RPR concept is designed to overcome conventional limitations by embracing
a non-orthogonal scheme of information representation. Subsequently, every entry of
information can be encoded by numerous distinct system configurations, including the
conventional one. These configurations are referred as the system’s microstates. If the
microstates are designed such that their response to error are non-homologous, in any
given instance, provided a sufficient number of distinct microstates, there exist with
asymptotic certainty one or more microstates that have a smaller error than the conventional representation. Therefore, an overall RPR-system would have a theoretical
accuracy almost always superior to the conventional counterpart with similar structure.
Entangled redundancy: The number of microstates represents the information capacity - an abstract property of the design that is not necessarily proportional to its
physical size. In order to effectively deploy a RPR-system in practice, the microstates
must be designed so that they do not incur excessive resource overhead. As a result,
the statistical distribution of the microstates with respect to error cannot be independent, but partially correlated or entangled. This concept is known as ETR. The level
of entanglement should be engineered just sufficient to create excessive redundancy
6
without trading off large amounts of resource. ETR should be differentiated from the
conventional method of creating redundancy by replication where the distribution of
repeated instances are independent of each other and the resource utilization is linearly
proportional to the level of redundancy.
Figure 1 illustrates the distinction between a conventional orthogonal system (COS),
a conventional redundant system (CRS) and a proposed RPR+ETR system (RES). An
entry of information in this illustration is a processing pathway that takes an input xi
and produce a corresponding output yi (i = 1, 2, ...). In the COS, every input/output
(I/O) pair (xi , yi ) is represented by a unique pathway which has a determined error that
cannot be easily removed. The pathways in the CRS are partially or entirely replicated,
which requires a proportional resource overhead. Although in practice, the replication
is mostly used for fault-tolerance, a marginal accuracy gain is feasible by selecting the
pathway with the least error for each input instance. The RES incorporates redundancy
by having the pathways of different I/O pairs share certain elements. Each (xi , yi ) pathways can now be represented by multiple system’s pathways, i.e. microstates, which
number increases exponentially with the number of shared elements. The RES is superior compared to COS because there almost always exists a pathway with a lower
error for any given I/O pair. The RES is also superior compared to CRS because an
exponential level of redundancy can be achieved with minimal additional resources.
Challenges: A proper implementation of RPR and ETR in the same architecture is
essential to achieve the performance boost. The goal is to create an excessive number
of microstates while utilizing their entanglement to allow the microstates to co-exist in
superposition thus requiring minimal additional resources as shown in Figure 1(c). Un-
7
࢞
࢞
= ݔଵ , ݕଵ elements
= ݔଶ , ݕଶ elements
= shared elements
Inputs
Outputs
࢞
࢞
࢟
࢟
࢞
࢞
࢞
࢞
(a) Conventional orthogonal system (ROS)
࢞
࢞
࢟
࢟
(b) Conventional redundant system (CRS)
࢞
࢞
࢟
࢟
Optimal pathway
for (ݔଵ , ݕଵ )
࢟
࢟
࢟
࢟
Optimal pathways
for (ݔଶ , ݕଶ )
࢟
࢟
(c) Proposed representational & entangled
redundancy system (RES)
Figure 1: Illustration of the differences among a conventional orthogonal system (COS),
a conventional redundant system (CRS) and a proposed RPR+ETR system (RES). An
entry of information is a processing pathway that takes an input xi and produce a corresponding output yi (i = 1, 2, ...). (a) Every input/output (I/O) pair in the COS is
represented by a unique pathway. (b) The pathways in the CRS are partially or entirely
replicated which give the system fault-tolerance properties and a marginal accuracy
gain. (c) The proposed RES achieves redundancy by having different pathways share
certain elements. The entanglement allows exponential level of redundancy with minimal additional resources. For each I/O pair, the pathway with the least error can be
selected resulting in major accuracy enhancement.
fortunately, there is no universal solution that can be applied to all types of system. In
our proof-of-concept system described in (Nguyen, 2015, 2016), redundancy is realized
by integrating two similar binary-weighted arrays, which structure is speculated to re8
semble the human’s binocular vision. In the subsequent sections, the HMS and ResNet
provide additional examples where redundancy elegantly emerges in entirely different
manners.
Furthermore, provided a redundant non-orthogonal structure of information representation, there is no universal solution to identify the optimal microstate given a particular input. In fact, in almost all examples of RES, it appears to be an NP-optimization
problem that can only be resolved by the mean of approximation. Biological processes
such as the visual and musculoskeletal system overcome this challenge by harnessing
the computational capacity of the nervous system, which is exceptionally adequate at
approximation. A similar mechanism could be utilized by the ResNet which itself is a
neural network. For engineering systems such as our redundant sensor (Nguyen, 2015),
an approximation method need to be derived, which consists of a one-shot unsupervised
error estimation and a simplified calibration algorithm.
3 From Biological to Artificial Systems
3.1 Redundant Sensing
The work of (Nguyen, 2015) shows a proof-of-concept implementation of a RPR+ETR
system: a sensor that converts analog to digital signals. A entry of information is a
digital code xD ∈ {0, 1, ..., 2N − 1} (N = resolution) representing an input analog
voltage. In practice, each code is generated by assembling a set of components which
are miniature capacitors embedded on a silicon chip which number is proportional to the
required physical resources and cost. The random error occurred during the fabrication
9
Component
Set
Conventional
Redundant
Sensing
{ , ,
Digital Codes
1
2
3
4
5
6
7
}
{ , , , , }
= Physical unit-component
Figure 2: Illustration of the RPR and ETS properties of the proposed redundant sensing
(RS) architecture (Nguyen, 2015). The device converts an analog input to a digital
output by assembling a set of physical components. While utilizing the same number
of unit-components (2N − 1 = 7, N: sensor’s resolution) as a conventional design, the
component set of a RS architecture allows each digital code to be created by multiple
different assemblies, i.e. microstates. By selecting the microstate with the least error for
every code, a significant boost of the accuracy can be achieved. This example illustrates
a simplified case of N = 3. The RS is most effective for high resolution because the
number of microstates increases exponentially with N.
process of these capacitors, i. e. mismatch error, has been shown to be a major factor
limiting the device’s accuracy.
Figure 2 compares the differences between a conventional and a RS architecture in
a simplified case of N = 3. The conventional system utilizes a binary-weighted set of
components which is the most efficient encoding scheme yet vulnerable to mismatch
10
error. On the other hand, the proposed RS architecture employs a non-orthogonal component set which satisfies both RPR and ETR requirements. With the name number of
unit-component (2N − 1 = 7), RS allows each digital code to be generated by multiple
different component assemblies, i.e. microstates. The number of microstates increases
exponentially with N resulting in major accuracy enhancement of high-resolution devices.
Interestingly, the RS component set resembles exchanging and integrating the information between two smaller conventional binary-weighted sub-arrays, which is inspired by the binocular structure of the human visual system. Thus, we ask the question:
whether RPR and ETR are fundamental properties that facilitate visual acuity? The spatial distribution of photoreceptors on the retina is notably irregular, which echoes the
impact of mismatch error. It is therefore conjectured that binocular vision effectively
creates a form of static redundancy allowing the brain to collect sufficient information
to remedy the error and produce the images we perceive. In fact, the binocular vision
has been shown to help differentiation of fine details, even exceeding the diffraction
limit of the photoreceptors - a phenomenon known as hyperacuity (Beck, 1979). Our
theory is further supported by the fact that human and many higher-order animals only
have two eyes. During the development of our RS sensor, we found out that the amount
of computational power required to process redundant information increases rapidly
with the number of sub-arrays. Two sub-arrays or two eyes is the minimum number
necessary to create a redundant structure. Despite dedicating 30-60% of its mass for
visual processing, the brain simply lacks the capacity to process information from three
or more eyes.
11
Furthermore, as a complement to the binocular structure, we conjecture that eyes’
micro-fixational movement or microsaccade (Martinez-Conde, 2013) creates a form of
dynamic redundancy. During microsaccades, the field of vision of each eye is sampled
multiple times by different spatial configurations of photoreceptors, which resemble entangled redundant microstates and facilitate visual acuity. This observation is supported
by both experiments with human subjects (Hicheur, 2013) and mathematical modeling
(Hennig, 2004) where microsaccades have been shown to play an important role in the
visual precision and could lead to hyperacuity.
3.2 Muscle Redundancy
The HMS has more muscles and joints than the necessary mechanical degrees-offreedom even though are energetically expensive to produce and maintain. This paradoxical phenomenon of muscle redundancy (MS) presents a long-standing problem in
human kinesiology of understanding how and why the human brain coordinates all
muscles and joints to achieve complex movements with precision (Bernstein, 1967).
By examining this biological process from the perspective of our model, we hope to
unravel the principles underlying the behavior of MS.
A conventional interpretation would suggest that redundancy contributes to the reliability of the HMS allowing compensation for the loss or dysfunction of individual
muscles. However, emerging empirical evidence suggests this is not true. Even a mild
dysfunction of a few critical muscles due to disorders, injuries or aging can significantly weaken the force production and overall functions of the whole HMS (Forssberg,
1991; Schreuders, 2006). The results are supported by Kutch & Valero-Cuevas works
12
1. glmed/min
2. glmax
3. hamstr
4. bfsh
5. gastroc
6. tibpost
7. soleus
8. perbrev
9. tibant
10. vasti
11. tensfl
12. rectfem
13. addlong
14. iliacus
14
13
1
12
11
2
3
10
4
5
5
14
6
12
3
7
8
9
2
10
(a)
(b)
(c)
Figure 3: (a, b) A sagittal view of a human leg’s mechanical model consisting of 14
muscles and muscle groups (Kutch & Valero-Cuevas, 2011). (c) Any specific movement trajectory and force can be achieved by multiple distinct muscles and joints combinations. Also, different muscles have overlapping but not exclusive mechanical functions, and all contribute with different degrees in generating force and movement. These
characteristics resemble a RPR+ETR system.
(Kutch & Valero-Cuevas, 2011; Valero-Cuevas, 2015). Using both computational models and empirical experiments with cadaver specimens, the authors point out that less
than 5% of the feasible forces and movements in their models are robust to a loss of any
muscle, so it is clear that reliability is not an inherited characteristic of MS.
Figure 3 presents a sagittal view of a human leg’s mechanical model used by Kutch
& Valero-Cuevas (Kutch & Valero-Cuevas, 2011) which consists of 14 muscles and
13
muscle groups1 . At the kinematic and muscular level, any specific movement trajectory and force can be achieved by virtually infinite combinations of muscles and joints.
While at the control level, each muscle consists of numerous units that can be activated by different motor neurons and patterns while resulting in the same behavior.
Furthermore, the muscles have overlapping but not exclusive mechanical functions, and
all contribute with different degrees in generating force and movement. Clearly, these
characteristics of the HMS resemble ones of a RPR+ETR system, thus our model predicts that MR plays a major role in enhancing the accuracy and precision of muscle
movements.
Indeed, the hypothesis is supported by a number of studies. In (Cleather, 2010),
the authors examine two different muscular models, namely Delp and Horsman, in predicting the patellofemoral force during standing, jumping, and weightlifting. They conclude that higher level of redundancy in the Horsman model contributes to its higher
predictive accuracy and closer realistic approximation in all activities. The authors’
conjecture is consistent with our theory which implies redundancy effectively increases
the variability and number of independent musculoskeletal movements (i.e. microstates),
so an optimal solution is more likely to be found. The argument is further strengthened
in (Moissenet, 2016) where an increased level of redundancy correlates to a better pre1 List of 14 muscles/muscle groups and their abbreviation: (1) gluteus medialis and minimus
(glmed/min); (2) gluteus maximus (glmax); (3) semimembranoseus, semitendenosis and biceps femoris
long head (hamstr); (4) biceps femoris short head (bfsh); (5) medial and lateral gastrocnemius (gastroc); (6) tibialis posterior (tibpost); (7) soleus (soleus); (8) peroneus brevis (perbrev); (9) tibialis anterior
(tibant); (10) vastus intermedius, lateralis and medialis (vasti); (11) tensor facia lata (tensfl); (12) rectus
femoris (rectfem); (13) adductor longus (addlong); (14) iliacus (iliacus).
14
dicting accuracy of tibiofemoral contact forces in all gait patterns.
3.3 Deep Residual Networks
Input
Layer 1
…
Layer 2
Layer ܰ
Output
Layer ܰ
Output
(a) Conventional feedforward network
Input
Layer 1
…
Layer 2
= skip connection/identity mapping
(b) Deep residual network
Layer 1
Layer 1
No. of distinct paths ൎ 2ே
Layer 2
Layer 1
Input
Layer 1
…
Layer 2
Layer ܰ
Output
(c) “Unraveled” deep residual network
Figure 4: (a) A conventional feedforward DNN produces a prediction by convoluting
the input through multiple feature layers. (b) A ResNet achieves superior predictive accuracy with the incorporation of skip connections/identity mappings which allow information to bypass an entire layer. (c) It has been shown that the behavior of the ResNet
is similar to a collection of shallower networks, resembling a RPR+ETR system (Veit,
2016).
There exist several prominent uses of redundancy for enhancing classification accuracy in machine learning. In a “committee machine” or “ensemble learning”, multiple
predictions are generated simultaneously by a collection of discrete instances that are
based on the same or distinct predictive models. Because each instance produces a re15
sult with a different degree of error, an appropriate integration of these outcomes could
lead to a higher overall accuracy (Bishop, 2006). Another approach involves replication of individual neurons or sub-circuits of an artificial neural network (ANN). Using
mathematical models, (Izui, 1990; Tanaka, 1988) conclude that replication can fundamentally alter the computation carried out by an ANN resulting in quantitative enhancement of convergence speed, solution accuracy, interconnection stability. The findings
were utilized to design redundant ANNs simulating a robotic arm grasping an object
in 2D space and a pattern-classification task with improved accuracy and convergence
time (Medler, 1994b,a). These are prime examples of CRS as shown in Figure 1(b).
Even though a marginal accuracy boost can be accomplished, the replication-based implementation prevents these systems from effectively employing redundancy without
incurring excessive resource overhead.
Recently, deep learning has emerged as a leading field of machine learning (LeCun,
2015). A feedforward deep neural network (DNN) produces a prediction by convoluting the inputs through various feature layers encoding the acquired knowledge (Figure
4(a)). One of the breakthroughs in DNN design - the ResNet (He, 2016a,b) - modifies
the conventional structure by including “skip connections” or “identity mapping” that
allow information to occasionally bypass an entire layer (Figure 4(b)). Empirical experiments have demonstrated the superior predictive accuracy of ResNet compared to
conventional networks with the same number of layers and parameters (He, 2016a,b;
Huang, 2016; Wu, 2016; Zagoruyko, 2017).
Although the advantage of ResNet is evident, many are baffled by how a subtle yet
critical modification of the DNN could fundamentally alter its properties. It becomes
16
clear as Veit et al. (Veit, 2016) show that the ResNet’s behaviors resemble characteristics of an ensemble of shallower networks. As illustrated in Figure 4(c), the network can
be “unraveled” as a sum of smaller sub-circuits where information can flow through any
one of the 2N distinct pathways (N: number of layers) and is integrated at the last step.
The structure resembles a proposed RPR+ETR system where each of pathway corresponds to a microstate. Because of the entanglement among microstates, an excessive
level of redundancy which is exponentially proportional to the number of layers can be
formulated without compromising the size of the network. Therefore, we argue that by
including the skip connections, the conventional DNN has been transformed into a redundant system with both RPR and ETR properties which leads to major enhancement
of performance.
Discussion & Conclusion
Although redundancy is no doubt an essential property of many biological processes,
there are reasons to believe that its functions have not been fully appreciated resulting
in the absence in artificial designs. While the conventional interpretation ties redundancy with fault-tolerance, we propose a new model arguing that it can be engineered
to advance the performance regarding accuracy and precision. Our theory highlights
two fundamental mechanisms enabling such function: (i) RPR facilitates redundant encoding of information, and (ii) ETR facilitates practical implementation of redundancy.
Besides suggesting the presence of these mechanisms in biological processes such as the
human visual and musculoskeletal systems, we present two state-of-the-art man-made
17
designs, the RS sensor (Nguyen, 2016) and the ResNet (He, 2016a), where redundancy
is successfully employed.
Clearly, there are future works needed to be done to demonstrate the feasibility of
such redundant architectures. Firstly, under the guidelines of our framework, new engineering solutions should be derived to integrate redundancy into other designs for
accuracy and precision enhancement. Although the principles of RPR and ETR are
universal, their actual implementations vary drastically. The examples in electrical engineering and computer pointed out in this paper are merely the tip of the iceberg.
Secondly, a new technique should be investigated to evaluate the information capacity
of redundant systems which correlates to its upper bound of performance. A brute force
approach utilized in the RS design (Nguyen, 2016) certainly cannot be applied to more
complex systems such as the ResNet. Finally, new methods should be developed to harness the full capacity of redundant systems. Redundant representation of information
is irrelevant without an effective way to extract the optimal configuration. In almost all
examples shown in this work, they present NP-optimization problems which solutions
can be adequately obtained by the mean of approximation.
References
Beck, J. and Schwartz, T. (1979) Vernier Acuity with Dot Test Objects. Vision Research,
19(3), 313-9
Bernstein, N. A. (1967) The Co-ordination and Regulation of Movements. Psyccritiques.
18
Bishop, C. M. (2006) Pattern Recognition And Machine Learning. Springer.
Cleather, D. J. and Bull, A. M. (2010) Lower-extremity Musculoskeletal Geometry Affects the Calculation of Patellofemoral Forces in Vertical Jumping and Weightlifting.
Journal of Engineering in Medicine, 224(9), 1073-83.
Forssberg, H., Eliasson, A., Kinoshita, H., Johansson, R. and Westling, G. (1991) Development of Human Precision Grip I: Basic Co-ordination of Force. Experimental
Brain Research, 85(2), 451-7.
Glassman, R. B. (1987) An Hypothesis About Redundancy And Reliability In The
Brains Of Higher Species: Analogies With Genes, Internal Organs, And Engineering
Systems. Neuroscience and Biobehavioral Reviews, 11(3), 275-85.
He, K., Zhang, X., Ren, S. and Sun, J. (2016a) Deep Residual Learning For Image
Recognition. IEEE Conference on Computer Vision and Pattern Recognition, 770778.
He, K., Zhang, X., Ren, S. and Sun, J. (2016b) Identity Mappings In Deep Residual
Networks. European Conference on Computer Vision, 630-45.
Hennig, M. H. and Worgotter, F. (2004) Eye Micro Movements Improve Stimulus Detection Beyond The Nyquist Limit In The Peripheral Retina. Advances in Neural
Information Processing Systems, 1475-82.
Hicheur, H., Zozor, S., Campagne, A. and Chauvin, A. (2013) Microsaccades Are Modulated By Both Attentional Demands Of A Visual Discrimination Task And Background Noise. Journal of Vision, 13(13), 18.
19
Huang, G., Sun, Y., Liu, Z., Sedra, D. and Weinberger, K. Q. (2016) Deep Networks
With Stochastic Depth. European Conference on Computer Vision, 646-61.
Izui, Y. and Pentland, A. (1990) Analysis Of Neural Networks With Redundancy. Neural Computation, 2(2), 226-38.
Kafri, R., Springer, M. and Pilpel, Y. (2009) Genetic Redundancy: New Tricks for Old
Genes. Cell, 136(3), 389-92.
Kutch, J. J. and Valero-Cuevas, F. J. (2011) Muscle redundancy does not imply robustness to muscle dysfunction. Journal of Biomechanics, 44(7), 1264-70.
LeCun, Y., Bengio, Y. and Hinton, G. (2015) Deep Learning. Nature Neuroscience,
521(7553), 436-44.
Martinez-Conde, S., Otero-Millan, J. and Macknik, S. L. (2013) The Impact Of Microsaccades On Vision: Towards A Unified Theory Of Saccadic Function. Nature
Reviews Neuroscience, 14(2), 83-96.
Medler, D. A. and Dawson, M. R. W. (1994a) Training Redundant Artificial Neural
Networks: Imposing Biology on Technology. Psychological Research, 57(1), 54-62.
Medler, D. A. and Dawson, M. R. W. (1994b) Using Redundancy to Improve The Performance of Artificial Neural Networks. Biennial Conference of the Canadian Society for Computational Studies of Intelligence on Advances in Artificial Intelligence,
131-8.
Moissenet, F., Cheze, L. and Dumas, R. (2016) Influence of the Level of Muscular
20
Redundancy on the Validity of a Musculoskeletal Model. Journal of Biomechanical
Engineering, 138(2), 021019.
Nguyen, A. T., Xu, J. and Yang, Z. (2015) A 14-Bit 0.17mm2 SAR ADC In 0.13µm
CMOS for High Precision Nerve Recording. IEEE Custom Integrated Circuits Conference, 1-4.
Nguyen, A. T., Xu, J. and Yang, Z. (2016) A Bio-Inspired Redundant Sensing Architecture. Advances in Neural Information Processing Systems, 2379-87.
Nowak, M. A., Boerlijst, M. C., Cooke, J. and Smith, J. M. (1997) Evolution Of Genetic
Redundancy. Nature Neuroscience, 388(6638), 167-171.
Schreuders, T. A., Selles, R. W., Roebroeck, M. E. and Stam, H. J. (2006) Strength
Measurements of the Intrinsic Hand Muscles: A Review of the Development and
Evaluation of the Rotterdam Intrinsic Hand Myometer. Journal of Hand Therapy,
19(4), 393-402.
Tanaka, H., Matsuda, S., Ogi, H., Izui, Y., Taoka, H. and Sakaguchi, T. (1988) Redundant Coding for Fault Tolerant Computing on Hopfield Network. Neural Networks,
1, 141.
Tautz, D. (1992) Redundancies, Development and The Flow of Information. Bioessays,
14(4), 263-6.
Valero-Cuevas, F. J., Cohn, B. A., Yngvason, H. F. and Lawrence, E. L. (2015) Exploring the High-Dimensional Structure of Muscle Redundancy via Subject-Specific and
Generic Musculoskeletal Models. Journal of Biomechanics, 48(11), 2887-96.
21
Veit, A., Wilber, M. and Belongie, S. (2016) Residual Networks Behave Like Ensembles Of Relatively Shallow Networks. Advances in Neural Information Processing
Systems, 550-8.
Wu, Z., Shen, C. and Hengel, A. V. D. (2016) Wider Or Deeper: Revisiting The Resnet
Model For Visual Recognition. arXiv preprint arXiv:1611.10080.
York, R. A. and Fernald, R. D. (2017) The Repeated Evolution Of Behavior. Frontiers
in Ecology and Evolution, 4(143), 1-10.
Zagoruyko, S. and Komodakis, N. (2017) Wide Residual Networks. arXiv preprint
arXiv:1605.07146.
22
| 7cs.IT
|
arXiv:1511.03171v3 [math.QA] 16 Mar 2016
SKEW BRACES AND THE YANG–BAXTER EQUATION
L. GUARNIERI AND L. VENDRAMIN
Abstract. Braces were introduced by Rump to study non-degenerate involutive set-theoretic solutions of the Yang–Baxter equation. We generalize
Rump’s braces to the non-commutative setting and use this new structure to
study not necessarily involutive non-degenerate set-theoretical solutions of the
Yang–Baxter equation. Based on results of Bachiller and Catino and Rizzo, we
develop an algorithm to enumerate and construct classical and non-classical
braces of small size up to isomorphism. This algorithm is used to produce a
database of braces of small size. The paper contains several open problems,
questions and conjectures.
Introduction
The Yang–Baxter equation first appeared in theoretical physics and statistical
mechanics in the works of Yang [42] and Baxter [4, 5] and it has led to several
interesting applications in quantum groups and Hopf algebras, knot theory, tensor categories and integrable systems, see for example [27], [30] and [39]. In [14],
Drinfeld posed the problem of studying this equation from the set-theoretical perspective.
Recall that a set-theoretical solution of the Yang–Baxter equation is a pair (X, r),
where X is a set and
r : X × X → X × X,
r(x, y) = (σx (y), τy (x)),
x, y ∈ X,
is a bijective map such that
(r × id)(id × r)(r × id) = (id × r)(r × id)(id × r).
Such a map r is usually called a braiding.
A solution (X, r) is said to be non-degenerate if the maps σx and τx are bijective for each x ∈ X, and (X, r) is said to be involutive if r2 = idX×X . The
seminal works of Etingof, Schedler and Soloviev [15], and Gateva-Ivanova and Van
den Bergh [24], discussed algebraic and geometrical interpretations and introduced
several structures associated with the class of non-degenerate involutive solutions.
Such solutions have been intensively studied, see for example [17, 18, 19], [21, 22,
23], [25, 26], [20], [32, 34], [6], [10], [11], [13], [28], and [40].
It was in studying involutive solutions that Rump introduced in [34] the brace
structure. In [12], Cedó, Jespers and Okniński, defined a left brace as an abelian
group (A, +) with another group structure, defined via (a, b) 7→ ab, such that the
compatibility condition
a(b + c) + a = ab + ac
holds for all a, b, c ∈ A. This definition is equivalent to that of Rump.
This work is partially supported by CONICET, PICT-2014-1376, MATH-AmSud and ICTP.
1
2
L. GUARNIERI AND L. VENDRAMIN
Many of the problems related to involutive solutions can be restated in terms of
braces. Two prominent examples are the following:
• Is every finite solvable group an involutive Yang–Baxter group? Recall
that an involutive Yang–Baxter group is a group isomorphic to the group
generated by the set {σx : x ∈ X}, where
r : X × X → X × X,
r(x, y) = (σx (y), τy (x)),
is a non-degenerate involutive solution of the Yang–Baxter equation. Based
on a sketch of proof of Rump [36], Bachiller [2] found a solvable finite group
that is not an involutive Yang–Baxter group.
• Are there good methods to contruct all finite non-degenerate involutive
solutions to the Yang–Baxter equation? Brute force seems not to be good
enough. In [3], Bachiller, Cedó and Jespers, give a method to construct all
finite solutions of a given size. For it to work, one needs the classification
of left braces.
Non-involutive solutions were studied by Soloviev [38] and Lu, Yan and Zhu [29].
Such solutions have applications in knot theory, since they produce powerful knot
and virtual knots invariants, see for example [31] and the references therein. The
following question naturally arises: Is there an algebraic structure similar to the
brace structure useful for studying non-involutive solutions? This paper introduces
the notion of skew brace and provides an affirmative answer to the above question.
Remarkably, this new structure provides the right algebraic framework to study
involutive and non-involutive braidings and allows us to restate the main results
of [29], [38] and [41].
As in the case of involutive solutions, the classification of finite skew braces is
one of the main steps needed for constructing finite solutions of the Yang–Baxter
equation. One of the main results of this paper is an explicit classification of classical
and skew braces of small size. An algorithm to construct all non-isomorphic classical
and skew braces of a given size is described. This heavily depends on results of
Bachiller [2] and Catino and Rizzo [9]. This algorithm was used to build a database
of classical and skew braces, a good source of examples that gives an explicit and
direct way to approach some of the problems related to the Yang–Baxter equation.
The database is available as a library for GAP [16] and Magma [8] immediately from
the authors on request.
The paper is organized as follows. In Section 1 we extend braces to the noncommutative setting by defining skew braces, and state their main properties. We
prove in Proposition 1.11 that skew braces are equivalent to bijective 1-cocycles.
Section 2 is devoted to a study of quotients of skew braces. It is worth mentioning
that the proofs in Section 1 and 2 are basically the same as for classical braces.
In Section 3 the connection between skew braces and the Yang–Baxter equation is
explored. In Theorem 3.1 we generalize a result of Rump and produce a canonical
solution for each skew left brace. Some reconstruction theorems similar to those
of Etingof, Schedler and Soloviev [15], Lu, Yan and Zhu [29] and Soloviev [38]
are given at the end of this section. The method for constructing classical and
skew braces is given in Section 4. Section 5 discusses the algorithm that produces
and enumerates classical and skew left braces and some consequences. Problems,
questions and conjectures are discussed in Section 6.
SKEW BRACES AND THE YANG–BAXTER EQUATION
3
1. Skew left braces
Braces were introduced by Rump in [34] to study set-theoretical involutive solutions of the Yang–Baxter equation. The following definition generalizes braces to
the non-commutative setting.
Definition 1.1. A skew left brace is a group A (written multiplicatively) with an
additional group structure given by (a, b) 7→ a ◦ b such that
(1.1)
a ◦ (bc) = (a ◦ b)a−1 (a ◦ c)
holds for all a, b, c ∈ A, where a−1 denotes the inverse of a with respect to the group
structure given by (a, b) 7→ ab.
Of course Rump’s left braces are examples of skew braces. These are braces
where the group (A, ·) is abelian.
Definition 1.2. A homomorphism between two skew left braces A and B is a map
f : A → B such that f (ab) = f (a)f (b) and f (a ◦ b) = f (a) ◦ f (b) for all a, b ∈ A.
The kernel of f is
ker f = {a ∈ A : f (a) = 1},
where 1 denotes the identity of the group (A, ·) with multiplication a · b = ab for all
a, b ∈ A.
Example 1.3. Let (A, ·) be a group. Then A is a skew left brace with a ◦ b = ab
for all a, b ∈ A. Similarly, a ⋆ b = ba defines a skew left brace structure over A.
These braces are isomorphic if and only if (A, ·) is abelian.
Example 1.4. Let A and B be groups and let α : A → Aut(B) be a group homomorphism. Then A × B has a skew left brace structure given by
(a, b)(a′ , b′ ) = (aa′ , bb′ ),
(a, b) ◦ (a′ , b′ ) = (aa′ , bαa (b′ )),
where a, a′ ∈ A and b, b′ ∈ B.
Example 1.5. Let A and B be groups and let α : A → Aut(B) be a group homomorphism. Assume that A is abelian. Then A × B has a skew left brace structure
given by
(a, b)(a′ , b′ ) = (aa′ , bαa (b′ )),
(a, b) ◦ (a′ , b′ ) = (aa′ , bb′ ),
where a, a′ ∈ A and b, b′ ∈ B.
Example 1.6. This example is motivated by the paper of Weinstein and Xu on the
Yang–Baxter equation, see [41]. Let A be a group and A+ , A− be subgroups of A
such that A admits a unique factorization as A = A+ A− . Thus each a ∈ A can be
written in a unique way as a = a+ a− for some a+ ∈ A+ and a− ∈ A− . The map
A+ × A− → A,
(a+ , a− ) 7→ a+ (a− )−1 ,
is bijective. Using this map we transport the group structure of the direct product
A+ × A− into the set A. For a = a+ a− ∈ A and b = b+ b− ∈ A let
a ◦ b = a+ ba− .
Then (A, ◦) is a group. Furthermore, A is a skew left brace.
4
L. GUARNIERI AND L. VENDRAMIN
Lemma 1.7. Let A be a skew left brace. Then the following properties hold:
(1) 1 = 1◦ , where 1◦ denotes the unit of the group (A, ◦).
(2) a ◦ (b−1 c) = a(a ◦ b)−1 (a ◦ c) for all a, b, c ∈ A.
(3) a ◦ (bc−1 ) = (a ◦ b)(a ◦ c)−1 a for all a, b, c ∈ A.
Proof. The first claim follows from (1.1) with c = 1◦ . To prove the second claim
let d = bc. Then (1.1) becomes a ◦ d = (a ◦ b)a−1 (a ◦ b−1 d) and the claim follows.
The third claim is proved similarly.
Remark 1.8. Let A be a skew left brace. For each a ∈ A the map
λa : A → A,
b 7→ a−1 (a ◦ b),
is bijective with inverse λ−1
a : A → A, b 7→ a ◦ (ab), where a is the inverse of a with
respect to ◦. It follows that
a ◦ b = aλa (b),
ab = a ◦ λ−1
a (b)
hold for all a, b ∈ A.
The following proposition extends results of Rump [34] and Gateva-Ivanova into
the non-commutative setting, see [17, Proposition 3.3].
Proposition 1.9. Let A be a set and assume that A has two operations such
that (A, ·) and (A, ◦) are groups. Assume that λ : A → SA , a 7→ λa , is given by
λa (b) = a−1 (a ◦ b). The following are equivalent:
(1) A is a skew left brace.
(2) λa◦b (c) = λa λb (c) for all a, b, c ∈ A.
(3) λa (bc) = λa (b)λa (c) for all a, b, c ∈ A.
Proof. Let us first prove that (1) =⇒ (2). Let a, b, c ∈ A. Since A is a brace and
a ◦ b−1 = a(a ◦ b)−1 a by Lemma 1.7,
λa λb (c) = a−1 (a ◦ λb (c)) = a−1 (a ◦ (b−1 (b ◦ c)))
= a−1 (a ◦ b−1 )a−1 (a ◦ b ◦ c) = (a ◦ b)−1 (a ◦ b ◦ c) = λa◦b (c).
Now we prove (2) =⇒ (3). Since ab = a ◦ λ−1
a (b) for all a, b ∈ A,
−1
λa (bc) = λa (b ◦ λ−1
(a ◦ b ◦ λ−1
b (c)) = a
b (c))
= a−1 (a ◦ b)(a ◦ b)−1 (a ◦ b ◦ λ−1
b (c))
−1
= λa (b)λa◦b λ−1
b (c) = λa (b)λa λb λb (c) = λa (b)λa (c).
Finally we prove that (3) =⇒ (1). Let a, b, c ∈ A. Then
a−1 (a ◦ (bc)) = λa (bc) = λa (b)λa (c) = a−1 (a ◦ b)a−1 (a ◦ c),
and hence a ◦ (bc) = (a ◦ b)a−1 (a ◦ c).
Corollary 1.10. Let A be a skew left brace and
λ : (A, ◦) → Aut(A, ·),
a 7→ λa (b) = a−1 (a ◦ b).
Then λ is a group homomorphism.
Proof. It follows immediately from Proposition 1.9.
SKEW BRACES AND THE YANG–BAXTER EQUATION
5
Let A and G be groups and assume that G× A → A, (g, a) 7→ g ·a, is a left action
of G on A by automorphisms. A bijective 1-cocyle is a bijective map π : G → A
such that
(1.2)
π(gh) = π(g)(g · π(h))
for all g, h ∈ G.
Proposition 1.11. Over any group (A, ·) the following data are equivalent:
(1) A group G and a bijective 1-cocycle π : G → A.
(2) A skew left brace structure over A.
Proof. Consider on A a second group structure given by
a ◦ b = π(π −1 (a)π −1 (b))
for all a, b ∈ A. Since π is a 1-cocycle and G acts on A by automorphisms,
a ◦ (bc) = π(π −1 (a)π −1 (bc)) = a(π −1 (a) · (bc))
= a((π −1 (a) · b)(π −1 (a) · c)) = (a ◦ b)a−1 (a ◦ c)
holds for all a, b, c ∈ A.
Conversely, assume that A is a skew left brace. Set G = A with the multiplication
(a, b) 7→ a◦ b and π = id. By Corollary 1.10, a 7→ λa , is a group homomorphism and
hence G acts on A by automorphisms. Then (1.2) holds and therefore π : G → A
is a bijective 1-cocycle.
Remark 1.12. The construction of Proposition 1.11 is categorical.
2. Ideals and quotients
Definition 2.1. Let A be a skew left brace. A normal subgroup I of (A, ◦) is said
to be an ideal of A if Ia = aI and λa (I) ⊆ I for all a ∈ A.
Example 2.2. Let f : A → B be a skew brace homomorphism. Then ker f is an
ideal of A since
f (λa (x)) = λf (a) (f (x)) = 1
for all x ∈ ker f and a ∈ A.
Lemma 2.3. Let A be a skew left brace and I ⊆ A be an ideal. Then the following
properties hold:
(1) I is a normal subgroup of (A, ·).
(2) a ◦ I = aI for all a ∈ A.
(3) I and A/I are skew braces.
Proof. Let a, b ∈ I. Then a−1 b = λa (a ◦ b) ∈ I and hence I is a subgroup of (A, ·).
Remark 1.8 implies
aI = a ◦ I = I ◦ a = Ia
for all a ∈ A. Thus I is a normal subgroup of (A, ·) and hence it follows that I is
a skew left brace. Since the quotient groups A/I for both operations are the same,
A/I is a skew left brace.
Definition 2.4. Let A be a skew left brace. The socle of A is
Soc(A) = {a ∈ A : a ◦ b = ab, b(b ◦ a) = (b ◦ a)b for all b ∈ A}.
6
L. GUARNIERI AND L. VENDRAMIN
Lemma 2.5. Let A be a skew left brace. Then Soc(A) is an ideal of A contained
in the center of (A, ·).
Proof. Let us first prove that Soc(A) is a subgroup of (A, ◦). Clearly 1 ∈ Soc(A).
Let a, a′ ∈ A and b ∈ A. Then a ◦ a′ ∈ Soc(A) since
(a ◦ a′ ) ◦ b = a ◦ (a′ ◦ b) = a ◦ (a′ b) = a(a′ b) = (aa′ )b = (a ◦ a′ )b.
Now since a = a−1 ∈ Soc(A) and b = (aa−1 ) ◦ b = a ◦ (a−1 ◦ b) = a(a−1 ◦ b), it
follows that ab = a−1 b = a−1 ◦ b = a ◦ b. Hence Soc(A) is a subgroup of (A, ◦).
A direct calculation proves that
(2.1)
λb (a) = b ◦ a ◦ b for all a ∈ Soc(A) and b ∈ A.
Then it follows that Soc(A) ⊆ {a ∈ A : a ◦ b = ab, λb (a) ◦ b = b ◦ a for all b ∈ A}.
Let a ∈ Soc(A) and b, c ∈ A. Then
λc λb (a) = λc◦b (a) = (c ◦ b) ◦ a ◦ c ◦ c = c ◦ λb (a) ◦ c,
λb (a)c = b−1 (b ◦ a)c = (b ◦ a)b−1 c = b ◦ (a(b ◦ c)) = b ◦ a ◦ b ◦ c = λb (a) ◦ c.
Hence λb (Soc(A)) ⊆ Soc(A) for all b ∈ A and Soc(A) is a normal subgroup of (A, ◦)
by (2.1).
Now we prove that Soc(A) is central in (A, ·). Let a ∈ Soc(A), b ∈ A and c = b.
Since
c ◦ (ba) = (c ◦ b)c−1 (c ◦ a) = c−1 (c ◦ a) = (c ◦ a)c−1 = c ◦ (ab),
it follows that ba = ab.
3. Braces and the Yang–Baxter equation
We turn our attention to the connection between skew left braces and settheoretic solutions of the Yang–Baxter equation. The following theorem generalizes
a result of Rump to the non-commutative setting, see [12, Lemma 2].
Theorem 3.1. Let A be a skew left brace. Then
(3.1)
rA : A × A → A × A,
−1
rA (a, b) = (λa (b), λ−1
a(a ◦ b)),
λa (b) ((a ◦ b)
is a non-degenerate solution of the Yang–Baxter equation. Furthermore, rA is involutive if and only if ab = ba for all a, b ∈ A.
Remark 3.2. Recall from [29] that a braiding operator over a group (A, ◦) with
multiplication m : (a, b) 7→ a ◦ b is a bijective map r : A × A → A × A such that
(1) r(a ◦ b, c) = (id × m)r12 r23 (a, b, c) for all a, b, c ∈ A,
(2) r(a, b ◦ c) = (m × id)r23 r12 (a, b, c) for all a, b, c ∈ A,
(3) r(a, 1) = (1, a) and r(1, a) = (a, 1) for all a ∈ A, and
(4) mr(a, b) = a ◦ b for all a, b ∈ A.
Braiding operators are equivalent to bijective 1-cocycles by [29, Theorem 2], and
bijective 1-cocycles are equivalent to skew left braces by Proposition 1.11. One can
prove that (3.1) is the braiding operator corresponding to the skew left brace A
under this equivalence.
Proof of Theorem 3.1. Every braiding operator is a non-degenerate solution of the
Yang–Baxter equation by [29, Corollary 3]. Thus it is enough to prove that rA is a
braiding operator on (A, ◦). Set r = rA . Since λa−1 (b) = a ◦ (ab) for all a, b ∈ A,
λλ−1
((a ◦ b)−1 a(a ◦ b)) = λa (b) ◦ (λa (b)(a ◦ b)−1 a(a ◦ b)) = λa (b) ◦ (a ◦ b)
a (b)
SKEW BRACES AND THE YANG–BAXTER EQUATION
7
holds for all a, b ∈ A. Thus mr(a, b) = a ◦ b for all a, b ∈ A. Clearly r(a, 1) = (1, a)
and r(1, a) = (a, 1) for all a ∈ A. Let a, b, c ∈ A. By Corollary 1.10 one obtains
(id × m)r12 r23 (a, b, c) = (id × m)r12 (a, λb (c), λb (c) ◦ b ◦ c)
= (id × m)(λa λb (c), λa λb (c) ◦ a ◦ λb (c), λb (c) ◦ b ◦ c)
= (λa λb (c), λa λb (c) ◦ a ◦ b ◦ c) = r(a ◦ b, c).
From Remark 1.8 and Proposition 1.9 one obtains that
λa (b ◦ c) = λa (b)λa◦b (c)
holds for all a, b, c ∈ A. From this formula one deduces that
λa (b) ◦ λλa (b)◦a◦b (c) = λa (b) ◦ λ−1
λa (b) λa λb (c) = λa (b)λa◦b (c) = λa (b ◦ c).
holds for all a, b, c ∈ A. Then
(m × id)r23 r12 (a, b, c) = (m × id)r23 (λa (b), λa (b) ◦ a ◦ b, c)
= (m × id)(λa (b), λλa (b)◦a◦b (c), λλa (b)◦a◦b ◦ λa (b) ◦ a ◦ b ◦ c)
= (λa (b) ◦ λλa (b)◦a◦b (c), λλa (b)◦a◦b ◦ λa (b) ◦ a ◦ b ◦ c)
= (λa (b ◦ c), λa (b ◦ c) ◦ a ◦ b ◦ c) = r(a, b ◦ c).
for all a, b, c ∈ A.
Corollary 3.3. Let A be a skew left brace and X ⊆ A be a subset of A. Assume
bλa (x)b−1 ∈ X for all x ∈ X and a, b ∈ A. Then rA |X×X is a non-degenerate
solution of the Yang–Baxter equation.
Proof. Clearly λa (x) ∈ X and bxb−1 ∈ X for all a, b ∈ A and x ∈ X. Then it
−1
follows that λ−1
x(x ◦ y)) ∈ X for all x, y ∈ X. Now Theorem 3.1
λx (y) ((x ◦ y)
implies the claim.
Example 3.4. Let A and B be groups and α : A → Aut(B) be a group homomorphism. The skew left brace of Example 1.4 yields the following solution:
r : (A × B) × (A × B) → (A × B) × (A × B),
−1
r((a1 , b1 )(a2 , b2 )) = ((a2 , αa1 (b2 )), (a−1
2 a1 a2 , b2 αa−1 (b1 αa1 (b2 )))).
2
Example 3.5. Let A be an abelian group, B be a group and α : A → Aut(B) be
a group homomorphism. The skew left brace of Example 1.5 yields the following
solution:
r : (A × B) × (A × B) → (A × B) × (A × B),
r((a1 , b1 )(a2 , b2 )) = ((a2 , αa−1 (b2 )), (a1 , αa−1 (b−1
2 )b1 b2 )).
1
1
Example 3.6. Let A be the skew left brace constructed in Example 1.6. The solution of Theorem 3.1 is similar to the solution constructed by Weinstein and Xu in
terms of factorizable Poisson groups [41, Theorem 9.2]. The latter is τ rA τ , where
rA is the solution of Theorem 3.1 and τ (x, y) = (y, x) for all x, y.
Based on [29], for each skew left brace A we relate the solution r given by
Theorem 3.1 to the so-called Venkov solution, i.e.
s(a, b) = (b, b−1 ab),
a, b ∈ A.
8
L. GUARNIERI AND L. VENDRAMIN
Proposition 3.7. Let A be a skew left brace. For each n ∈ N the map Tn given by
Tn (a1 , . . . , an−1 , an ) = (a1 , λa1 (a2 ), λa1 ◦a2 (a3 ), . . . , λa1 ◦···◦an−1 (an ))
is invertible and satisfies
(3.2)
Tn ri,i+1 = si,i+1 Tn
for all n ≥ 2 and i ∈ {1, . . . , n− 1}, where ri,i+1 and si,i+1 denote the actions of the
braid group Bn on An = A × · · · × A (n-times) induced from r and s respectively.
Proof. A direct calculation shows that Tn is invertible with inverse
−1
−1
Tn−1 (a1 , . . . , an ) = (a1 , λ−1
a1 (a2 ), λa1 a2 (a3 ), . . . , λa1 ...an−1 (an )).
To prove (3.2) we proceed by induction on n. The case n = 2 follows from a
direct calculation since
−1
T2 r12 (a, b) = T2 (λa (b), λ−1
a(a ◦ b))) = (λa (b), (a ◦ b)−1 a(a ◦ b)),
λa (b) ((a ◦ b)
= (λa (b), λa (b)−1 aλa (b)) = s12 (a, λa (b)) = s12 T2 (a, b)
holds for all a, b ∈ A. So assume that the claim holds for n−1. Since Tn r1,2 = s1,2 Tn
is the same as T2 r = sT2 , we need to prove (3.2) for all i ∈ {2, . . . , n − 1}. Write
Tn = Un (id × Tn−1 ),
where
Un (a1 , . . . , an−1 , an ) = (a1 , λa1 (a2 ), . . . , λa1 (an−1 ), λa1 (an )).
Since each λa is an automorphism of (A, ·), it follows that Un si,i+1 = si,i+1 Un for
i ≥ 2 and hence (3.2) holds for all i ≥ 2.
Remark 3.8. Proposition 3.7 also follows from [28, Proposition 6.2]. The map Tn
is the so-called guitar map, see for example [28, §6].
The universal construction of Lu, Yan and Zhu, given in [29, Theorem 9] can be
restated in the language of skew left braces. This was done by Rump in the case of
involutive solutions, see [34]. Recall that the enveloping (or structure) group of a
solution (X, r) is the group G(X, r) generated by the elements of X with relations
x ◦ y = σx (y) ◦ τy (x),
x, y ∈ X.
Let ι : X → G(X, r) be the canonical map.
Theorem 3.9. Let X be a set, r : X × X → X × X, r(x, y) = (σx (y), τy (x)) be a
non-degenerate solution of the Yang–Baxter equation. Then there exists a unique
skew left brace structure over G(X, r) such that its associated solution rG satisfies
rG (ι × ι) = (ι × ι)r.
Furthermore, if B is a skew left brace and f : X → B is a map such that (f × f )r =
rB (f × f ), then there exists a unique group homomorphism φ : G(X, r) → B such
that f = φι and (φ × φ)rG = rB (φ × φ).
Proof. The claim follows from the universal construction of [29, Theorem 9] and
the equivalence between braiding operators and skew braces, see Remark 3.2.
The following corollary is essentially [38, Theorem 2.6].
Corollary 3.10. Let (X, r) be a finite non-degenerate solution of the Yang–Baxter
equation. Then G(X, r)/Soc(G(X, r)) is a finite skew left brace.
SKEW BRACES AND THE YANG–BAXTER EQUATION
Proof. It follows from Theorem 3.9 and Lemma 2.5.
9
4. Constructing skew braces
Let A be a group. The holomorph of A is the group Hol(A) = Aut(A)⋉ A, where
the product is given by
(f, a)(g, b) = (f g, af (b))
for all a, b ∈ A and f, g ∈ Aut(A). Any subgroup H of Hol(A) acts on A
(4.1)
(f, x) · a = π2 ((f, x)(id, a)) = xf (a),
a, x ∈ A, f ∈ Aut(A),
where π2 : Hol(A) → A, (f, a) 7→ a. In particular Hol(A) acts transitively on A and
the stabilizer of any a ∈ A is isomorphic to Aut(A).
Recall that a subgroup H of Hol(A) is regular if for each a ∈ A there exists a
unique (f, x) ∈ H such that xf (a) = 1. The following result is well-known.
Lemma 4.1. Let A be a group and H be a regular subgroup of Hol(A). Then
π2 |H : H → A, (f, a) 7→ a, is bijective.
Proof. We first prove that π2 |H is injective. Let (f, a), (g, b) ∈ H be such that
π2 (f, a) = π2 (g, b). Then a = b. Since H is a subgroup,
(f, a)−1 = (f −1 , f −1 (a−1 )) ∈ H,
(g, a)−1 = (g −1 , g −1 (a−1 )) ∈ H,
and hence f = g since f −1 (a)f −1 (a−1 ) = g −1 (a)g −1 (a−1 ) = 1 and H is a regular
subgroup.
Now we prove that π2 |H is surjective. Let a ∈ A. The regularlity of H implies
the existence of an automorphism f ∈ Aut(A) such that (f, f (a−1 )) ∈ H. Then
(f −1 , a) ∈ H and the claim follows.
The following theorem goes back to Bachiller [2]. The proof in our case is the
same as for braces. It is a generalization of a result of Catino and Rizzo [9].
Theorem 4.2. Let A be skew left brace. Then {(λa , a) : a ∈ A} is a regular
subgroup of Hol(A, ·). Conversely, if (A, ·) is a group and H is a regular subgroup
of Hol(A, ·), then A is a skew left brace with (A, ◦) ≃ H, where
a ◦ b = af (b)
−1
and (π2 |H )
(a) = (f, a) ∈ H.
Proof. Since λ is a group homomorphism and aλa (b) = a ◦ b for all a, b ∈ A, it
follows that {(λa , a) : a ∈ A} is a subgroup of Hol(A, ·). Since (A, ◦) is a group,
the regularlity also follows.
Assume now that H is a regular subgroup. By Lemma 4.1, π2 |H is bijective.
Use the bijection π2 |H to transport the product of H into A:
a ◦ b = π2 |H (π2 |H )−1 (a)(π2 |H )−1 (b) = af (b),
where a, b ∈ A and (π2 |H )−1 (a) = (f, a) ∈ H. Then (A, ◦) is a group and A is a
skew left brace since
a ◦ (bc) = af (bc) = af (b)f (c) = af (b)a−1 af (c) = (a ◦ b)a−1 (a ◦ c)
holds for all a, b, c ∈ A.
10
L. GUARNIERI AND L. VENDRAMIN
Proposition 4.3. Let A be a group. There exists a bijective correspondence between
skew left brace structures over A and regular subgroups of Hol(A). Moreover, isomorphic skew braces structures over A correspond to conjugate subgroups of Hol(A)
by elements of Aut(A).
Proof. Assume that the group A has two skew left brace structures given by (a, b) 7→
a ◦ b and (a, b) 7→ a × b and that φ ∈ Aut(A, ·) satisfies φ(a ◦ b) = φ(a) × φ(b) for
all a, b ∈ A. We claim that {(λa , a) : a ∈ A} and {(µa , a) : a ∈ A}, where
λa (b) = a−1 (a ◦ b) and µa (b) = a−1 (a × b), are conjugate by φ. Since
φλa φ−1 (b) = φ(a−1 (a ◦ φ−1 (b)) = φ(a)−1 (φ(a) × b) = µφ(a) (b),
one obtains that φ(λa , a)φ−1 = (µφ(a) , φ(a)) and hence the claim follows.
Conversely, let H and K be regular subgroups of Hol(A) and asssume that
there exists φ ∈ Aut(A, ·) such that φ−1 Hφ = K. Let (f, a) = (π2 |H )−1 (a) ∈ H,
(g, a) = (π2 |K )−1 (a) ∈ K and write a ◦ b = af (b) and a × b = ag(b). Since
φ(f, a)φ−1 = (φf φ−1 , φ(a)) ∈ K, it follows that (π2 |K )−1 (φ(a)) = (φf φ−1 , φ(a)).
Then, since φ ∈ Aut(A, ·),
φ(a) × φ(b) = φ(a)(φf φ−1 )(φ(b)) = φ(a)φ(f (b)) = φ(af (b)) = φ(a ◦ b)
and hence the skew left braces corresponding to H and K are isomorphic.
5. Computational results
We first present the algorithm used to enumerate skew left brace structures over
a given group A. The algorithm uses Theorem 4.2.
Algorithm 5.1. Let A be a finite group. To construct all skew left brace structures
over A we proceed as follows:
(1) Compute the holomorph Hol(A) of A.
(2) Compute the list of regular subgroups of Hol(A) of order |A| up to conjugation by elements of Aut(A).
(3) For each representative H of regular subgroups of Hol(A) construct the map
p : A → H given by a 7→ (f, f (a)−1 ), where (f, f (a)−1 ) ∈ H. The triple
(H, A, p : A → H) yields a skew left brace structure over A with multiplication given by a ◦ b = p−1 (p(a)p(b)) for all a, b ∈ A.
Remark 5.2. To enumerate all skew left brace structures over A the third step of
Algorithm 5.1 is not needed.
Remark 5.3. Recall that a left brace is an abelian group (A, +) with another group
structure, defined via the multiplication (a, b) 7→ ab such that a(b + c) + a = ab + ac
holds for all a, b, c ∈ A. Left braces with additive group isomorphic to a given
group A can be constructed applying Algorithm 5.1 to the abelian group A. In the
second step of Algorithm 5.1 it is enough to compute the list of regular solvable
subgroups of Hol(A), since the multiplicative group of a left brace is solvable by [15,
Proposition 2.5].
Algorithm 5.1 was implemented both in GAP and Magma with different performances and were run on a Intel(R) Core(TM) i5-4440 CPU @3.10GHz with 16gb
of RAM, under Linux.
SKEW BRACES AND THE YANG–BAXTER EQUATION
11
5.1. Skew left braces. For n ∈ N let c(n) be the number of non-isomorphic skew
left braces of size n.
The number of skew left braces of size n ≤ 30 has been determined using Algorithm 5.1. Table 5.1 shows some values of c(n). The calculation took about twenty
minutes.
Table 5.1. The number of non-isomorphic skew left braces.
n
1
2 3 4 5 6 7 8
9 10 11 12 13 14 15
c(n)
1
1 1 4 1 6 1 47 4
6 1 38 1 6 1
n
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
c(n) 1605 1 49 1 43 8 6 1 855 4 6 101 29 1 36
5.2. Left braces. For n ∈ N let b(n) be the number of non-isomorphic left braces
of size n.
The number of left braces (up to isomorphism) of size n ≤ 120 has been determined using Algorithm 5.1. Table 5.3 shows some values of b(n) and Table 5.2 gives
runtimes for our Magma implementation for some examples. The construction of
left braces requires considerably more CPU time, see Table 5.4 for some examples.
Table 5.2. Some runtimes for enumerating left braces of size n.
n CPU time b(n)
16
1 hour
357
48
18 hours 1708
54 5 minutes
80
72
1 hour
489
80
17 hours 1985
100
15 secs
51
108
28 hours 494
112
12 hours 1671
With current computational resources, we were not able to compute the number
of non-isomorphic left braces of orders 32, 64, 81 and 96.
5.3. Two-sided left braces (radical rings). Recall that a brace B is a two-sided
brace if (a + b)c + c = ac + bc holds for all a, b, c ∈ B. Two-sided braces are in
bijective correspondence with radical rings [33]. Recall that a non-zero radical ring
is a ring R without identity such that for each x ∈ R there is y ∈ R such that
x + y + xy = 0. Assume that R is a radical ring. Then the circle operation,
a ◦ b = ab + a + b,
a, b ∈ R,
makes (R, +, ◦) into a two-sided brace. Conversely, if A is a two-sided brace, the
operation a ∗ b = ab − a − b, a, b ∈ A makes (A, +, ∗), into a radical ring.
To test whether a left brace is a two-sided brace one has the following lemma of
Gateva-Ivanova, see [17, Corollary 3.5].
Lemma 5.4 (Gateva-Ivanova). Let A be a left brace. Then A is a two-sided brace
if and only if
−1
bcλabc
(c) = cλ−1
ac (λa (b)c)
12
L. GUARNIERI AND L. VENDRAMIN
Table 5.3. The number of non-isomorphic left braces.
n
1
2
3
4
5
6
7
8
9
10 11
12
b(n) 1
1
1
4
1
2
1
27
4
2
1
10
n
13 14 15
16
17 18 19
20
21 22 23
24
b(n) 1
2
1
357
1
8
1
11
2
2
1
96
n
25 26 27
28
29 30 31
32
33 34 35
36
b(n) 4
2
37
9
1
4
1
?
1
2
1
46
n
37 38 39
40
41 42 43
44
45 46 47
48
b(n) 1
2
2
106
1
6
1
9
4
2
1 1708
n
49 50 51
52
53 54 55
56
57 58 59
60
b(n) 4
8
1
11
1
80
2
91
2
2
1
28
n
61 62 63
64
65 66 67
68
69 70 71
72
b(n) 1
2
11
?
1
4
1
11
1
4
1
489
n
73 74 75
76
77 78 79
80
81 82 83
84
b(n) 1
2
5
9
1
6
1 1985 ?
2
1
34
n
85 86 87
88
89 90 91
92
93 94 95
96
b(n) 1
2
1
90
1
16
1
9
2
2
1
?
n
97 98 99 100 101 102 103 104 105 106 107 108
b(n) 1
8
4
51
1
4
1
106
2
2
1
494
n
109 110 111 112 113 114 115 116 117 118 119 120
b(n) 1
6
2 1671 1
6
1
11
11
2
1
395
Table 5.4. Some runtimes for constructing left braces of size n.
n CPU time b(n)
16
3 hours 357
54 40 minutes
80
72
24 hours 489
112
5 days 1671
for all a, b, c ∈ A.
For n ∈ N let t(n) be the number of non-isomorphic two-sided braces of size n.
Using the database of left braces constructed with Algorithm 5.1 and Lemma 5.4
one computes t(n). Table 5.5 shows the value of t(n) for n ≤ 24.
Table 5.5. The number of non-isomorphic two-sided braces.
n
1 2 3
4
5 6 7 8 9 10 11 12
t(n) 1 1 1
4
1 1 1 22 4 1 1 4
n
13 14 15 16 17 18 19 20 21 22 23 24
t(n) 1 1 1 221 1 4 1 4 1 1 1 22
Remark 5.5. For information on square-free two-sided braces, see [12]. These braces
are defined by nilpotent groups of class ≤ 2.
SKEW BRACES AND THE YANG–BAXTER EQUATION
13
6. Further questions
In this section we collect some questions and conjectures that appear naturally
after inspecting Table 5.3.
6.1. Left braces. We first collect some problems and conjectures related to the
number of left braces.
Problem 6.1. Compute b(32), b(64), b(81) and b(96).
Table 5.3 suggests the following conjectures.
Conjecture 6.2. Let p > 3 be a prime number. Then
(
11 if m ≡ 1 mod 4,
b(4p) =
9
if m ≡ 3 mod 4.
Conjecture 6.3. Let p > 3 be a prime. Then
14 if p ≡ 1 mod 9,
b(9p) = 4
if p ≡ 2, 5 mod 9,
11 if p ≡ 4, 7 mod 9.
Conjecture 6.4. Let p, q be prime numbers such that p < q and q 6≡ 1 mod p.
Then b(p2 q) = 4.
We have used [7] and computer calculations to show that Conjectures 6.2 and 6.3
are true up to p = 997. In [37], Agata Smoktunowicz proved that Conjecture 6.4 is
true.
6.2. Quaternionic braces. We now consider an important family of braces. Recall that for m ∈ N the generalized quaternion group is the group
Q4m = ha, b : am = b2 , a2m = 1, b−1 ab = a−1 i.
Definition 6.5. A brace is a quaternion brace if its multiplicative group is isomorphic to some quaternion group.
Conjecture 6.6. For m ∈ N let q(4m) be the number of isomorphism classes of
quaternion braces of size 4m. Then for m > 2
2 if m is odd,
7 if m ≡ 0 mod 8,
q(4m) =
9 if m ≡ 4 mod 8,
6 if m ≡ 2 mod 8 or m ≡ 6 mod 8.
We have checked Conjecture 6.6 for all m ≤ 512. It seems natural to ask the
following questions.
Question 6.7. Which finite abelian groups appear as the additive group of a quaternion brace?
For m ∈ {2, . . . , 512} the additive group of a quaternion brace of size m is
isomorphic to one of the following groups:
Z4m , Z2m × Z2 , Zm × Z2 × Z2 , Zm × Z4 , Zm/2 × Z2 × Z2 × Z2 .
By inspection, one sees that the groups Zm ×Z22 appear whenever m ≡ 2, 4, 6 mod 8
and the groups Zm × Z4 and Zm/2 × Z32 appear whenever m ≡ 4 mod 8.
14
L. GUARNIERI AND L. VENDRAMIN
Question 6.8. For m > 2 let A be a finite abelian group size 4m. Compute the
number of isomorphism classes of quaternion braces of size 4m with additive group
isomorphic to A.
In [3, §5] quaternion braces of size 2k are mentioned as an important class of
braces which could be useful to classify a certain family of involutive non-degenerate
solutions of the Yang–Baxter equation. Conjecture 6.6 implies the following:
Conjecture 6.9. There are seven classes of isomorphism of quaternion braces of
size 2k for k > 4.
Conjecture 6.9 was verified for all k ∈ {5, 6, 7, 8, 9}. Table 6.1 sums up our
findings related to this important subclass of braces.
Table 6.1. Number of braces with multiplicative group isomorphic to the quaternion group Q2k with k > 4.
Additive Group Number of Braces
Z2k
1
Z2k−1 × Z2
6
Remark 6.10. The classification of left braces over cyclic groups was done by Rump
in [35]. He proved that if a left brace A has additive group isomorphic to Z/pk ,
where p > 2 is a prime number, then (A, ·) ≃ Z/pk . According to [3], the converse
holds for all p. In [1], Bachiller classified left braces of size p2 and p3 , where p is a
prime number. The techniques used in these papers might prove useful to address
the questions, problems and conjectures in this section.
Acknowledgements
We thank Leandro del Pezzo for the computer where some calculations were
performed, David Bachiller for his suggestions, and Victoria Lebed for several useful
comments and corrections. We also thank the referees for their careful review of
our manuscript and valuable suggestions.
References
[1] D. Bachiller. Classification of braces of order p3 . J. Pure Appl. Algebra, 219(8):3568–3603,
2015.
[2] D. Bachiller. Counterexample to a conjecture about braces. J. Algebra, 453:160–176, 2016.
[3] D. Bachiller, F. Cedo, and E. Jespers. Solutions of the Yang-Baxter equation associated with
a left brace. arXiv:1503.02814.
[4] R. J. Baxter. Partition function of the eight-vertex lattice model. Ann. Physics, 70:193–228,
1972.
[5] R. J. Baxter. Exactly solved models in statistical mechanics. Academic Press, Inc. [Harcourt
Brace Jovanovich, Publishers], London, 1989. Reprint of the 1982 original.
[6] N. Ben David and Y. Ginosar. On groups of central type, non-degenerate and bijective cohomology classes. Israel J. Math., 172:317–335, 2009.
[7] H. U. Besche, B. Eick, and E. A. O’Brien. A millennium project: constructing small groups.
Internat. J. Algebra Comput., 12(5):623–644, 2002.
[8] W. Bosma, J. Cannon, and C. Playoust. The Magma algebra system. I. The user language. J.
Symbolic Comput., 24(3-4):235–265, 1997. Computational algebra and number theory (London, 1993).
SKEW BRACES AND THE YANG–BAXTER EQUATION
15
[9] F. Catino and R. Rizzo. Regular subgroups of the affine group and radical circle algebras.
Bull. Aust. Math. Soc., 79(1):103–107, 2009.
[10] F. Cedó, E. Jespers, and Á. del Rı́o. Involutive Yang-Baxter groups. Trans. Amer. Math.
Soc., 362(5):2541–2558, 2010.
[11] F. Cedó, E. Jespers, and J. Okniński. Retractability of set theoretic solutions of the YangBaxter equation. Adv. Math., 224(6):2472–2484, 2010.
[12] F. Cedó, E. Jespers, and J. Okniński. Braces and the Yang-Baxter equation. Comm. Math.
Phys., 327(1):101–116, 2014.
[13] P. Dehornoy. Set-theoretic solutions of the Yang–Baxter equation, RC-calculus, and Garside
germs. Adv. Math., 282:93–127, 2015.
[14] V. G. Drinfel′ d. On some unsolved problems in quantum group theory. In Quantum groups
(Leningrad, 1990), volume 1510 of Lecture Notes in Math., pages 1–8. Springer, Berlin, 1992.
[15] P. Etingof, T. Schedler, and A. Soloviev. Set-theoretical solutions to the quantum YangBaxter equation. Duke Math. J., 100(2):169–209, 1999.
[16] The GAP Group. GAP – Groups, Algorithms, and Programming, Version 4.7.8, 2015.
[17] T. Gateva-Ivanova. Set-theoretic solutions of the Yang-Baxter equation, Braces, and Symmetric groups. arXiv:1507.02602.
[18] T. Gateva-Ivanova. A combinatorial approach to the set-theoretic solutions of the YangBaxter equation. J. Math. Phys., 45(10):3828–3858, 2004.
[19] T. Gateva-Ivanova. Quadratic algebras, Yang-Baxter equation, and Artin-Schelter regularity.
Adv. Math., 230(4-6):2152–2175, 2012.
[20] T. Gateva-Ivanova and P. Cameron. Multipermutation solutions of the Yang-Baxter equation.
Comm. Math. Phys., 309(3):583–621, 2012.
[21] T. Gateva-Ivanova and S. Majid. Set-theoretic solutions of the Yang-Baxter equation, graphs
and computations. J. Symbolic Comput., 42(11-12):1079–1112, 2007.
[22] T. Gateva-Ivanova and S. Majid. Matched pairs approach to set theoretic solutions of the
Yang-Baxter equation. J. Algebra, 319(4):1462–1529, 2008.
[23] T. Gateva-Ivanova and S. Majid. Quantum spaces associated to multipermutation solutions
of level two. Algebr. Represent. Theory, 14(2):341–376, 2011.
[24] T. Gateva-Ivanova and M. Van den Bergh. Semigroups of I-type. J. Algebra, 206(1):97–112,
1998.
[25] E. Jespers and J. Okniński. Monoids and groups of I-type. Algebr. Represent. Theory,
8(5):709–729, 2005.
[26] E. Jespers and J. Okniński. Noetherian semigroup algebras, volume 7 of Algebras and Applications. Springer, Dordrecht, 2007.
[27] C. Kassel. Quantum groups, volume 155 of Graduate Texts in Mathematics. Springer-Verlag,
New York, 1995.
[28] V. Lebed and L. Vendramin. Homology of left non-degenerate set-theoretic solutions to the
yang-baxter equation. arXiv:1509.07067.
[29] J.-H. Lu, M. Yan, and Y.-C. Zhu. On the set-theoretical Yang-Baxter equation. Duke Math.
J., 104(1):1–18, 2000.
[30] S. Majid. Foundations of quantum group theory. Cambridge University Press, Cambridge,
1995.
[31] S. Nelson. The combinatorial revolution in knot theory. Notices Amer. Math. Soc.,
58(11):1553–1561, 2011.
[32] W. Rump. A decomposition theorem for square-free unitary solutions of the quantum YangBaxter equation. Adv. Math., 193(1):40–55, 2005.
[33] W. Rump. Modules over braces. Algebra Discrete Math., (2):127–137, 2006.
[34] W. Rump. Braces, radical rings, and the quantum Yang-Baxter equation. J. Algebra,
307(1):153–170, 2007.
[35] W. Rump. Classification of cyclic braces. J. Pure Appl. Algebra, 209(3):671–685, 2007.
[36] W. Rump. The brace of a classical group. Note Mat., 34(1):115–144, 2014.
[37] A. Smoktunowicz. A note on set-theoretic solutions of the Yang-Baxter equation.
arXiv:1512.95542.
[38] A. Soloviev. Non-unitary set-theoretical solutions to the quantum Yang-Baxter equation.
Math. Res. Lett., 7(5-6):577–596, 2000.
16
L. GUARNIERI AND L. VENDRAMIN
[39] M. Takeuchi. Survey on matched pairs of groups—an elementary approach to the ESS-LYZ
theory. In Noncommutative geometry and quantum groups (Warsaw, 2001), volume 61 of
Banach Center Publ., pages 305–331. Polish Acad. Sci., Warsaw, 2003.
[40] L. Vendramin. Extensions of set-theoretic solutions of the Yang–Baxter equation and a conjecture of Gateva-Ivanova. J. Pure Appl. Algebra, 220(5):2064–2076, 2016.
[41] A. Weinstein and P. Xu. Classical solutions of the quantum Yang-Baxter equation. Comm.
Math. Phys., 148(2):309–343, 1992.
[42] C. N. Yang. Some exact results for the many-body problem in one dimension with repulsive
delta-function interaction. Phys. Rev. Lett., 19:1312–1315, 1967.
Departamento de Matemática – FCEN, Universidad de Buenos Aires, Pab. I – Ciudad
Universitaria (1428) Buenos Aires – Argentina
E-mail address: [email protected]
E-mail address: [email protected]
| 4math.GR
|
Hypothesis Testing for High-Dimensional Multinomials:
A Selective Review
Sivaraman Balakrishnan†
Larry Wasserman†
arXiv:1712.06120v1 [stat.ML] 17 Dec 2017
Department of Statistics†
Carnegie Mellon University,
Pittsburgh, PA 15213.
{siva,larry}@stat.cmu.edu
In memory of Stephen E. Fienberg.
December 19, 2017
Abstract
The statistical analysis of discrete data has been the subject of extensive statistical
research dating back to the work of Pearson. In this survey we review some recently
developed methods for testing hypotheses about high-dimensional multinomials. Traditional tests like the χ2 -test and the likelihood ratio test can have poor power in the
high-dimensional setting. Much of the research in this area has focused on finding tests
with asymptotically Normal limits and developing (stringent) conditions under which tests
have Normal limits. We argue that this perspective suffers from a significant deficiency: it
can exclude many high-dimensional cases when — despite having non-Normal null distributions — carefully designed tests can have high power. Finally, we illustrate that taking
a minimax perspective and considering refinements of this perspective can lead naturally
to powerful and practical tests.
1
Introduction
Steve Fienberg was a pioneer in the development of theory and methods for discrete data. His
textbook (Bishop et al. (1995)) remains one of the main references for the topic. Our focus
in this review is on high-dimensional multinomial models where the number of categories d
can grow with, and possibly exceed the sample-size n. Steve’s paper (Fienberg and Holland
(1973)), written with with Paul Holland, was one of the first to consider multinomial data in the
high-dimensional case. In Fienberg (1976), Steve provided strong motivation for considering
the high-dimensional setting:
“The fact remains . . . that with the extensive questionnaires of modern-day sample-surveys,
and the detailed and painstaking inventory of variables measured by biological and socialscientists, the statistician is often faced with large sparse arrays full of 0’s and 1’s, in need of
careful analysis”.
In this review we focus on hypothesis testing for high-dimensional multinomials. In the
context of hypothesis testing, several works (see for instance Read and Cressie (1988); Holst
(1972) and references therein) have considered the high-dimensional setting. Hoeffding (1965)
(building on an unpublished result of Stein) showed that for testing goodness-of-fit, in sharp
contrast to the fixed-d setting, in the high-dimensional setting the likelihood ratio test can
be dominated by the χ2 test. In traditional asymptotic testing theory, the power of tests are
1
often investigated at local alternatives which approach the null as the sample-size grows. In
the high-dimensional setting, considering local alternatives, Ivchenko and Medvedev (1978)
showed that neither the χ2 or the likelihood ratio test are uniformly optimal. These results
show some of the difficulties of using classical theory to identify optimal tests in the highdimensional regime.
Morris (1975) studied the limiting distribution of a wide-range of multinomial test statistics and gave relatively stringent conditions under which these statistics have asymptotically
Normal limiting distributions. In general, as we illustrate in our simulations, carefully designed
tests can have high power under much weaker conditions, even when the null distribution of
the test statistics are not Gaussian or χ2 . In many cases, understanding the limiting distribution of the test statistic under the null is important to properly set the test threshold, and
indeed this often leads to practical tests. However, in several problems of interest, including
goodness-of-fit, two-sample testing and independence testing we can determine practical (nonconservative) thresholds by simulation. In the high-dimensional setting, rather than rely on
asymptotic theory and local alternatives, we advocate for using the minimax perspective and
developing refinements of this perspective to identify and study optimal tests.
Such a minimax perspective has been recently developed in a series of works in different
fields including statistics, information theory and theoretical computer science and we provide
an overview of some important results from these different communities in this paper.
2
Background
Suppose that we have data of the form Z1 , . . . , Zn ∼ P where P is a d-dimensional multinomial
and Zi ∈ {1, . . . , d}. We denote the probability mass function for P by p ∈ Rd . The set of all
multinomials is then denoted by
X
M = p = (p(1), . . . , p(d)) : p(j) ≥ 0, for all j,
p(j) = 1 .
(1)
j
We are interested in the case where d can be large, possibly much larger than n. In this paper,
we focus on two hypothesis testing problems, and defer a discussion of other testing problems
to Section 6. The problems we consider are:
1. Goodness-of-fit testing: In its most basic form, in goodness-of-fit testing we are
interested in testing the fit of the data to a fixed distribution P0 . Concretely, we are
interested in distinguishing the hypotheses:
H0 : P = P0
versus H1 : P 6= P0 .
2. Two-sample testing: In two-sample testing we observe
Z1 , . . . , Zn1 ∼ P,
W1 , . . . , Wn2 ∼ Q.
In this case, the hypotheses of interest are
H0 : P = Q versus H1 : P 6= Q.
Notation: Throughout this paper, we write an bn if both an /bn is bounded away from 0
and ∞ for all large n.
2
2.1
Why Hypothesis Testing?
As with any statistical problem, there are many inferential tasks related to multinomial models:
estimation, constructing confidence sets, Bayesian inference, prediction and hypothesis testing,
among others.
Our focus on testing in this paper is not meant to downplay the importance of these other
tasks. Indeed, many would argue that hypothesis testing has received too much attention:
over-reliance on hypothesis testing is sometimes cited as one of the causes of the reproducibility
crisis. However, there is a good reason for studying hypothesis testing. When trying to
understand the theoretical behavior of statistical problems in difficult cases — such as in high
dimensional models — hypothesis testing provides a very clean, precise framework.
Hypothesis testing is a good starting point for theoretical investigations into difficult statistical models. As an example, in Section 3.2 we will see that the power of goodness-of-fit
tests can vary drastically depending on where the null sits in the simplex, a phenomenon that
we refer to as local minimaxity. This local minimax phenomenon is very clear and easy to
precisely capture in the hypothesis testing framework.
2.2
Minimax and Local Minimax Testing
In traditional asymptotic testing theory, local measures of performance are often used to assess
the performance of various hypothesis tests. In the local approach, one typically examines
√
the power at a sequence of alternatives at θ0 + C/ n where θ0 denotes the null value of
the parameter. The local approach is well-suited to well-behaved, low dimensional models. It
permits very precise power comparisons for distributions which are close to the null hypothesis.
Generally, the tools for local analysis are tied to ideas like contiguity and asymptotic Normality
and in the high dimensional setting these tools often break down. Furthermore, as we discussed
earlier, results of Ivchenko and Medvedev (1978) suggest that the local perspective does not
provide a clear overall picture in the high-dimensional case.
For these reasons, we will use the minimax perspective. For goodness-of-fit testing, we
can refine the minimax perspective to obtain results that also have a local nature, but in a
very different sense than the local results described above. Formally, a test is a map from the
samples to {0, 1}. We let Φn denote all level α tests, i.e. φ ∈ Φn if φ(Z1 , . . . , Zn ) ∈ {0, 1} and
sup P n (φ = 1) ≤ α,
P ∈P0
where P0 denotes the (possibly composite) collection of possible null distributions. Let
n
o
M = p : d(P0 , p) ≥
where P0 is the set of null distributions, d(P0 , p) = inf q∈P0 d(q, p) and d(p, q) is some distance.
The maximum type II error of a test φ ∈ Φn is
Rn, (φ, P0 ) = sup P n (φ = 0).
P ∈M
The minimax risk is
†
Rn,
(P0 ) = inf Rn, (φ, P0 ).
φ∈Φn
†
A test φ ∈ Φn is minimax optimal if Rn, (φ) = Rn,
(P0 ). It is common to study the minimax
risk via a coarse lens by studying instead the minimax separation, also called the critical radius.
3
The minimax separation n is defined by
n
o
†
n (P0 ) = inf : Rn,
(P0 ) ≤ 1/2
which is the smallest such that the power is non-trivial. The choice of 1/2 is not important
and any sufficiently small, non-zero number will suffice.
We need to choose a distance d. We will focus on the total variation (TV) distance defined
by
TV(P, Q) = max |P (A) − Q(A)|
A
where the maximum is over all events A. The reason we use total variation distance is because
it has a clear probabilistic meaning and is invariant to natural transformations (Devroye and
Gyorfi, 1985): if TV(P, Q) = then |P (A) − Q(A)| ≤ for every event A. The total variation
distance is equivalent to the L1 distance:
TV(P, Q) =
1X
1
|p(j) − q(j)| ≡ kp − qk1
2
2
j
where p and q are the probability functions corresponding to the distributions P and Q. Other
distances, such as L2 , Hellinger and Kullback-Leibler can be used too, but in some cases can
be less interpretable and in other cases lead to trivial minimax rates. We revisit the choice of
metric in Section 3.4.
Typically we characterize the minimax separation by providing upper and lower bounds
on it: upper bounds are obtained by analyzing the minimax separation for practical tests,
while lower bounds are often obtained via an analysis of the likelihood ratio test for carefully
constructed pairs of hypotheses (see for instance the pioneering work of Ingster and co-authors
Ingster and Suslina (2003); Ingster (1997)).
The Local Minimax Separation:
Focusing on the problem of goodness-of-fit testing
with a simple null, we observe that the minimax separation n (p0 ) is a function of the null
distribution. Classical work in hypothesis testing Ingster and Suslina (2003); Ingster (1997) has
focused on characterizing the global minimax separation, i.e. in understanding the quantity:
n
o
†
n = inf : sup Rn,
(p0 ) ≤ 1/2 ,
(2)
p0 ∈M
where M is defined in (1). In typical non-parametric problems, the local minimax risk and
the global minimax risk match up to constants and this has led researchers in past work to
focus on the global minimax risk.
Recent work by Valiant and Valiant (2017) showed that for goodness-of-fit testing in the
TV metric for high-dimensional multinomials, the critical radius can vary considerably as a
function of the null distribution p0 . In this case, the local minimax separation, i.e. n (p0 )
provides a much more refined notion of the difficulty of the goodness-of-fit problem. Valiant
and Valiant (2017) further provided a locally minimax test, i.e. a test that is nearly-optimal
for every possible null distribution.
Developing refinements to the minimax framework in problems beyond goodness-of-fit
testing is an active area of research. For the problem of two-sample testing for multinomials,
a recent proposal appears in the work of Acharya et al. (2012). Other works developing a local
perspective in testing and estimation include Donoho and Johnstone (1994); Goldenshluger
and Lepski (2011); Cai and Low (2015); Chatterjee et al. (2015); Wei and Wainwright (2017).
4
3
Goodness-of-Fit
Let Z1P
, . . . , Zn ∼ P where Zi ∈ {1, . . . , d}. Define the vector of counts (X1 , . . . , Xd ) where
Xj = ni=1 I(Zi = j). We consider testing the simple null hypothesis:
versus H1 : P 6= P0 .
H0 : P = P0
The most commonly used test statistics are the chi-squared statistic,
Tχ2
d
X
(Xj − np0 (j))2
=
,
np0 (j)
(3)
j=1
and the likelihood ratio test (LRT)
TLRT =
d
X
pb(j) log
j=1
pb(j)
p0 (j)
,
(4)
where pb(j) = Xj /n. See the book of Read and Cressie (1988) for a variety of other popular
multinomial goodness-of-fit tests. As we show in Section 5, when d is large, these tests can
have poor power. In particular, they are not minimax optimal. Much of the research on tests
in the large d setting has focused on establishing conditions under which these statistics have
convenient limiting distributions (see for instance Morris (1975)). Unfortunately, these conditions are generally not testable, and they rule out many interesting cases, when test statistics
despite not having a convenient limiting distribution can have high power (see Section 5).
3.1
Globally Minimax Optimal Tests
The global minimax separation rate was characterized in the works of Paninski (2008); Valiant
and Valiant (2017). In particular, these works show that the global minimax separation rate
(see (2)) is given by:
d1/4
n √ .
n
(5)
This implies, surprisingly,
√ that we can have non-negligible power even when n d. In
the regime when n d most categories of the multinomial are unobserved, but we can
still distinguish any multinomial from alternatives separated in `1 with high power.
p In start
contrast, it can be shown that the minimax estimation rate in the `1 distance, is d/n which
is much slower than the testing rate. This is a common phenomenon: hypothesis testing is
often easier than estimation.
An important fact, elucidated by the minimax perspective, is that none of the traditional
tests are minimax. A very simple minimax test, from Balakrishnan and Wasserman (2017) is
the truncated χ2 test defined by
Ttrunc =
X (Xj − np0 (j))2 − Xj
.
1
max{p
(j),
}
0
d
j
The α level critical value tα is defined by
n
o
tα (p0 ) = inf t : P0n (T > t) ≤ α .
5
(6)
Normal approximations or χ2 approximations cannot be used to find tα since the asymptotics
are not uniformly valid over M. However, the critical value tα for the test can easily be found
by simulating from P0 . In Section 5 we report some simulation studies that illustrate the gain
in power by using this test.
3.2
Locally Minimax Optimal Tests
In goodness-of-fit testing, some nulls are easier to test than others. For example, when p0 is
uniform the local minimax risk is quite large and scales as in (5). However, when p0 is sparse
the problem effectively behaves as a much lower dimensional multinomial testing problem
and the minimax separation can be much smaller. This observation has important practical
consequences. Substantial gains in power can be achieved, by adapting the test to the shape
of the null distribution p0 .
Roughly, Valiant and Valiant (2017) showed that the local minimax rate is given by
r
V (p0 )
n (p0 )
n
for a functional V that depends on p0 as,
V (p0 ) ≈ kp0 k2/3
3/2
d
X
2/3
=
p0 (j) .
j=1
We provide a more precise statement in the Appendix. The fact that the local minimax rate
depends on the 2/3 norm is certainly not intuitive and is an example of the surprising nature
of the results in the world of high dimensional multinomials.
When p0 is uniform the 2/3-rd
√
norm is maximal and takes the value kp0 k2/3 = d whereas when p0 = (1, 0, . . . , 0) the 2/3rd norm is much smaller, i.e. kp0 k2/3 = 1. This means that a test tailored to p0 can have
dramatic gains in power.
Valiant and Valiant (2017) constructed a test that achieves the local minimax bound. To
describe the test we need a few definitions. First, without loss of generality, assume that
p0 (1) ≥ p0 (2) ≥ · · · ≥ p0 (d). Let σ ∈ [0, 1] and define the tail and bulk by
d
o
n
X
Qσ (p0 ) = i :
p0 (j) ≤ σ
j=i
and
n
o
Bσ (p0 ) = i > 1 : i ∈
/ Qσ (p0 ) .
The test is φ = φ1 ∨ φ2 where φ1 = I(T1 (σ) > t1 ), φ2 = I(T2 (σ) > t2 ),
r
X
nP0 (Qσ )
T1 (σ) =
(Xj − np0 (j)),
t1 (α, σ) =
,
α
j∈Qσ
sP
2/3
2
X (Xj − np0 (j))2 − Xj
j∈Bσ 2n p0 (j)
T2 (σ) =
,
t
(α,
σ)
=
.
2
2/3
α
p (j)
j∈Bσ
0
The test may appear to be somewhat complicated but all the quantities are easy to compute.
Furthermore, in practice the thresholds are easily computed by simulation. Other near-local
6
minimax tests for testing multinomials appear in Diakonikolas and Kane (2016); Balakrishnan
and Wasserman (2017).
A problem with the above test is that there is a tuning parameter σ. Valiant and Valiant
(2017) suggested using σ = /8. While this choice is useful for theoretical analysis it is
not useful in practice as it would require knowing how far P is from P0 if H0 is false. In
Balakrishnan and Wasserman (2017) we propose ways to select the tuning parameter σ in
a data-driven fashion. For instance, one might consider a Bonferroni corrected test. More
precisely, let Σ = {σ1 , . . . , σN } be a grid of values for σ. Let φj be the test using tuning
parameter σj and significance level α/N . We then use the Bonferroni corrected test φ =
maxj {φj }. It can be shown that, if Σ is chosen carefully, there is only a small loss of power
from the Bonferroni correction.
3.3
Implications for Continuous Data
Although the focus of this paper is on discrete data, we would like to briefly mention the
fact that these results have implications for continuous data. This discussion is based on
Balakrishnan and Wasserman (2017).
Suppose that X1 , . . . , Xn ∼ p where p is a density on [0, 1]. We want to test H0 : p = p0 . As
shown in LeCam (1973) and Barron (1989), the power of any test over the set {p : TV(p, p0 ) >
} is trivial unless we add further assumptions. For example, suppose we restrict attention to
densities p that satisfy the Lipschitz constraint:
|p(y) − p(x)| ≤ L|x − y|.
In this case, Ingster (1997) showed that, when p0 is the uniform density, the minimax separation rate is n n−2/5 . The optimal rate can be achieved by binning the data and using a
χ2 test. However, if p0 is not uniform and the Lipschitz constant L is allowed to grow with n,
Balakrishnan and Wasserman (2017) showed that the local minimax rate is
√
2/5
Ln T (p0 )
n (p0 ) ≈
n
Rp
where T (p0 ) ≈
p0 (x)dx. This rate can be achieved by using a very careful, adaptive
binning procedure, and then invoking the test from Section 3.2. The proofs make use of some
of the tools for high dimensional multinomials described in the previous sections.
The main point here is that the theory for multinomials has implications for continuous
data. It is worth noting that Fienberg and Holland (1973) was one of the first papers to
explicitly link the high dimensional multinomial problem to continuous problems.
3.4
Testing in Other Metrics
A natural question is to characterize the dependence of the local minimax separation and the
local minimax test on the choice of metric. We have focused thus far on the TV metric. The
paper Daskalakis et al. (2018) considers goodness-of-fit testing in other metrics. They provide
results on the global minimax separation for the Hellinger, Kullback-Leibler and χ2 metric. In
particular, they show that while the global minimax separation is identical for Hellinger and
TV, this separation is infinite for the Kullback-Leibler and χ2 distance because these distances
can be extremely sensitive to small perturbations of small entries of the multinomial.
In forthcoming work (Balakrishnan and Wasserman, 2018) we characterize the local minimax rate in the Hellinger metric for high-dimensional multinomials as well as for continuous
7
distributions. The optimal choice of test, as well as the local minimax separation can be
sensitive to the choice of metric and (in the continuous case) to the precise nature of the
smoothness assumptions.
Despite this progress, developing a comprehensive theory for minimax testing of highdimensional multinomials in general metrics, which provides useful practical insights, remains
an important open problem.
3.5
Composite Nulls and Imprecise Nulls
Now we briefly consider the problem of goodness-of-fit testing for a composite null. Let
P0 ⊂ M be a subset of multinomials and consider testing
H0 : P ∈ P0
verus H1 : P ∈
/ P0 .
A complete minimax theory for this case is not yet available, but many special cases have
been studied. In particular, the work of Acharya et al. (2015), provides some results for
testing monotonicity, unimodality and log-concavity of a high-dimensional multinomial. Here,
we briefly outline a general approach due to Acharya et al. (2015).
A natural approach to hypothesis testing with a composite null is to split the sample, estimate the null distribution using one sample, and then to test goodness-of-fit to the estimated
null distribution using the second sample. In the high-dimensional setting, we cannot assume
that our estimate of the null distribution is very accurate (particularly in the TV metric).
However, as highlighted in Acharya et al. (2015) even in the high-dimensional setting we can
often obtain sufficiently accurate estimates of the null distribution in the χ2 distance. This
observation motivates the study of the following two-stage approach. Use half the data to get
an estimate pb0 assuming H0 is true. Now use the other half to test the imprecise null of the
form:
H0 : d1 (p, pb0 ) ≤ θn verus H1 : d(p, pb0 ) ≥ n .
Here, pb0 is treated as fixed. Acharya et al. (2015) refer to this as “robust testing” but they are
not using the word robust in the usual sense. This imprecise null testing problem has been
studied for a variety of metric choices in Valiant and Valiant (2011); Acharya et al. (2015) and
Daskalakis et al. (2018).
Towards developing practical tests for general composite nulls, an important open problem
is to provide non-conservative methods for determining the rejection threshold for imprecise
null hypothesis tests. In the high-dimensional setting we can no longer rely on limiting distribution theory, and it seems challenging to develop simulation based methods in general. Some
alternative proposals have been suggested for instance in Berger and Boos (1994), but warrant
further study in the high-dimensional setting.
More specific procedures can be constructed based on the structure of P0 , and important
special cases of composite null testing such as two-sample testing and independence testing
have been studied in the literature. We turn our attention towards two-sample testing next.
4
Two Sample Testing
In this case the data are
Z1 , . . . , Zn1 ∼ P,
W 1 , . . . , W n2 ∼ Q
8
and the hypotheses are
H0 : P = Q versus H1 : P 6= Q.
Let X and Y be the corresponding vectors of counts. First, suppose that n1 = n2 := n. In
this case, Chan et al. (2014) showed that the minimax rate is
)
(
d1/2 d1/4
n max
,
.
(7)
n3/4 n1/2
The second term in the maximum is identical to the goodness-of-fit rate. In the high1/2
dimensional case d ≥ n, the minimax separation rate is nd 3/4 , which is strictly slower than
the goodness-of-fit rate. This highlights that, in contrast to the low-dimensional setting, there
can be a price to pay for testing when the null distribution is not known precisely (i.e. for
testing with a composite null).
Chan et al. (2014) also showed that the (centered) χ2 statistic
T =
X (Xj − Yj )2 − Xj − Yj
Xj + Yj
(8)
j
is minimax optimal. The critical value tα can be obtained using the usual permutation procedure (Lehmann and Romano, 2006). It should be noted, however, that the theoretical results
do not apply to the data-based permutation cutoff but, rather, to a cutoff based on a loose
upper bound on the variance of T . In the remainder of this section we discuss some extensions:
1. Unequal Sample Sizes: The problem of two-sample testing with unequal sample sizes
has been considered in Diakonikolas and Kane (2016) and Bhattacharya and Valiant
(2015), and we summarize some of their results here. Without loss of generality, assume
that n1 ≥ n2 . The minimax rate is
(
)
d1/2
d1/4
n max
(9)
, 1/2 .
1/4 1/2
n1 n2
n2
Note that this rate is identical to the minimax goodness of fit rate from Section 3 when
n1 ≥ d. This makes sense since, when n1 is very large, P can be estimated to high
precision and we are essentially back in the simple goodness-of-fit setting.
From a practical perspective, while the papers (Diakonikolas and Kane, 2016; Bhattacharya and Valiant, 2015) propose tests that are near-minimax they are not tests in
the usual statistical sense. They contain a large number of unspecified constants and it
is unclear how to choose the test threshold in such a way that the level of the test is α.
We could choose the constants somewhat loosely then use the permutation distribution
to get a level α test, but it is not known if the resulting test is still minimax, highlighting
an important gap between theory and practice.
2. Refinements of Minimaxity: In two-sample testing, unlike in goodness-of-fit testing,
it is less clear how precisely to define the local minimax separation. Roughly, we would
expect that distinguishing whether two samples are drawn from the same distribution
or not should be more difficult if the distributions of the samples are nearly uniform,
than if the distributions are concentrated on a small number of categories. Translating
this intuition into a satisfactory refined minimax notion is more involved than in the
goodness-of-fit problem.
9
One such refinement appears in the work of Acharya et al. (2012), who restrict attention to so-called “symmetric” test statistics (roughly, these statistics are invariant to
re-labelings of the categories of the multinomial). Their proposal is in the spirit of classical notions of adaptivity and oracle inequalities in statistics (see for instance van de
Geer (2015); Donoho et al. (1996) and references therein), where they benchmark the
performance of a test/estimator against an oracle test/estimator which is provided some
side-information about the local structure of the parameter space. Concretely, they compare the separation achieved by χ2 -type tests to the minimax separation achieved by an
oracle that knows the distributions P, Q up to a permutation.
3. Testing Continuous Distributions: Finally, we note that results for two-sample
testing of high-dimensional multinomials have implications for two-sample testing in the
continuous case. The recent paper of Arias-Castro et al. (2016) building on work by Ingster (1997), considers two-sample testing of smooth distributions by reducing the testing
problem to an appropriate high-dimensional multinomial testing problem. Somewhat
surprisingly, they observe that at least under sufficiently strong smoothness assumptions, the minimax separation rate for two-sample testing matches that of goodness-of-fit
testing.
5
Simulations
In this section, we report some simulation studies performed to illustrate that tests can have
high power even when their limiting distributions are not Gaussian and to illustrate the gains
from using careful modifications to classical tests for testing high-dimensional multinomials.
5.1
Limiting distribution of test statistics
One of the main messages of our paper is that tests can have high-power even in regimes
where their null distributions are not Gaussian (or more generally well-behaved). As a result,
restricting attention to regimes where the null distribution of a test statistic is well-behaved
can be severely limiting.
As an illustration, we consider goodness-of-fit testing where the null distribution is a power
law, i.e. we take p0 (i) ∝ 1/i. We will consider a high-dimensional setting where d = 1000
and n = 400. In this setting, as we will show in Section 5.2, the χ2 -statistic performs poorly
but the truncated statistic in (6) has high power. However, as illustrated in Figure 1 in this
high-dimensional regime the limiting distributions of both statistics are quite far from Normal.
We also observe that the classical χ2 statistic has a huge variance, which explains its poor
power and motivates our introduction of the truncated χ2 statistic. The truncated χ2 statistic
has a high-power, and achieves the minimax rate for goodness-of-fit testing, and as illustrated
in this simulation has a much better behaved distribution under the null.
5.2
Testing goodness-of-fit
In this section, we compare the performance of several goodness-of-fit test statistics. Throughout we take n = 400 and d = 1000. Closely related simulations appear in our prior work Balakrishnan and Wasserman (2017). In particular, we compare the classical χ2 statistic in (3),
10
6
5
4
3
2
1
0
-1
-2
-3
-4
-2
0
2
4
-2
0
2
4
5
4
3
2
1
0
-1
-2
-3
-4
-4
Figure 1: A plot of the distribution of the classical and truncated χ2 test statistics under a
power law null distribution, in the high-dimensional setting where n = 400, d = 1000, obtained
via simulation.
the likelihood-ratio test in (4), the truncated χ2 statistic in (6) and the two-stage locally
minimax 2/3rd and tail test described in Section 3.2, with the `1 and `2 tests given as
T`1 =
d
X
|Xi − np0 (i)|,
and T`2 =
i=1
d
X
(Xi − np0 (i))2 .
i=1
We examine the power of these tests under various alternatives:
1. Minimax Alternative: We perturb each entry by an amount proportional to p0 (i)2/3
with a randomly chosen sign. This is close to the worst-case perturbation used in Valiant
and Valiant (2017) in their proof of the local-minimax lower bound.
2. Uniform Dense Alternative: We perturb each entry of the null distribution by a
scaled Rademacher random variable. (A Rademacher random variable takes values +1
and −1 with equal probability.)
3. Sparse Alternative: In this case, we essentially only perturb the first two entries of
the null multinomial. We increase the two largest entries of the multinomial and then
11
1
1
0.9
0.9
0.8
0.8
0.7
0.7
Chi-sq.
L2
L1
Truncated chi-sq
LRT
2/3rd and tail
0.6
0.5
0.4
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
0.2
0.4
0.6
0.8
Chi-sq.
L2
L1
Truncated chi-sq
LRT
2/3rd and tail
0.6
0
1
0
0.2
0.4
0.6
0.8
Figure 2: A comparison between the truncated χ2 test, the 2/3rd + tail test (Valiant and
Valiant, 2017), the χ2 -test, the likelihood ratio test, the `1 test and the `2 test. The null is
chosen to be uniform, and the alternate is either a dense or sparse perturbation of the null.
The power of the tests are plotted against the `1 distance between the null and alternate.
Each point in the graph is an average over 1000 trials. Despite the high-dimensionality (i.e.
n = 200, d = 2000) the tests have high-power, and perform comparably.
re-normalize the resulting distribution, this results in a large perturbation to the two
largest entries and a relatively small perturbation to the other entries of the multinomial.
4. Alternative Proportional to Null: We perturb each entry of the null distribution
by an amount proportional to p0 (i), with a randomly chosen sign.
We observe that the truncated χ2 test and the 2/3rd + tail test from Valiant and Valiant (2017)
are remarkably robust. All tests are comparable when the null is uniform but the two-stage
2/3rd + tail test suffers a slight loss in power due to the Bonferroni correction. The distinctions
between the classical tests and the recently proposed modified tests are clearer for the power
law null. In particular, from the simulation testing a power-law null against a sparse alternative
it is clear that the χ2 and likelihood ratio test can have very poor power in the high-dimensional
setting. The `2 test appears to have high-power against sparse alternatives but performs poorly
against dense alternatives suggesting potential avenues for future investigation.
5.3
Two-sample testing
Finally, we turn our attention to the problem of two-sample testing for high-dimensional
multinomials. We compare three different test statistics, the two-sample χ2 statistic (8), the
`1 and `2 statistics:
T`1
d
X
Xi
Yi
=
−
,
n1
n2
T`2 =
i=1
d
X
Xi
i=1
12
Yi
−
n1
n2
2
,
1
1
Chi-sq.
L2
L1
Truncated chi-sq
LRT
2/3rd and tail
0.9
0.8
0.7
0.9
0.8
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0
0.6
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
1
Chi-sq.
L2
L1
Truncated chi-sq
LRT
2/3rd and tail
0.9
0.8
0.7
0.8
0.7
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0.1
0.2
Chi-sq.
L2
L1
Truncated chi-sq
LRT
2/3rd and tail
0.9
0.6
0
Chi-sq.
L2
L1
Truncated chi-sq
LRT
2/3rd and tail
0.3
0.4
0.5
0.6
0
0.7
0
0.2
0.4
0.6
0.8
1
Figure 3: A comparison between the truncated χ2 test, the 2/3rd + tail test (Valiant and
Valiant, 2017), the χ2 -test, the likelihood ratio test, the `1 test and the `2 test. The null is
chosen to be a power law with p0 (i) ∝ 1/i. We consider four possible alternates, the first
uniformly perturbs the coordinates, the second is a sparse perturbation only perturbing the
first two coordinates, the third perturbs each co-ordinate proportional to p0 (i)2/3 and the final
setting perturbs each coordinate proportional to p0 (i). The power of the tests are plotted
against the `1 distance between the null and alternate. Each point in the graph is an average
over 1000 trials.
and an oracle goodness-of-fit statistic. The oracle has access to the distribution P (and ignores
the first sample) and tests goodness-of-fit using the second sample. In our simulations, we use
the truncated χ2 test for goodness-of-fit.
Motivated by our simulations in the previous section, we consider the following pairs of
distributions in the two-sample setup:
1. Uniform P , Dense Perturbation Q: We take the distribution P to be uniform and
the distribution Q to be the distribution where we perturb each entry of P by a scaled
Rademacher random variable.
2. Power-Law P , Sparse Perturbation Q: Noting the difficulty faced by classical tests
for goodness-of-fit testing of a power law versus a sparse perturbation (see Figure 3) we
13
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
Chi-Sq.
L1
Oracle GoF
L2
0.2
0.1
0
Chi-Sq.
L1
Oracle GoF
L2
0
0.2
0.4
0.6
0.8
0.2
0.1
0
1
0
0.2
0.4
0.6
0.8
1
1
Chi-Sq.
L1
Oracle GoF
L2
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Figure 4: A comparison between the χ2 test, the `1 test, the `2 test, and the (oracle) goodnessof-fit test. In the three settings, the two distributions are chosen as described in the text (the
distribution P is chosen to be either uniform, power-law or minimax). The power of the tests
are plotted against the `1 distance between P and Q. The sample sizes from P and Q are
taken to be balanced and are each equal to 200. Each point in the graph is an average over
1000 trials.
consider a similar setup in the two-sample setting. We take each entry p(i) ∝ 1/i and
take q(i) to be the sparse perturbation described previously (where the two largest entries
are perturbed by a relatively large magnitude).
3. Minimax P, Q: This construction is inspired by the work of Batu et al. (2000), and
is used in their construction of a minimax lower bound for two-sample testing in the
high-dimensional setting (i.e. when d max{n1 , n2 }).
For a prescribed separation , with probability n1 /2d we choose p(i) = q(i) = 1/n1 and
with probability 1 − n1 /2d we choose p(i) = 1/(2d) and q(i) = 1/(2d) + Ri /(2d), where
the Ri denote independent Rademacher random variables. Both distributions are then
normalized.
Roughly, the two distributions contain a mixture of heavy elements of mass 1/n1 and
light elements of mass close to 1/d. The two distributions have `1 distance close to
and the insight of Batu et al. (2000) is that in the two-sample setting is quite difficult
to distinguish between variations in observed frequencies due to the perturbation of the
14
entries and due to the random mixture of heavy and light entries.
In each case, the cut-off for the tests is determined via the permutation method.
The balanced case: In this case, we set n1 = n2 = 200 and d = 400. We observe in
Figure 4 that there is a clear and significant loss in power relative to the goodness-of-fit oracle,
indicating the increased complexity of two-sample testing in the high-dimensional setup. We
note that, as is clear from the minimax separation rate we do not expect any loss in power in
the low-dimensional setting. The loss in power is exacerbated when we consider the minimax
two-sample pair (P, Q) described above. We also note that the loss in power is negligible for
the power law pair of distributions: due to the rapid decay of their entries these multinomials
can be well estimated from a small number of samples.
The imbalanced case: In this case, we set n1 = 2000, n2 = 200 and d = 400. The χ2
statistic for two-sample testing in the imbalanced case, is given by:
T =
d
X
(n2 Xi − n1 Yi )2 − n2 Xi − n2 Yi
2
i=1
1
Xi + Yi
where we follow Bhattacharya and Valiant (2015) and use a slightly modified centering of
the usual χ2 statistic. Since the χ2 -statistic is minimax optimal in the balanced case, one
might conjecture that this continues to be the case in the imbalanced case. However, as our
simulations suggest this is not the case. Somewhat surprisingly, the performance of the χ2
statistic can degrade when one of the sample sizes is increased (see Figure 5). The `1 statistic
on the other hand appears to be perform as expected, i.e. its power is close to that of the
oracle goodness-of-fit test in the case when one of the sample sizes is very large, and we believe
that this statistic warrants further study.
6
Discussion
Despite the fact that discrete data analysis is an old subject, it is still a vibrant area of
research and there is still much that we don’t know. Steve Fienberg showed prescience in
drawing attention to one of the thorniest issues: high-dimensional multinomials.
Much of the statistical literature has dealt with the high-dimensional case by imposing
assumptions on the distribution so that simple limiting distributions can be obtained. Doing
so gives up the most appealing property of multinomial inference: it is completely distributionfree. As we have seen in this paper, recent work by a variety of communities has developed
new and rather surprising theoretical results. What is often missing in the recent literature is
the appreciation that statisticians want tests with precise control of the type I error rate. As
a result, there remain gaps between theory and practice.
We have focused on goodness of fit and two sample problems. There is a rich literature
on other problems such as independence testing and testing shape constraints (Diakonikolas
and Kane, 2016; Acharya et al., 2015). As we discussed earlier, Balakrishnan and Wasserman
(2017) showed that these new results for high-dimensional discrete data have implications for
continuous data. There is much more to say about this and this is a direction that we are
actively pursuing. Finally, we have restricted attention to hypothesis testing. In future work,
we will report results on high dimensional inference using confidence sets and point estimation.
15
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
Chi-Sq.
L1
Oracle GoF
L2
0.3
Chi-Sq.
L1
Oracle GoF
L2
0.2
0.1
0
0
0.2
0.4
0.6
0.8
0.2
0.1
0
1
0
0.2
0.4
0.6
0.8
1
1
0.9
0.8
0.7
0.6
Chi-Sq.
L1
Oracle GoF
L2
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Figure 5: A comparison between the χ2 test, the `1 test, the `2 test, and the (oracle) goodnessof-fit test. In the three settings, the two distributions are chosen as described in the text (the
distribution P is chosen to be either uniform, power-law or minimax). The power of the tests
are plotted against the `1 distance between P and Q. The sample sizes from P and Q are
taken to be imbalanced, i.e. we take n1 = 2000 and n2 = 200. Each point in the graph is an
average over 1000 trials.
7
Acknowledgements
This research was supported by NSF grant DMS1713003. All of our simulation results and
figures can be reproduced using code made available at (http://www.stat.cmu.edu/~siva/
multinomial).
References
J. Acharya, H. Das, A. Jafarpour, A. Orlitsky, S. Pan, and A. Suresh. Competitive classification
and closeness testing. In Proceedings of the 25th Annual Conference on Learning Theory,
volume 23, pages 22.1–22.18. 2012.
J. Acharya, C. Daskalakis, and G. Kamath. Optimal testing for properties of distributions. In
Advances in Neural Information Processing Systems, pages 3591–3599. 2015.
E. Arias-Castro, B. Pelletier, and V. Saligrama. Remember the curse of dimensionality: The
16
case of goodness-of-fit testing in arbitrary dimension. arXiv preprint arXiv:1607.08156,
2016.
S. Balakrishnan and L. Wasserman. Hypothesis testing for densities and high-dimensional
multinomials: Sharp local minimax rates. arXiv preprint arXiv:1706.10003, 2017.
S. Balakrishnan and L. Wasserman. Hypothesis testing for densities and high-dimensional
multinomials II: Sharp local minimax rates. Forthcoming, 2018.
A. R. Barron. Uniformly powerful goodness of fit tests. Ann. Statist., 17(1):107–124, 1989.
T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White. Testing that distributions are
close. In 41st Annual Symposium on Foundations of Computer Science, FOCS 2000, 12-14
November 2000, Redondo Beach, California, USA, pages 259–269. 2000.
R. L. Berger and D. D. Boos. P values maximized over a confidence set for the nuisance
parameter. Journal of the American Statistical Association, 89(427):1012–1016, 1994.
B. Bhattacharya and G. Valiant. Testing closeness with unequal sized samples. In Advances
in Neural Information Processing Systems, pages 2611–2619. 2015.
Y. M. Bishop, S. E. Fienberg, and W. Paul. Holland. 1975. Discrete Multivariate Analysis:
Theory and Practice, pages 57–122, 1995.
T. Cai and M. Low. A framework for estimation of convex functions. Statistica Sinica, 2015.
S.-O. Chan, I. Diakonikolas, G. Valiant, and P. Valiant. Optimal algorithms for testing
closeness of discrete distributions. In Proceedings of the twenty-fifth annual ACM-SIAM
symposium on Discrete algorithms, pages 1193–1203. Society for Industrial and Applied
Mathematics, 2014.
S. Chatterjee, A. Guntuboyina, and B. Sen. On risk bounds in isotonic and other shape
restricted regression problems. Ann. Statist., 43(4):1774–1800, 2015.
C. Daskalakis, G. Kamath, and J. Wright. Which distribution distances are sublinearly
testable? Symposium on Discrete Algorithms, 2018.
L. Devroye and L. Gyorfi. Nonparametric Density Estimation: The L1 View. Wiley Interscience Series in Discrete Mathematics. Wiley, 1985.
I. Diakonikolas and D. M. Kane. A new approach for testing properties of discrete distributions.
In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on, pages
685–694. IEEE, 2016.
D. L. Donoho and I. M. Johnstone. Ideal spatial adaptation by wavelet shrinkage. Biometrika,
81:425–455, 1994.
D. L. Donoho, I. M. Johnstone, G. Kerkyacharian, and D. Picard. Density estimation by
wavelet thresholding. Ann. Statist., 24(2):508–539, 1996.
S. E. Fienberg. The analysis of cross-classified categorical data. Springer, 1976.
S. E. Fienberg and P. W. Holland. Simultaneous estimation of multinomial cell probabilities.
Journal of the American Statistical Association, 68(343):683–691, 1973.
17
A. Goldenshluger and O. Lepski. Bandwidth selection in kernel density estimation: Oracle
inequalities and adaptive minimax optimality. Ann. Statist., 39(3):1608–1632, 2011.
W. Hoeffding. Asymptotically optimal tests for multinomial distributions. Ann. Math. Statist.,
36(2):369–401, 1965.
L. Holst. Asymptotic normality and efficiency for certain goodness-of-fit tests. Biometrika,
59(1):137–145, 1972.
Y. Ingster and I. Suslina. Nonparametric Goodness-of-Fit Testing Under Gaussian Models.
Lecture Notes in Statistics. Springer, 2003.
Y. I. Ingster. Adaptive chi-square tests. Zapiski Nauchnykh Seminarov POMI, 244:150–166,
1997.
G. Ivchenko and Y. Medvedev. Decomposable statistics and hypotheses testing: The case of
small samples. Teoriya Veroyatnostej i Ee Primeneniya, 23, 1978.
L. LeCam. Convergence of estimates under dimensionality restrictions. Ann. Statist., 1(1):38–
53, 1973.
E. Lehmann and J. Romano. Testing Statistical Hypotheses. Springer Texts in Statistics.
Springer New York, 2006.
C. Morris. Central limit theorems for multinomial sums. Ann. Statist., 3(1):165–188, 1975.
L. Paninski. A coincidence-based test for uniformity given very sparsely sampled discrete data.
IEEE Transactions on Information Theory, 54(10):4750–4755, 2008.
T. R. C. Read and N. A. C. Cressie. Goodness-of-fit statistics for discrete multivariate data.
Springer-Verlag Inc, 1988.
G. Valiant and P. Valiant. Estimating the unseen: An n/log(n)-sample estimator for entropy
and support size, shown optimal via new clts. In Proceedings of the Forty-third Annual
ACM Symposium on Theory of Computing, STOC ’11, pages 685–694. ACM, New York,
NY, USA, 2011.
G. Valiant and P. Valiant. An automatic inequality prover and instance optimal identity
testing. SIAM Journal on Computing, 46(1):429–455, 2017.
S. van de Geer. Estimation and Testing Under Sparsity: Saint-Flour XLV - 2015. Springer,
1st edition, 2015.
Y. Wei and M. J. Wainwright. The local geometry of testing in ellipses: Tight control via
localized Kolomogorov widths. arxiv preprint arXiv:1712.00711, 2017.
Appendix
Here we describe the local minimax results for goodness of fit testing more precisely. Without
loss of generality we assume that the entries of the null multinomial p0 are sorted so that
18
p0 (1) ≥ p0 (2) ≥ . . . ≥ p0 (d). For any 0 ≤ σ ≤ 1 we denote σ-tail of the multinomial by:
d
X
(10)
Qσ (p0 ) = i :
p0 (j) ≤ σ .
j=i
The σ-bulk is defined to be
Bσ (p0 ) = {i > 1 : i ∈
/ Qσ (p0 )}.
(11)
Note that i = 1 is excluded from the σ-bulk. The minimax rate depends on the functional:
3/2
Vσ (p0 ) =
X
p0 (i)2/3
.
(12)
i∈Bσ (p0 )
Define, `n and un to be the solutions to the equations:
)
( r
)
( r
V`n (p0 ) (p0 )
Vun (p0 )/16 (p0 )
1
1
,
, un (p0 ) = max
,
.
`n (p0 ) = max
n
n
n
n
(13)
With these definitions in place, we are now ready to state the result of Valiant and Valiant
(2017). We use c1 , c2 , C1 , C2 > 0 to denote positive universal constants.
Theorem 1 (Valiant and Valiant (2017)). The local critical radius n (p0 , M) for multinomial
testing is upper and lower bounded as:
c1 `n (p0 ) ≤ n (p0 , M) ≤ C1 un (p0 ).
Furthermore, the global critical radius n (M) is bounded as:
c2 d1/4
C2 d1/4
√
≤ n (M) ≤ √ .
n
n
19
(14)
| 7cs.IT
|
A macro placer algorithm for chip design
Endre Csóka, Attila Deák
December 13, 2017
arXiv:1509.01867v1 [cs.DS] 6 Sep 2015
Abstract
There is a set of rectangular macros with given dimensions, and there are wires connecting some pairs (or
sets) of them. We have a placement area where these macros should be placed without overlaps in order to
minimize the total length of wires. We present a heuristic algorithm which utilizes a special data structure for
representing two dimensional stepfunctions. This results in fast integral computation and function modification
over rectangles. Our heuristics, especially our data structure for two-dimensional functions, may be useful in
other applications, as well.
1
Introduction
A chip is composed of basic elements called cells, circuits, boxes or modules. They usually have a rectangular shape,
contain several transistors and internal connections, and have (at least two) fixed pins. There is a netlist describing
which pin should be connected to which other pins. The goal is to place the cells legally – without overlaps – in
the chip area so as to minimize the total (weighted) length of the wires connecting the pins. This problem is also
called the VLSI placement problem.
Finding the optimum is NP-hard, therefore, we present a heuristic algorithm based on primal-dual optimization
inspired by the Hungarian Algorithm [6] for the minimum weight maximum matching problem. Namely, we use a
cost function as dual function on the placement area, and we are looking for a placement minimizing the sum of the
total netlength and the total costs of the areas covered by the macros. We try to find a non-negative cost function
by which an almost optimal placement is legal even if we allow overlaps, and costs are counted with multiplicity.
We will use an iterated algorithm on the space of primal-dual pairs based on the following two steps:
1. For every overlap, we increase the cost function under intersecting areas.
2. We try to find a better placement with respect to the new cost function.
We tried to focus on typical instances in practice, and we found that these have the following properties.
1. The placement area is not very large compared to the total size of the macros, but it is still easy to find a
legal placement.
2. There are about a few hundreds of macros and every macro is contained in at most 10 nets.
3. Most of the nets connect two, sometimes three, and rarely more than three macros to each other.
Our method is optimized for such inputs.
This paper is organized as follows. In Section 2, we introduce some notations and give a formal definition of
the macro placement problem. In Section 3, we describe the basic idea behind our algorithm and in Section 4, we
present the algorithm. In Section 5, we describe some additional heuristics used in our placer.
2
The macro placement problem
Now we give a formal definition of the simplified macro placement problem. Let us denote by M the set of macros.
We assume that all pins of each macro are in the center of the macro. The place of a macro is identified with the
1
place of its center pin. For a macro M , denote its horizontal and vertical size by sizex(M ), sizey (M ), respectively.
For a macro M at (x, y), we denote the area occupied by M by
sizex(M )
sizey (M )
sizex(M )
sizey (M )
, x+
, y+
S(M, (x, y)) = x −
× y−
.
2
2
2
2
A net N is a subset of the macros that are connected.
Definition 1. A netlist is a pair (M, N ) where M is a finite set of macros and N ⊆ P(M) is a set of subsets of
M.
One can think of N as a hypergraph on M, where each N ∈ N is a hyperedge. We assume |N | ≥ 2 for each
N ∈ N.
Definition 2. The placement area is a rectangle denoted by A. This contains a set of rectangular blockages B
(where ∀B ∈ B, B ⊂ A). The sides of all rectangles are parallel to the axis.
A blockage is a part of the placement area where no macro can be placed.
Definition 3. A placement is a map p : M 7→ A. The placement p is legal if all of the followings hold.
• Every macro M ∈ M is placed in the placement area:
S M, p(M ) ⊆ A.
• The places of any two macros M, M ′ ∈ M are disjoint:
S M, p(M ) ∩ S M ′ , p(M ′ ) = ∅.
• None of the macros M ∈ M are placed on a blockage B ∈ B:
S M, p(M ) ∩ B = ∅.
The macros have to be placed in the given orientation, these cannot be rotated. Let (M, N ) be a netlist
and p a legal placement to the placement area A with blockages B.
Define p on the set of nets N as follows. For
N = {M1 , M2 , . . . , Mk } ∈ N , let p(N ) = p(M1 ), p(M2 ), . . . , p(Mk ) . We have a function L : A2 ∪A3 ∪A4 ∪. . . 7→ R+
which evaluates the length of a net. L is also called the net (or netlength) model. One commonly used net model
is the bounding-box model:
BB (x1 , y1 ), (x2 , y2 ), . . . , (xk , yk ) = max{xi } − min{xi } + max{yi } − min{yi }
(1)
i
i
i
i
This is the half perimeter of the smallest rectangle with sides parallel to the axis, containing all pins of the
macros contained in the net N .
The Simplified Placement Problem:
Given a netlist (M, N ), a placement area A, the set of blockages B and a net model L, find a legal placement
p : M 7→ A which minimizes the total netlength:
X
L p(N ) .
N ∈N
3
Basic tools of the placer
The initial problem is to place macros in the placement area (avoiding the blocked areas) so that the total netlength
is minimal (or close
to the minimum). As finding the optimum is NP hard, we present a heuristic algorithm with
O log(n) log(m)s running time, where the placement area is a discrete n × m grid and we run the algorithm for s
rounds.
We introduce our algorithm in several steps.
2
Problem 1. We have a set of macros M and disjoint slots A. There is a cost function c : M × A 7→ R+ which
assigns costs to every possible macro-slot assignment. Find an injective assignment p : M 7→ A with minimum
total cost
X
c M, p(M ) .
M∈M
Solution. This scenario can be represented by a bipartite weighted graph. The two set of points are M and A and
the cost of an edge (M, A) is c(M, A). The task is to cover M with a minimum cost maximum matching. This is
a well known optimisation problem, and it is usually solved by primal-dual methods (e.g. the Hungarian Method
[6]). However, we choose the following method instead, because we will generalize this in the later steps. We try to
find a primal-dual solution by the following market simulation: if there is an area which is the best possible choice
for at least two macros, then we increase the cost of that area.
Problem 2. For a macro M ∈ M and a given netlist N , let E(M ) be the set of nets containing M :
E(M ) = N ∈ N M ∈ N ,
and let N (M ) be the set of its neighbors:
N (M ) = M ′ ∈ M
∃N ∈ E(M ) : M ′ ∈ N .
Let L(N ) be the netlength model. Given a set of macros M, disjoint slots A and netlist N , find an injective
assignment p : M 7→ A which minimizes
X
L p(N ) .
N ∈N
Solution. We try to use the solution of Problem 1, where the cost of the placement of one macro is replaced by
its marginal contribution to the total cost. Now, the cost of one macro depends on the placement of its neighbors,
but there are not too many neighbors, therefore, we expect the cost function to change rather slowly. This allows
us to use a variation of the above described market simulation:
1. Take an arbitrary macro-area assignment p.
2. For each macro M ∈ M, fix the other macros at their current position. For every A ∈ A, we get a placement
p(M,A) from p by changing the assignment of the macro M to A. Let us define the marginal contribution of
M placed at A by
X
L p(M,A) (N ) .
cp (M, A) = Cost(A) +
N ∈E(M)
3. Use the method described in the solution of Problem 1 with cost function cp for M, A to get a better
assignment p′ .
4. Continue with step 2 using the assignment p′ given in 3.
We run this procedure for a number of rounds.
Remark 1. With no further adjustment, the algorithm can easily result in an infinite loop as the following example
shows:
Consider two macros and let the netlist be one single net connecting the two macros. During the run of the algorithm
if the two macros are at different position, the macro with the higher cost would move to the area where the other
macro is, because this decreases the total netlength and the total cost of the macros. The cost of this area increases
until one of them moves to another place. As before the net connecting this macro to the other causes the other
macro to move as well to the same area. Therefore the same process starts again. This shows that increasing the
cost under the overlaps alone is not enough.
After each round in the algorithm if there are at least two macros at the same area, we increase the cost of that
area. At the beginning of the algorithm we allow overlaps to get a better placement, basically allowing not only
better, but slightly worse placements to prevent the algorithm to get stuck early in some local minimum. Later,
we increasingly punish overlaps to prevent the loop in Remark 1. This is a kind of cooling process. If we set
the increment rate properly, the macros will have the time to distribute evenly in the placement area, with small
netlength. Later, this punishment goes to infinity, hereby enforcing a legal solution.
3
Problem 3. There is a set of macros M, a netlist N , a placement area A which is a discrete n × m grid and a
netlength model L given. Each edge length of each macro is the multiple of the edge length of the grid. A placement
p : M 7→ A is legal if S(M1 , p(M1 )) ∩ S(M2 , p(M2 )) = ∅, M1 6= M2 . During the run of the algorithm we allow
non-legal placements. We only require that the final placement p is legal. Our task is to find a legal placement
which minimizes
X
L p(N ) .
N ∈N
Solution. Here the places are not disjoint as in Problem 2, so during the run of the algorithm the macros can
overlap partially as well. In this case we increase the cost under the intersection proportional to its size.
Problem 4. In the general setting the sizes can be real numbers.
Solution. To use the solution of Problem 3, we divide the placement area to a sufficiently fine discrete grid and
use only natural numbers for approximation. We round up the edge length of each macro to the nearest multiple
of the edge length of the grid.
4
Our global placer
The algorithm receives an initial placement (e.g. random with many overlaps) and then refines it to a global
placement with minimized total netlength.
4.1
The structure of the algorithm
Our algorithm consists of 4 steps as follows: At the beginning, we generate an initial placement, or we use the given
one. Then, for a given number of rounds, we do the following.
1. We choose a macro M randomly with original position X0 .
2. We generate t possible new positions X1 , . . . , Xt around its original position, with move macro(M)
3. We move the macro M to the positon Xi that minimizes
weight (M , Xi ) + N etLength EXi (M ) + penalty(M, Xi),
(2)
where EXi (M ) is obtained by moving the macro M to Xi .
4. We increase the weights under the overlaps of M with every other macros.
4.2
Notations
To discretize the problem, we consider the placement area to be a finite n × m grid. We can assume that n =
2p , m = 2q . These parameters are free to choose according to the available computing resources. Denote the size of
the placement area by Ax and Ay (for the horizontal and vertical size). After we set n, m, we divide the placement
A
area to an n × m grid. Our grid will consist of nm squares of dimensions x × y = Anx × my . We record the cost as
a stepfunction on the placement area which is constant on the cells of the grid. In other words we define the cost
on the cells of the grid and not as a function on the placement area. Let Pi,j be the weight under the ith square
of the jth row. Denote by Mn,m the set of all n × m matrices. Define the inner product of two matrices (say
A, B ∈ Mn,m ) as follows:
m
n X
X
ai,j bi,j .
A ⋆ B :=
i=1 j=1
We represent a rectangle with its top-left and bottom-right corners. For a given rectangle R̃ = (x1 , y1 ), (x2 , y2 ) ,
x1 < x2 , y1 < y2 , we consider the slightly larger rectangle
j k j k l m l m !
x1
y1
x2
y2
R=
x,
y ,
x,
y
= (a1 x, b1 x), (a2 x, b2 x) .
x
y
x
y
4
This is the smallest rectangle of the grid covering R̃. From now on let every rectangle be given in the form:
R = (a1 x, b1 y), (a2 x, b2 y) = (a1 , b1 ), (a2 , b2 ) .
The characteristic function of R = (a1 , b1 ), (a2 , b2 ) is defined as the following n × m matrix.
1 if a1 < i ≤ a2 and b1 < j ≤ b2
AR = (ai,j )n,m
,
a
=
i,j
i,j=1
0 otherwise
4.3
Data structure for the weights
We introduce a data structure by which we can calculate (2) in O log(n) log(m) time, andalso, we can increase
the cost function by a constant under any rectangle R (as in Problem 4) in O log(n) log(m) time.
Remark 2. We have two operations on P .
• f (R, P ) = AR ⋆ P returns the total cost under a given rectangle R.
• g(R, P, w) increases each entry of P by w under the rectangle R. (P := P + wAR )
In our algorithm, we use these operations in every round, therefore, we need to compute them fast.
We construct an orthogonal basis {B1 , B2 , . . . , Bk } in this space (k = nm). For a given rectangle R, in order to
compute AR ⋆PP , we only need to know the products Bi ⋆ AR for every i = 1, . . . , k. Increasing the entries under R
by w in P = i αi Bi can be done by increasing the coefficients of the expansion αi = αi + wAR ⋆ Bi . We construct
a base such that for every rectangle R, there are only a few basis elements which are not orthogonal to R, and
hence the inner product can be computed in constant time. First, consider the one dimensional array P = (pi )ni=1
and let n = 2p . Define Bka as follows.
(2k − 2)2a < j ≤ (2k − 1)2a
1
a
Bk (j) =
−1 (2k − 1)2a < j ≤ 2k2a
0
else
where a = 0, . . . , p, k = 1, . . . , 2p−a−1 , j = 1 . . . , n. We also consider the basis element B1p = 1. For example, the
elements for n = 8 are
0
0
1
0
0
0
−1
0
0
0
0
1
0
, B20 = −1 , B30 = 0 , B40 = 0 ,
B10 =
1
0
0
0
−1
0
0
0
0
1
0
0
0
0
−1
0
1
B1 =
1
1
−1
−1
0
0
0
0
, B21 =
0
0
0
0
1
1
−1
−1
, B12 =
1
1
1
1
−1
−1
−1
−1
, B13 =
1
1
1
1
1
1
1
1
Lemma 1. Let s, t ∈ N, s < t. For a given Rs,t , there are at most 2 log(n) elements of the basis Bka for which
Rs,t ⋆ Bka 6= 0.
Proof: It is easy to check that Bka is an orthogonal basis in Rn . Let Rs,t ∈ Rn as in the Lemma:
1 s<j≤t
Rs,t (j) =
0 else
5
If Bka ⋆ Rs,t 6= 0 then either (3) or (4) or both holds.
(2k − 2)2a < s ≤ 2k2a
(3)
(2k − 2)2a < t ≤ 2k2a
(4)
For a given Rs,t , the number of Bka for which (3),(4) or both holds, is at most 2 log(n). This completes the proof
of the lemma.
It is easy to compute the scalar product of Rs,t and Bka :
Starx (Rs,t , Bka ) = − min s − (2k − 2)2a , 2k2a − s ,
Stary (Rs,t , Bka ) = min t − (2k − 2)2a , 2k2a − t .
Then the scalar product of Rs,t and Bka is:
if only (3) holds
Starx (Rs,t , Bka )
a
Stary (Rs,t , Bka )
if only (4) holds
Rx,y ⋆ Bk =
Starx (Rs,t , Bka ) + Stary (Rs,t , Bka ) if (3) and (4) holds
(5)
Now we can get a basis in Rn×m from the one dimensional case as follows. Let
a,b
Bk,l
(i, j) = Bka (i) · Blb (j),
a,b
where Bka corresponds to the basis in Rn and Blb to the basis in Rm . It is not hard to check that {Bk,l
} is a basis in
Mn,m . Following the argument of Lemma 1, for any rectangle R, there are at most O log(n) log(m) basis elements
a,b
a,b
(Bk,l
) not orthogonal to AR . Furthermore, from (5), the scalar product A(x1 ,x2 )×(y1 ,y2 ) ⋆ Bk,l
can be computed in
constant time.
a,b
It is easy to see that Bk,l
satisfies:
a,b
a,b
• ∀R rectangle, Bk,l
: AR ⋆Bk,l
6= 0 = O log(n) log(m) . Furthermore, we can find them in O log(n) log(m)
time.
a,b
• ∀R rectangle AR ⋆ Bk,l
can be computed in constant time.
4.4
Inflation
During the run of the algorithm, the cost of crowded areas may get too high, causing that all macros will avoid
that area. Rather than waiting for the costs of all the other places to increase, we implement a cost reducer. It will
reduce the differences between the high- and low-cost areas. We chose the method below because it can be easily
implemented without further computation time. The best rate of inflation should be adjusted.
4.5
The increase(R,value) subroutine
a,b
Let α = (αa,b
k,l ) be a global variable denoting the coefficients of the basis elements Bk,l in the expansion of P =
P
a,b a,b
a,b
k,l,a,b αk,l Bk,l . The increase(R,value) subroutine computes the scalar product of the basis elements Bk,l and AR ,
a,b
a,b
and increases the current coefficient of Bk,l
with this product multiplied by value. We repeat this for all Bk,l
:
Algorithm 1 increase(R,value)
for {a = 0, . . . , log(n)} do
for {b = 0, . . . , log(m)} do
a,b
for {k, l : Bk,l
⋆ AR 6= 0} do
a,b
a,b
αa,b
k,l = αk,l + scalar(AR , Bk,l ) ∗ value
end for
end for
end for
6
In line 3, we can find the pairs (k, l) in constant time as follows. For a given R there are at most 4 pairs (k, l)
a,b
such that scalar(AR , Bk,l
) 6= 0. The possible pairs (k, l) can be found easily from the coordinates of R. Fix a, b
and a rectangle R = ((x1 , y1 ), (x2 , y2 )), where x1 < x2 , y1 < y2 . Let ki , lj be such that
(2k1 − 2)2a < x1 ≤ 2k1 2a , (2k2 − 2)2a < x2 ≤ 2k2 2a , and
(2l1 − 2)2a < y1 ≤ 2l1 2a , (2l2 − 2)2a < y2 ≤ 2l2 2a
.
, Bka,b
, Bka,b
, Bka,b
holds. The basis elements (with fixed a, b) possibly not orthogonal to AR are Bka,b
2 ,l2
1 ,l2
2 ,l1
1 ,l1
4.6
The cost(R) subroutine
Here, Round ∈ N is a global variable denoting the current round of the algorithm. The cost(R) function receives a
rectangle R and returns the total cost of the cells inside this rectangle. This routine uses the basis expansion for
the cost matrix P in order to compute the scalar product as follows:
Algorithm 2 cost(R)
cost = 0;
for {a = 0, . . . , log(n)} do
for {b = 0, . . . , log(m)} do
a,b
for {k, l : Bk,l
⋆ AR 6= 0} do
a,b
cost = cost + αa,b
k,l ∗ scalar(AR , Bk,l )
end for
end for
end for
cost = cost + penalty(Round, R);
return cost;
5
Heuristics
In this section, we discuss further parameters of the algorithm. We make suggestions for all parameters, but these
should be experimentally adjusted.
5.1
The move macro(M) subroutine
This routine returns a new possible place for M . As before, let Ax , Ay denote the horizontal and vertical size of the
placement area A. The location of a macro M is given by its placement coordinates (x, y). For a macro M ∈ M,
let us denote the largest and smallest possible x coordinates for the macro M by
xmax (M ) = Ax −
sizex(M )
,
2
sizex(M )
.
2
We define ymin (M ), ymax (M ) analogously. Let γ(x) = exp log(x) · U [0, 1] where U [0, 1] is a uniformly distributed random variable in [0, 1]. This distribution is our heuristic choice. The subroutine:
xmin (M ) =
Algorithm 3 move macro(M )
a = Rand{-1,1}
b = Rand{-1,1}
if
x − γ(x + 1)
xnew =
x + γ(xmax − x) if
if
y − γ(y + 1)
ynew =
y + γ(ymax − y) if
Return xnew , ynew
a=1
a = −1
b=1
b = −1
7
5.2
The penalty(Step,R) function
Algorithm 4 penalty(Step, R)
cost = 0
for {M ∈ M, M 6= R} do
cost = cost + c ∗ δStep ∗ Circ(M ∩ R)
end for
Return cost
Here, Circ(R) denotes the circumference of the rectangle R, c is a constant and δStep is a parameter.
5.3
The smooth edge(E) function
We will consider the bounding-box model only. During the earlier stages of the optimization, when we compare
different positions of a macro and we calculate the total distance of the wires, then we should consider that the
positions of the neighbors are still rough. Therefore, it turns out to be useful to consider the positions of the
neighboring pins with some uncertainty, namely, as distributions around their present positions. This can be
expressed by using a smoothed version of the absolute value function of the difference in each coordinate. This tool
was already used in the literature, it is common to approximate the bounding box model (1) with strictly convex
functions which converges to the bounding-box netlength. One of them is the log-sum-exp function (see [2], [3],
[4]):
X
X
LSEx (N ) := α log
exp x(p)/α + α log
exp − x(p)/α ,
p∈N
p∈N
and LSE(N ) := LSEx (N ) + LSEy (N ). It is easy to see that LSE(N ) → BB(N ), as α → 0.
An alternative way is to approximate with Lp norms (see [5]):
1/p
X
p
LPx (N ) :=
x(p) − x(q) + α
,
p,q∈N
and LP (N ) := LPx (N ) + LPy (N ). LP (N ) → BB(N ) holds again, if α1 → ∞, p → ∞.
We used exponential functions, in a way similar to the log-sum-exp model, as follows:
1 X
N Lx (N ) =
log exp βx(p) + exp − βx(p) ,
β
p∈N
and N L(N ) = N Lx (N ) + N Ly (N ). It is clear that N L(N ) → BB(N ) holds if β → ∞. We use
β=
M axRounds
,
M axRounds − Round + 1
where M axRounds is the number of rounds for which we want to run the algorithm. Formally the code of this
subroutine is as follows:
Algorithm 5 smooth edge(E)
Cx = β1 log(exp(βx(E)) + exp(β(−x(E))))
Cy = β1 log(exp(βy(E)) + exp(β(−y(E))))
Return Cx + Cy
Notice that after many rounds, the edge length tends to the actual Bounding-box netlength.
5.4
Possible remaining overlaps
It is usually useful to stop the global placement before it removes all the overlaps. Our placer is ineffective in the
very final stages of the algorithm, when the actual placement is almost legal, and only a few small overlaps should
be eliminated. Therefore, we can get slightly better results if we stop the algorithm before the very final steps, and
we use some other final legalization method, even a simple naive one. In our case, these final minor modifications
were performed by hand.
8
6
Conclusions
In this paper we gave a heuristic algorithm for the NP-hard macro placement problem. The design of the algorithm
is based on a primal-dual approach to a matching problem (see Section 3, Problem 1).
First, we implemented a special data structure to handle the dual (cost) function efficiently during the algorithm.
This can records a multidimensional (in our case, 2-dimensional) discrete function, and performs efficiently the
following two operations. It returns with the sum (integral) of the values in any rectangle, and it can increase the
function with any constant in any rectangle. This data structure can also be useful for other purposes.
The second part includes the heuristics (see Section 5) inspired by the Hungarian Algorithm. We suggest an
algorithm that iteratively revises the primal and the dual functions. Despite a pair of optimal primal-dual solutions
do not exists, this causes problems only around the finalization of the placement. Our this heuristics seemed to
perform well for finding good rough positions for the macros. Therefore, we used a natural continuous transition
of the primal-dual method to a simple algorithm which just enforces disjointness. There were many minor details
where we found nontrivial solutions which can be used in other problems, as well. All these together provide a
flexible and robust algorithm for the VLSI placement problem, which can be easily optimized for different scenarios.
References
[1] B. Korte, D. Rautenbach, J. Vygen, BonnTools: Mathematical Innovation for Layout and Timing Closure of
Systems on a Chip, Proceedings of the IEEE (2007), 95 (3), 555-572
[2] T.F. Chan, J. Cong, J.R. Shinnerl, K. Sze, M. Xie, mPL6: enhanced multilevel mixed-size placement, Proceedings
of the International Symposium on Physical Design (2006), 212-214.
[3] J. Cong, G. Luo, Highly efficient gradient computation for desity-constrained analytical placement methods,
Proceedings of the International Symposium on Physical Design (2008), 39-46.
[4] A.R. Agnihotri, P.H. Madden, Fast Analytic Placement using Minimum Cost Flow, Proceedings of the Asia and
South Pacific Design Automation Conference (2007), 128-134.
[5] C.J. Alpert, T.F Chan, D.J. Huang, A.B Kahng, I.L. Markov, P. Mulet, K. Yan, Faster minimization of linear
wirelength for global placement, Proceedings of the International Symposium on Physical Design (1997), 4-11.
[6] H. W. Kuhn, The Hungarian Method for the assignment problem, Naval Research Logistic Quarterly, 2 (1955)
83-97.
9
| 8cs.DS
|
Markov Chain Lifting and Distributed ADMM
Guilherme França∗
Boston College, Computer Science Department and
Johns Hopkins University, Center for Imaging Science
arXiv:1703.03859v1 [stat.ML] 10 Mar 2017
José Bento†
Boston College, Computer Science Department
Abstract
The time to converge to the steady state of a finite Markov chain can be greatly reduced by
a lifting operation, which creates a new Markov chain on an expanded state space. For a class
of quadratic objectives, we show an analogous behavior where a distributed ADMM algorithm
can be seen as a lifting of Gradient Descent algorithm. This provides a deep insight for its faster
convergence rate under optimal parameter tuning. We conjecture that this gain is always present,
as opposed to the lifting of a Markov chain which sometimes only provides a marginal speedup.
∗
[email protected]
†
[email protected]
1
I.
INTRODUCTION
Let M and M̂ be two finite Markov chains with states V and V̂, of sizes |V| < |V̂|, and
with transition matrices M and M̂ , respectively. Let their stationary distributions be π and
π̂. In some cases it is possible to use M̂ to sample from the stationary distribution of M.
A formal set of conditions under which this happen is known as lifting. We say that M̂ is
a lifting of M if there is a row stochastic matrix S ∈ R|V̂|×|V| with elements Sij ∈ {0, 1}
and a single nonvanishing element per line, where 1|V| = S > 1|V̂| , and 1n is the all-ones
n-dimensional vector, such that
π = S > π̂,
Dπ M = S > Dπ̂ M̂ S.
(1)
We denote S > the transpose of S, and for any vector v ∈ Rn , Dv = diag(v1 , . . . , vn ).
Intuitively, M̂ contains copies of the states of M and transition probabilities between the
extended sates V̂ such that it is possible to collapse M̂ onto M. This is the meaning of
relation (1). See Fig. 1 for an illustration. (We refer to [1] for more details on Markov chain
lifting.)
The mixing time H is a measure of the time it takes for the distribution of a Markov
chain M to approach stationarity. We follow the definitions of [1] but, up to multiplicative
factors and slightly loser bounds, the reader can think of
H = min{t : max
|pti − πi | < 1/4},
0
{i,p }
(2)
where pti is the probability of being on state i after t steps, starting from the initial distribution p0 . Lifting is particularly useful when the mixing time Ĥ of the lifted chain is much
√
smaller than H. There are several examples where Ĥ ≈ C H, for some constant C ∈ (0, 1)
which depends only on π. However, there is a limit on how much speedup can be achieved.
√
If M is irreducible, then Ĥ ≥ C H. If M and M̂ are reversible, then the limitation is even
stronger, Ĥ ≥ CH.
Consider the undirected and connected graph G = (V, E), with vertex set V and edge set
E. Let z ∈ R|V| with components zi , and consider the quadratic problem
n
o
1 X
min f (z) =
qij (zi − zj )2 .
2
z∈R|V|
(3)
(i,j)∈E
We also write qij = qe for e = (i, j) ∈ E. There is a connection between solving (3)
by Gradient Descent (GD) algorithm and the evolution of a Markov chain. To see this,
2
M̂
M
x
y
x̂
ŷ
x̂
collapse
ẑ
ẑ
ŷ
FIG. 1. The base Markov chain M over the cycle graph C4 and its lifted Markov chain M̂ with
duplicate states. Each state in V (green nodes) is expanded into two states in V̂ (blue and red
nodes above each other).
consider qe = 1 for simplicity. The GD iteration with step-size α > 0 is given by
z t+1 = (I − α∇f ) z t = (I − αDd (I − W )) z t
(4)
where W = Dd−1 A is the transition matrix of a random walk on G, A is the adjacency matrix,
and Dd is the degree matrix, where d = diag(d1 , . . . , d|V| ).
This connection is specially clear for d-regular graphs. Choosing α = 1/d, equation (4)
simplifies to z t+1 = W z t . In particular, the convergence rate of GD is determined by the
spectrum of W , which is connected to the mixing time of M. More precisely, when W is
irreducible and aperiodic, and denoting λ2 (W ) its second largest eigenvalue in absolute value,
the mixing time of the Markov chain and the convergence time of GD are both equal to
H=
C
C
≈
log(1/|λ2 |)
1 − |λ2 |
(5)
where the constant C comes from the tolerance error, which in (2) is 1/4. In the above
approximation we assumed λ2 ≈ 1. Therefore, at least for GD, we can use the theory of
Markov chains to analyze the convergence rate when solving optimization problems. For
this example, and whenever there is linear convergence, the convergence rate τ and the
convergence time H are related by τ H = Θ(1). For an introduction on Markov chains,
mixing times, and transition matrix eigenvalues, we refer the reader to [2].
The main goal of this paper is to extend the above connection to the over-relaxed Alternating Direction Method of Multipliers (ADMM) algorithm, and the concept of lifting will
play an important role. Specifically, for problem (3), we show that a distributed implementation of over-relaxed ADMM can be seen as a lifting of GD, in the same way that M̂ is a
lifting of the Markov chain M. More precisely, there is a matrix MA with stationary vector
3
Lifted
Markov Chain
ADMM
lifting
lifting
GD
Markov Chain
FIG. 2.
GD is the analogue of a Markov chain, while distributed ADMM is the analogue of a
lifted version of this Markov chain, which mixes faster.
vA associated to distributed ADMM, and a matrix MG with stationary vector vG associated
to GD, such that relation (1) is satisfied. This duality is summarized in Fig. 2. In some
cases, MA might have a few negative entries preventing it from being the transition matrix
of a Markov chain. However, it always satisfies all the other properties of Markov matrices.
As explained in the example preceding (5), the convergence time of an algorithm can be
related to the mixing time of a Markov chain. Let HA be the convergence time of ADMM,
and HG the convergence time of GD. The lifting relation between both algorithms strongly
suggest that, for problem (3) and optimally tuned parameters, ADMM is always faster than
GD. Since lifting can speed mixing times of Markov chains up to a square root factor, we
?
?
conjecture that the optimal convergence times HA
and HG
are related as
∗
HA
≤C
p
∗
HG
,
(6)
where C > 0 is a universal constant. Moreover, we conjecture that (6) holds for any
connected graph G. Note that (6) is much stronger than the analogous relation for lifted of
Markov chains, where for some graphs, e.g. with low conductance, the gain is marginal.
The outline of this paper is the following. After mentioning related works in Section II,
we state our main results in Section III, which shows that distributed implementations of
over-relaxed ADMM and GD obey the lifting relation (1). The proofs can be found in the
Appendix. In Section IV we support conjecture (6) with numerical evidence. We present
our final remarks in Section V.
II.
RELATED WORK AND AN OPEN PROBLEM
We state conjecture (6) for the relatively simple problem (3), but, to the best of our
knowledge, it cannot be resolved through the existing literature. We compare the exact
4
asymptotic convergence rates after optimal tuning of ADMM and GD, while the majority
of previous papers focus on upper bounding the global convergence rate of ADMM and, at
best, optimize such an upper bound.
Furthermore, to obtain linear convergence, strong convexity is usually assumed [3], which
does not hold for problem (3). Most results not requiring strong convexity focus on the
convergence rate of the objective function, as opposed to this paper which focus on the
convergence rate of the variables; see [4] for example.
Few papers consider a consensus problem with an objective function different than (3).
P P
For instance, [5] considers f (z) = i∈V kzi − ci k2 , subject to zi = zj if (i, j) ∈ E, where
ci > 0 are constants. This problem is strongly convex and does not reduce to (3), and
P
vice-versa. Other branch of research consider f (z) = i fi (z) with ADMM iterations that
are agnostic to whether or not fi (z) depends on a subset of the components of z; see [6] and
references therein. These are in contrast with our setting where decentralized ADMM is a
message-passing algorithm [7], and the messages between agents i and j are only associated
to the variables shared by functions fi and fj .
For quadratic problems, there are explicit results on the convergence rate and optimal
parameters of ADMM [8–10]. However, their assumptions do not hold for the non strongly
convex distributed problem considered in this paper. Moreover, there are very few results
comparing the optimal convergence rate of ADMM as a function of the optimal convergence
rate of GD. For a centralized setting, an explicit comparison is provided in [11], but it
assumes strong convexity.
Finally, and most importantly, there is no prior result connecting GD and ADMM to lifted
Markov chains. Lifted Markov chains were previously employed to speedup convergence time
of distributed averaging and gossip algorithms [12–14], but these do not involve ADMM
algorithm.
III.
ADMM AS A LIFTING OF GRADIENT DESCENT
We now show that the lifting relation (1) holds when distributed implementations of
over-relaxed ADMM and GD are applied to problem (3) defined over the graph G = (V, E).
ˆ
Let us introduce the extended set of variables x ∈ R|E| , where
Ê = {(e, i) : e ∈ E, i ∈ e, and i ∈ V}.
5
(7)
Note that |Ê| = 2|E|. Each component of x is indexed by a pair (e, i) ∈ Ê. For simplicity
we denote ei = (e, i). We can now write (3) as
n
o
1 X
min f (x) =
qe (xei − xej )2
x,z
2
(8)
e=(i,j)∈E
subject to xei = zi , xej = zj , for all e = (i, j) ∈ E. The new variables are defined according
to the following diagram:
e = (i, j)
zi
zj
1 >
x Qx
2
Notice that we can also write f (x) =
edge e = (i, j), in the form Qe = qe
zi
+1 −1
−1 +1
x ei
Qe
x ej
zj
where Q is block diagonal, one block per
. Let us define the matrix S ∈ R|Ê|×|V| with
components
Sei ,i = Sej ,j =
1 if e = (i, j) ∈ E,
(9)
0 otherwise.
The distributed over-relaxed ADMM is a first order method that operates on five variables: x and z defined above, and also u, m and n introduced below. It depends on the
relaxation parameter γ ∈ (0, 2), and several penalty parameters ρ ∈ R|Ê| . The components
of ρ are ρei > 0 for ei ∈ Ê; see [7, 11] and also [15] for details on multiple ρ’s. We can now
write ADMM iterations as
xt+1 = Ant ,
mt+1 = γxt+1 + ut ,
st+1 = (1 − γ)st + Bmt+1 ,
(10)
ut+1 = ut + γxt+1 + (1 − γ)st − st+1 ,
nt+1 = st+1 − ut+1 ,
where st = Sz t , B = S(S > Dρ S)−1 S > Dρ , and A = (I + Dρ−1 Q)−1 . The next result shows
that these iterations are equivalent to a linear system in |Ê| dimensions. (The proofs of the
following results are in the Appendix.)
Theorem 1 (Linear Evolution of ADMM). Iterations (10) are equivalent to
nt+1 = TA nt ,
TA = I − γ(A + B − 2BA),
with st = Bnt and ut = −(I − B)nt . All the variables of ADMM depend only on nt .
6
(11)
We can also write GD update z t+1 = (I − α∇f )z t to problem (3) as
TG = I − αS > QS.
z t+1 = TG z t ,
(12)
In the following, we establish lifting relations between distributed ADMM and GD in
terms of matrices MA and MG , which are very closely related but not necessarily equal to
TA and TG . They are defined as
MG = (I − DG )−1 (TG − DG ),
(13)
MA = (I − DA )−1 (TA − DA ),
(14)
where DG 6= I and DA 6= I are, for the moment, arbitrary diagonal matrices. Let us also
introduce the vectors
vG = (I − DG )1,
(15)
vA = (I − DA )ρ.
(16)
As shown below, these matrices and vectors satisfy relation (1). Moreover, MG can be
interpreted as a probability transition matrix, and the rows of MA sum up to one. We only
lack the strict non-negativity of MA , which in general is not a probability transition matrix.
Thus, in general, we do not have a lifting between Markov chains, however, we still have
lifting in the sense that MA can be collapsed onto MG according to (1).
Theorem 2. For (DG )ii < 1 and sufficiently small α, MG in (13) is a doubly stochastic
matrix.
Lemma 3. The rows of MG and MA sum up to one, i.e. MG 1 = 1 and MA 1 = 1. Moreover,
>
>
vG
MG = vG
and vA> MA = vA> . These properties are shared with Markov matrices.
Theorem 4 (ADMM as a Lifting of GD). MA and MG defined in (13) and (14) satisfy
relation (1), namely,
vG = S > vA ,
DvG MG = S > DvA MA S,
(17)
provided DG , DA , α, γ, and ρ are related according to
S > Dρ (I − DA )S = I − DG ,
γ qe ρei ,i ρej ,j
,
α=
ρei ,i ρej ,j + qe ρei ,i + ρej ,j
(18)
(19)
for all e = (i, j) ∈ E. Equation (19) restricts the components of ρ, and (17) is an equation
for DA and DG .
7
Theorem 5 (Negative Probabilities). There exists a graph G such that, for any diagonal
matrix DA , ρ and γ, the matrix MA has at least one negative entry. Thus, in general, MA
is not a probability transition matrix.
For concreteness, let us consider some explicit examples illustrating Theorem 4.
Regular Graphs. Let us consider the solution to equations (18) and (19) for d-regular
graphs. Fix qe = 1 and ρ = ρ1 for simplicity. Equation (18) is satisfied with
DA = (1 − (ρ|Ê|)−1 )I,
DG = (1 − |V|−1 )I,
(20)
since d|V| = |Ê| = 2|E|, while (19) requires
α=
γρ
.
2+ρ
(21)
Notice that (DG )ii < 1 for all i, so choosing γ or ρ small enough we can make MG positive.
>
1=
Moreover, the components of (15) and (16) are non-negative and sum up to one, i.e. vG
vA> 1 = 1, thus these vectors are stationary probability distributions of MG and MA .
Cycle Graph. Consider solving (3) over the 4-node cycle graph G = C4 shown in
Fig. 1. By direct computation and upon using (20) we obtain
x̂ 0 0 0 0
0 x̂ ẑ ŷ 0
ŷ ẑ x̂ 0 0
x y 0 y
0 0 0 x̂ ẑ
y x y 0
,
M
=
MG =
A
0 0 ŷ ẑ x̂
0 y x y
0 0 ŷ 0 0
y 0 y x
0 0 0 0 ŷ
ẑ ŷ 0 0 0
0 ŷ ẑ
0 0
0 0
ŷ 0
0 0
x̂ ẑ
ẑ x̂
0
0
0
,
0
ŷ
0
(22)
0 0 x̂
where the probabilities of MG are given by x = 1 − 8α and x + 2y = 1, and the probabilities
of MA are x̂ = 1 − 4γρ, ŷ = 8γρ/(2 + ρ) and x̂ + ŷ + ẑ = 1. The stationary probability
vectors are π = 41 1 and π̂ = 18 1. Now (17) holds provided the parameters are related as
(21). Moreover, in this particular case the matrix MA is strictly non-negative, thus ADMM
is a lifting of GD in the Markov chain sense.
Based on the above theorems we propose conjecture (6). The convergence rate τ is related
to the convergence time, for instance H ∼ (1 − τ )−1 if τ ≈ 1. Thus, let τG? and τA? be the
8
optimal convergence rates of GD and ADMM, respectively. Then, at least for objective (3),
and for any G, we conjecture that there is some universal constant C > 0 such that
1 − τA? ≥ C
IV.
p
1 − τG? .
(23)
NUMERICAL EVIDENCE
For many graphs, we observe very few negative entries in MA , which can be further
reduced by adjusting the parameters ρ and γ. Nonetheless, in general, the lack of strict
non-negativity of MA prevents us from directly applying the theory of lifted Markov chains
to prove (23). However, there is compelling numerical evidence to (23) as we now show.
Consider a sequence of graphs {Gn }, where n = |V|, such that τG? → 1 and τA? → 1 as
n → ∞. Denote RG (n) = (1 − τG? )−1 and RA (n) = (1 − τA? )−1 . We look for the smallest
β such that RA (n) ≤ C RG (n)β , for some C > 0, and all large enough n. If (23) was
false, there would exist sequences {Gn } for which β > 1/2. For instance, if {Gn } have low
conductance it is well-known that lifting does not speedup the mixing time and we could
find β = 1.
To numerically find β, we plot
β̂1 =
log RA (n)
log RG (n)
and β̂2 =
RG (n) ∆RA (n)
RA (n) ∆RG (n)
(24)
against n, where ∆h(n) = h(n + 1) − h(n) for any function h(n). The idea behind this is
very simple. Let f (x) = Cg(x)β , and f, g → ∞ as x → ∞. Then, log f /log g → β and
also ∂x log f /∂x log g = (g ∂x f )/(f ∂x g) → β as x → ∞. Thus, we analyze (24) which are
their discrete analogue. Given a graph Gn , from (12) and (11) we numerically compute τ =
maxj {|λj (T )| : |λj (T )| < 1}. The optimal convergence rates are thus given by τG? = minα τG
and τA? = min{γ,ρ} τA , where we consider ρ = ρ1 with ρ > 0.
In Fig. 3 we show the three different graphs considered in the numerical analysis contained
in the respective plots of Fig. 4. We show the values of (24) versus n, and also the curves
log RA and log RG against n, which for visualization purposes are scaled by the factor −0.03.
In Fig. 4a we see that (6), or equivalently (23), holds for the cycle graph G = Cn . The same
is true for the periodic grid, or torus grid graph G = Tn , as shown in Fig. 4b. Surprisingly,
as shown in Fig. 4c, we get the same square root speedup for a barbell graph, whose random
9
(a)
(b)
(c)
We consider the following graphs for numerical analysis. (a) Cycle graph, G = Cn . (b)
FIG. 3.
Torus or periodic grid graph, G = Tn = C√n × C√n . (c) Barbel graph, obtained by connecting two
complete graphs Kn by a bridge.
0.5
0.5
0.5
β̂2
0.4
β̂1
0.4
−0.03 log RG
0.3
0.2
β̂2
β̂1
0.3
0.3
0.2
−0.03 log RA
0.1
0.0
0.2
500
1000
1500
2000
2500
(a)
3000
3500
−0.03 log RG
−0.03 log RG
0.1
0.1
−0.03 log RA
0.0
0
β̂2
β̂1
0.4
0
2000
4000
6000
(b)
8000
10 000
−0.03 log RA
0.0
0
50
100
150
200
(c)
FIG. 4. Plot of (24) versus n, and also log RA and log RG versus n, scaled by −0.03 for visualization
purposes only. The sequence of graphs {Gn } in each plot are of the types indicated in Fig. 3, in
the same respective order. Notice that β̂1 , β̂2 ≤ 1/2, and β̂1 and βˆ2 gets very close to 1/2 for large
n. (a) Cycle graph. (b) Torus grid graph. The two green curves occur because odd and even n
behave differently. (c) Barbell graph. A Markov chain over this graph does not speedup via lifting,
however, (23) still holds.
walk is known to not speedup via lifting. We find similar behavior for several other graphs
but we omit these results due to the lack of space.
V.
CONCLUSION
For a class of quadratic problems (3) we established a duality between lifted Markov chains
and two important distributed optimization algorithms, GD and over-relaxed ADMM; see
Fig. 2. We proved that precisely the same relation defining lifting of Markov chains (1), is
satisfied between ADMM and GD. This is the content of Theorem 4. Although the lifting
10
relation holds, in general, we cannot guarantee that the matrix MA associated to ADMM
is a probability transition matrix, since it might have a few negative entries. Therefore, in
general, Theorem 4 is not a Markov chain lifting, but it is a lifting in a graph theoretical
sense.
These negative entries actually make this parallel even more interesting since (6), or
equivalently (23), do not violate theorems of lifted Markov chains where the square root
improvement is a lower bound, thus the best possible, and in (6) it is an upper bound.
For graphs with low conductance, the speedup given by Markov chain lifting is negligible.
On the other hand, the lifting between ADMM and GD seems to always give the best
possible speedup achieved by Markov chains, even for graphs with low conductance. This is
numerically confirmed in Fig. 4c.
Due to the strong analogy with lifted Markov chains and numerical evidence, we conjectured the upper bound (6), or (23), which was well supported numerically. However, its
formal explanation remains open.
Finally, although we considered a simple class of quadratic objectives, when close to the
minimum the leading order term of more general convex functions is usually quadratic. In
the cases where the dominant term is close to the form (3), the results presented in this
paper should still hold. An attempt to prove our conjecture (23) is under investigation1 .
ACKNOWLEDGMENT
We would like to thank Andrea Montanari for his guidance and useful discussions. We also
would like to thank the anonymous referees for careful reading of the manuscript and useful
suggestions. This work was partially supported by an NIH/NIAID grant U01AI124302.
Appendix A: Proof of Main Results
In the main part of the paper we introduced the extended edge set Ê which essentially
duplicates the edges of the original graph, |Ê| = 2|E|. This is the shortest route to state our
results concisely but it complicates the notation in the following proofs. Therefore, we first
introduce the notion of a factor graph for problem (3).
1
Note added: soon after the acceptance of this paper, we found a proof of (23) for a class of quadratic
problems. These results will be presented elsewhere.
11
x34
z4
f3
z3
x53
x44
f4
x23
f5
x41
z1
x33
f2
x22
x51
f1
x11
z2
x12
FIG. 5. Example of a factor graph Ḡ for problem (3), and (8), where G is the complete graph K4
with one edge removed.
1.
Factor Graph
The factor graph Ḡ = (F̄, V̄, Ē) for problem (3) is a bipartite graph that summarizes how
different variables are shared across different terms in the objective. This is illustrated in
Fig. 5, where for this case (3) is given by
f (z) = q 1 (z1 − z2 )2 + q 2 (z2 − z3 )2 + q 3 (z3 − z4 )2
(A1)
+ q 4 (z4 − z1 )2 + q 5 (z4 − z2 )2
while (8) is given by
f (x) = q 1 (x12 − x11 )2 + q 2 (x23 − x22 )2
3
2
4
2
5
(A2)
2
+ q (x34 − x33 ) + q (x41 − x44 ) + q (x45 − x52 )
where x11 = x41 , x12 = x22 , x23 = x33 , and x44 = x34 .
The factor graph Ḡ has two sets of vertices, F̄ and V̄. The circles in Fig. 5 represent the
nodes in V̄ = V, and the squares represent the nodes in F̄ = E, where G = (V, E) is the
original graph. Note that each a ∈ F̄ is uniquely associated to one edge e ∈ E, and uniquely
associated to one term in the sum of the objective. In equation (8) we referred to each term
as fe = qe (xei − xej )2 , but now we refer to it by fa . With a slightly abuse of notation we
indiscriminately write a ∈ F̄ or fa ∈ F̄. Each node b ∈ V̄ is uniquely associated to one node
i ∈ V, and uniquely associated to one component of z. Before we referred to this variable
by zi , but now we refer to it by zb , and indiscriminately write b ∈ F̄ or zb ∈ F̄. Each edge
12
(a, b) ∈ Ē must have a ∈ F̄ and b ∈ V̄, and its existence implies that the function fa depends
on variable zb . Moreover, each edge (a, b) ∈ Ē is also uniquely associated to one component
of x in the equivalent formulation (8). In particular, if a ∈ Ē is associated to e ∈ E, and
b ∈ V̄ is associated to i ∈ V, then (a, b) ∈ Ē is associated to xei . Here, we denote xei by xab .
Thus, we can think of Ē as being the same as Ê. Another way of thinking of Ē and x is as
follows. If (a, b) ∈ Ē then xab = zb appears as a constraint in (8).
Let us introduce the neighbor set of a given node in Ḡ. For a ∈ F̄, the independent
variables of fa are in the set
Na = {b ∈ V̄ : (a, b) ∈ Ḡ}.
(A3)
Analogously, for b ∈ V̄, the functions that depend on zb are in the set
Nb = {a ∈ F̄ : (a, b) ∈ Ḡ}.
(A4)
In other words, N• denotes the neighbors of either circle or square nodes in Ḡ. For a ∈ F̄
we define
Ia = {e ∈ Ē : e is incident on a}.
(A5)
Ib = {e ∈ Ē : e is incident on b}.
(A6)
For b ∈ V̄ we define
If we re-write problem (8) using this new notation, which indexes variables by the position
they the take on Ḡ, the objective function takes the form
1X > a
1
x Q xa
f (x) = x> Qx =
2
2 a∈F a
(A7)
where Q ∈ RĒ×Ē is block diagonal and each block, now indexed by a ∈ F̄, takes the form
−1
a
>
a
Qa = q a +1
−1 +1 , where q > 0, and xa = (xab , xac ) for (a, b) ∈ Ē and (a, c) ∈ Ē. Here, q
is the same as qe in the main text. We also have the constraints xab = xa0 b = zb for each
a, a0 ∈ Nb and b ∈ V̄. The row stochastic matrix S introduced in the ADMM iterations is
now expressed as S ∈ R|Ē|×|V̄| and has a single 1 per row such that Seb = 1 if and only if
edge e ∈ Ē is incident on b ∈ V̄. Notice that S > S = Dd is the degree matrix of the original
graph G.
With this notation at hands, we now proceed to the proofs.
13
2.
Proof of Theorem 1
Recall that B = S(S > Dρ S)−1 S > Dρ , thus B 2 = B is a projection operator, and B ⊥ =
I − B its orthogonal complement. Consider updates (10). Substituting xt+1 and mt+1 into
the other variables we obtain
t+1
I 0 0
s
(1 − γ)I B γBA
st
I I 0 ut+1 = (1 − γ)I I γA ut
nt+1
nt
−I I I
0
0 0
which can be easily inverted yielding
st+1 = (1 − γ)st + But + γBAnt ,
(A8)
ut+1 = B ⊥ ut + γB ⊥ Ant ,
(A9)
nt+1 = (1 − γ)st + (B − B ⊥ )ut + γ(B − B ⊥ )Ant .
(A10)
Note the following important relations:
Bnt = st ,
B ⊥ nt = −ut ,
(A11)
Bst = st ,
B ⊥ st = 0,
(A12)
B ⊥ ut = ut ,
But = 0.
(A13)
Equation (A12) is a simple consequence of the definition of B, i.e.
B st = S(S > Dρ S)−1 (S > Dρ S)z t = st ,
(A14)
which also implies B ⊥ st = 0. Since BB ⊥ = 0, acting with B over (A9) implies But = 0 for
every t, and also B ⊥ ut = ut , which shows (A13). Now (A11) follows from these facts and
the own definition nt = st − ut . Finally, applying (A11) on (A10) we obtain nt+1 = TA nt
where TA = I − γ(A + B − 2BA).
3.
Proof of Theorem 2
Write Q = Q+ + Q− where Q+ is diagonal and has only positive entries, and Q− only
has off-diagonal and negative entries. First, notice that (S > Q+ S) is also diagonal. Indeed,
14
for b ∈ V̄ and c ∈ V̄, (S > Q+ S)bc =
P
e∈Ē
Seb Q+
ee Sec = δbc
P
e∈Ib
Q+
ee where δ is the Kronecker
delta. By a similar argument, S > Q− S is off-diagonal. Hence, if b 6= c we have
(TG )bc = −α
XX
e∈Ib e0 ∈Ic
Q−
ee0 ≥ 0.
(A15)
Recall that MG = (I − DG )−1 (TG − DG ), where DG 6= I is diagonal. For MG to be nonnegative we first impose that (DG )bb < 1 for all b ∈ V̄. Then, since the off-diagonal elements
of TG are automatically positive by (A15), we just need to consider the diagonal elements
of TG − DG . Thus we require that for every b ∈ V̄,
1−α
Denoting Qmax = maxb∈V̄
P
e∈Ib
X
e∈Ib
Qee + (DG )bb ≥ 0.
(A16)
Qee and DG,min the smallest element of DG , the matrix MG
will be non-negative provided α ≤ (1 + DG,min )/Qmax .
Notice that S1|V̄| = 1|Ē| and Q1 = 0. Thus S > QS1 = 0, implying TG 1 = 1, and
1> TG = 1> . From this we have MG 1 = 1 and 1> MG = 1> , so all the rows and columns of
MG sum up to one.
4.
Proof of Lemma 3
We proved above that MG is a doubly stochastic matrix. Now let us consider MA . Recall
the definition of B = S > (S > Dρ S)−1 S > Dρ . Note that the action of B on a vector v ∈ R|Ē|
is to take a weighted average of its components, namely, if (a, b) ∈ Ē then
P
c∈N ρcb vcb
(Bv)ab = P b
.
c∈Nb ρcb
(A17)
Therefore, B1 = 1. Recall that Q1 = 0, thus A1 = 1, where A = (I + Dρ−1 Q)−1 , which
implies TA 1 = 1, and in turn MA 1 = 1. Now the other relations follow trivially.
5.
Proof of Theorem 4
Due to the block diagonal structure of Q it is possible write A explicitly as
A = I − F Q,
15
(A18)
where F is a block diagonal matrix with |F̄| blocks. Each block F a , for a ∈ F̄, is of the
form
a
Fa =
ρ
0
q
ac
,
a
ρab ρac + q (ρab + ρac )
0 ρab
(A19)
where b, c ∈ Na . Now by the definition of B we have S > Dρ B = S > Dρ . Hence,
S > DvA MA S = S > Dρ (I − DA )S − γS > Dρ F QS,
DvG MG = (I − DG ) − αS > QS.
(A20)
(A21)
Equating the first term of (A20) to the first term of (A21), and also the second terms to
each other, on using (A18) we obtain
S > Dρ (I − DA )S = I − DG ,
(A22)
α=
(A23)
γ q a ρab ρac
,
ρab ρac + q a (ρab + ρbc )
where (A23) must hold for all a ∈ F̄ and b, c ∈ Na . This gives the second equality in (17)
together with relations (18) and (19). Finally, since diagonal matrices commute, S > vA =
S > (I − DA )Dρ S1|V̄| = (I − DG )1|V̄| = vG , which gives the first relation in (17).
6.
Proof of Theorem 5
It suffices to show one example with at least one negative entry. Let G be the complete
graph K4 with one edge removed, as shown in Fig. 5. By direct inspection one finds the
following sub-matrix of TA :
T (S) =
(TA )21 (TA )24
(TA )31 (TA )34
(A24)
whose elements are explicitly given by
(TA )21 =
(TA )24 =
(TA )31 =
γ ρ11 (ρ12 −ρ22 )
,
(ρ12 +ρ22 )(ρ11 +ρ12 +ρ11 ρ12 )
2γ
−1
(ρ12 +ρ22 ) 1+ρ−1
22 +ρ23
2γ
−1
(ρ12 +ρ22 ) 1+ρ−1
11 +ρ12
,
(A26)
,
(A27)
23 (ρ12 −ρ22 )
.
(TA )34 = − (ρ12 +ργ22ρ)(ρ
22 +ρ23 +ρ22 ρ23 )
16
(A25)
(A28)
First notice that subtracting DA from TA does not affect T (S) . Now recall that all components
of ρ must be strictly positive. The elements (A25) and (A28) have opposite signs, so one
of them is negative. Since (A26) and (A27) are both positive, one cannot remove the
negative entries of an entire row of TA by multiplying TA by the diagonal matrix (I − DA )−1 .
Therefore, MA = (I − DA )−1 (TA − DA ) has at least one negative entry.
[1] F. Chen, L. Lovász, and L. Pak. Lifting Markov Chains to Speed up Mixing. In Proceedings
of the thirty-first annual ACM symposium on Theory of computing, pages 275–281, 1999.
[2] J. Norris. Markov Chains. Cambridge University Press, Cambridge, 1998.
[3] W. Shi, Q. Ling, K. Yuan, G. Wu, and W. Yin. On the Linear Convergence of the ADMM in
Decentralized Consensus Optimization. IEEE Transactions on Signal Processing, 62(7):1750–
1761, 2014.
[4] D. Davis and W. Yin. Convergence Rate Analysis of Several Splitting Schemes. arXiv preprint
arXiv:1406.4834, 2014.
[5] T. Erseghe, D. Zennaro, E. Dall’Anese, and L. Vangelista. Fast Consensus by the Alternating
Direction Multipliers Method. IEEE Transactions on Signal Processing, 59(11):5523–5537,
2011.
[6] E. Wei and A. Ozdaglar. Distributed Alternating Direction Method of Multipliers. In 2012
IEEE 51st IEEE Conference on Decision and Control (CDC), pages 5445–5450. IEEE, 2012.
[7] N. Derbinsky, J. Bento, V. Elser, and J. Yedidia. An Improved Three-Weight Message Passing
Algorithm. arXiv:1305.1961v1 [cs.AI], 2013.
[8] André Teixeira, Euhanna Ghadimi, Iman Shames, Henrik Sandberg, and Mikael Johansson.
Optimal Scaling of the ADMM Algorithm for Distributed Quadratic Programming. In 52nd
IEEE Conference on Decision and Control, pages 6868–6873. IEEE, 2013.
[9] E. Ghadimi, A. Teixeira, I. Shames, and M. Johansson. Optimal Parameter Selection for the
Alternating Direction Method of Multipliers (ADMM): Quadratic Problems. IEEE Transactions on Automatic Control, 60(3):644–658, 2015.
[10] F. Iutzeler, P. Bianchi, P. Ciblat, and W. Hachem. Explicit Convergence Rate of a Distributed Alternating Direction Method of Multipliers. IEEE Transactions on Automatic Control, 61(4):892–904, 2016.
17
[11] G. França and J. Bento. An Explicit Rate Bound for Over-Relaxed ADMM. In IEEE International Symposium on Information Theory, ISIT 2016, Barcelona, Spain, July 10-15, 2016,
pages 2104–2108, 2016.
[12] K. Jung, D. Shah, and J. Shin. Fast Gossip Through Lifted Markov Chains. In Proc. Allerton
Conf. on Comm., Control, and Computing, Urbana-Champaign, IL, 2007.
[13] W. Li, H. Dai, and Y. Zhang. Location-Aided Fast Distributed Consensus in Wireless Networks. IEEE Transactions on Information Theory, 56(12):6208–6227, 2010.
[14] K. Jung, D. Shah, and J. Shin. Distributed averaging via lifted markov chains. IEEE Transactions on Information Theory, 56(1):634–647, 2010.
[15] J. Bento, N. Derbinsky, J. Alonso-Mora, and J. Yedidia. A Message-Passing Algorithm for
Multi-Agent Trajectory Planning. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani,
and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, pages
521–529. Curran Associates, Inc., 2013.
18
| 8cs.DS
|
arXiv:1711.02233v1 [cs.IT] 7 Nov 2017
RUDIN-SHAPIRO-LIKE POLYNOMIALS WITH MAXIMUM
ASYMPTOTIC MERIT FACTOR
DANIEL J. KATZ AND STANISLAV A. TRUNOV
Abstract. Borwein and Mossinghoff investigated the Rudin-Shapirolike polynomials, which are infinite families of Littlewood polynomials,
that is, polynomials whose coefficients are all in {−1, 1}. Each family of
Rudin-Shapiro-like polynomials is obtained from a starting polynomial
(which we call the seed) by a recursive construction. These polynomials can be regarded as binary sequences. Borwein and Mossinghoff
show that the asymptotic autocorrelation merit factor for any such family is at most 3, and found the seeds of length 40 or less that produce
the maximum asymptotic merit factor of 3. The definition of RudinShapiro-like polynomials was generalized by Katz, Lee, and Trunov to
include polynomials with arbitrary complex coefficients, with the sole
condition that the seed polynomial must have a nonzero constant coefficient. They proved that the maximum asymptotic merit factor is 3 for
this larger class. Here we show that a family of such Rudin-Shapiro-like
polynomials achieves asymptotic merit factor 3 if and only if the seed is
the interleaving of a pair of Golay complementary sequences.
In this paper, we assume that all the definitions from the introduction
and from Section 2 (up to Lemma 2.2) of the paper of Katz, Lee, and
Trunov [3] are in force, so one should read that first. Recall that a sequence
(f0 , f1 , . . . , fℓ−1 ) ∈ Cℓ of length ℓ is identified with the polynomial f (z) =
f0 + f1 z + · · · + fℓ−1 z ℓ−1 ∈ C[x].
Recall from the introduction of [3] that if f is a sequence, then ADF(f )
is the autocorrelation demerit factor of f , which is the sum of the squared
magnitudes of the autocorrelation values of f at all nonzero shifts divided
by the squared length of f . The autocorrelation merit factor of f is the
reciprocal of its demerit factor, and our goal is to obtain sequences with
high merit factor, or equivalently, low demerit factor, which indicates low
mean square magnitude autocorrelation.
Recall from [3, Theorem 1.2] that if (f0 , f1 , . . .) is a sequence of RudinShapiro-like polynomials generated from the seed f0 ∈ C[z] with nonzero
constant coefficient, then
(1)
lim ADF(fn ) = −1 +
n→∞
2 kf0 k44 + kf0 fe0 k22
1
·
≥ .
4
3
3
kf0 k2
Date: 06 November 2017.
This paper is based on work of the two authors supported in part by the National
Science Foundation under Grant DMS 1500856.
1
2
DANIEL J. KATZ AND STANISLAV A. TRUNOV
This is a generalization of the result of Borwein-Mossinghoff [1, Corollary 1],
who prove the special case when the Rudin-Shapiro polynomials are Littlewood polynomials, that is, polynomials whose coefficients are all in {−1, 1}.
We are interested when the inequality becomes an equality, that is, when the
asymptotic autocorrelation demerit factor is precisely 1/3, or equivalently,
when the asymptotic autocorrelation merit factor is precisely 3. To give a
criterion for this, we need to explain two more concepts: interleaving and
Golay complementary pairs.
If f (z), g(z) ∈ C[z] are a pair of polynomials, then the interleaving of f
and g is f (z 2 )+zg(z 2 ). If f (z) and g(z) both represent sequences of length ℓ,
then their interleaving represents the sequence (f0 , g0 , f1 , g1 , . . . , fℓ−1 , gℓ−1 )
of length 2ℓ.
We say that a pair of polynomials f (z), g(z) ∈ C[z] with nonzero constant
coefficients is a Golay complementary pair when f (z)f (z) + g(z)g(z) is a
constant. These were first used by Golay in [2]. In terms of autocorrelation,
this means that Cf,f (s) + Cg,g (s) = 0 for every nonzero shift s. In our
notation from Section 2 of [3], the condition for (f, g) to be a Golay pair is
that |f (z)|2 + |g(z)|2 be a constant polynomial, which means that |f (z)|2 +
|g(z)|2 is constant for z on the complex unit circle. Note that if f and g
have different degrees, then (f, g) cannot be a Golay pair: the polynomial
of higher degree d will have a nonzero correlation value at shift d, while the
shorter one will have a zero correlation value at that shift, and so the sum
of these correlations cannot be zero.
Now we can state our theorem.
Theorem 1. Let (f0 , f1 , . . .) be a sequence of Rudin-Shapiro polynomials
generated from seed f0 ∈ C[z]. Then limn→∞ ADF(fn ) = 1/3 if and only
either f0 is constant or is the interleaving of a Golay complementary pair.
Proof. By (1), we
autocorrelation demerit factor is 1/3
see that the limiting
4
2
if and only if kf0 k4 + kf0 fe0 k2 /kf0 k42 = 2. If f0 is a constant, this is
clearly true, so we shall assume that f0 is nonconstant henceforth. Recall
that since the definitions of the introduction of [3] are in force, f0 has a
nonzero constant coefficient. We select polynomials g(z), h(z) ∈ C[z] such
that f is the interleaving of g and h, that is, f (z) = g(z 2 ) + zh(z 2 ).
For any a(z) ∈ C[z], we note that kak22 is the sum of the squared magnitudes of the coefficients of a(z). Therefore, since the coefficients of f0 (z)
are just the coefficients of g and h interleaved, we have
kf0 k22 = kgk22 + khk22 .
(2)
Also
kf0 k44
+ kf0 fe0 k22 =
Z
|f0 (z)|2 |g(z 2 ) + zh(z 2 )|2 + |g(z 2 ) − zh(z 2 )|2
Z
= 2 |f0 (z)|2 |g(z 2 )|2 + |zh(z 2 )|2 .
RUDIN-SHAPIRO-LIKE POLYNOMIALS
3
Now note that |zh(z 2 )|2 = |h(z 2 )|2 since |z| = 1 on the complex unit circle.
We set a(z) = |g(z)|2 + |h(z)|2 , and so
Z
4
2
e
kf0 k4 + kf0 f0 k2 = 2 |g(z 2 ) + zh(z 2 )|2 a(z 2 )
Z
a(z 2 ) + g(z 2 )z −1 h(z 2 ) + g(z 2 )zh(z 2 ) a(z 2 ),
=2
R
and since z j = 0 when j 6= 0, we can drop the terms that have odd degree
to obtain
Z
4
2
e
kf0 k4 + kf0 f0 k2 = 2 a(z 2 )2
Z
(3)
= 2 a(z)2
= 2k|g(z)|2 + |h(z)|2 k22 ,
R
R
where we have used the fact that z 2j = z j for every j ∈ Z in the second
equality.
Note that kgk22 + khk22 is the constant term of |g(z)|2 + |h(z)|2 , and since
k|g(z)|2 + |h(z)|2 k22 is the sum of the squared magnitudes of all the coefficients of |g(z)|2 + |h(z)|2 , this means that
2
k|g(z)|2 + |h(z)|2 k22 ≥ kgk22 + khk22 ,
with equality if and only if |g(z)|2 + |h(z)|2 is a constant, that is, if and only
if (g, h) is a Golay pair. We use our equalities (2) and (3) above to express
the left and right hand sides in terms of f0 , and see that
2
1
kf0 k44 + kf0 fe0 k22 ≥ kf0 k22 ,
2
with equality if and only if (g, h) is a Golay pair. Or equivalently (since
kf0 k22 6= 0 because it is the sum of the squared magnitudes of the coefficients
of f0 , which is nonzero),
kf0 k44 + kf0 fe0 k22
≥ 2,
kf0 k42
with equality if and only if (g, h) is a Golay pair. So by (1), our nonconstant
seed f0 gives rise to a sequence (f0 , f1 , . . .) of Rudin-Shapiro-like polynomials
with asymptotic autocorrelation demerit factor 1/3 if and only if it is an
interleaving of a Golay pair.
This tells us something about the possible lengths of optimal seeds for
binary Rudin-Shapiro-like sequences.
Corollary 2. There exists a Littlewood polynomial f0 (z) of length ℓ giving
rise to a sequence (f0 , f1 , . . .) of Rudin-Shapiro-like polynomials with asymptotic autocorrelation merit factor 3 if and only if either ℓ = 1 or else ℓ = 2m
for an m such that there exists a Golay complementary pair of binary sequences, each of length m.
4
DANIEL J. KATZ AND STANISLAV A. TRUNOV
In particular, there are Littlewood seeds of length ℓ = 1 and ℓ = 2a 10b 26c
for every a ≥ 1 and every b, c ≥ 0.
The last statement is a consequence of the Corollary to Lemma 5 in [4],
which gives values of m such that there are known to be Golay pairs of
binary sequences of length m.
References
[1] P. Borwein and M. Mossinghoff. Rudin-Shapiro-like polynomials in L4 . Math. Comp.,
69(231):1157–1166, 2000.
[2] M. J. E. Golay. Static multislit spectrometry and its application to the panoramic
display of infrared spectra. J. Opt. Soc. Am., 41(7):468–472, 1951.
[3] D. J. Katz, S. Lee, and S. A. Trunov. Crosscorrelation of Rudin-Shapiro-like polynomials. arXiv:1702.07697, 2017.
[4] R. J. Turyn. Hadamard matrices, Baumert-Hall units, four-symbol sequences, pulse
compression, and surface wave encodings. J. Combinatorial Theory Ser. A, 16:313–333,
1974.
Department of Mathematics, California State University, Northridge,
United States
| 7cs.IT
|
Published as a conference paper at ICLR 2015
T ECHNIQUES FOR L EARNING B INARY
S TOCHASTIC F EEDFORWARD N EURAL N ETWORKS
arXiv:1406.2989v3 [stat.ML] 9 Apr 2015
Tapani Raiko & Mathias Berglund
Department of Information and Computer Science
Aalto University
Espoo, Finland
{tapani.raiko,mathias.berglund}@aalto.fi
Guillaume Alain & Laurent Dinh
Department of Computer Science and Operations Research
Université de Montréal
Montréal, Canada
[email protected], [email protected]
A BSTRACT
Stochastic binary hidden units in a multi-layer perceptron (MLP) network give at
least three potential benefits when compared to deterministic MLP networks. (1)
They allow to learn one-to-many type of mappings. (2) They can be used in structured prediction problems, where modeling the internal structure of the output is
important. (3) Stochasticity has been shown to be an excellent regularizer, which
makes generalization performance potentially better in general. However, training stochastic networks is considerably more difficult. We study training using M
samples of hidden activations per input. We show that the case M = 1 leads to
a fundamentally different behavior where the network tries to avoid stochasticity.
We propose two new estimators for the training gradient and propose benchmark
tests for comparing training algorithms. Our experiments confirm that training
stochastic networks is difficult and show that the proposed two estimators perform
favorably among all the five known estimators.
1
I NTRODUCTION
Feedforward neural networks, or multi-layer perceptron (MLP) networks, model mappings from
inputs x to outputs y through hidden units h. Typically the network output defines a simple (unimodal) distribution such as an isotropic Gaussian or a fully factorial Bernoulli distribution. In case
the hidden units are deterministic (using a function x → h as opposed to a distribution P (h|x)), the
conditionals P (y|x) belong to the same family of simple distributions.
Stochastic feedforward neural networks (SFNN) (Neal, 1990; 1992) have the advantage when the
conditionals P (y|x) are more complicated. While each configuration of hidden units h produces
a simple output, the mixture over them can approximate any distribution, including multimodal
distributions required for one-to-many type of mappings. In the extreme case of using empty vectors
as the input x, they can be used for unsupervised learning of the outputs y.
Another potential advantage of stochastic networks is in generalization performance. Adding noise
or stochasticity to the inputs of a deterministic neural network has been found useful as a regularization method (Sietsma & Dow, 1991). Introducing multiplicative binary noise to the hidden units
(dropout, Hinton et al., 2012) regularizes even better.
Binary units have additional advantages in certain settings. For instance conditional computations
require hard decisions (Bengio et al., 2013). In addition, some harwdare solutions are restricted to
binary outputs (e.g. the IBM SyNAPSE, Esser et al., 2013).
The early work on SFNNs approached the inference of h using Gibbs sampling (Neal, 1990; 1992) or
mean field (Saul et al., 1996), which both have their downsides. Gibbs sampling can mix poorly and
1
Published as a conference paper at ICLR 2015
the mean-field approximation can both be inefficient and optimize a lower bound on the likelihood
that may be too loose. More recent work proposes simply drawing samples from P (h|x) during
the feedforward phase (Hinton et al., 2012; Bengio et al., 2013; Tang & Salakhutdinov, 2013). This
guarantees independent samples and an unbiased estimate of P (y|x).
We can use standard back-propagation when using stochastic continuous-valued units (e.g. with
additive noise or dropout), but back-propagation is no longer possible with discrete units. There
are several ways of estimating the gradient in that case. Bengio et al. (2013) proposes two such
estimators: an unbiased estimator with a large variance, and a biased version that approximates
back-propagation.
Tang & Salakhutdinov (2013) propose an unbiased estimator of a lower bound that works reasonably
well in a hybrid network containing both deterministic and stochastic units. Their approach relies on
using more than one sample from P (y|x) for each training example, and in this paper we provide
theory to show that using more than one sample is an important requirement. They also demonstrate
interesting applications such as mapping the face of a person into varying expressions, or mapping
a silhouette of an object into a color image of the object.
Tang & Salakhutdinov (2013) argue for the choice of a hybrid network structure based on the finite
(and thus limited) number of hidden configurations in a fully discrete h. However, we offer an
alternate hypothesis: It is much easier to learn a deterministic network around a small number of
stochastic units, so that it might not even be important to train the stochastic units properly. In an
extreme case, the stochastic units are not trained at all, and the deterministic units do all the work.
In this work, we take a step back and study more rigorously the training problem with fully stochastic
networks. We compare different methods for estimating the gradient and propose two new estimators. One is an approximate back-propagation with less bias than the one by Bengio et al. (2013),
and the other is a modification of the estimator by Tang & Salakhutdinov (2013) with less variance.
We propose a benchmark test setting based on the well-known MNIST data and the Toronto Face
Database.
2
S TOCHASTIC FEEDFORWARD NEURAL NETWORKS
We study a model that maps inputs x to outputs y through stochastic binary hidden units h. The
equations are given for just one hidden layer, but the extension to multiple layers is easy1 . The activation probability is computed just like the activation function in deterministic multilayer perceptron
(MLP) networks:
P (hi = 1 | x) = σ(ai ) = σ(Wi: x + bi ),
(1)
where Wi: denotes the ith row vector of matrix W and σ(·) is the sigmoid function. For classification problems, we use softmax for the output probability
exp(Vi: h + ci )
P (y = i | h) = P
.
(2)
j exp(Vj: h + cj )
For predicting binary vectors y, we use a product of Bernoulli distributions
P (yi = 1 | h) = σ(Vi: h + ci ).
(3)
The probabilistic training criterion for deterministic MLP networks is log P (y|x). Its gradient with
respect to model parameters θ = {W, V, b, c} can be computed using the back-propagation algorithm, which is based on the chain rule of derivatives. Stochasticity brings difficulties in both
estimating the training criterion and in estimating the gradient. The training criterion of the stochastic network
X
X
C = log P (y | x) = log
P (y, h | x) = log
P (y | h)P (h | x)
(4)
h
h
= log EP (h|x) P (y | h)
(5)
requires summation over an exponential number of configurations of h. Also, derivatives with
respect to discrete variables cannot be directly defined. We will review and propose solutions to
both problems below.
1
Apply P (h|x) separately for each layer such that x denotes the layer below instead of the original input.
2
Published as a conference paper at ICLR 2015
2.1
P ROPOSED ESTIMATOR OF THE TRAINING CRITERION
We propose to estimate the training criterion in Equation (4) by
M
1 X
CˆM = log
P (y | h(m) )
M m=1
h(m) ∼ P (h | x).
(6)
(7)
This can be interpreted as the performance of a finite mixture model over M samples drawn from
P (h|x).
One could hope that using just M = 1 sample just like in many other stochastic networks (e.g. Hinton et al., 2012) would work well enough. However, here we show in that case the network always
prefers to minimize the stochasticity, for instance by increasing the input weights to a stochastic
sigmoid unit such that it behaves as a deterministic step-function nonlinearity.
Theorem 1. When maximizing the expectation of Cˆ1 in Equation (6) using M = 1, a hidden unit
hi never prefers a stochastic output over a deterministic one. However, when maximizing the expectation of C in Equation (4), the hidden unit hi may prefer a stochastic output over any of the
deterministic ones.
Proof. The expected Cˆ1 over the data distribution can be upper-bounded as
h i
EPd (x,y) EP (h|x) Cˆ1 = EPd (x,y) EP (h|x) log P (y | h)
h
i
= EPd (x) EP (hi |x) EPd (y|x) EP (h\i |x,hi ) log P (y | h)
(8)
(9)
= EPd (x) EP (hi |x) f (hi , x)
(10)
≤ EPd (x) max f (hi , x),
(11)
hi
where Pd denotes the data distribution. The value in the last inequality is achievable by selecting the
distribution of P (hi |x) to be a Dirac delta around the value hi which maximizes the deterministic
function f (hi , x). This can be done for every x under the expectation in an independent way.
This is analogous to a idea from game theory: since the performance achieved with P (hi |x) is a
linear combination of the performances f (hi , x) of the deterministic choices hi , any mixed strategy
P (hi |x) cannot be better than the best deterministic choice.
Let us now look at the situation for the expectation C and see how it differs from the case of C1
that we had with one particle. We can see that the original training criterion can be written as the
expectation of a KL-divergence.
EPd (x,y) [C] = EPd (x) EPd (y|x) log P (y | x)
= EPd (x) [−KL (Pd (y | x)kP (y | x)) + const]
(12)
(13)
The fact that this expression features a negative KL-divergence means that the maximum is achieved
when the conditionals match exactly. That is, it it maximized when we have that P (y|x) = Pd (y|x)
for each value of x.
We give a simple example in which (x, h, y) each take values in {0, 1}. We define the following
conditions on Pd (y|x) and P (y|h), and we show how any deterministic P (h|x) is doing a bad job
at maximizing (12).
0.5 0.5
0.9 0.1
Pd (y | x) =
, P (y | h) =
0.5 0.5
0.1 0.9
a
b
A deterministic P (h | x) =
is one in which a, b take values in {0, 1}.
1−a 1−b
Criterion (12) is maximized by (a, b) = (0.5, 0.5), regardless of the distribution Pd (x). For the
purposes of comparing solutions, we can simply take Pd (x) = [0.5 0.5]. In that case, we get that
the expected C takes the value 0.5 log(0.5) + 0.5 log(0.5) ≈ −0.30. On the other hand, all the
deterministic solutions yield a lower value 0.5 log(0.9) + 0.5 log(0.1) ≈ −0.52.
3
Published as a conference paper at ICLR 2015
2.2
G RADIENT FOR TRAINING P (y|h)
We will be exploring five different estimators for the gradient of a training criterion wrt. parameters
θ. However, all of them will share the following gradient for training P (y|h).
Let o = Vh + c be the incoming signal to the activation function φ(·) in the final output layer. For
training P (y|h), we compute the gradient of the training criterion CˆM in Equation (6)
M
M
1 X
1 X
CˆM = log
P (y | h(m) ) = log
φ(o(m) )
M m=1
M m=1
G(o(m) ) :=
φ0 (o(m) )
φ(o(m) )
∂ log φ(o(m) )
∂ CˆM
=
=
P
P
M
M
(m0 ) )
(m0 ) )
∂o(m)
∂o(m)
m0 =1 φ(o
m0 =1 φ(o
P (y | h(m) )
∂ log φ(o(m) )
∂ log φ(o(m) )
w(m)
= PM
= PM
,
0
0
(m)
(m ) )
(m )
∂o
∂o(m)
m0=1 P (y | h
m0 =1 w
(14)
(15)
(16)
where w(m) = P (y|h(m) ) are unnormalized weights. In other words, we get the gradient in the
mixture by computing the gradient of the individual contribution m and multiplying it with normalPM
0
ized weights w̄(m) = w(m) / m0 =1 w(m ) . The normalized weights w̄(m) can be interpreted as
responsibilities in a mixture model (see e.g. Bishop et al., 2006, Section 2.3.9).
The gradients G(V), G(h), and G(c) are computed from G(o) using the chain rule of derivatives
just like in standard back-propagation.
2.3
F IRST ESTIMATORS OF THE GRADIENT FOR TRAINING P (h|x)
Bengio et al. (2013) proposed two estimators of the gradient for P (h|x). The first one is unbiased
but has high variance. It is defined as
G1 (ai ) := (hi − σ(ai ))(L − L̄i )
E (hi − σ(ai ))2 L
,
L̄i =
E [(hi − σ(ai ))2 ]
(17)
(18)
where we plug in L = CˆM as the training criterion. We estimate the numerator and the denominator
of L̄i with an exponential moving average.
The second estimator is biased but has lower variance. It is based on back-propagation where we set
∂hi
∂ai := 1 resulting in
G2 (ai ) := G(hi ) = (V:i )T G(o).
2.4
(19)
P ROPOSED BIASED ESTIMATOR OF THE GRADIENT FOR TRAINING P (h|x)
We propose a new way of propagating the gradient of the training criterion through discrete hidden
units hi . Let us consider hi continuous random variables with additive noise i
hi = σ(ai ) + i
1 − σ(ai )
i ∼
−σ(ai )
(20)
with probability σ(ai )
with probability 1 − σ(ai )
(21)
Note that hi has the same distribution as in Equation (1), that is, it only gets values 0 and 1. With
this formulation, we propose to back-propagate derivatives through hi by
G3 (ai ) := σ 0 (ai )G(hi ) = σ 0 (ai )(V:i )T G(o)
(22)
This gives us a biased estimate of the gradient since we ignore the fact that the structure of the noise
i depends on the input signal ai . One should note, however, that the noise is zero-mean with any
input ai , which should help keep the bias relatively small.
4
Published as a conference paper at ICLR 2015
2.5
VARIATIONAL TRAINING
Tang & Salakhutdinov (2013) use a variational lower bound L(Q) on the training criterion C as
X
P (y, h | x) X
P (y, h | x)
C = log P (y | x) =
P (h | y, x) log
≥
Q(h) log
=: L(Q).
P (h | y, x)
Q(h)
h
h
(23)
The above inequality holds for any distribution Q(h), but we get more usefulness out of it by choosing Q(h) so that it serves as a good approximation of P (h | y, x).
We start by noting that we can use importance sampling to express P (h | y, x) in terms of a proposal
distribution R(h|y, x) from which we can draw samples.
P (h | y, x) ∝ P (y | h)P (h | x) =
P (y | h)P (h | x)
R(h | y, x)
R(h | y, x)
(24)
Let δ(h) be the Dirac delta function centered at h. We construct Q(h) based on this expansion:
Q(h) =
M
X
w̄(m) δ(h(m) )
(25)
m=1
h(m) ∼ R(h | y, x)
(m)
w(m) =
(26)
(m)
P (y | h )P (h
| x)
R(h(m) | y, x)
w(m)
w̄(m) = PM
(m0 )
m0 =1 w
(27)
(28)
where w(m) and w̄(m) are called the unnormalized and normalized important weights.
It would be an interesting line of research to train an auxiliary model for the proposal distribution
R(h|x, y) following ideas from Kingma & Welling (2013); Rezende et al. (2014); Mnih & Gregor
(2014) that call the equivalent of R the recognition model or the inference network. However, we
do not pursue that line further in this paper and follow Tang & Salakhutdinov (2013) who chose
R(h|x, y) := P (h|x), in which case the importance weights simplify to w(m) = P (y|h(m) ).
Tang & Salakhutdinov (2013) use a generalized EM algorithm, where they compute the gradient for
the lower bound L(Q) given that Q(h) is fixed
∂ X
P (y, h | x)
G4 (θ) :=
Q(h) log
(29)
∂θ
Q(h)
h
=
∂
∂θ
M
X
h
i
w̄(m) log P (y | h(m) ) + log P (h(m) | x) .
(30)
m=1
Thus, we train P (h(m) |x) using h(m) as target outputs.
It turns out that the resulting gradient for P (y|h) is exactly the same as in Section 2.2, despite the
rather different way of obtaining it. The importance weights w̄(m) have the same role as responsibilities w̄(m) in the mixture model, so we can use the same notation for them.
Proposed Unbiased Estimator of the Gradient We propose a new gradient estimator by applying
a variance reduction technique (Weaver & Tao, 2001; Mnih & Gregor, 2014) to the estimator by
Tang & Salakhutdinov (2013). First we note that
Z
Z
∂
∂
∂
∂
∂θ P (h | x)
EP (h|x)
log P (h | x) = P (h | x)
dh =
1 = 0.
P (h | x)dh =
∂θ
P (h | x)
∂θ
∂θ
(31)
That is, when training P (h|x) with samples h(m) ∼ P (h|x) drawn from the model distribution, the
gradient is on average zero. Therefore we can change the estimator of P (h|x) by subtracting any
5
30
−50
25
−60
20
−70
ĈM
norm of gradient
Published as a conference paper at ICLR 2015
15
−80
10
−90
5
0
−0.5
0
0.5
cM
1
−100
1.5
0
20
40
60
Number of particles M
80
100
Figure 1: Left: The norm of the gradient for the weights of the first hidden layer as a function of cM
1
where the proposed c = M
in Equation (32) corresponds to cM = 1. The norm is averaged over
a mini-batch, after {1,7,50} epochs of training (curves from top to bottom) with G5 in the MNIST
classification experiment (see Appendix A). Varying c only changes the variance of the estimator, so
the minimum norm corresponds to the minimum variance. Right: CˆM as a function of the number
of particles used during test time for the MNIST structured prediction task for the two proposed
models trained with M = 20.
1
constant c from the weights w̄(m) without introducing any bias. We choose c = E w̄(m) = M
which is empirically shown to be sufficiently close to the optimum (see Figure 1 (left)). Finally, the
proposed estimator becomes
M
∂ X
1
(m)
(m)
(m)
(m)
G5 (θ) :=
log P (h
| x) .
w̄
log P (y | h ) + w̄
−
∂θ m=1
M
3
(32)
E XPERIMENTS
We propose two experiments as benchmarks for stochastic feedforward networks based on the
MNIST handwritten digit dataset (LeCun et al., 1998) and the Toronto Face Database (Susskind
et al., 2010). In both experiments, the output distribution is likely to be complex and multimodal.
In the first experiment, we predicted the lower half of the MNIST digits using the upper half as inputs. The MNIST dataset used in the experiments was binarized as a preprocessing step by sampling
each pixel independently using the grey-scale value as its expectation. In the second experiment, we
followed Tang & Salakhutdinov (2013) and predicted different facial expressions in the Toronto
Face Database (Susskind et al., 2010). As data, we used all individuals with at least 10 different facial expression pictures, which we do not binarize. We set the input to the mean of these images per
subject, and as output predicted the distribution of the different expressions of the same subject2 . We
randomly chose 100 subjects for the training data (1372 images) and the remaining 31 subjects for
the test data (427 images). As the data in the second problem is continuous, we assumed unit variance Gaussian noise and thus trained the network using the sum of squares error. We used a network
structure of 392-200-200-392 and 2304-200-200-2304 in the first and second problem, respectively.
Before running the experiments, we did a simple viability check of the gradient estimators by training a network to do MNIST classification. Based on the results, we kept G3 to G5 that performed
significantly better than G1 and G2 . The results of the viability experiment can be found in Appendix A.
For comparison, we also trained four additional networks (labeled A-D) in addition to the stochastic
feedforward networks. Network A is a deterministic network (corresponding to G3 with i = 0
in Equation (21)). In network B, we used the weights trained to produce deterministic values for
the hidden units, but instead of using these deterministic values at test time we use their stochastic
equivalent. We therefore trained the network in the same way as network A, but ran the tests as
2
We hence discarded the expression labels
6
Published as a conference paper at ICLR 2015
the network would be a stochastic network. Network C is a hybrid network inspired by Tang &
Salakhutdinov (2013), where each hidden layer consists of 40 binary stochastic neurons and 160
deterministic neurons. However, the stochastic neurons have incoming connections from the deterministic input from the previous layer, and outgoing connections to the deterministic neurons in
the same layer. As in the original paper, the network was trained using the gradient estimator G4 .
Network D is the same as the hybrid network C with one difference: the stochastic neurons have
a constant activation probability of 0.5, and do hence not have any incoming weights or biases to
learn.
In all of the experiments, we used stochastic gradient descent with a mini-batch size of 100 and
momentum of 0.9. We used a learning rate schedule where the learning rate increases linearly from
zero to maximum during the first five epochs and back to zero during the remaining epochs. The
maximum learning rate was chosen among {0.0001, 0.0003, 0.001, . . . , 1} and the best test error
for each method is reported.3 The models were trained with M ∈ {1, 20}, and during test time we
always used M = 100.
As can be seen in Table 1, excluding the comparison methods, the proposed biased estimator G3
performs the best in both tasks. It is notable that the performance of G3 increased significantly
when using more than M = 1 particles, as could be predicted from Theorem 1. In Figure 1 (right)
we plot the objective CM at test time based on a number of particles M = 1, . . . , 100. In theory, a
larger number of particles M is always better (if given infinite computational resources), but here
Figure 1 (right) shows how the objective CM is estimated very accurately with only M = 20 or
M = 40.
Of all the networks tested, the best performing network in both tasks was however comparison
network D, i.e. the deterministic network with added binary stochastic neurons that have a constant
activation probability of 0.5. It is especially interesting to note that this network also outperformed
the hybrid network C where the output probabilities of the stochastic neurons are learned. Network D
seems to gain from being able to model stochasticity without the need to propagate errors through
the binary stochastic variables. The results give some support to the hypothesis that a hybrid network
outperforms a stochastic network because it is easier to learn a deterministic network around a small
number of stochastic units than learning a full stochastic network, although the stochastic units are
not trained properly.
The results could possibly be improved by making the networks larger and continuing training longer
if given enough computational capacity. This might be the case especially in the experiments with
the Toronto Face Dataset, where the deterministic network A outperforms some of the stochastic
networks. However, the critical difference between the stochastic networks and the deterministic
network can be be observed in Figure 2, where the stochastic networks are able to generate reconstructions that correspond to different digits for an ambiguous input. Clearly, the deterministic
network cannot model such a distribution.
4
D ISCUSSION
In the proposed estimator of the gradient for P (h|x) in Equation (32), there are both positive and
negative weights for various particles h(m) . Positive weights can be interpreted as pulling probability mass towards the particle, and negative weights as pushing probability mass away from the
particle. Although we showed that the variance of the gradient estimate is smaller when using both
positive and negative weights (G5 vs. G4 ), the difference in the final performance of the two estimators was not substantial
One challenge with structured outputs y is to find samples h(m) that give a reasonably large probability P (y|h(m) ) with a reasonably small sample size M . Training a separate R(h|x, y) 6= P (h|x)
as a proposal distribution looks like a promising direction for addressing that issue. It might still
be useful to use a mix of particles from R and P (h|x), and subtract a constant from the weights of
the latter ones. This approach would yield both particles that explain y well, and particles that have
negative weights.
3
In the MNIST experiments we used a separate validation set to select the learning rate. However, as we
chose just one hyperparameter with a fairly sparse grid, we only report the best test error in the TFD experiments
without a separate validation set.
7
Published as a conference paper at ICLR 2015
Figure 2: Samples drawn from the prediction of the lower half of the MNIST test data digits based
on the upper half with models trained using G3 (left), G5 (middle), and for the deterministic network
(right). The leftmost column is the original MNIST digit, followed by the masked out image and ten
samples. The figures illustrate how the stochastic networks are able to model different digits in the
case of ambiguous inputs.
Table 1: Results obtained on MNIST and TFD structured prediction using various number of
samples M during training and various estimators of the gradient Gi . Error margins are ± two
standard deviations from 10 runs.
MNIST Neg. test
TFD test Sum of
Squared Errors
log-likelihood (Cˆ100 )
M =1
M = 20
M =1
M = 20
G3 (proposed biased)
59.8 ± 0.1 53.8 ± 0.2 31.7 ± 0.7 26.3 ± 3.7
na
64.0 ± 1.7
na
51.4 ± 0.1
G4 (Tang et al., 2013)
G5 (proposed unbiased)
na
63.2 ± 1.2
na
51.3 ± 0.1
deterministic (A)
68.4 ± 0.1
na
35.3 ± 0.4
na
deterministic as stochastic (B) 59.1 ± 0.2
na
48.3 ± 7.5
na
hybrid (C)
na
58.4 ± 0.8
na
35.5 ± 1.0
deterministic, binary noise (D) 67.9 ± 1.1 52.0 ± 0.2 33.4 ± 0.6 21.4 ± 0.6
5
C ONCLUSION
Using stochastic neurons in a feedforward network is more than just a computational trick to train
deterministic models. The model itself can be defined in terms of stochastic particles in the hidden
layers, and we have shown many valid alternatives to the usual gradient formulation.
These proposals for the gradient involve particles in the hidden layers with normalized weights
that represent how well the particles explain the output targets. We showed both theoretically and
experimentally how involving more than one particle significantly enhances the modeling capacity.
We demonstrated the validity of these techniques in three sets of experiments: we trained a classifier on MNIST that achieved a reasonable performance, a network that could fill in the missing
information when we deleted the bottom part of the MNIST digits, and a network that could output
individual expressions of face images based on the mean expression.
We hope that we have provided some insight into the properties of stochastic feedforward neural
networks, and that the theory can be applied to other contexts such as the study of Dropout or other
important techniques that give a stochastic flavor to deterministic models.
ACKNOWLEDGEMENTS
The authors would like to acknowledge NSERC, Nokia Labs and the Academy of Finland as sources
of funding, in addition to the developers of Theano (Bastien et al., 2012; Bergstra et al., 2010)
8
Published as a conference paper at ICLR 2015
R EFERENCES
Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Bergstra, James, Goodfellow, Ian J., Bergeron,
Arnaud, Bouchard, Nicolas, and Bengio, Yoshua. Theano: new features and speed improvements.
Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop, 2012.
Bengio, Yoshua, Léonard, Nicholas, and Courville, Aaron. Estimating or propagating gradients
through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.
Bergstra, James, Breuleux, Olivier, Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guillaume, Turian, Joseph, Warde-Farley, David, and Bengio, Yoshua. Theano: a CPU
and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), June 2010. Oral Presentation.
Bishop, Christopher M et al. Pattern recognition and machine learning, volume 1. Springer New
York, 2006.
Esser, Steve K, Andreopoulos, Alexander, Appuswamy, Rathinakumar, Datta, Pallab, Barch, Davis,
Amir, Arnon, Arthur, John, Cassidy, Andrew, Flickner, Myron, Merolla, Paul, et al. Cognitive
computing systems: Algorithms and applications for networks of neurosynaptic cores. In Neural
Networks (IJCNN), The 2013 International Joint Conference on, pp. 1–10. IEEE, 2013.
Hinton, Geoffrey E, Srivastava, Nitish, Krizhevsky, Alex, Sutskever, Ilya, and Salakhutdinov, Ruslan R. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint
arXiv:1207.0580, 2012.
Kingma, Diederik P and Welling, Max.
arXiv:1312.6114, 2013.
Auto-encoding variational Bayes.
arXiv preprint
LeCun, Yann, Bottou, Léon, Bengio, Yoshua, and Haffner, Patrick. Gradient-based learning applied
to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Mnih, Andriy and Gregor, Karol. Neural variational inference and learning in belief networks. arXiv
preprint arXiv:1402.0030, 2014.
Neal, Radford M. Learning stochastic feedforward networks. Department of Computer Science,
University of Toronto, 1990.
Neal, Radford M. Connectionist learning of belief networks. Artificial intelligence, 56(1):71–113,
1992.
Raiko, Tapani, Valpola, Harri, and LeCun, Yann. Deep learning made easier by linear transformations in perceptrons. In International Conference on Artificial Intelligence and Statistics, pp.
924–932, 2012.
Rezende, Danilo Jimenez, Mohamed, Shakir, and Wierstra, Daan. Stochastic back-propagation and
variational inference in deep latent Gaussian models. arXiv preprint arXiv:1401.4082, 2014.
Saul, Lawrence K, Jaakkola, Tommi, and Jordan, Michael I. Mean field theory for sigmoid belief
networks. arXiv preprint cs/9603102, 1996.
Sietsma, Jocelyn and Dow, Robert JF. Creating artificial neural networks that generalize. Neural
Networks, 4(1):67–79, 1991.
Susskind, Joshua, Anderson, Adam, and Hinton, Geoffrey. The Toronto face database. Technical
report, University of Toronto, UTML TR 2010001, 06 2010.
Tang, Yichuan and Salakhutdinov, Ruslan. Learning stochastic feedforward neural networks. In
Advances in Neural Information Processing Systems, pp. 530–538, 2013.
Weaver, Lex and Tao, Nigel. The optimal reward baseline for gradient-based reinforcement learning.
In Proceedings of the Seventeenth conference on Uncertainty in artificial intelligence, pp. 538–
545. Morgan Kaufmann Publishers Inc., 2001.
9
Published as a conference paper at ICLR 2015
A
C LASSIFICATION EXPERIMENT
MNIST classification is a well studied problem where performances of a huge variety of approaches
are known. Since the output y is just a class label, the advantage of being able to model complex output distributions is not applicable. Still, the benchmark is useful for comparing training algorithms
against each other, and was used in this paper to test the viability of the gradient estimators.
We used a network structure with dimensionalities 784-200-200-10. The input data was first scaled
to the range of [0, 1], and the mean of each pixel was then subtracted. As a regularization method,
Gaussian noise with standard deviation 0.4 was added to each pixel separately in each epoch (Raiko
et al., 2012). The models were trained for 50 epochs.
Table 2 gives the test set error rate for each method. As can be seen from the table, deterministic
networks give the best results. Excluding the comparison networks, the best result is obtained with
the proposed biased gradient G3 followed by the proposed unbiased gradient G5 . Based on the
results, gradient estimators G1 and G2 were left out from the structured prediction experiments.
Table 2: Results obtained on MNIST classification using various number of samples M during
training and various estimators of the gradient Gi .
Test error (%)
G1 (Bengio et al., 2013, unbiased)
G2 (Bengio et al., 2013, biased)
G3 (proposed biased)
G4 (Tang et al., 2013)
G5 (proposed unbiased)
deterministic (A)
deterministic as stochastic (B)
hybrid (C)
deterministic, binary noise (D)
10
M =1
7.85
7.97
1.82
na
na
1.51
1.80
na
1.80
M = 20
11.30
7.86
1.63
3.99
2.72
na
na
2.19
1.92
| 9cs.NE
|
Image matting with normalized weight and semisupervised learning
Ping Li, Tingyan Duan, Yongfeng Cao*
Big data and computer science school
Guizhou Normal University
Guizhou China
[email protected], [email protected], [email protected]
Abstract—Image matting is an important vision problem. The
main stream methods for it combine sampling-based methods and
propagation-based methods. In this paper, we deal with the combination with a normalized weighting parameter, which could
well control the relative relationship between information from
sampling and from propagation. A reasonable value range for
this parameter is given based on statistics from the standard
benchmark dataset[1]. The matting is further improved by introducing semi-supervised learning iterations, which automatically
refine the trimap without user’s interaction. This is especially
beneficial when the trimap is coarse. The experimental results on
standard benchmark dataset have shown that both the normalized weighting parameter and the semi-supervised learning iteration could significantly improve the matting performance.
Keywords—image
matting;
sampling-based
propagation-based method; semi-supervised learning
method;
I. INTRODUCTION
A. Background and related work
Matting is an important image processing technology for
accurately estimating foreground objects in images and videos.
It is often used in image processing software, virtual studio,
film post-production and so on[2]. Mathematically, for a given
image I , any pixel can be expressed as a linear combination
of foreground color F and background color B ：
I z z F 1 z B ,
(1)
where z ( x, y ) represents the pixel coordinates in the image,
and z [0,1] is the foreground opacity of the pixel at z[3]. If
z 1 , the pixel is the foreground; if z 0 , then the pixel is
the background; when 0 z 1 , the pixel is a mixed pixel,
which means the pixel is affected by the foreground pixel and
the background pixel at the same time. Usually the most pixels
of a natural image are foreground and background, and only a
small number of them are mixed pixels[2]. Most matting
algorithms need a trimap as input[2]. In a trimap, there are
three areas: the foreground area F ( z =1 ), the background
area B ( z =0 ) and the unknown area U ( 0 z 1 ). The
main purpose of the matting is to accurately classify the pixels
in the unknown area.
Recently, many deep learning methods were used for image matting[4, 5]. Xu etc[4]. first predicted the initial alpha
matte with an deep convolutional encoder-decoder neural network, then made a further refinement for the initial alpha matte
with an small convolution neural network. Shen etc[5]. Proposed an automatic matting method for portrait using convolution neural network.
Except for above deep-learning-based ones, most of the
image matting methods could be categorized into samplingbased methods[3, 6], propagation-based methods[7] and combination of the two methods[8, 9]. Sampling-based methods
need to collect sample pairs similar to the unknown pixels
from the foreground area and the background area based on
pixel color similarity. If the input image has no obvious foreground color and background color or has highly textured regions, this kind of methods are less effective. The latest work
[6] makes up for this shortcoming by applying local smoothing
as a post-processing step to further improve the quality of alpha matte. Propagation-based methods propagate alpha values
of the known pixels to the unknown pixels through local
smoothing. They could work on texture images, but still not on
images with complex structures.
More effective methods are those combining sampling and
propagation. The robust matting method[8] selects high confident samples to estimate alpha value and propagates alpha values of known pixels to unknown pixels with a local window.
The local and nonlocal smooth priors method[9] adds nonlocal
smoothing information, except for the information from sampling and propagation. These methods have achieved good
matting effects. But they did not give a reasonable way to balance the data term from sampling and the local smooth term
from propagation, setting only an empirical weight on the data
term.
B. The main contributions of this paper
There are two points accounting for the main contributions
of this paper:
A normalized weight parameter is used to well control
the relative role of the data term and the local smooth
term in matting, and a reasonable value range for
setting the parameter is given based on experiments on
the standard data set[1].
from propagation-based methods. We calculate these two term
as in[8]. In contrast to the Laplacian construction formula
L W W lap of the robust matting method[8], it can be seen
that our normalized parameter can more clearly control the
relative weight between the two items. The experiment section
will suggest a range for setting .
Original Image
Trimap Image
Local Smooth Term
W lap
Data Term W
Constructing Laplacian Matrix
L=λW+(1-λ) W lap
Spectral Clustering
Optimization
Semi-Supervised
Learning
B. Optimization method
Image matting can be treated as a graph partition problem
which aims at optimally dividing a weighted graph into two or
more sub graphs. The spectral clustering method[10] solve this
problem as :
1
arg min( qT Lq ) ,
q
2
(3)
where L is the Laplacian matrix constructed in Section A, the
vector q records the alpha values of all pixels in the image
(where the foreground is 1, the background is 0, and the others
to be solved). Rewrite matrix L and vector q as
Result Alpha
Lk
L T
R
R
,
Lu
(4)
q qk
qu ,
(5)
Fig.1. Method flow chart
Semi-supervised learning iterations are introduced into
matting to increase the labeled pixels in the trimap
incrementally. It could improve matting effect without
increasing the trimap-making burden of users.
C. Paper content arrangement
The contents of this paper are arranged as follows: The first
section introduces the research background, related works and
the main contributions of this paper. The second section is the
method part of this paper. In the third section, the method is
experimented and analyzed based on the standard data set[1].
The fourth section summarizes this paper and points out the
future research direction.
II. MATTING METHOD
Our method mainly includes three steps as in Fig.1. First,
the Laplacian Matrix L is constructed by combining data term
W and local smooth term W lap with a normalized weight parameter. Second, alpha matte is got with spectral clustering
optimization based on L . Third, based on current alpha matte,
semi-supervised learning is used to refine the trimap. This
makes a loop in the process and enables our method iterating
many times to achieve good resulting matte.
A. Normalized weight parameter
In order to well control the relatively relationship between
data term and local smooth term in the matting process, this
paper constructs the Laplacian matrix with a normalized
weight parameter [0,1] as follow:
L W (1 )W lap ,
(2)
where W is the data item that is contributed by sampling-based
methods, and W lap is the local smooth term that is contributed
where Lk is the Laplacian matrix of the known region
(foreground area and background area), Lu is the Laplacian
matrix of the unknown region, qk is the alpha vector of the
known region and qu is the alpha vector of the unknown region.
By substituting (4) (5) into (3) and expanding it, it can be seen
that qu can be got by solving following linear optimization
problem：
Lu qu RT qk ,
(6)
in this paper, we use the conjugate gradient method to obtain
qu
C. Semi-supervised learning
In the trimap given by users, the more detailed the foreground and background be delineated, the better the result of
matting is. However, users want to pay as little effort as possible, so we usually obtain very rough trimap, where a large
number of pixels are unknown. In this section, we will introduce semi-supervised learning[11] to automatically increase
the number of pixels with known labels in the trimap. The specific process is shown in Fig.2. Firstly, based on the current
alpha matte, some pixels with high confidence are chosen from
the unknown area and labeled automatically. For example, if
the alpha value of a selected pixel is close to zero, it will be
labeled as the background; if the alpha value is close to one, it
will be labeled as the foreground. Then we update the trimap
with these newly-labeled pixels and do the matting process
again to produce a new alpha matte. Because there are more
pixels with known labels in the updated trimap, the resulting
alpha matte will be improved this time. To get the maximum i-
Alpha matte
trimap
Label
prediction
unknown area
pixels
(a)
(b)
(c)
Fig.3 Example images from the online benchmark for image matting. (a)
original image; (b) trimap of coarse level 1; (c) trimap of coarse level 2. In the
trimaps, the foreground is white, the background is black, and the unknown
area is gray.
Pixels label
Fig.2. Semi-supervised learning for matting
mprovement, this semi-supervised learning process can be
iterated several times. It does not increase workload of users
but computers.
The wrongly labeled pixels will bring error information into data term and local smooth term, and thus make the alpha
matte go wrong. Considering this error, we select pixels from
the unknown area with three strict constraints, trying to make
sure the predicted labels of these pixels are correct. Namely,
for each pixel x that will be updated its label in the trimap by
semi-supervised learning, it needs to meet the following conditions at the same time.
space constraint:
x U and y K , makes x y ,
(7)
where U is the set of unknown pixels in current trimap, K is
the set of foreground and background pixels in current trimap,
and y is the spatial neighborhood of pixel y (In this paper,
we choose the 4-connected pixels).
confidence constraint:
x t _ ,
(8)
where x is the alpha value of pixel x and t _ is a threshold.
proportion constraint:
x U t _ percent ,
(9)
where t _ percent is a proportion threshold (In this paper, we
use 10%), U t _ percent is the set formed by the top t _ percent
pixels of a list that was got by sorting all pixels in current unknown area in descending order according to 0.5 x .
III. EXPERIMENT AND ANALYSIS
The online benchmark for image matting on
www.alphamatting.com is used here to evaluate our methods.
We choose the low resolution image set which includes 27
input images of size from 490×800 to 719×800, their ground
truth and two set of trimaps of different coarse levels. Fig.3
shows sample images in the online benchmark.
The mean square error(MSE) is used as the indicator of m-
Fig.4. The MSE of proposed matting method ( =0.001 ) and robust matting.
atting performance in all experiments. The local window used
for calculating the local smooth term is set to 3×3.
A． Normalized weight parameter
Fig.4 shows the MSE of our proposed matting method with
the normalized parameter 0.001 and the robust matting
method[8] with parameter 0.1 . It can be seen that our
method outperforms the robust matting method in almost all
test images.
It is important to know how this normalized weight parameter affect matting performance. Fig.5 shows how MSE indicator changes with parameter . It can be seen that MSE indicator is quite stable when parameter is bigger than 0.05 and all
best MSEs for testing images are got with in [0,0.01]. A
rule that the smaller the is, the better the matting performance is works for all values of but near zero where the
matting performance goes bad when approaching zero. So
we suggest a value range [0.001, 0.01] for setting parameter
in practice.
B. Semi-supervised learning
In this section, we analyze how semi-supervised learning
affect the matting performance and how to choose a good
number of iterations. In all experiments, the normalized weight
parameter is set to 0.001. Because semi-supervised learning
is especially worth doing when labeled pixels are not enough,
for experiments in this section, we use trimaps of coarse level
2(see Fig.3 for the example).
Fig.6 shows the MSE of our proposed matting method
without and with semi-supervised learning(iterating 4 times). It
can be seen that semi-supervised iterations could improve the
matting
Fig.7. The percentage increase of matting performace(PIMP) changes with the
number of iterations
(a)
of matting results increase first and then begin to decrease after
a certain number of iterations.
In the forepart of iterations, those automatically given labels to originally unknown pixels are correct, so that the matting results are improved by the additional information. When
the number of iterations become big, those left unknown pixels
are less and tend to be mixed pixels (near the boundary of
foreground and background). This makes predicting their correct label very difficult and once some pixels are wrongly labeled, the matting results began to go bad.
(b)
Fig.5 MSE indicator changes with parameter
level 1. (b) using trimaps of coarse level 2.
. (a) using trimaps of coarse
This observation enlightens us not using too big iteration
numbers and the best iteration number for an image is related
not only with the coarseness of its trimap but also some other
characteristics (we need to find out in future research).
IV. CONCLUSION
This paper proposed a matting method based on normalized
weight and semi-supervised learning.
The normalized parameter can well control the relative
weight between data term and local smooth term in matting. A
experimental value range has been suggested for setting this
parameter. Semi-supervised learning iterations could significantly reduce users’ burden to delineate a refined trimap and
get good matting result from a coarse trimap. But the best
number of iterations depend not only on the roughness of the
trimap, and so is not easy to set. Generally, the more rough the
trimap, the more semi-supervised leaning iterations could be
taken.
Fig.6.The MSE of proposed matting methos with and without semi-supervised
learning.
Fig.7 shows the percentage increase of matting performance(PIMP), on each image by the use of semi-supervised
learning iterations. The specific formula for defining PIMP is,
MUSL
PIMP 0
1
PIMP MWUSL
,
0
PIMP 0
(10)
where MUSL is the MSE of using semi-supervised learning,
MWUSL is the MSE of without using semi-supervised learning. It can be seen that with the increase of iterations, the PIMP
Our future research will focus on adaptively selecting the
optimal weight coefficient and the number of semi-supervised
learning iterations.
ACKNOWLEDGMENT
This work was partially supported by NSF of China
(41161065), NSF of Guizhou (GZKJ [2017]1128) and Educational Commission of Guizhou (KY[2016]027)
REFERENCES
[1]
[2]
http://www.alphamatting.com.
G. L. Yao, "A Survey on Pre-Processing in Image Matting," JCST, vol.
32, pp. 122-138, 2017.
[3]
[4]
[5]
[6]
[7]
L. Karacan, A. Erdem, and E. Erdem, "Alpha Matting with KLDivergence Based Sparse Sampling," in ICCV 2016, pp. 424-432.
N. Xu, B. Price, S. Cohen, and T. Huang, "Deep Image Matting," arXiv
preprint arXiv:1703.03872, 2017.
X. Shen, X. Tao, H. Gao, C. Zhou, and J. Jia, "Deep automatic portrait
matting," in ECCV, 2016, pp. 92-107.
J. Johnson, D. Rajan, and H. Cholakkal, "Sparse codes as alpha matte,"
in BMVC, 2014.
Y. Shi, O. C. Au, J. Pang, K. Tang, W. Sun, H. Zhang, et al., "Color
clustering matting," in in ICME, 2013, pp. 1-6.
[8]
J. Wang and M. F. Cohen, "Optimized color sampling for robust
matting," in CVPR, 2007, pp. 1-8.
[9] X. Chen, D. Zou, S. Zhiying Zhou, Q. Zhao, and P. Tan, "Image matting
with local and nonlocal smooth priors," in CVPR, 2013, pp. 1902-1907.
[10] D. R. DeFord and S. D. Pauls, "Spectral Clustering Methods for
Multiplex Networks," arXiv preprint arXiv:1703.05355, 2017.
[11] Z. Yu, Y. Lu, J. Zhang, J. You, H. S. Wong, Y. Wang, et al, "Progressive
Semisupervised Learning of Multiple Classifiers," IEEE Trans.Cybern,
in press.
| 7cs.IT
|
1
Neural Component Analysis for Fault Detection
Haitao Zhao
arXiv:1712.04118v1 [cs.LG] 12 Dec 2017
Key Laboratory of Advanced Control and Optimization for
Chemical Processes of Ministry of Education, School of
Information Science and Engineering, East China University
of Science and Technology
Abstract—Principal component analysis (PCA) is largely
adopted for chemical process monitoring and numerous PCAbased systems have been developed to solve various fault detection
and diagnosis problems. Since PCA-based methods assume that
the monitored process is linear, nonlinear PCA models, such as
autoencoder models and kernel principal component analysis
(KPCA), has been proposed and applied to nonlinear process
monitoring. However, KPCA-based methods need to perform
eigen-decomposition (ED) on the kernel Gram matrix whose
dimensions depend on the number of training data. Moreover,
prefixed kernel parameters cannot be most effective for different
faults which may need different parameters to maximize their
respective detection performances. Autoencoder models lack the
consideration of orthogonal constraints which is crucial for PCAbased algorithms. To address these problems, this paper proposes
a novel nonlinear method, called neural component analysis
(NCA), which intends to train a feedforward neural work with
orthogonal constraints such as those used in PCA. NCA can
adaptively learn its parameters through backpropagation and
the dimensionality of the nonlinear features has no relationship
with the number of training samples. Extensive experimental
results on the Tennessee Eastman (TE) benchmark process show
the superiority of NCA in terms of missed detection rate (MDR)
and false alarm rate (FAR). The source code of NCA can be found
in https://github.com/haitaozhao/Neural-Component-Analysis.git.
Note to Practitioner: Online monitoring for chemical process
has been considered as a critical and hard task in real industrial
applications. In this paper, an innovative method called neural
component analysis (NCA) is proposed for fault detection. NCA
is a unified model including a nonlinear encoder and a linear
decoder. Due to its simple and intuitive format, NCA has superior
performance in both computational efficiency and fault detection
which makes it suitable for process monitoring in real industrial
applications. Moreover, the experimental results presented can
be reproduced effortlessly.
Index Terms—Process monitoring, Fault detection, Feedforward neural network, Autoencoder
I. I NTRODUCTION
Monitoring process conditions is crucial to its normal operation [1]. Over last decades, data-driven multivariate statistical
process monitoring (MSPM) has been widely applied to fault
diagnosis for industrial process operations and production
results [2], [3]. Due to the data-based nature of MSPM, it is
relatively convenient to apply to real processes of large scale
comparing to other methods based on theoretical modelling or
rigorous derivation of process systems [4], [5].
The task of MSPM is challenging mainly due to the “curse
of dimensionality” problem and the “data rich but information poor” problem. Many methods have been proposed
to transform original high dimensional process data into a
lower dimensional feature space and then performing fault
detection or fault diagnosis in that feature space [6]. Principal
component analysis (PCA) [7], [8], [9] is one of the most
widely used linear techniques for fault detection. Due to
orthogonal linear projection, PCA separates data information
into two subspaces: a significant subspace which contains
most variation in training data and a residual subspace which
includes noises or outliers in training data.
PCA-based methods to inherently nonlinear processes may
lead to unreliable and inefficient fault detection, since a linear
transformation is hard to tackle the nonlinear relationship
between different process variables [10], [11]. To deal with
this problem, various nonlinear extensions of PCA have been
proposed for fault detection. These extensions can be divided
into 3 categories.
The first category is kernel approaches. Kernel PCA
(KPCA) is one of the mostly used kernel approaches for fault
detection [12]. KPCA implicitly maps the data from an input
space into some high dimensional nonlinear feature sapce,
where linear PCA can be applied. KPCA need to perform
eigen decomposition (ED) on the kernel Gram matrix whose
size is the square of the number of data points. When there are
too many data points, the calculation of ED becomes hard to
perform [13]. Moreover, KPCA need to determine the kernel
and the associated parameters in advance.
The second category is based on linear approximation of
nonlinear process. In the linear approximation, several local
linear models are constructed and then integrated by Bayesian
inference [14]. Linear approximation is simple and easy to
realize, but it may not be able to handle strong nonlinearities
in the process.
The third category is neural-network-based models, such
as robust autoencoder (RAE) [13] and autoassociative neural
network [15]. These models train a feedforward neural network to perform the identity encoding, where the inputs
and the outputs of the network are the same. The network
contains an internal “bottleneck” layer (containing fewer nodes
than the output and input layers) for feature extraction. In
autoassociative neural network [15], the mapping from the
input layer to the “bottleneck” layer can be considered as
encoding, while de-mapping from the “bottleneck” layer to
the output layer can be considered as decoding. Although the
encoding can deal with the nonlinearities present in the data,
it has no consideration of the orthogonal constraints used in
PCA.
Recent years, a new trend in neural-network-based techniques known as deep learning has become popular in artificial
intelligence and machine learning [16]. Deep-learning-based
models are widely used in unsupervised training to learn the
representation of original data. Although these models are
often derived and viewed as extensions of PCA, all of them
2
'HFRGHU
n output nodes
p hidden nodes
(a) Illustration of Vector X on two
orthogonal directions.
n input nodes
(b) Illustration of Vector X on two
non-orthogonal directions.
(QFRGHU
Figure 1: Projections of Vector X on non-orthogonal and
orthogonal directions respectively.
Figure 2: The autoencoder model.
lack the consideration of the orthogonal constraints used in
PCA. The orthogonal constraints are quite important, since
they can largely reduce the correlations between extracted
features. Figure 1 shows simple plots of features of Vector
V obtained by orthogonal projections and non-orthogonal
projections respectively. From Figure 1b it is easy to find PC1
and PC2 are largely correlated. It means that the extracted
features contain redundant information and may distort the
reconstruction of the original vector [17].
Motivated by the above analysis, this paper proposes a novel
unified model, called neural component analysis (NCA), for
fault detection. NCA firstly utilizes a nonlinear neural network
as an encoder to extract features. Then linear orthogonal
transformation is adopted to decode the features to the original
data space. Finaly, this unified model is trained by minimizing
the reconstruction error between original data and the decoded
data. After training, NCA can be used as an unsupervised
learning method to extract the key features of process data. In
this paper, Hotelling T 2 statistic and the squared prediction
error (SPE) statistic are used for fault detection. The merits of
the proposed NCA method is demonstrated by both theoretical
analysis and case studies on the Tennessee Eastman (TE)
benchmark process.
II. AUTOENCODER AND PCA
An autoencoder model is an artificial neural network
adopted for unsupervised feature extraction [18]. An autoencoder model tries to learn a representation (encoding) for
original data, specifically for the purpose of dimensionality reduction. Recently, due to the research works in deep learning,
the autoencoder concept has be widely accepted for generative
models of data [19].
We assume that there are N input samples xi =
[xi1 , xi2 , · · · , xin ] ∈ R1×n (i = 1, 2, · · · , N ). The simplest
structure of an autoencoder model is a feedforward neural
network which consists of one input layer with n inputs, one
hidden layer with p (p < n) units and one output layer with
the same number of nodes as the input layer (see Figure 2).
The purpose of this structure is to reconstruct its own inputs.
Therefore, an autoencoder model belongs to unsupervised
learning.
An autoencoder model includes both the encoder part and
the decoder part, which can be defined as transitions α and β,
such that:
α : R1×n → F
β : F → R1×n
α, β = arg min
α,β
N
X
2
kxi − β (α (xi ))k .
(1)
i=1
In this paper, we only consider the case F =R1×p . The
encoder part takes the input x and maps it to z ∈ R1×p :
z = α(x) , σ (W x + b) .
This feature z is usually referred to as latent code, latent
feature or latent representation. Here, σ(·) is an element-wise
activation function such as a sigmoid function or a hyperbolic
tangent function. Matrix W is a parameter matrix and b is a
bias vector.
The decoder part maps the feature vector z to the reconstruction x̃ of the same dimensionality as x:
x̃ = β(z) , σ̃ W̃ z + b̃
where σ̃, W̃ and b̃ for the decoder part may be different from
the corresponding σ, W and b for the encoder part, depending
on the applications of the autoencoder model.
In order to learn the parameters in the activation functions,
an autoencoder model is often trained to minimize reconstruction error:
N
X
W, b; W̃ , b̃ = arg min
xi − σ̃ W̃ σ (W x + b) + b̃
W,b;W̃ ,b̃
i=1
If linear activation functions are used, the optimal solution
to an autoencoder is strongly related to PCA [20]. In this case,
Equation (1) can be written as
A, B = arg min
A,B
N
X
xi − xi AB T
2
i=1
or
A, B = arg min X − XAB T
A,B
2
F
(2)
2
.
3
where X =
x1
x2
..
.
n output nodes
and k·kF is the Frobenius norm.
xN
A ∈ Rn×p and B ∈ Rn×p are two linear transformation
matrices. Baldi and Hornik [21] showed that, if the covariance
matrix Σx associated with the data X is invertible, a unique
local and global minimum to Equation (2) corresponding
to an orthogonal projection onto the subspace spanned by
the first principal eigenvectors of the covariance matrix Σx .
More precisely, the optimal solutions to (2) can be obtain by
A = B = Un×p , where Un×p = [u1 , u2 , · · · , up ] are the p
eigenvectors corresponding to the first p largest eigenvalues of
the covariance matrix Σx . In PCA, u1 , u2 , · · · , up are defined
as loading vectors or principal components and T = XU is
the score matrix corresponding to the loading matrix U .
Although there is no orthogonal constraints of the transformation matrices A or B, the solutions to Equation (2)
are composed by orthogonormal bases. Due to the orthogonal
decomposition, PCA can transform an original data space into
two orthogonal subspaces, a principal subspace which contains
most variation in original data and a residual subspace which
includes noises or outliers. T 2 statistic and SPE statistic are
often adopted as indicators for fault detection corresponding to
the principal subspace and the residual subspace respectively.
The orthogonal decomposition minimizes the correlations between these two subspaces and makes the linear reconstruction
of original data with the least distortion [22], [23]. Because
of this orthogonal property, PCA is widely used in process
monitoring and many other fault detection methods can be
considered as the extensions of PCA [1].
III. N EURAL C OMPONENT A NALYSIS
A nonlinear autoencoder model can be trained to extract
latent features. However, due to the lacking of the orthogonal
property, the significant information of original data and the
information of noises or outliers are largely combined in this
model. An autoencoder model turns to overfit original data
and learns to capture as much information as possible rather
than reducing correlations in original data and extracting the
significant information. Because of this problem, nonlinear
autoencoder models are not used as widely as PCA.
In the section, we propose a novel unified model, called
neural component analysis (NCA), for fault detection. NCA
firstly utilizes a nonlinear neural network as a encoder to extract features. Then linear orthogonal transformation is adopted
to decode the features to the original data space. In this way,
NCA can be considered as a combination of nonlinear and
linear models.
In NCA, we use linear transformation B instead of nonlinear
transition β used in Equation (1). The optimization problem
of NCA is
W, b, B = arg min
W,b,B
N
X
xi − g (xi ; W, b) B T
i=1
subject to B T B = Ip×p
2
(3)
%
n input nodes
Figure 3: Illustration of neural component analysis (NCA).
where g (xi ; W, b) is a neural network with p outputs. Assume
B = [b1 , b2 , · · · , bp ] where bi ∈ Rn×1 , then the orthogonormal constraint, B T B = I, means
(
1 i=j
bTi bj =
(i, j = 1, 2, · · · , p).
0 otherwise
Figure 3 illustrates the structure of NCA. Here B contains
three orthogonal bases (plotted in red, blue, and green), i.e.
B = [b1 , b2 , b3 ].
Equation (3) shows the difference between NCA and autoencoder. Firstly, NCA is a unified model of nonlinear encoding and linear decoding. For the linear decoding, it will
be shown later that the computation of the transformation
matrix B is quite simple and no gradient-descent-based optimization is needed. Secondly, the orthogonormal constraints
B T B = Ip×p are added to NCA. It means that the decoding
from latent features is an orthogonal reconstruction, which can
largely reduce the correlation of different variables.
Let
g (x1 ; W, b)
g (x2 ; W, b)
G=
(4)
∈ RN ×p ,
..
.
g (xN ; W, b)
the optimization problem of NCA in Equation (3) can also be
written as
W, b, B = arg min
W,b,B
X − GB T
2
F
(5)
subject to B T B = Ip×p .
Matrix G, corresponding to the score matrix of PCA, is
the key features which we want to obtain for further analysis,
such as fault detection and diagnosis. However, it is difficult
to compute the optimal W , b and B simultaneously since the
4
optimization problem in Equation (5) is nonconvex. In this
paper, we compute W , b and B iteratively as follows. We
firstly fix W , b and obtain G, then B can be computed by
optimizing
B = arg min X − GB T
B
2
F
(6)
subject to B T B = Ip×p .
variable; Then NCA is performed to obtain the neural network
g (x; W, b) for nonlinear feature extraction. Finally Hotelling
T 2 and the squared prediction error (SPE) statistics are used
for fault detection.
Let gi = g (xi ; W, b) be the nonlinear latent features of xi
(i = 1, 2, · · · , N ) and Σg is the covariance matrix associated
T T
with the features G = g1T , g2T , · · · , gN
. T 2 statistic of gi
is computed as follows:
T
Ti2 = gi Σ−1
g gi .
Once B is obtained, W and b can be updated by solving
the following optimization problem:
W, b = arg min X − GB T
W,b
2
F
.
Theorem 1. [26] Reduced Rank Procrustes Rotation. Let
MN ×n and Nn×p be two matrices. Consider the constrained
minimization problem
Ĥ = arg min M − N H T
2
H
subject to
SPE statistic of feature gi can be calculated as follows:
SP Ei = xi − gi B T
(7)
The solution to the optimization problem in Equation (7)
can be obtained by the backpropagation algorithm [24] which
is widely adopted in training feedforward neural networks.
The solution to Equation (6) is to determine an orthogonal
matrix, that rotates G to fit original data matrix X. In linear
algebra and statistics [25], Procrustes analysis is a standard
technique for geometric transformations between two matrices.
The orthogonal Procrustes problem can be viewed as a matrix
approximation problem which tries to find the optimal rotation
or reflection for the transformation of a matrix with respect
to the other. Theorem 1 shows how to solve the reduced rank
Procrustes rotation problem.
H T H = Ip×p .
Suppose the singular value decomposition (SVD) of M T N is
U DV T , then Ĥ = U V T .
According to Theorem 1, we can design iterative procedures
of NCA to obtain W , b and B as follows:
1) Perform PCA on original data X to obtain the loading
matrix U ∈ Rn×p and let B = U .
2) Fix B, solve the optimization problem in Equation (7)
by the backpropagation algorithm.
3) Form G by Equation (4) and perform SVD on X T G,
i.e. X T G = U DV T .
4) Compute B̂ = U V T .
5) If B − B̂ < , break;
else, let B = B̂ and go to Step 2.
6) Output W , b and B.
After training, g (x; W, b) can be used for further feature
extraction.
IV. FAULT DETECTION BASED ON NCA
A novel fault detection method based on NCA is developed
in this section. The implementation procedures are given as
follows. Firstly in the modeling stage, the process data are
collected under normal process conditions and scaled by each
(8)
2
(9)
Because of no prior information available about the distribution of gi , we compute the confidence limit for T 2 and SP E
statistics approximately by kernel density estimation (KDE)
[27]. Let T12 , T22 , · · · , TN2 with an unknown density ρ(·) are the
T 2 statistics of g1 , g2 , · · · , gN . The kernel density estimator
of T 2 statistic is
2
N
N
T − Ti2
1 X
1 X
2
2
2
Kh (T − Ti ) =
K
ρ̂h (T ) =
N i=1
hN i=1
h
where K(·) is a non-negative function that integrates to one
and has zero mean and h > 0 is a bandwidth parameter. In this
paper, we take the RBF kernel for density estimation, which
is given by
2 !
N
X
T 2 − Ti2
1
2
exp
.
ρ̂h (T ) = √
2h2
2πhN i=1
2
After estimating ρ̂h (T 2 ), for the testing statistic Tnew
, the
2
) < τ then xnew
following condition is checked: If ρ̂h (Tnew
is normal else xnew is abnormal. The threshold τ is assigned
globally and could be adjusted in order to lower the percentage
of false alarm. Practically, τ is often equal to 0.01.
Similarly, we also can obtain
!
N
2
X
1
(SP E − SP Ei )
%̂h (SP E) = √
exp
2h2
2πhN i=1
where SP E1 , SP E2 , · · · , SP EN with an unknown density
%(·) are the SP E statistics of X1 , X2 , · · · , XN . For the
testing statistic SP Enew , the following condition is checked:
If %̂h (SP Enew ) < τ then xnew is normal else xnew is
abnormal.
The offline modeling and online monitoring flow charts are
shown in Figure 4. The procedures of offline modeling and
online monitoring are as follows:
Offline modeling:
1) Collect normal process data as the training data.
2) Normalize the training data by each variable with zero
mean and unit variance.
3) Initialize B as the loading matrix which contains the first
p eigenvectors of the covariance matrix of the training
data.
•
5
Training samples
Scale the testing
g
sample
Next
Scale each variable
Obtain the feature
Compute B
according to Eq. (5)
No
Compute aand
according to Eq. (6)
Whether
convergent?
Compute
and SPE of
the feature
Whether exceed the
control limit˛
No
Yes
Yes
Fault alarm
Determine the control limit of
and SPE using KDE
Offline modeling
Figure 5: A diagram of the TEP simulator.
Online monitoring
Figure 4: The steps of the proposed NCA method for fault
detection.
4) Fix B to compute W and b by training a neural network
optimizing Equation (7).
5) Fix W and b to compute B by Reduced Rank Procrustes
Rotation solving Equation (6).
6) Compute T 2 and SPE statistics of xi by Equation (8)
and (9) respectively.
7) Determine the control limit of T 2 and SPE by KDE
respectively.
• Online monitoring:
1) Sample a new testing data point xnew . Normalize it
according to the parameters of the training data.
2) Extract the feature gi = g (xnew ; W, b).
3) Compute T 2 and SPE statistics of the feature gi .
4) Alarm if T 2 (or SPE) of the extracted feature exceed the
control limit; Otherwise, view xnew as a normal data.
V. S IMULATION AND DISCUSSION
The Tennessee Eastman process (TEP) has been widely
used by process monitoring community as a source of
publicly available data for comparing different algorithms.
The simulated TEP is mainly based on an practical industrial
process in which the kinetics, operation and units have been
altered for specific reasons. The data generated by TEP are
nonlinear, strong coupling and dynamic [28], [29]. There are
five major units in TEP: a chemical reactor, condenser, recycle
compressor, vapor/liquid separator, and stripper. A flow sheet
of TEP with its implemented control structure is shown in
Figure 5. The MATLAB codes can be downloaded from
http://depts.washington.edu/control/LARRY/TE/download.html.
Besides normal data, the simulator of TEP can also generate
21 different types of faults in order to test process monitoring
algorithms.
A total of 52 variables including 22 continuous process
measurements, 19 compositions and 11 manipulated variables1
1 the
agitation speed was not included because it was not manipulated
Table I: TEP fault modes.
Fault
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Description
A/C Feed ratio, B composition constant (Stream 4)
B composition, A/C ratio constant (Stream 4)
D feed temperature (Stream 2)
Reactor cooling water inlet temperature
Condenser cooling water inlet temperature
A feed loss (Stream 1)
C header pressure loss (Stream 4)
A, B, C feed composition (Stream 4)
D feed temperature (Stream 2)
C feed temperature (Stream 4)
Reactor cooling water inlet temperature
Condenser cooling water inlet temperature
Reaction kinetics
reactor cooling water valve
Condenser cooling water valve
Unknown
Unknown
Unknown
Unknown
Unknown
Valve (Stream 4)
Type
Step
Step
Step
Step
Step
Step
Step
Random variation
Random Variation
Random Variation
Random Variation
Random Variation
Slow drift
Sticking
Sticking
Unknown
Unknown
Unknown
Unknown
Unknown
Constant position
were selected as the monitoring variables in our experiments.
The training data set contained 500 normal data. Twenty-one
different faults were generated and 960 data for each fault
were chosen for testing, in which the fault happened from
161th data to the end of the data. The 21 fault modes are
listed in Table I.
In this paper, we compare our proposed NCA method
with PCA, KPCA and autoencoder. For KPCA, we use
the most widely used
Gaussian kernel k(xi , xj ) =
2
exp − kxi − xj k /c and select the kernel parameter c as
10nδ̄, where δ̄ is the mean of standard deviations of different
variables [30].
A. Visualization of data of different methods
In order to intuitively visualize the features of different
methods, we extract 2 features for each method and plot them
in Figure 6. In Figure 6, the blue “∗” indicates normal data,
while the red “◦” indicates fault data from Fault 1. It can be
found that normal data and fault data are largely overlapped
in Figure 6a, 6b, and 6c. In this case, PCA, KPCA and
6
600
40
0.025
Normal data
Fault data
Normal data
Fault data
0.02
30
500
0.015
20
0.005
0
0
-0.005
-10
-0.01
-20
-60
-40
-20
0
20
40
-0.015
-0.03
-0.02
-0.01
0
PC 1
0.02
0.03
400
300
200
PC 1
(a) Visualization plot of PCA.
(b) Visualization plot of KPCA.
20
100
3
Normal data
Fault data
15
0
Normal data
Fault data
2
10
5
PC 2
-5
-2
-3
-15
-4
-15
-10
-5
0
5
10
15
20
20
25
30
-1
-10
-20
-20
15
Figure 8: Convergence plot of NCA.
0
0
10
Iteration No.
1
5
PC 2
0.01
Reconstruction error
PC 2
PC 2
0.01
10
-5
-5
0
5
10
PC 1
15
20
25
30
PC 1
(c) Visualization plot of autoencoder.
(d) Visualization plot of NCA.
Figure 6: Visualization of the normal samples and samples of
Fault 1 on the first two dimensions of 4 different methods.
the training stage. It can be found that NCA converges very
fast. Thanks for GPU-accelerated computing, the total training
time of NCA is 2.58 seconds. We perform the experiment on a
computer with Intel Core i7 3.4GHz, 16G RAM, and NVIDIA
GeForce GTX 1080Ti. For comparison, KPCA is performed
on the same computer and the training time is 1.53 seconds.
B. Case studies
1.5
0.6
0.4
1
0.2
0.5
0
0
-0.2
-0.5
-0.4
20
20
20
10
PC No.
10
0
0
PC No.
(a) Illustration of B T B.
20
10
PC No.
10
0
0
PC No.
(b) Illustration of B T U .
Figure 7: Illustration of the correlations of the columns of B
and the correlations between the columns of B and those of
U.
autoencoder cannot find the significant information for fault
detection. However, in Figure 6d, the overlapping of normal
samples and fault samples is much less than those in Figure
6a, 6b, and 6c. Through the nonlinear feature extraction based
on orthogonal constraints, NCA directly gives out the key
features in a two-dimensional space which include significant
information for fault detection.
Figure 7 shows the difference between B of NCA and the
loading matrix U of PCA. Figure 7a shows the result of B T B.
It is easy to find that B contains orthogonormal columns, i.e.
bTi bj = 1 if i = j, otherwise, bTi bj = 0. Moreover, the
correlations between the columns of B and those of P are
plotted in Figure 7b. Obviously with the nonlinear feature
extract, the optimum solution B of NCA is utterly different
from the loading matrix U of PCA.
2
Figure 8 plots the reconstruction errors X − GB T F in
In this subsection, we investigate the performance of our
proposed NCA. The performance is compared with PCA,
KPCA, and autoencoder. The source codes of NCA and other
methods can be found in https://github.com/haitaozhao/NeuralComponent-Analysis.git.
According to the cumulative percentage variance (CPV)
rule, the reduced dimensionality, p, was determined as 27
for PCA such that 85% of the energy in the eigenspetrum
(computed as the sum of eigenvalues) was retained. In order
to give a fair comparison, the same value p = 27 was used
for KPCA, autoencoder and NCA.
Missed detection rate (MDR) refers to the rate of the
abnormal events being falsely identified as normal events in
the monitoring process, which is only applied to the fault
detection situation. For testing all the 21 faults, MDR is
recorded together in Table II, where smaller values indicate
better performances. False alarm rate (FAR) which refers to
the normal process monitoring results of PCA, KPCA, autoencoder and NCA are shown in parentheses. The small FARs
also indicate better performances. In Table II, the best achieved
performance for each fault is highlighted in bold. In this study,
we only consider the fault cases where MDR< 50% and
FAR≤ 5%. Because if MDR≥ 50%, the detection performance
would be even worse than random guess whose MDR is 50%.
Moreover, we adopt the threshold value for FAR as 5% which
is commonly used in fault detection.
NCA outperforms PCA, KPCA and autoencoder in 11 cases
with lower MDRs. Autoencoder gives the best performance
with 5 cases, KPCA gives the best performance with 7 cases
and PCA gives the best performance with 4 cases. Autoencoder, KPCA, and NCA provide better results than PCA
7
Table II: Missed Detection Rate (%) and False Alarm Rate (%) (shown in parentheses) of PCA, LPP, LPP_Markov and DGE
in TEP
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
PCA
T2
SPE
0.50(4.38)
0.13(20.6)
1.63(2.50)
0.75(18.1)
92.0(0.63)
71.9(30.0)
39.1(2.50)
0.00(24.4)
73.8(0.63)
50.3(22.5)
1.00(0.00)
0.00(10.6)
0.00(0.00)
0.00(16.3)
2.50(0.63)
1.75(17.5)
96.4(5.00)
76.9(23.8)
58.4(0.00)
24.13(15)
47.9(0.63)
19.0(20.0)
1.25(0.63)
1.13(22.5)
4.88(0.00)
3.75(12.5)
0.13(0.63)
0.0(23.13)
95.1(0.00)
74.4(16.9)
76.8(6.88)
30.8(22.5)
19.9(1.25)
2.50(25.6)
10.9(0.0)
6.25(20.6)
93.25(0.0)
41.8(14.4)
62.63(0.0)
26.9(15.6)
62.3(1.25)
33.5(31.3)
KPCA
T2
SPE
0.5(7.50)
0.63(5.63)
1.63(5.00)
1.75(5.00)
88.5(1.25)
93.8(0.63)
51.1(5.00)
81.8(3.75)
70.5(5.00)
79.3(3.75)
0.88(2.50)
2.88(3.13)
0.00(0.63)
2.0(0.63)
2.50(2.50)
4.75(3.13)
89.0(15.0)
94.1(5.63)
48.9(1.88)
81.6(0.63)
38.6(4.38)
52.4(1.88)
1.00(3.13)
10.4(0.63)
4.63(1.25)
18.0(0.63)
0.00(1.25)
9.88(0.63)
86.9(3.13)
94.1(2.50)
64.4(20.6)
86.9(6.25)
13.5(0.63)
28.0(0.00)
10.0(2.50)
14.4(2.50)
82.5(1.88)
87.3(0.63)
52.1(2.50)
90.0(2.50)
59.5(5.63)
71.0(3.75)
in more fault cases indicates that nonlinear extensions can
generally obtain better fault detection results than linear PCA.
Although both KPCA and NCA include orthogonal constraints
in their feature extraction, the performances of NCA are much
better than KPCA. The reason is that NCA adaptively learns
the parameters of a neural network, while KPCA uses prefixed
kernel and the associated parameter. NCA obtains different
parameters for different variables with nonlinear combination
through backpropagation strategy and should be much more
suitable for nonlinear feature extraction and the following fault
detection tasks.
Figure 9, 10 and 11 illustrate the detailed fault detection
results of Fault 2, 4 and 7. In these figures, the blue points
indicate the first 160 normal samples, while the red points
represent the following 800 fault samples. The black dash lines
are the control limits according to the threshold τ . The blue
points above control limits lead to false alarm, while the red
points below control limits cause missed detection.
According to Table II, all 4 methods can successfully detect
Fault 2. For PCA, the FARs are 2.5% and 18.1% for T 2
statistic and SPE statistic respectively. The FARs of KPCA
are 5% for both T 2 statistic and SPE statistic. Autoencoder
achieves no false alarm for SPE statistic and its FAR for T 2
statistic is 1.88%. However, as for our proposed NCA, the
FARs are zero for both two statistics. Figure 9 shows the plots
the results of four methods on Fault 2. For each method, the
subplot above is the results of T 2 statistic and the subplot
below is the results of SPE statistic. The small overlay plots
clearly show the results of the first 160 normal samples in each
subplot. Based on Table II and Figure 9, for Fault 2, we can
found that NCA have the lowest MDR with no false alarm.
Figure 10 illustrates the results of Fault 4. In this experiment, neither KPCA nor autoencoder can detect this fault. the
MDR of PCA is 39.1% using T 2 statistic. It means that PCA,
KPCA, and autoencoder are not suitable for the detection of
autoencoder
T2
SPE
0.50(3.13)
0.75(0.00)
1.50(1.88)
2.00(0.00)
96.4(1.88)
99.9(0.00)
55.8(0.63)
96.3(0.00)
73.3(0.63)
77.9(0.00)
1.00(0.00)
1.13(0.00)
0.00(0.00)
30.1(0.63)
2.50(1.88)
4.38(0.00)
96.0(5.00)
99.38(1.88)
64.3(0.63)
77.38(0.00)
49.6(1.25)
71.9(0.00)
1.00(1.88)
2.63(0.00)
5.75(0.63)
5.88(1.25)
0.38(0.63)
4.88(0.00)
94.3(0.63)
98.1(0.00)
82.1(5.00)
91.0(1.88)
20.1(1.88)
27.4(0.00)
11.0(3.75)
11.63(0.63)
91.8(0.00)
99.8(0.00)
64.1(0.00)
78.1(0.00)
61.6(1.25)
79.6(0.63)
NCA
T2
SPE
0.50(0.00)
0.75(0.00)
1.75(0.00)
1.50(0.00)
98.4(0.00)
97.9(0.63)
63.3(1.25)
5.75(0.63)
75.4(1.25)
71.1(1.88)
0.50(0.00)
0.88(0.00)
0.00(0.00)
0.00(1.88)
2.75(0.00)
2.50(2.50)
97.9(2.50)
94.0(5.63)
66.3(0.00)
72.8(1.88)
52.0(0.63)
30.1(8.13)
1.00(1.25)
5.50(7.5)
5.13(1.25)
4.88(3.75)
0.38(0.63)
0.00(5.00)
97.5(0.00)
90.9(3.75)
85.6(1.88)
76.6(1.25)
15.4(0.00)
53.3(0.00)
9.88(0.63)
12.1(0.00)
100(0.00)
99.4(0.00)
69.1(0.00)
52.0(0.00)
64.8(0.00)
78.0(0.00)
Fault 4. However, using SPE statistic, NCA obtains the MDR
of 5.75% and the FAR of 0.63%. Obviously NCA is much
more appropriate for detecting Fault 4.
Figure 11 shows the detection results of four methods for
Fault 7. For SPE statistic, the MDR of PCA is 0.00%, but
the FAR is 16.3% which is greater than 5%. The reason
for the high FAR is that the linear decomposition of PCA
cannot obtain the appropriate residual subspace to determine
the right control limit. KPCA can detect this fault with the
MDR of 2.0% for SPE statistic and the FAR is 0.63%. As
for autoencoder, the MDR for SPE statistic is 30.1% which is
much higher than that of KPCA. NCA outperforms these two
nonlinear methods with no missed detection for SPE statistic
and the FAR is 1.88%. The cause for the high MDR of autoencoder for SPE statistic should be that autoencoder overfits
the training data and becomes more prone to accept noises and
outliers in the normal data to construct the significant subspace
and residual subspace. Under this condition, the features of
the residual subspace may not contain enough information to
detect the fault samples. Thanks for orthogonal constraints in
KPCA and NCA, both of them can successfully detect the
fault data.
We summarize the case studies below:
1) Although no single method gives optimal performance
for all fault cases consisting of diverse numbers of fault
conditions. NCA outperforms PCA, KPCA, and autoencoder with regard to the number of best performances
and emerges as the clear winner.
2) Due to the incorporation of orthogonal constraints, NCA
is less prone to overfit training data and performs better
than autoencoder.
3) Since NCA considers both nonlinear feature extraction
and orthogonal constraints for feature extraction, NCA
becomes effective in fault detection. Although NCA is
an iterative method and need more time for training,
4000
8
3000
6
T2
T2
8
2000
1000
4
2
0
0
0
200
400
600
800
1000
0
200
400
Sample
SPE
SPE
600
400
1000
600
800
1000
600
800
1000
600
800
1000
504
502
200
0
500
0
200
400
600
800
0
1000
200
400
Sample
Sample
(a) PCA
(b) KPCA
1500
4000
3000
1000
T2
T2
800
506
800
2000
500
1000
0
0
0
200
400
600
800
1000
0
200
400
Sample
10
8
Sample
-10
4000
3000
SPE
6
SPE
600
Sample
4
2000
1000
2
0
0
0
200
400
600
800
1000
Sample
(c) Autoencoder
0
200
400
Sample
(d) NCA
Figure 9: Monitoring results of 4 different methods for Fault 2.
given the superior performance of NCA, the trade-off
between training time and performance seems justified.
VI. C ONCLUSION
In this paper, we propose a nonlinear feature extraction
method, nerual component analysis (NCA), for fault detection.
NCA is a unified model which includes a nonlinear encoder
part and a linear decoder part. Orthogonal constraints are
adopted in NCA which can alleviate the overfitting problem
occurred in autoencoder and improve performances for fault
detection. NCA takes the advantages of the backpropogation
technique and the eigenvalue-based techniques. The convergence of the iteration scheme of NCA is very fast. The idea
behind NCA is general and can potentially be extended to
other detection or diagnosis problems in process monitoring.
We compare NCA with other linear and nonlinear fault
detection methods, such as PCA, KPCA, and autoencoder.
Based on the case studies, it is clear that NCA outperforms
PCA, KPCA, and autoencoder. NCA can be considered as
an alternative to the prevalent data driven fault detection
techniques.
Future works will be contributed to the design of new
regularization terms to the optimization problem of NCA in
Equation (5). Reconstruction-based feature extraction methods, such as PCA, KPCA, autoencoder, and NCA, are mainly
focus on the global Euclidean structure of process data and
overlook the latent local correlation of process data. In the
future, we will try to design new constraints to ensure the local
information is considered in nonlinear feature extraction.
ACKNOWLEDGMENT
This research is sponsored by National Natural Science
Foundation of China (61375007) and Basic Research Programs of Science and Technology Commission Foundation of
Shanghai (15JC1400600).
R EFERENCES
[1] S. J. Qin, “Survey on data-driven industrial process monitoring and
diagnosis,” Annual Reviews in Control, vol. 36, no. 2, pp. 220–234,
2012.
[2] J. MacGregor and A. Cinar, “Monitoring, fault diagnosis, fault-tolerant
control and optimization: Data driven methods,” Computers & Chemical
Engineering, vol. 47, pp. 111–120, 2012.
9
200
0.4
T2
0.6
T2
300
100
0.2
0
0
0
200
400
600
800
1000
0
200
400
600
800
1000
600
800
1000
600
800
1000
600
800
1000
Sample
80
500.4
60
500.3
SPE
SPE
Sample
40
500.2
500.1
20
0
500
0
200
400
600
800
0
1000
200
400
Sample
Sample
(a) PCA
(b) KPCA
100
200
T2
T2
150
100
50
50
0
0
0
200
400
600
800
1000
0
200
400
Sample
10
200
150
4
SPE
SPE
6
Sample
-11
100
2
50
0
0
0
200
400
600
800
1000
0
Sample
200
400
Sample
(c) Autoencoder
(d) NCA
Figure 10: Monitoring results of 4 different methods for Fault 4.
[3] S. Yin, S. X. Ding, A. Haghani, H. Hao, and P. Zhang, “A comparison
study of basic data-driven fault diagnosis and process monitoring methods on the benchmark tennessee eastman process,” Journal of Process
Control, vol. 22, no. 9, pp. 1567–1581, 2012.
[4] Z. Q. Ge, Z. H. Song, and F. Gao, “Review of recent research on
data-based process monitoring,” Industrial and Engineering Chemistry
Research, vol. 52, no. 10, pp. 3543–3562, 2013.
[5] T. Feital, U. Kruger, J. Dutra, J. C. Dutra, and E. L. Lima, “Modeling and
performance monitoring of multivariate multimodal processes,” AIChE
Journal, vol. 59, no. 5, pp. 1557 – 1569, 2013.
[6] M. Askarian, G. Escudero, M. Graells, R. Zarghami, F. Jalali-Farahani,
and N. Mostoufi, “Fault diagnosis of chemical processes with incomplete
observations: A comparative study,” Computers & chemical engineering,
vol. 84, pp. 104–116, 2016.
[7] X. Gao and J. Hou, “An improved svm integrated gs-pca fault diagnosis
approach of tennessee eastman process,” Neurocomputing, vol. 174,
pp. 906–911, 2016.
[8] Y. Wang, F. Sun, and B. Li, “Multiscale neighborhood normalizationbased multiple dynamic pca monitoring method for batch processes
with frequent operations,” IEEE Transactions on Automation Science
and Engineering, vol. PP, no. 99, pp. 1–12, 2017.
[9] T. J. Rato, J. Blue, J. Pinaton, and M. S. Reis, “Translation-invariant
multiscale energy-based pca for monitoring batch processes in semiconductor manufacturing,” IEEE Transactions on Automation Science and
Engineering, vol. 14, pp. 894–904, April 2017.
[10] L. Luo, S. Bao, J. Mao, and D. Tang, “Nonlinear process monitoring
based on kernel global-local preserving projections,” Journal of Process
Control, vol. 38, no. Supplement C, pp. 11–21, 2016.
[11] N. Sheng, Q. Liu, S. J. Qin, and T. Chai, “Comprehensive monitoring
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
of nonlinear processes based on concurrent kernel projection to latent
structures,” IEEE Transactions on Automation Science and Engineering,
vol. 13, pp. 1129–1137, April 2016.
M. Mansouri, M. Nounou, H. Nounou, and N. Karim, “Kernel pca-based
glrt for nonlinear fault detection of chemical processes,” Journal of Loss
Prevention in the Process Industries, vol. 40, no. Supplement C, pp. 334
– 347, 2016.
L. Jiang, Z. Song, Z. Ge, and J. Chen, “Robust self-supervised model and
its application for fault detection,” Industrial & Engineering Chemistry
Research, vol. 56, no. 26, pp. 7503–7515, 2017.
Z. Ge, M. Zhang, and Z. Song, “Nonlinear process monitoring based
on linear subspace and bayesian inference,” Journal of Process Control,
vol. 20, no. 5, pp. 676 – 688, 2010.
M. A. Kramer, “Nonlinear principal component analysis using autoassociative neural networks,” AIChE Journal, vol. 37, no. 2, pp. 233–243,
1991.
G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507,
2006.
D. Cai, X. He, J. Han, and H. J. Zhang, “Orthogonal laplacianfaces
for face recognition,” IEEE Transactions on Image Processing, vol. 15,
pp. 3608–3614, Nov 2006.
C.-Y. Liou, W.-C. Cheng, J.-W. Liou, and D.-R. Liou, “Autoencoder
for words,” Neurocomputing, vol. 139, no. Supplement C, pp. 84 – 96,
2014.
D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,”
ArXiv e-prints, Dec. 2013.
D. Chicco, P. Sadowski, and P. Baldi, “Deep autoencoder neural networks for gene ontology annotation predictions,” in Proceedings of the
10
3
1500
2
T2
T2
1000
1
500
0
0
0
200
400
600
800
0
1000
200
400
800
1000
600
800
1000
600
800
1000
600
800
1000
501
300
200
SPE
SPE
600
Sample
Sample
500.5
100
500
0
0
200
400
600
800
0
1000
200
400
Sample
Sample
(a) PCA
(b) KPCA
800
800
600
T2
T2
600
400
400
200
200
0
0
0
200
400
600
800
0
1000
200
400
Sample
Sample
10-10
8
300
200
SPE
SPE
6
4
100
2
0
0
0
200
400
600
800
1000
Sample
(c) Autoencoder
0
200
400
Sample
(d) NCA
Figure 11: Monitoring results of 4 different methods for Fault 7.
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
5th ACM Conference on Bioinformatics, Computational Biology, and
Health Informatics, BCB ’14, (New York, NY, USA), pp. 533–540,
ACM, 2014.
P. Baldi and K. Hornik, “Neural networks and principal component analysis: Learning from examples without local minima,” Neural Networks,
vol. 2, no. 1, pp. 53 – 58, 1989.
I. Jolliffe, Principal Component Analysis. Springer Verlag, 1986.
I. T. Jolliffe, “A note on the use of principal components in regression,”
Journal of the Royal Statistical Society, vol. 31, no. 3, pp. 300 – 303,
1982.
Y. Li, Y. Fu, H. Li, and S. W. Zhang, “The improved training algorithm
of back propagation neural network with self-adaptive learning rate,”
in 2009 International Conference on Computational Intelligence and
Natural Computing, vol. 1, pp. 73–76, June 2009.
D. G. Kendall, “A survey of the statistical theory of shape,” Statistical
Science, vol. 4, no. 2, pp. 87–99, 1989.
J. M. F. Ten Berge, “Orthogonal procrustes rotation for two or more
matrices,” Psychometrika, vol. 42, pp. 267–276, Jun 1977.
R. T. Samuel and Y. Cao, “Nonlinear process fault detection and
identification using kernel pca and kernel density estimation,” Systems
Science & Control Engineering, vol. 4, no. 1, pp. 165–174, 2016.
L. H. Chiang, R. D. Braatz, and E. L. Russell, Fault detection and
diagnosis in industrial systems. Springer Science & Business Media,
2001.
P. Lyman and C. Georgakis, “Plant-wide control of the tennessee
eastman problem,” Computers and Chemical Engineering, vol. 19, no. 3,
pp. 321 – 331, 1995.
[30] H. Zhao, P. C. Yuen, and J. T. Kwok, “A novel incremental principal
component analysis and its application for face recognition,” IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics),
vol. 36, pp. 873–886, Aug 2006.
| 9cs.NE
|
Optimal Causal Imputation for Control
arXiv:1703.07049v1 [cs.SY] 21 Mar 2017
Roy Dong, Eric Mazumdar, and S. Shankar Sastry
Abstract— The widespread applicability of analytics in cyberphysical systems has motivated research into causal inference
methods. Predictive estimators are not sufficient when analytics
are used for decision making; rather, the flow of causal effects
must be determined. Generally speaking, these methods focus
on estimation of a causal structure from experimental data. In
this paper, we consider the dual problem: we fix the causal
structure and optimize over causal imputations to achieve
desirable system behaviors for a minimal imputation cost. First,
we present the optimal causal imputation problem, and then
we analyze the problem in two special cases: 1) when the causal
imputations can only impute to a fixed value, 2) when the causal
structure has linear dynamics with additive Gaussian noise.
This optimal causal imputation framework serves to bridge the
gap between causal structures and control.
I. I NTRODUCTION
Recently, data analytics have achieved amazing levels of
success. As analytics penetrate more and more industrial
applications, they are increasingly used for decision-making
and planning. In these applications, it is important to use
estimators that are not only predictive, but estimate the causal
structure of the underlying processes.
Correlation is not the same as causation. However, in
practice, it is not always easy to apply this principle. In many
real-life applications, machine learning is used to determine
the relationship between two variables. This analysis is
often used as the basis for determining which actions to
take. However, an algorithm with low test error does not
necessarily mean that the causal effect has been estimated.
For example, one may train a classifier to estimate the
energy consumption of a household given the presence
and absence of eco-friendly devices, and this may provide
guidelines for which devices should be discounted through
rebate programs. Unless the causal structures are explicitly
accounted for, there could easily be confounding variables
or incorrect causal relationships that change the behavior of
the system under consideration.
This has motivated new interest in causal inference techniques. Generally speaking, these techniques take experimental data and attempt to uncover the causal structure. (We defer
a literature review of these methods to Section III, when
a more formal model of causality has been developed.) In
this paper, we consider the dual problem: we fix the causal
structure and attempt to determine what causal actions will
lead to system behaviors we desire at a minimal cost.
R. Dong, E. Mazumdar, and S. S. Sastry are with the
Department of Electrical Engineering and Computer Sciences,
University of California, Berkeley, Berkeley, CA, 94707, USA
{roydong,emazumdar,sastry}@eecs.berkeley.edu
A. Outline
The rest of the paper is organized as follows. We discuss
the main paradigms for modeling causality in Section II. In
Section III, we outline the mathematical formulation of a
causal structure, discuss relevant literature in causal estimation, and define the problem of optimal causal imputation.
In Section IV, we provide theoretical analysis of two special
cases of the optimal causal imputation problem: the case
where imputation can only be done to a single value, and the
case where the dynamics are linear and the noise is Gaussian.
Finally, we present closing remarks in Section V.
II. BACKGROUND
There are three main paradigms for the mathematical
modeling of causality:
1) Rubin causality
2) Granger causality
3) Pearl’s structural equation modeling (SEM)
Each of these paradigms has a vast literature in its own right;
we will try to present a few representative samples from each
field here. Note that each paradigm uses its own notation,
so we will change notation as we switch from approach to
approach.
It should be noted that these paradigms are not mutually
exclusive: for example, a problem that is modeled using
Granger causality can be put into Pearl’s SEM if the underlying processes operate in discrete time. Rubin causality can
often be phrased as an SEM problem, but in applications this
will require more structural assumptions to learn the causal
structure. A full exposition of the intersections and nonintersections of these three paradigms is outside the scope
of this paper, but we note that these paradigms can often
model the same phenomena and shed different insights on
the causal behaviors observed.
Rubin causality was first introduced in [1]. In the basic
formulation of Rubin causality, we are given some control
variable X taking values in {0, 1}. There are also two distinct
random variables Y0 and Y1 . If X = 0, then we observe
Y0 and not Y1 . If X = 1, then we only observe Y1 , and
not Y0 . Another way to write this notationally is that we
observe YX but do not observe Y1−X , which is often called
the counterfactual. The fact that we can only observe one or
the other, but not both, is the fundamental misery of causality.
One of the key results that the Rubin causality paradigm
provides is that if X is independent of Y0 and Y1 , then
randomly assigning X ∈ {0, 1} yields a dataset that can
provide valid estimates of the counterfactuals; thus, Rubin
causality provides the theoretical foundation for randomized
control trials. This paradigm has also been extended to
consider many covariates [2], handle confounding variables
and incorporate instrumental variables [2], and incorporate
some machine learning approaches [3]. Sample applications
include estimating the causal effect of residential demand
response in the Western United States [4] or the causal
effects of providing money, healthcare and education to the
very poor in Ethiopia, Ghana, Honduras, India, Pakistan, and
Peru [5].
Granger causality was first introduced in [6]. In this
paradigm, we are given data from two stationary random
processes X and Y , both indexed by time. First, let Ut
denote all the information available in the universe at time
t, and let (U − X)t denote all the information available
at time t except for X. Then, let σ 2 (Y |U ) denote the
error variance of the unbiased, least-squares estimator of Yt
using Ut , and similarly let σ 2 (Y |U − X) denote the error
variance of the unbiased, least-squares estimator of Yt using
(U − X)t . Then, X Granger-causes (or G-causes, for short)
Y if σ 2 (Y |U ) < σ 2 (Y |U −X), i.e. the estimator that utilizes
X has lower variance on its error than the one that cannot.
In other words, X has explanatory power for Y .
Granger causality essentially relies on the relationship
between causal effects and the arrow of time to distinguish it
from general correlations. Although this framework does not
address many of the more pernicious philosophical aspects
of causality, oftentimes prior knowledge allows us to make
the inductive leap from time-lagged correlations to causality.
This paradigm is particularly appealing because it is easy
to calculate in practice. Sample applications include determining which neuron assemblies Granger-cause other neuron
assemblies to fire synapses [7] or finding that exchange rates
Granger-cause stock market prices in Asia [8].
Pearl’s SEM approach to causality models the statistical
relationship between random elements with a Bayesian network [9]. Bayesian networks are directed acyclic graphs,
such that the distribution of a random element at node i
only depends on the values taken at the parent nodes. This
is meant to model causal relationships between nodes in the
graph. Pearl defines the imputation operator as follows: if one
imputes at a node i, one disconnects i from all its parents and
deterministically sets its value to some fixed, predetermined
constant. We will be building on this approach in this paper,
so we will defer the formal development of Pearl’s SEM
until Section III.
At a high level, the imputation operator captures a lot of
our intuitions about how the subjunctive conditional should
function. When one says If it had rained today, I would have
brought my umbrella, what does one mean? Intuitively, one
often means: ‘If everything else were the same, only it is
the case that it is raining today instead of sunny, these are
the actions I would have taken.’ One does not mean that the
world is structured in a way such that the necessary processes
to induce rain today were instead the case. In other words:
causal imputation does not travel upstream, e.g. backwards
through time. This is captured in Pearl’s SEM.
More practically, consider the question: What are the
causal effects of this medication? If we wish to estimate
this, we should ‘set’ medication taken to TRUE, and see
the consequences of this imputation. If we do not explicitly
‘set’ this value, then the decision to take medication is a
consequence of preceding factors. This makes it difficult
to determine if the observed effects are a result of the
medication or some other confounding variables1 . Again, this
will be more formally discussed in Section III.
Thus, we can think of these paradigms in terms of the
central phenomenon it is designed to model. In summary:
1) Rubin causality is focused on the estimation of the
counterfactual.
2) Granger causality is focused on the explanatory power
one process provides over another process.
3) Pearl’s SEM is focused on the causal effects of the
imputation operator.
Throughout this paper, we use Pearl’s SEM. However, we
note again that oftentimes problems framed in the Rubin
causality or Granger causality paradigm often can be translated to an equivalent formulation in SEM.
A. Notation
For any set A, we denote the powerset of A as 2A , which
can also be thought of as the set of functions mapping
A → {0, 1}. For a collection
of sets {Ai }i∈I , we denote
Q
the Cartesian product as i∈I Ai .
Also, I will denote the identity matrix, where context will
often be sufficient to determine its dimensions.
We let U [a, b] denote the uniform distribution on the interval [a, b] and N (µ, Σ) to denote the multivariate Gaussian
distribution with mean µ and covariance matrix Σ.
III. C AUSAL F RAMEWORK
In this section, we introduce our framework for modeling
causal effects, and then define the problem of optimal causal
imputation.
A. Causal structure
We build on the structural equation modeling framework
presented in [9]. First, we will introduce Bayesian networks.
Definition 1. A directed graph G = (V, E) is a set of nodes
V and a set of edges E ⊂ V × V . Throughout this paper
we will assume V is at most countably infinite.
A path from v0 ∈ V to vN ∈ V is a finite sequence of
edges (v0 , v1 ), (v1 , v2 ), . . . , (vN −1 , vN ) ∈ E.
We define the parents of node i as pa(i) = {j : (j, i) ∈ E}.
We can iterate this relationship to define the ancestor
relationship: let pan (i) = {j : k ∈ pan−1 (i), (j, k) ∈ E},
where pa1 (i) = pa(i) defined above. Then, the ancestors of
n
a node i are given by anc(i) = ∪∞
n=1 pa (i).
We say j is a descendant of i if i ∈ anc(j).
A directed graph is acyclic if i ∈
/ anc(i) for every i ∈
V . We will refer to such graphs as directed acyclic graphs
(DAGs).
1 We note that similar reasoning can be done in the Rubin causality
formulation as well.
Definition 2. A random process X indexed by a set V is
a collection of random elements (Xi )i∈V . Q
We will let Xi
denote the possible values of Xi , and X = i∈V Xi .
When there is an associated graph G = (V, E), we will
use the notation pa(Xi ) to denote the tuple (Xj )j∈pa(i) .
Definition 3. A random process X indexed by V is Markov
relative to a DAG G = (V, E) if its distribution factorizes:
Y
P (X) =
P (Xi |pa(Xi ))
i∈V
We can also say that X and G are compatible, or G
represents X.
This formalization will serve as our model for causality.
The interpretation is that if there is an edge going from i to
j, then Xi causes Xj .
Throughout this paper, we will treat the causal structure
G = (V, E) as given. Estimation of this causal structure
is a non-trivial task, and an active topic of research. Some
approaches to the task of causal inference include: using
metrics like directed information to estimate the causal
strength between random variables [10], [11], graphicalmodel based methods for estimating structure between random variables [12], [13], [14], [15], and regression based
approaches [16], [17], [18]. Again, this list is far from
exhaustive as an extensive literature review of this general
field is outside the scope of this paper. For a broader
overview of various approaches to the problem of causal
inference, see [12], [9].
Although the estimation of causal structures is never a
simple task, the growing field of research promises more
and more applications in which accurate estimation of causal
structures is feasible.
Previous work has focused on the estimation of causal
structures. In contrast, our contribution is to consider the
problem of control of causal structures. In other words, once
we are given a causal structure, how can we impute causal
effects to drive the overall system into a desirable state?
For example, once we can estimate the causal effects of
issuing rebates for energy-efficiency appliances, how do we
best distribute these rebates to induce more energy-efficient
consumption patterns? To the best of our knowledge, this is
the first paper to consider the problem of when and where
to impute on a causal structure.
There is an equivalent formulation of the condition in Definition 3 which utilizes disintegration results in probability
theory. This is referred to as the structural equation modeling
framework in [9].
Proposition 1. [19], [9] A random process X indexed by V
is Markov relative to G = (V, E) if and only if there exists
a collection of functions (fi )i∈V and independent random
elements (ξi )i∈V such that:
Xi = fi (pa(Xi ), ξi )
(1)
Furthermore, if Xi are Borel spaces2 , then ξi can be taken
2A
measurable space S is Borel if there exists a measurable function
S → [0, 1] with a measurable inverse.
to be U [0, 1].
We note that Borel spaces are a very general category
of measurable spaces: they include Polish spaces equipped
with the Borel σ-algebra3 . This includes finite sets, R, Rn ,
Lp (Rn ), the set of p-integrable functions defined on Rn .
Additionally, the space of probability distributions on any
Borel space is also a Borel space.
Assumption 1. Throughout the rest of this paper, we will
always use X to denote a random process indexed by V
that is Markov relative to a DAG G = (V, E), where Xi
takes values in Xi . Similarly, fi shall denote the functions
as specified in Equation 1, and similarly ξi .
B. Causal imputation
In this section, we will formally define the causal imputation operation. Intuitively, imputation of X produces a new
random process Y . This random process Y is equal to X
prior to the causal imputation, is forced to some value at the
node of imputation, and experiences causal effects after the
node of imputation. This is formally defined below.
Definition 4. [9] A random process Y indexed by V is the
imputation of X at i ∈ V to a constant xi ∈ Xi if:
• Yi = xi .
• For any j that is not a descendant of i, Yj = Xj .
• For any j that is a descendant of i, Yj = fj (pa(Yj ), ξj ).
If this is the case, we will write Y = do(X; i, xi ).
The imputation operator produces a copy of the original
process that is exactly equal at all nodes that do not causally
depend on the node of imputation Xi . At the point of
imputation, the node is disconnected from its parents and
forced a constant value xi . The nodes Xj that causally
depend on Xi are replaced with new values that depend
on the causal effects of Xi , keeping the innovation terms
ξ constant throughout.
Referring back to the discussions in Section II, this can be
thought of as manually setting the value of Xi to xi . This
should be something that is done exogenously, as a control
variable, rather than as a consequence of endogenous factors:
this is why Yi is disconnected from pa(Yi ).
From this definition, it immediately follows that the imputation operator commutes.
Proposition 2. Let i, j ∈ V such that i 6= j and
xi ∈ Xi and xj ∈ Xj . Then do(do(X; i, xi ); j, xj ) =
do(do(X; j, xj ); i, xi ) almost surely.
This allows us to define imputation on any set of nodes,
rather than just at a single node.
Q
Definition 5. For any I ⊂ V and xI ∈
i∈I Xi , we
define the imputation Y = do(X; I, xI ) as the sequential
application of element-wise do operations. This is almost
surely unique by Proposition 2.
3 A topological space T is Polish if it is separable and completely
metrizable. The Borel σ-algebra of a topological space is the smallest σalgebra containing all the open sets.
C. Optimal causal imputation
In the previous section, we defined the causal imputation
operator. We can think of our system designer as having the
capacity of issuing control commands that have causal effects
on the system downstream. When we can define the cost of
imputation as well as a control objective, we can formulate
the optimal causal imputation problem.
We suppose we are
Q given a collection of functions (cI )I⊂V
where each cI :
i∈I Xi → R. These functions can be
interpreted as the cost of imputation at a set of nodes I ⊂ V .
Drawing on our running example, c represents the cost
of issuing rebates for eco-friendly refrigerators at a set of
households.
Furthermore, we suppose we are given an operational
objective in the form of a cost function g : X → R. For
example, g can be a penalty on energy-wasting consumption
patterns.
Definition 6. The problem of optimal causal imputation is
given by:
min
min
Q
I⊂V xI ∈
i∈I
Xi
subject to
cI (xI ) + EY [g(Y )]
(2)
Y = do(X; I, xI )
(3)
IV. A PPLICATIONS
In Section III, we defined the optimal causal imputation
problem in its full generality. In this section, we shall provide
methods to solve the optimal causal imputation problem in
special cases. In particular, we consider two contexts: 1)
situations where imputation is only allowed to a single value,
2) situations where the dynamics are linear-Gaussian. In both
instances, we shall assume Xi = Rni for some ni .
A. Single-value case
In many applications where we can causally impute values,
we can only impute to one particular value. For example,
when issuing incentives, we may be able to only offer one
form of rebate to consumers. Motivated by this context,
we consider situations where the optimal causal imputation
problem can be reduced to one of submodular optimization.
Assumption 2. In this section, we assume V is a finite
set and that for each I ⊂ V , there exists an xI such that
cI (xI ) < ∞ and cI (x′I ) = ∞ for any x′I 6= xI . We shall
refer to this as the single-value case.
In the single-value case, we use the shorthand F (I) =
c(I) + E[g(do(X; I))], where we drop dependencies on x as
it can only take a single value.
1) Submodular minimization:
Definition 7. The set mapping F : 2V → R is submodular
if for any I1 ⊂ I2 ⊂ V and i ∈ G \ I2 , we have:
F (I1 ∪ {i}) − F (I1 ) ≥ F (I2 ∪ {i}) − F (I2 )
(4)
Intuitively, this definition is motivated by economies of
scale. We often expect economies of scale from these imputations, e.g. the per-customer cost of a rebate is non-increasing
as the number of customers increases, due to bulk-purchase
discounts. In our running example, the additional cost of
issuing a rebate to customer i is higher when you have issued
few rebates than when you have issued a lot of rebates.
(When I1 ⊂ I2 , then I2 corresponds to the situation where
you have issued more rebates than I1 .)
From a combinatorial optimization perspective, submodularity is a very well-behaved property that makes optimization, or approximate optimization, very tractable. We shall
quickly outline the details now, but we refer the interested
reader to [20] for more details.
First, note that there is a very direct correspondence
between a subset I ⊂ V and a tuple in {0, 1}V . For example,
if V = {0, 1, 2}, then (0, 1, 1) corresponds to the subset
{1, 2}. Thus, we can think of F : {0, 1}V → R. Now, we
define the Lovász extension [21].
Definition 8. Let λ ∼ U [0, 1]. Then, for any set mapping F :
{0, 1}V → R, we define the Lovász extension f : [0, 1]V →
R as:
f (z) = Eλ [F ({i : zi > λ})]
For the rest of this section, an unindexed f will denote the
Lovász extension of F .
We note two nice properties of the Lovász extension
immediately.
Proposition 3. [21] For any z ∈ {0, 1}V , we have f (z) =
F (z).
Proposition 4. [21] F is submodular if and only if f is
convex.
Note that the optimal causal imputation problem can be
written as:
min F (z)
z∈{0,1}V
The Lovász extension provides us with an easy solution to
the problem.
Proposition 5. [21] If F is submodular, then the following
is a convex optimization program.
min f (z)
z∈[0,1]V
(5)
Furthermore, there exist minimizers of (5) in {0, 1}V .
In other words, the combinatorial optimization problem
can be solved tractably with convex optimization if F is submodular. Thus, we are motivated in searching for conditions
under which F (I) = c(I) + E[g(do(X; I))] is submodular.
We provide a common sufficient condition for submodularity
of F in the following theorem:
Theorem 1. If:
2
• g(Y ) = kYi − EYi k2 for some i ∈ V .
ξ
• There exists functions fj such that, if Xj has no parents,
ξ
Xj = fj (ξj ) and otherwise Xj = pa(Xj ) + fjξ (ξj ).
• For each j ∈ anc(i), there exists one unique path from
j to i.
c(I) is submodular.
Then F (I) = c(I) + E[g(do(X; I))] is submodular.
•
Note here that we treat pa(Xi ) as a vector in Rni ,
where ni is the appropriate dimension. These assumptions
encompass many graphical models where a node’s parents set
a location parameter, and the control objective is the second
moment of some feature.
Proof. Note that the desired result will follow if we show
that the set mapping G : I 7→ E[g(do(X; I))] is submodular, since the sum of submodular functions is submodular.
Throughout this proof, we use i to refer to the index i pulled
out by the function g.
We can see that G(∅) = E[g(X)]. By the independence of
the (ξi )i∈V and the form ofP
the (Xi )i∈V relationships, we
can write this as E[g(X)] = j∈anc(i) kfjξ (ξj )− Efjξ (ξj )k22 .
(Note that the unique path assumption ensures that each
variance is only counted once in this sum.)
More generally, we can write an expression for G(I). Note
that if we impute at a node j, all the uncertainty due to
node j, and the ancestorsPof j, is zeroed out. Thus, we can
write G(I) = E[g(X)]− j∈(I∪anc(I)) kfjξ (ξj )−Efjξ (ξj )k22 ,
where we define anc(I) = ∪j∈I anc(j).
Now, we can verify the submodularity condition on G.
Pick I1 ⊂ I2 and i′ ∈ V \ I2 . Then:
G(I1 ∪ {i′ }) − G(I1 ) =
X
kfjξ (ξj ) − Efjξ (ξj )k22 −
j∈(I1 ∪anc(I1 ))
X
kfjξ (ξj ) − Efjξ (ξj )k22 =
j∈(I1 ∪{i′ }∪anc(I1 ∪{i′ }))
−
X
kfjξ (ξj ) − Efjξ (ξj )k22
j∈{i′ }∪(anc(i′ )\anc(I1 ))
In words, the change in G due to adding i′ to I1 is the
variances due to the terms related to i′ and the ancestors of
i′ that have not already been zeroed out due to imputation,
i.e. the ancestors of i′ that are not already ancestors of I1 .
A similar derivation can be done for I2 .
Thus, we can verify that G(I1 ∪ {i′ }) − G(I1 ) ≥ G(I2 ∪
′
{i }) − G(I2 ) by noting that anc(i′ ) \ anc(I2 ) ⊂ anc(i′ ) \
anc(I1 ), so the right-hand side of the inequality adds more
negative terms. This concludes our proof.
2) Submodular maximization: Alternatively, suppose we
are attempting to maximize a submodular function subject
to a constraint, i.e. F (I) = c(I) + E[g(do(X; I))] subject to
a constraint that I ∈ S ⊂ 2V and our objective is to solve
maxI∈S F (I).4
First, consider the greedy method for submodular maximization. This is presented as Algorithm 1. At each iteration,
it simply adds an element to I which maximizes F (I ∪ {i}),
if one exists. If one does not exist, it terminates and returns
4 Strictly speaking, to remain consistent with the problem in Section III,
we should be solving minI∈S −F (I), but we express it as a maximization
for clarity of presentation.
Algorithm 1 The greedy approach for combinatorial maximization.
I←∅
while maxi:I∪{i}∈S F (I ∪ {i}) − F (I) ≥ 0 do
Pick i∗ ∈ arg maxi:I∪{i}∈S F (I ∪ {i})
I ← I ∪ {i∗ }
end while
return I
I. Under certain structural conditions, this algorithm yields
approximate optimizers.
Definition 9. A set mapping F : 2V → R is nondecreasing
if F (S) ≤ F (T ) whenever S ⊂ T .
The monotonicity condition effectively prevents the algorithm from straying too far from the optimum when taking
the greedy approach, as shown in [22]. Note that if F is nondecreasing, then the condition maxi:I∪{i}∈S F (I ∪ {i}) −
F (I) ≥ 0 is equivalent to the existence of i ∈ V such that
I ∪ {i} ∈ S.
Proposition 6. [22] If F is nondecreasing and submodular,
then the greedy method presented in Algorithm 1 will return
I ∗ ∈ S such that F (I ∗ ) ≥ e−1
maxI∈S F (I).
e
We now present a quick corollary of Theorem 1, which
provides conditions under which we can leverage the existing results for maximization of nondecreasing submodular
functions.
Corollary 1. If:
′
2
• g (Y ) = −kYi − EYi k2 for some i ∈ V .
ξ
• There exists functions fj such that, if Xj has no parents,
Xj = fjξ (ξj ) and otherwise Xj = pa(Xj ) + fjξ (ξj ).
• For each j ∈ anc(i), there exists one unique path from
j to i.
• c(I) is nondecreasing and submodular.
Then F (I) = c(I) + E[g ′ (do(X; I))] is nondecreasing and
submodular.
Proof. This follows from Theorem 1 if we can show that G′ :
I 7→ E[g(do(X; I))] is nondecreasing. Let Y = do(X; I),
and note that adding elements to I can only decrease the
variance of Yi . This can be formalized by noting, similar
to the arguments
in the proof of Theorem 1, G′ (I) =
P
′
E[g (X)]+ j∈(I∪anc(I)) kfjξ (ξj )−Efjξ (ξj )k22 . Thus, G′ , the
additive inverse of the variance of Yi , is nondecreasing.
Note the minus sign in g ′ in Corollary 1: in most instances
where you are maximizing a submodular cost, you would still
wish to reduce uncertainty, i.e. have a lower variance.
B. Linear-Gaussian case
In this section, we consider causal imputation on a
discrete-time linear dynamical system with Gaussian noise.
That is, we analyze the special case of a random process
with the form:
Xt+1 = AXt + ǫt
Where Xt ∈ Rn , ǫt ∼ N (0, σ 2 I) independently for
t = 0, ..., T , and A ∈ Rn×n is a matrix representing the
dependencies.
This process can be represented as a causal graph in the
form of a trellis, where the random variables are all Gaussian.
More specifically, each node has its expected value equal to a
linear combination of their parents, as described by a matrix
A, and additive noise of the distribution N (0, σ 2 ).
To analyze our optimal casual imputation problem, we first
redefine the indices for this problem. Since our causal graph
represents a process over time, we index into the process by
state k, for k = 1, ...n as well as a time t for t = 0, ..., T .
Thus Xkt indicates the value of state k at time t, and our
graph has vertices V = {1, . . . , n} × {0, . . . , T }. As before,
Xt represents the value of the vector of all the states of X
at time t, and we can think of X as a vector in RnT . We
assume that the cost of imputation cI (xI ) has the following
form for some parameters δi , qi ≥ 0:
X
cI (xI ) =
δi + qi x2i
i∈I
Further, we look at the case where the system cost of
interest is minimizing the expected distance of the the
random process from some target trajectory ȳ. Thus g(Y ) =
kY − ȳk22 .
Our optimal causal imputation problem in this case is thus:
X
min min
δi + qi x2i + E kY − ȳk22
S⊂V xS ∈RS
(6)
i∈S
subject to Y = do(X; S, xS )
The summation term can be thought of as a cost of issuing
control commands and the expectation term can be thought
of as a trajectory tracking objective.
Given our structure on the random process, we can rewrite
this optimization problem more concretely.
We first define Q ∈ RnT ×nT to be diagonal matrix with
the qi ’s on the diagonal. We define δ ∈ RnT to be the vector
of δi ’s. Further, let 1nT denote the column vector of all ones
in RnT . Lastly, we define diag(S) to be the square matrix
with the elements of S on the diagonal, and zeros everywhere
else.
The optimization in (6) now becomes:
min
S∈{0,1}nT
x̄∈RnT
subject to
x̄⊤ (Q + D)x̄ + σ 2 trace(DIS ) + δ ⊤ S − 2ȳ ⊤ x̄
(Si − 1)x̄i = 0 for all i = 1, . . . , nT
P = (1 − Ã)−1
D = P ⊤P
IS = I − diag(S)
0 0 0 ... 0 0
A 0 0 . . . 0 0
0 A 0 . . . 0 0
à = IS
0 0 A . . . 0 0
.
.. .. . .
.. ..
..
.
. .
. .
0 0 0 ... A 0
We note that for any matrix A and any S, the matrix I − Ã,
with à as defined above, is invertible, so P will always be
well-defined.
Additionally, for a fixed S, the optimization across x̄ is
easy to solve. That is, D is entirely determined by S. If we
let (Q + D)S denote the submatrix of (Q + D) indexed by
the non-zero elements of S, and similarly x̄S and ȳS , then
the optimizer is given by x̄∗S = (Q + D)−1
S ȳS , with the other
entries of x̄∗ equal to 0.
Thus, we can easily calculate a set mapping F (S) such
that optimal causal imputation in the linear-Gaussian case
is simply minS⊂V F (S). We can solve this when nT is
relatively small, and are currently investigating properties
of F (S) which would allow us to apply combinatorial
optimization techniques [20].
V. C ONCLUSION
AND
F UTURE W ORK
The previous literature on mathematical formulations of
causality has been focused on the estimation of causal
structures. In this paper, we presented the problem of control
of causal structures. We formally defined the problem of
optimal causal imputation, and formulate solutions for it in
two cases: where imputation is allowed to only a single value,
and the case where the dynamics are linear and the noise is
Gaussian.
In future work, we hope to apply this framework to
real situations which allow both the estimation of causal
structures, as well as verification of the consequences and
costs of imputation. Additionally, we hope to generalize
our results to consider dynamical systems whose behavior
are influenced by different features. For example, we can
consider the dynamics of the power grid, but also account
for frequently used machine learning features as well, such
as the zip code of different energy consumers and the age
of deployed assets.
We believe that considering the control aspects of causality
is increasingly more relevant. In many smart infrastructure
applications, we no longer have control commands that
directly affect the dynamics, but rather our control actions act
more like causal imputations. The optimal causal imputation
framework is a promising direction to model these interactions between machine learning and control, and provides
a model for closing the loop on analytics in cyber-physical
systems.
R EFERENCES
[1] D. B. Rubin, “Estimating causal effects of treatments in randomized
and nonrandomized studies,” Journal of Educational Psychology, vol.
66(5), pp. 688–701, 1974.
[2] G. W. Imbens and D. B. Rubin, Causal Inference for Statistics, Social,
and Biomedical Sciences: An Introduction. Cambridge University
Press, 2015.
[3] S. Athey and G. W. Imbens, “Machine learning methods for estimating
heterogeneous causal effects,” arXiV, 2015.
[4] D. Zhou, M. Balandat, and C. J. Tomlin, “Residential demand response
targeting using machine learning with observational data,” in 55th
IEEE Conference on Decision and Control (CDC), 2016.
[5] A. Banerjee, E. Duflo, N. Goldberg, D. Karlan, R. Osei, W. Parienté,
J. Shapiro, B. Thuysbaert, and C. Udry, “A multifaceted program
causes lasting progress for the very poor: Evidence from six
countries,” Science, vol. 348, no. 6236, 2015. [Online]. Available:
http://science.sciencemag.org/content/348/6236/1260799
[6] C. W. J. Granger, “Investigating causal relations by
econometric models and cross-spectral methods,” Econometrica,
vol. 37, no. 3, pp. 424–438, 1969. [Online]. Available:
http://www.jstor.org/stable/1912791
[7] A. Brovelli, M. Ding, A. Ledberg, Y. Chen, R. Nakamura, and S. L.
Bressler, “Beta oscillations in a large-scale sensorimotor cortical
network: Directional influences revealed by Granger causality,”
Proceedings of the National Academy of Sciences of the United States
of America, vol. 101, no. 26, pp. 9849–9854, 06 2004. [Online].
Available: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC470781/
[8] C. W. Granger, B.-N. Huang, and C.-W. Yang, “A bivariate
causality between stock prices and exchange rates: Evidence
from recent Asian flu,” The Quarterly Review of Economics and
Finance, vol. 40, no. 3, pp. 337 – 354, 2000. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S1062976900000429
[9] J. Pearl, Causality: Models, Reasoning and Inference. Cambridge
University Press, 2009.
[10] C. Gourieroux, A. Monfort, and E. Renault, “Kullback causality
measures,” Annales d’conomie et de Statistique, no. 6/7, pp. 369–410,
1987. [Online]. Available: http://www.jstor.org/stable/20075662
[11] P. Amblard and O. J. J. Michel, “The relation between
granger
causality
and
directed
information
theory:
a
review,” CoRR, vol. abs/1211.3169, 2012. [Online]. Available:
http://arxiv.org/abs/1211.3169
[12] J. Pearl, Graphical Models for Probabilistic and Causal Reasoning.
Dordrecht: Springer Netherlands, 1998, pp. 367–389.
[13] S. L. Lauritzen, “Causal inference from graphical models,” 2001.
[14] J. Li, S. Ma, T. D. Le, L. Liu, and J. Liu, “Causal decision
trees,” CoRR, vol. abs/1508.03812, 2015. [Online]. Available:
http://arxiv.org/abs/1508.03812
[15] S. Athey and G. Imbens, “Recursive Partitioning for Heterogeneous
Causal Effects,” ArXiv e-prints, Apr. 2015.
[16] A. P. Dawid, “Causal inference without counterfactuals,” Journal of
the American Statistical Association, vol. 95, no. 450, pp. 407–424,
2000. [Online]. Available: http://www.jstor.org/stable/2669377
[17] D. Heckerman, C. Meek, and G. Cooper, A Bayesian Approach to
Causal Discovery. Berlin, Heidelberg: Springer Berlin Heidelberg,
2006, pp. 1–28.
[18] P. O. Hoyer, D. Janzing, J. M. Mooij, J. Peters, and
P. B. Schölkopf, “Nonlinear causal discovery with additive
noise models,” in Advances in Neural Information Processing
Systems 21, D. Koller, D. Schuurmans, Y. Bengio, and L. Bottou,
Eds. Curran Associates, Inc., 2009, pp. 689–696. [Online]. Available:
http://papers.nips.cc/paper/3548-nonlinear-causal-discovery-with-additive-noise-models.pdf
[19] O. Kallenberg, Foundations of Modern Probability. Springer, 2002.
[20] A. Schrijver, Combinatorial Optimization: Polyhedra and Efficiency.
Springer-Verlag Berlin Heidelberg, 2003.
[21] L. Lovász, Submodular functions and convexity. Springer, 1983, pp.
235–257.
[22] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, “An analysis of
approximations for maximizing submodular set functions,” Mathematical Programming, vol. 14, no. 1, pp. 265–294, 1978.
| 3cs.SY
|
Block-Diagonal Coding for Distributed Computing
With Straggling Servers
Albin Severinson† ‡ , Alexandre Graell i Amat† , and Eirik Rosnes‡
arXiv:1701.06631v3 [cs.IT] 15 Sep 2017
†Department of Electrical Engineering, Chalmers University of Technology, Gothenburg, Sweden
‡Simula@UiB, Bergen, Norway
Abstract—We consider the distributed computing problem of
multiplying a set of vectors with a matrix. For this scenario, Li
et al. recently presented a unified coding framework and showed
a fundamental tradeoff between computational delay and communication load. This coding framework is based on maximum
distance separable (MDS) codes of code length proportional to
the number of rows of the matrix, which can be very large. We
propose a block-diagonal coding scheme consisting of partitioning
the matrix into submatrices and encoding each submatrix using a
shorter MDS code. We show that the assignment of coded matrix
rows to servers to minimize the communication load can be
formulated as an integer program with a nonlinear cost function,
and propose an algorithm to solve it. We further prove that, up
to a level of partitioning, the proposed scheme does not incur
any loss in terms of computational delay (as defined by Li et al.)
and communication load compared to the scheme by Li et al..
We also show numerically that, when the decoding time is also
taken into account, the proposed scheme significantly lowers the
overall computational delay with respect to the scheme by Li et
al.. For heavy partitioning, this is achieved at the expense of a
slight increase in communication load.
I. I NTRODUCTION
Distributed computing has emerged as one of the most
effective ways of tackling increasingly complex computation
problems. One of the main application areas is large-scale
machine learning and data analytics. Google routinely performs computations using several thousands of servers in their
MapReduce clusters [1]. Distributed computing systems bring
significant challenges. Among them, the problems of straggling servers and bandwidth scarcity have recently received
significant attention. The straggler problem is a synchronization problem characterized by the fact that a distributed
computing task must wait for the slowest server to complete its
computation. On the other hand, distributed computing tasks
typically require that data is moved between servers during the
computation, the so-called data shuffling, which is a challenge
in bandwidth-constrained networks.
Coding for distributed computing to reduce the computational delay and the communication load between servers has
recently been considered in [2], [3]. In [2], a structure of
repeated computation tasks across servers was proposed, enabling coded multicast opportunities that significantly reduce
the required bandwidth to shuffle the results. In [3], the authors
showed that maximum distance separable (MDS) codes can be
This work was funded by the Swedish Research Council under grant 201604253 and the Research Council of Norway under grant 240985/F20.
applied to a linear computation task (e.g., multiplying a vector
with a matrix) to alleviate the effects of straggling servers
and reduce the computational delay. In [4], a unified coding
framework was presented and a fundamental tradeoff between
computational delay and communication load was identified.
The ideas of [2], [3] can be seen as particular instances of the
framework in [4], corresponding to the minimization of the
communication load or the computational delay.
The distributed computing problem addressed in [4] is a matrix multiplication problem where a set of vectors x1 , . . . , xN
are multiplied by a matrix A. In particular, as in [3], the
authors suggest the use of MDS codes, whose dimension is
equal to the number of rows of A, to generate some redundant
computations. In practice, the size of A can be very large.
For example, Google performs matrix-vector multiplications
with matrices of dimension of the order of 1010 × 1010 when
ranking the importance of websites [5]. Since the decoding
complexity of MDS codes on the packet erasure channel is
quadratic (for Reed-Solomon (RS) codes) in the code length
[6], for very large matrix sizes the decoding complexity may
be prohibitively high.
In this paper, we introduce a block-diagonal encoding
scheme for the distributed matrix multiplication problem. The
proposed encoding is equivalent to partitioning the matrix and
applying smaller MDS codes to each submatrix separately. The
storage design for the proposed block-diagonal encoding can
be cast as an integer optimization problem with a nonlinear
objective function, whose computation scales exponentially
with the problem size. We propose a heuristic solver for
efficiently solving the optimization problem, and a branchand-bound approach for improving on the resulting solution
iteratively. We exploit a dynamic programming approach to
speed up the branch-and-bound operations. We prove that up
to a certain partitioning level, partitioning does not increase the
computational delay (as defined in [4]) and the communication
load with respect to the scheme in [4]. Interestingly, when
the decoding time is taken into account, the proposed scheme
achieves an overall computational delay significantly lower
than the one of the scheme in [4]. This is due to the fact
that the proposed scheme allows using significantly shorter
MDS codes, hence reducing the decoding complexity and the
decoding time. Also, numerical results show that a high level
of partitioning can be applied at the expense of only a slight
increase in the communication load.
II. S YSTEM M ODEL
We consider the problem of multiplying a set of vectors
with a matrix. In particular, given a matrix A ∈ Fm×n and
N vectors x1 , . . . , xN ∈ Fn , where F is some field, we want
to compute the N vectors y1 = Ax1 , . . . , yN = AxN . The
computation is performed in a distributed fashion using K
servers, S1 , . . . , SK , where each server stores µm matrix rows,
1
≤ µ ≤ 1. Prior to distributing the rows among
for some K
the servers, A is encoded by an r × m encoding matrix Ψ,
resulting in the coded matrix C = ΨA, of size r × n, i.e.,
the rows of A are encoded using an (r, m) linear code with
r ≥ m.
Let q = K m
r , where we assume that r divides Km and
hence q is an integer.
cr,
The r coded rows of C, c1 , . . . , K
K
are divided into µq
disjoint batches, each containing r/ µq
coded rows. Each batch is assigned to µq servers. Correspondingly, a batch B is labeled by a unique set T ⊂ {S1 , . . . , SK },
of size |T | = µq, denoting the subset of servers that store that
batch, and we write BT . Server Sk , k = 1, . . . , K, stores the
coded rows of BT if and only if Sk ∈ T .
A. Probabilistic Runtime Model
We adopt the probabilistic model of the computation runtime of [3]. We assume that running a computation on a single
server takes a random amount of time according to the shiftedexponential cumulative probability distribution
(
t
1 − e−( σ −1) , for t ≥ σ
F (t) =
,
0,
otherwise
where σ is the number of multiplications and divisions required to complete the computation. We do not take addition
and subtraction into account as those operations are orders of
magnitude faster [7].
When the computation is split into K parallel subtasks
running on separate servers, we assume that the runtimes
of the subtasks are independent and identically distributed
random variables with distribution F (Kt) [3]. Furthermore,
the runtime of the g-th, g = 1, . . . , K, fastest server to
complete its subtask is given by the g-th order statistic, F(g) ,
with expectation [8]
K
X
1
f (σ, K, g) , E F(g) = σ 1 +
.
j
j=K−g+1
B. Distributed Computing Model
We consider the MapReduce framework described in [4],
where we assume that the input vectors x1 , . . . , xN are known
to all servers. The overall computation then proceeds in three
phases: map, shuffle, and reduce.
1) Map Phase: In the map phase, we compute in a distributed fashion coded intermediate values, which will be later
used to obtain vectors y1 , . . . , yN . Server S multiplies the
input vectors xj , j = 1, . . . , N , by all the coded rows of
matrix C it stores, i.e., it computes
(S)
Zj
, {cxj : c ∈ {BT : S ∈ T }}, j = 1, . . . , N.
The map phase terminates when a set of servers G ⊆
{S1 , . . . , SK } that collectively store enough values to decode
the output vectors have finished their map computations. We
denote the cardinality of G by g. The (r, m) linear code
proposed in [4] is an MDS code for which y1 , . . . , yN can
be obtained from any subset of q servers, i.e., g = q.
We define the computational delay of the map phase
as its average runtime per source row and vector y, i.e.,
σmap
1
f( K
, K, g). Dmap is referred to simply as the
Dmap = mN
computational delay in [4]. As all K servers compute µm
inner products, each requiring n multiplications for each of
the N input vectors, we have σmap = KµmnN .
After the map phase, the computation of y1 , . . . , yN proceeds using only the servers in G. We denote by Q ⊆ G the
set of the first q servers to complete the map phase. Each
of the q servers in Q is responsible to compute N/q of the
vectors y1 , . . . , yN . Let WS be the set containing the indices
of the vectors y1 , . . . , yN server S ∈ Q is responsible for. The
remaining servers in G assist the servers in Q in the shuffle
phase.
2) Shuffle Phase: In the shuffle phase, intermediate values
calculated in the map phase are exchanged between servers in
G until all servers in Q hold enough values to compute the
vectors they are responsible for. As in [4], we allow creating
and multicasting coded messages that are simultaneously useful for multiple servers. For a subset of servers S ⊂ Q and
S ∈ Q\S, we denote the set of intermediate values needed by
(S)
server S and known exclusively by the servers in S by VS .
More formally,
(S)
VS
, {cxj : j ∈ WS and c ∈ {BT : T ∩ Q = S}}.
K−q
Pµq
(q−1
j )(µq−j )
and sq , inf (s : l=s αl ≤ 1 − µ).
q K
K (µq )
For each j ∈ {µq, µq − 1, . . . , sq }, and every subset S ⊆ Q
of size j + 1, the shuffle phase proceeds as follows.
Let αj ,
(S)
1) For each S ∈ S, we evenly and arbitrarily split VS\S
(S)
(S)
into j disjoint segments VS\S = {VS\S,S̃ : S̃ ∈ S \ S},
(S)
and associate the segment VS\S,S̃ with server S̃ ∈ S \S.
2) Server S̃ ∈ S multicasts the bit-wise XOR of all the
segments associated with it in S. More precisely, it
(S)
multicasts ⊕S∈S\S̃ VS\S,S̃ to the other servers in S \ S̃.
For every pair of servers S, S̃ 0 ∈ S, since server S has
(S )
computed locally the segments VS\S 0 ,S̃ for all S 0 ∈ S \{S̃, S},
it can cancel them from the message sent by server S̃, and
recover the intended segment. We finish the shuffle phase by
either unicasting any remaining needed values until all servers
in Q hold enough intermediate values to decode successfully,
or by repeating the above two steps for j = sq − 1, selecting
the strategy achieving the lower communication load.
Definition 1. The communication load, denoted by L, is the
number of messages per source row and vector y exchanged
during the shuffle phase, i.e., the total number of messages
sent during the shuffle phase divided by mN .
Depending on the strategy, the P
communication
load after
Pµq
α
α
µq
completing the multicast phase is j=sq jj or j=sq −1 jj
[4]. For the scheme in [4], the total communication load is
LMDS
µq
µq
µq
X
X
X
α
α
j
j
.
= min
+1−µ−
αj ,
j
j
j=s
j=s
j=s −1
q
q
(1)
(1)
(1)
(1)
(2)
(2)
c1 c5 c1 c3 c 5
(1) (2) (3) (3) (3)
c2 c6 c2 c4 c6
Server S2
(1)
(3)
(4)
(5)
(5)
c1 c3 c3 c1 c5
(2) (3) (4) (5) (5)
c2 c4 c4 c2 c6
Server S5
c1 c3 c5 c 1 c3
(1) (1) (1) (2) (2)
c2 c4 c6 c 2 c4
Server S1
c5 c1 c1 c1 c3
(1) (3) (4) (5) (5)
c6 c2 c2 c2 c4
Server S4
(1)
(2)
(3)
(3)
(3)
c3 c5 c1 c3 c5
(1) (2) (4) (4) (4)
c4 c6 c2 c4 c6
Server S3
(2)
(3)
(4)
(5)
(5)
c3 c5 c5 c3 c5
(2) (3) (4) (5) (5)
c4 c6 c6 c4 c6
Server S6
(1)
Definition 2. The overall computational delay, D, is the sum
of the map and reduce phase delays, i.e., D = Dmap +Dreduce .
III. B LOCK -D IAGONAL C ODING
We introduce a block-diagonal encoding matrix of the form
ψ1
..
Ψ=
,
.
ψT
r m
where ψ1 , . . . , ψT are Tr × m
T encoding matrices of an ( T , T )
MDS code, for some integer T that divides m and r. Note that
the encoding given by Ψ amounts to partitioning the rows
of A into T disjoint submatrices A1 , . . . , AT and encoding
each submatrix separately. We refer to an encoding Ψ with
T disjoint submatrices as a T -partitioned scheme, and to
the submatrix of C = ΨA corresponding to ψi as the ith partition. We remark that all submatrices can be encoded
using the same encoding matrix, i.e., ψi = ψ, i = 1, . . . , T ,
reducing the storage requirements, and encoding/decoding can
be performed in parallel if many servers are available. We
further remark that the case Ψ = ψ (i.e., the number of
partitions is T = 1) corresponds to the scheme in [4], which
we will sometimes refer to as the unpartitioned scheme.
A. Assignment of Coded Rows to Batches
(t)
For a block-diagonal encoding matrix Ψ, we denote by ci ,
t = 1, . . . , T and i = 1, . . . , r/T , the i-th coded row of C
(2)
within partition t. For example, c1 denotes the first coded
row of the second partition. As described in Section II, the
K
coded rows are divided into µq
disjoint batches. To formally
describe
the
assignment
of
coded
rows to batches we use a
K
×
T
integer
matrix
P
=
[p
],
where pi,j describes the
i,j
µq
number of rows from partition j that are stored in batch i.
Note that, due to the MDS property, any set of m/T rows of
a partition is sufficient to decode the partition. Thus, without
loss of generality, we consider a sequential assignment of rows
of a partition into batches. For example, for the assignment P
(2)
(4)
(4)
(4)
(2)
(3)
(4)
(5)
(5)
Fig. 1. Storage design for m = 20, N = 4, K = 6, q = 4, µ = 1/2, and
T = 5.
q
As in [4], we consider the cost of a multicast message to be
equal to that of a unicast message. In real systems, however,
it may vary depending on the network architecture.
3) Reduce Phase: Finally, in the reduce phase, the vectors
y1 , . . . , yN are computed. More specifically, server S ∈ Q
(S)
(S)
uses the locally computed sets Z1 , . . . , ZN and the received messages to compute the vectors in {yj : j ∈ WS }.
The computational delay of the reduce phase is its average
runtime per source row and output vector y, i.e., Dreduce =
σreduce
1
mN f ( q , q, q), where σreduce is given in Section IV-B.
(1)
(1)
(1)
(2)
(2)
c1 c3 c5 c1 c3
(1) (1) (1) (2) (2)
c2 c4 c6 c2 c4
Server S1
(1)
(1)
c1 x3 ⊕ c3 x2
(1)
c2 x3
(1)
(2)
(2)
⊕ c5 x 1
(3)
(3)
(1)
(2)
c4 x2 ⊕ c6 x1
(1)
(3)
c1 c5 c1 c3 c5
(1) (2) (3) (3) (3)
c2 c6 c2 c4 c6
Server S2
(2)
(4)
(4)
(4)
c3 c5 c1 c3 c 5
(1) (2) (4) (4) (4)
c4 c6 c2 c4 c 6
Server S3
Fig. 2. Multicasting coded values between servers S1 , S2 , and S3 .
(1)
(1)
in Example 1 (see (2)), rows c1 and c2 are assigned to batch
(1)
(1)
1, c3 and c4 are assigned to batch 2, and so on. The rows
of P are labeled by the subset of servers the corresponding
batch is stored at, and the columns are labeled by its partition
index. We refer to the pair (Ψ, P ) as the storage design. The
assignment matrix P must satisfy the following conditions.
1) The entries
P must sumK to
the batch
PTof each row of
K
size, i.e., j=1 pi,j = r/ µq
, 1 ≤ i ≤ µq .
2) The entries of each column of P must sum to the
K
P(µq
)
number of rows per partition, i.e., i=1
pi,j = Tr , 1 ≤
j ≤ T.
Example 1 (m = 20, N = 4, K = 6, q = 4, µ = 1/2,
T = 5). For these parameters, there are r/T = 6 coded rows
per partition,
of which m/T = 4 are sufficient for
decoding,
K
K
and µq
= 15 batches, each containing r/ µq
= 2 coded
rows. We construct the storage design shown in Fig. 1 with
assignment matrix
1
(S1 , S2 ) 2
(S1 , S3 ) 2
(S1 , S4 ) 2
0
P = (S1 , S5 )
..
.
(S , S ) 0
4
6
(S5 , S6 )
2
0
0
0
2
3
0
0
0
0
..
.
4
0
0
0
0
5
0
0
0
0
,
0 0 0 2
0 0 0 0 2
(2)
where rows are labeled by the subset of servers the batch is
stored at, and columns are labeled by the partition index. For
this storage design, any g = 4 servers collectively store at least
4 coded rows from each partition. However, some servers store
more rows than needed to decode some partitions, suggesting
that this storage design is suboptimal.
Assume G = {S1 , S2 , S3 , S4 } is the set of g = 4
servers that finish their map computations first. Also, assign
vector yi to server Si , i = 1, 2, 3, 4. We illustrate the
coded shuffling scheme for S = {S1 , S2 , S3 } in Fig. 2.
(1)
(1)
S1 multicasts c1 x3 ⊕ c3 x2 to S2 and S3 . Since S2 and
(1)
(1)
S3 can cancel c1 x3 and c3 x2 , respectively, both servers
receive one needed intermediate value. Similarly, S2 multicasts
(1)
(2)
(1)
(2)
c2 x3 ⊕ c5 x1 , while S3 multicasts c4 x2 ⊕ c6 x1 . This
process is repeated for S = {S2 , S3 , S4 }, S = {S1 , S3 , S4 },
and S = {S1 , S2 , S4 }. After the shuffle phase, we have sent
12 multicast messages and 30 unicast messages, resulting in
a communication load of (12 + 30)/20/4 = 0.525, a 50%
increase from the load of the unpartitioned scheme (0.35,
given by (1)). In this case, S1 received additional intermediate
values from partition 2, despite already storing enough, further
indicating that the assignment in (2) is suboptimal.
IV. P ERFORMANCE OF THE B LOCK -D IAGONAL C ODING
In this section, we analyze the performance impact of
partitioning. We have the following theorem.
K
there exists an assignment matrix
Theorem 1. For T ≤ r/ µq
P such that the communication load and computational delay
of the map phase are equal to those of the unpartitioned
scheme.
A. Communication Load
For the unpartitioned scheme of [4], G = Q, and the
number of remaining values that need to be unicasted after
the multicast phase is constant, regardless which subset Q of
servers finish first their map computations. However, for the
block-diagonal (partitioned) coding scheme, both g and the
number of remaining unicasts may vary.
For a given assignment P and a specific Q, we denote
(S)
by UQ (P ) the number of remaining values needed after
the multicast phase by server S ∈ Q, and by UQ (P ) ,
P
(S)
S∈Q UQ (P ) the total number of remaining values needed
q
by the servers in Q. Let QP
denote the superset of all sets Q
1
1
and define LQ , mN
Q∈Qq UQ (P ). Then, for a given
|Qq |
storage design (Ψ, P ), the communication load of the blockdiagonal coding scheme is given by
µq
µq
X
X
α
α
j
j
+ LQ ,
+ LQ ,
LBDC (Ψ, P ) = min
j
j
j=s −1
j=s
q
(3)
q
where LQ depends on the shuffling scheme (see Section II-B2)
and is different in the first and second term of the minimization
(S)
in (3). To evaluate UQ , we count the total number of
intermediate values that need to be unicasted to server S until
it holds m/T intermediate values from each partition.
For a given Ψ, the assignment of rows into batches can be
formulated as an optimization problem, where one would like
to minimize LBDC over all assignments P . More precisely, the
optimization problem is minP ∈P LBDC (Ψ, P ), where P is the
set of all assignments P , and where the dependence of LBDC
on P is nonlinear. This is a computationally complex problem
since both the complexity of evaluating the performance of
a given assignment and the number of assignments scale
exponentially in the problem size.
B. Computational Delay
We consider the delay incurred by both the map and reduce
phases (see Definition 2). Note that in [4] only Dmap is
considered, i.e., D = Dmap . However, one should not neglect
the computational delay incurred by the reduce phase. Thus,
one should consider D = Dmap + Dreduce . The reduce phase
consists of decoding the N output vectors and hence the delay
it incurs depends on the code and decoding algorithm. We
assume that each partition is encoded using an RS code and is
decoded using the Berlekamp-Massey algorithm. We measure
the decoding complexity by its associated shifted-exponential
parameter σ (see Section II-A).
The number of field multiplications required to decode an
(r/T, m/T ) RS code is (r/T )2 ξ [7], where ξ is the fraction
q
(the map
of erased symbols. With ξ upperbounded by 1 − K
q
phase terminates when a fraction of at least K symbols from
each partition is available) the complexity of decoding the T
partitions for all N output vectors is at most
q
r2 (1 − K
)N
.
(4)
T
The decoding complexity of the scheme in [4] is given by
evaluating (4) for T = 1. By choosing T close to r, we can
thus significantly reduce the delay of the reduce phase.
σreduce =
V. A SSIGNMENT S OLVERS
We propose two solvers for the problem of assigning rows
into batches: a heuristic solver that is fast even for large
problem instances, and a hybrid solver combining the heuristic
solver with a branch-and-bound solver. The branch-and-bound
solver produces an optimal assignment but is significantly
slower, hence it can be used as stand-alone only for small
problem instances. We use a dynamic programming approach
(S)
to speed up the branch-and-bound solver by caching UQ for
(S)
all Q ∈ Qq , indexed by the batches each UQ is computed
(S)
from. This way we only need to update the affected UQ when
assigning a row to a batch. For all solvers, we first label the
batches lexiographically and then optimize LBDC in (3). The
solvers are available under the Apache 2.0 license [9].
A. Heuristic Solver
The heuristic solver creates an assignment jmatrix
k
P in two
K
·T ,
steps. We first set each entry of P to γ , r/ µq
K
thus assigning the first µq
γ rows of each partition to batches
K
such that each batch is assigned γT rows. Let d = r/ µq
−γT
be the number of rows
that
still
need
to
be
assigned
to
each
K
batch. The r/T − µq
γ rows per partition not assigned yet are
assigned in the second step as given in the algorithm below.
Input : P , d, K, T , and µq
K
for 0 ≤ a < d µq
do
i ← ba/dc + 1
j ← (a mod T ) + 1
pi,j ← pi,j + 1
end
return P
B. Branch-and-Bound Solver
The branch-and-bound solver finds an optimal solution by
recursively branching at each batch for which there is more
than one possible assignment and considering all options. For
each branch, we lowerbound the value of the objective function
of any assignment in that branch and only investigate branches
with possibly better assignments.
L
1.10
1.05
Heuristic
Hybrid
Random
L
1.15
1.2
1.1
1.0
0.9
0.8
0.7
0.6 0
10
Heuristic
Random
D
D
1.00
1.06
1.05
1.04
1.03
1.02
1.01
1.00
0.99
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0 0
10
101
102
103
104
T
101
102
103
K
Fig. 3. The tradeoff between partitioning and performance for m = 6000,
n = 6000, K = 9, q = 6, N = 6, and µ = 1/3.
Fig. 4. Performance dependence on system size for µq = 2, n = 10000,
µm = 2000, m/T = 10 rows per partition, and code rate m/r = 2/3.
1) Branch: For the first row of P with remaining assignments, branch on every available assignment for that row.
(S)
2) Bound: We keep a record of all nonzero UQ for all
Q and S, and index them by the batches they are computed
from. An assignment to a batch can at most reduce LBDC by
(S)
1/ (mN |Qq |) for each nonzero UQ indexed by that batch,
and we lowerbound LBDC for a subtree by assuming that no
(S)
UQ will drop to zero for any subsequent assignment.
partition, and code rate m/r = 2/3 as a function of the number
of servers, K. The results shown are averages over 1000
randomly generated realizations of G as it is computationally
unfeasible to evaluate the performance exhaustively in this
case. The heuristic solver outperforms the random assignments
for small K, but as K grows the performance of both solvers
converge. The delay is an order of magnitude lower than that
of the scheme in [4] for the largest system considered.
C. Hybrid Solver
We first find a candidate solution using the heuristic solver
and then iteratively improve it using the branch-and-bound
solver. In particular, we decrement by 1 a random set of
elements of P and then use the branch-and-bound solver to
reassign the corresponding rows optimally. We repeat this
process until the average improvement between iterations
drops below some threshold.
VI. N UMERICAL R ESULTS
In Figs. 3 and 4, we plot the performance of the proposed
block-diagonal coding scheme with assignment P given by
the heuristic and the hybrid solvers. We normalize the performance by that of the unpartitioned scheme of [4]. We also
give the average performance over 100 random assignments.
In Fig. 3, we plot the normalized communication load (L)
and overall computational delay (D) as a function of the
number of partitions T . The system parameters are m = 6000,
n = 6000,
K = 9, q = 6, N = 6, and µ = 1/3. For up
K
to r/ µq
= 250 partitions, the proposed scheme does not
incur any loss in Dmap and communication load with respect
to the unpartitioned scheme. However, the proposed scheme
yields a significantly lower D compared to the scheme in [4]
(about 40% speedup for T > 50). For heavy partitioning
(around T = 800) a tradeoff between partitioning level,
communication load, and map phase delay is observed. With
3000 partitions (the maximum possible), there is about a 10%
increase in communication load. Note that the gain in overall
computational delay saturates with the partitioning level, thus
there is no reason to partition beyond a given level.
In Fig. 4, we plot the normalized performance for a constant
µq = 2, n = 10000, µm = 2000, m/T = 10 rows per
VII. C ONCLUSION
We introduced a block-diagonal coding scheme for distributed matrix multiplication based on partitioning the matrix into smaller submatrices. Compared to earlier (MDS)
schemes, the proposed scheme yields a significantly lower
computational delay with no increase in communication load
up to a level of partitioning. For instance, for a square matrix
with 6000 rows and columns, the proposed scheme reduces
the computational delay by about 40% when the number of
partitions T > 50. We have also considered fountain codes,
and will present the results in an extension of this paper.
R EFERENCES
[1] J. Dean and S. Ghemawat, “MapReduce: Simplified data processing
on large clusters,” in Proc. 6th Symp. Operating Systems Design &
Implementation, San Francisco, CA, Dec. 2004, pp. 137–149.
[2] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, “Coded MapReduce,”
in Proc. Annual Allerton Conf. Commun., Control, and Computing,
Monticello, IL, Sep./Oct. 2015, pp. 964–971.
[3] K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran,
“Speeding up distributed machine learning using codes,” in Proc. IEEE
Int. Symp. Inf. Theory, Barcelona, Spain, Jul. 2016, pp. 1143–1147.
[4] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, “A unified coding
framework for distributed computing with straggling servers,” in Proc.
Workshop Network Coding and Appl., Washington, DC, Dec. 2016.
[5] H. Ishii and R. Tempo, “The PageRank problem, multiagent consensus,
and web aggregation: A systems and control viewpoint,” IEEE Control
Systems, vol. 34, no. 3, pp. 34–53, Jun. 2014.
[6] G. Garrammone, “On decoding complexity of Reed-Solomon codes on
the packet erasure channel,” IEEE Commun. Lett., vol. 17, no. 4, pp.
773–776, Apr. 2013.
[7] Z. Li, J. Higgins, and M. Clement, “Performance of finite field arithmetic
in an elliptic curve cryptosystem,” in Proc. Int. Symp. Model. Anal. Simul.
Comput. Telecommun. Syst., Cincinnati, OH, Aug. 2001, pp. 249–256.
[8] B. C. Arnold, N. Balakrishnan, and H. N. Nagaraja, A First Course in
Order Statistics, 2nd ed. Philadelphia, PA, USA: Society for Industrial
and Applied Mathematics, 2008.
[9] A. Severinson, “Coded Computing Tools in Python,” Aug. 2017.
[Online]. Available: https://doi.org/10.5281/zenodo.844866
| 7cs.IT
|
1
Estimating Dynamic Load Parameters from
Ambient PMU Measurements
arXiv:1703.03370v1 [cs.SY] 9 Mar 2017
Xiaozhe Wang, Member, IEEE.
Abstract—In this paper, a novel method to estimate dynamic
load parameters via ambient PMU measurements is proposed.
Unlike conventional parameter identification methods, the proposed algorithm does not require the existence of large disturbance to power systems, and is able to provide up-to-date
dynamic load parameters consistently and continuously. The
accuracy and robustness of the method are demonstrated through
numerical simulations.
Index Terms—dynamic load identification, phasor measurement units, parameter estimation.
I. I NTRODUCTION
Load modelling and identification are of great importance to
the security and stability of power systems. While the accurate
models are available for generators, transmission lines and
control devices, load modeling is still a challenging and open
subject due to the fact that electric load at each substation
is an aggregation of numerous individual loads with different
behaviors [1]- [3]. In addition, the poor measurements, modeling, exchange information, as well as the uncertainties in
customers behaviors/devices further result in load uncertainties
[2]. Indeed, load uncertainty is one of the main factors that
affect the accuracy of the power dynamic models implemented
by system operators over the world [3].
Generally speaking, the load uncertainty comes from both
model structure and parameter values. It has been shown in
previous literature [4]- [6] that the use of different load models
leads to different and even contradictory results for dynamic
stability studies. Even though the applied model structure is
verified, different parameter values may also yield different
damping performances in small signal stability [2] [7] [8].
For instance, different time constants of loads may lead to
either asymptotically stable system or systems experiencing
oscillations (i.e., Hopf bifurcation occurs) [2]. Both load modelling and parameter identifications are essential in studying
the dynamic behaviors of power systems. This paper mainly
focuses on parameter identification for a generic dynamic load
model that is suitable for small signal stability analysis.
Different methods for dynamic load parameter identification
have been proposed, which can be classified into two categories: component-based approach [9] and measurement-based
approach [10]- [16]. The latter approach is more commonly
applied because real-time load variations and dynamic characteristics can be taken into account [14]. Measurement-based
model identification is typically solved through optimization
Xiaozhe Wang is with the Department of Electrical and Computer Engineering, McGill University, Montreal, QC H3A 0G4, Canada. Email:
[email protected].
methods that minimize the error between the measured output
variables and the simulated ones. In particular, the nonlinear
least-square curve fitting method has been implemented in [1]
[10]- [12]. Genetic algorithms, neural network-based methods
and other heuristic techniques have been applied in [13]- [16].
However, the optimization-based methods are time consuming
and thus can not be implemented online [4]. More importantly,
all those methods require measurement data from dynamic behaviors of system under big disturbances (e.g., during faults),
which is not always available [17]. Indeed, the variation of
load parameters may be much faster than the occurrence rate
of natural disturbances [1].
In this paper, we propose a novel measurement-based
method for dynamic load identification in ambient conditions,
which does not require the existence of large disturbance.
Particularly, the method combines the statistical properties
extracted from PMU measurements and the inherent model
knowledge, and is able to provide fairly accurate estimations
for parameter values in near real-time. Note that a generic
dynamic load model is implemented in this paper which is
suitable for the purposes of small signal stability analysis and
damping performance [2] [7] [8] [18]. The proposed method
can be implemented in online security analysis to provide upto-date dynamic load parameters accurately.
The rest of the paper is organized as follows. Section
II introduces the power system stochastic dynamic model.
Particularly, the generic dynamic load model used in small
signal stability is presented. Section III elaborates the proposed
method for estimating parameters of dynamic loads. Section
IV presents the validation of the proposed method through
numerical simulations. The impact of measurement noise is
also investigated. Conclusions and perspectives are given in
Section V.
II.
POWER SYSTEM STOCHASTIC DYNAMIC MODEL
Although we focus on load models, generator models are
also incorporated to provide more realistic simulations. Specifically, the classical generator model which can reasonably
represent the dynamics of generator in ambient conditions
is implemented. The power system buses are numbered as
follows: load buses: k = 1, 2, ..., m, and generators: i =
m + 1, ..., N . Particularly, to include the effects of the loads,
the structure preserving model [19] [20] is used:
δ̇i
Mi ω̇i
=
=
ωi
Pmi − PGi (δi , θi , Vi ) − Di ωi
(1)
(2)
2
PGi (δi , θi , Vi )
=
N
X
|Vi ||Vk |(Gik cos θik + Bik sin θik ) (3)
k=1
QGi (δi , θi , Vi )
=
N
X
k=1
where
δi
ωi
Mi
Pmi
PGi (δi , θi , Vi )
QGi (δi , θi , Vi )
Di
N
θij
|Vi |
Gij
Bij
|Vi ||Vk |(Gik sin θik − Bik cos θik ) (4)
To incorporate load variation, we apply a similar approach
used in [21] [22] and modify the set of load equations (5)-(6)
as follows:
1
ġk = −
[Pk − Pks (1 + σkp ξkp )]
(9)
τgk
1
[Qk − Qsk (1 + σkq ξkq )]
(10)
ḃk = −
τbk
generator rotor angle
where the steady-state real and reactive load demands are
generator angular frequency
perturbed with independent Gaussian noise from their initial
inertial constant
values. Specifically, ξkp and ξkq are standard Gaussian noise,
mechanical power input
and σkp and σkq represent the noise intensities for static real
real power injection
and reactive power, respectively.
reactive power injection
As discussed in [2] [18], this dynamic load model can
damping coefficient
naturally represent the most common types of loads in ambient
total number of buses
conditions such as thermostatic load, induction motor, power
voltage angle difference between bus i and j electronic converter, aggregate effects of distribution load
voltage magnitude
tap changer (LTC) transformers, etc. However, the range of
line conductance between bus i and j
time constants is considerably large ranging from cycles to
line susceptance between bus i and j
several minutes, and even hours for different types of loads.
For industrial plants, such as aluminum smelters, the time
constants are in the range of 0.1s to 0.5s; for tap changers
The detailed expressions of PGi (δi , θi , Vi ) and QGi (δi , θi , Vi ) and other control devices, they are in the range of minutes; for
are neglected here for simplicity and can be found in many heating load, they may range up to hours [8]. As a result, the
books (e.g., [19]).
uncertainty of composition of different types of loads can be
Regarding dynamic loads, we use the following first-order aggregated in time constants τ and τ [2]. This is reasonable
g
b
load model proposed in [18] that can represent the common
in
the
situations
when
the
network
characteristics
are known,
types of loads (e.g., induction motors, thermostatically congenerator models are validated and static load characteristics
trolled loads) in ambient conditions:
are understood better than their dynamic response which is
the case in practical situations. In addition to a wide range of
1
s
(Pk − Pk )
(5)
ġk = −
time constants, the variation of τg and τb can also be fast. For
τgk
example, τb may change from 0.1s to 24.1s in one day (see
1
s
ḃk = −
(Qk − Qk )
(6)
Table I, II in [1]).
τbk
N
Because of wide range and fast variation of time constants
X
|Vk ||Vj |(−Gkj cos θkj − Bkj sin θkj ) (7) τg and τb , they need to be updated frequently to ensure the
Pk = gk Vk2 =
j=1
accuracy of dynamic load models used in online security and
N
X
stability analysis. Conventionally, τg and τb are estimated from
2
|Vk ||Vj |(−Gkj sin θkj + Bkj cos θkj ) (8)
Q k = bk V k =
dynamic data by perturbing the system, for example, through
j=1
changing the transformer tap [23]. However it is impractical
to perturb the system frequently for estimating parameter
where
values of loads. In this paper, we propose a novel method to
estimate τg and τb for the loads of interests from ambient PMU
gk
effective conductance of the load
measurements in daily operation. In particular, the estimation
bk
effective susceptance of the load
process does not require the existence of disturbance to the
τgk active power time constant of the load
system.
τbk reactive power time constant of the load
Pk real power demand of the load
III. M ETHODOLOGY
Qk reactive power demand of the load
s
Pk steady-state real power demand of the load
In ambient conditions, the stochastic dynamic load equaQsk steady-state reactive power demand of the load
tions (9)-(10) can be linearized as below:
0
−Tg −1 ∂P
ġ
g
∂g
=
b
ḃ
0
−Tb −1 ∂Q
The values Pks and Qsk describe the static (steady-state) power
∂b
p
−1 s p
characteristics of the loads achieved in equilibrium. The instant
0
Tg P Σ
ξ
+
−1 s q
real power and reactive power consumption can be characterξq
Q
Σ
0
T
b
p
ized by the effective conductance Pk = gk Vk2 and susceptance
g
ξ
Qk = bk Vk2 at any time. The time constants τgk and τbk that
= A
+B
(11)
b
ξq
typically depend on voltage and frequency represent the instant
relaxation rate of the load.
where
3
g = [g1 , ..., gm ]T ,
Tg = diag[τg1 , ..., τgm ],
P = [P1 , ..., Pm ]T ,
s
P s = diag[P1s , ..., Pm
],
p
p
p
Σ = diag[σ1 , ..., σm ],
p T
ξ p = [ξ1p , ..., ξm
] ,
b = [b1 , ..., bm ]T ,
Tb = diag[τb1 , ..., τbm ],
Q = [Q1 , ..., Qm ]T ,
Qs = diag[Qs1 , ..., Qsm ],
q
Σq = diag[σ1q , ..., σm
],
q
q
q T
ξ = [ξ1 , ..., ξm ] .
unbiased estimators—sample covariance matrixes Qgg and
Qbb in practice, each entry of which is calculated as below:
n
Qgk gj
It is observed that [g, b]T is a vector Ornstein-Uhlenbeck
process that is stationary, Gaussian and Markovian [24] [25].
Particularly, if the state matrix
A is stable,
the stationary
Cgg Cgb
can be shown to
covariance matrix Cxx =
Cbg Cbb
satisfy the following Lyapunov equation [24] [26]:
(12)
which nicely combines the model knowledge and the statistical
properties of state variables.
Since Pk = gk Vk2 , we have:
(
2
k
if j = k
Vk2 + 2gj Vk ∂V
∂Pk
∂gj ≈ Vk
(13)
=
∂Vk
if j 6= k
2gj Vk ∂gj ≈ 0
∂gj
under the assumption that △Vk ≈ 0 in ambient conditions.
k
Similar relation can be obtained for ∂Q
∂bj . As a result, the
Jacobian matrix A satisfies:
0
−Tg −1 V 2
A≈
(14)
0
−Tb −1 V 2
where V = diag[V1 , ..., Vm ]. Substituting (14) and the detailed
expression of Cxx and B into (12), and performing algebraic
simplification, we have:
1
Cgg = Tg−1 (P s )2 (Σp )2 V −2
2
1 −1 s 2 q 2 −2
Cbb = Tb (Q ) (Σ ) V
2
Cgb = Cbg = 0
bk
Ik
}
Vk
Ik
= Im{ }
Vk
= Re{
=
1 X
(bk (i) − b̄k )(bj (i) − b̄j )
n − 1 i=1
(15)
(16)
(17)
(18)
(19)
Regarding the covariance matrix Cgg = E[(g − E[g])(g −
E[g])T ] and Cbb = E[(b − E[b])(b − E[b])T ], we use their
(21)
where ḡk and b̄k denote the sample mean of gk and bk ,
respectively, and n is the sample size.
Therefore, the proposed algorithm can be summarized as
follows. We assume that PMUs are installed at the substations
that the (aggregated) loads of interests are connected to. We
also assume that the static characteristics of loads are well
understood such that P s , Qs , Σp and Σq are prior known,
which is reasonable as shown in [21] [22]. Then the following
algorithm provides an estimation of Tg and Tb for the dynamic
loads from ambient PMU measurements:
Step 1.
Step 2.
Step 3.
Compute the sample mean V̄ and estimate g and
b from PMU measurements by (18)-(19).
Calculate the sample covariance matrix Qgg and
Qbb by (20)-(21).
Approximate Tg and Tb as blow:
Tg
=
Tb
=
1 s 2 p 2 −2 −1
(P ) (Σ ) V̄ Qgg
2
1 s 2 q 2 −2 −1
(Q ) (Σ ) V̄ Qbb
2
(22)
(23)
Note that (22)-(23) are acquired by a simple algebraic manipulation of (15)-(16).
IV.
Particularly, we utilize the relations (15)-(16) that link the
measurements of stochastic load variation to the physical
model, and provide an ingenious way to estimate the dynamic
parameters Tg and Tb from measurements.
In practical applications, V , Cgg , and Cbb need to be
acquired or estimated from limited PMU measurements. A
window size of 1000s is used in the examples of this paper
where time constants are up to several seconds. Note that the
larger the time constants, the longer the sample window is
needed to ensure accuracy. First, the sample mean V̄ can be
used as an estimation of V , then g and b can be estimated
from PMU measurements (i.e., phasors Vk and Ik ) as follows:
gk
(20)
n
Qbk bj
ACxx + Cxx AT = −BB T
=
1 X
(gk (i) − ḡk )(gj (i) − ḡj )
n − 1 i=1
CASE STUDIES
In this section, the proposed algorithm to estimate time constants of dynamic loads are validated through numerical simulations. Furthermore, the robustness of the proposed method
to measurement noise is also demonstrated via simulation. All
case studies were done in PSAT-2.1.9 [27].
A. Validation of the Method
We consider the standard WSCC 3-generator, 9-bus
system model (see, e.g. [19]). The classical generator models (1)-(2) and the stochastic dynamic load
models (9)-(10) are implemented in the structure preserving framework. The system parameters are available online: https://github.com/xiaozhew/PES-load-parameterestimation. Particularly, there are three dynamic loads at buses
1, 2 and 3, the time constants of which are τg = 1, 3, 0.2s
and τb = 5, 7, 0.8s, respectively. The trajectories of some
state variables and algebraic variables are shown in Fig. 1,
from which we see that the state variables are fluctuating
around their nominal values in ambient conditions, yet larger
time constants lead to slower variations as expected (e.g., the
variations of g2 and b2 are slower than g3 and b3 ).
By the proposed algorithm, we firstly compute the sample
mean V̄ = diag[0.9952, 1.0126, 1.0155]. Then we estimate the
4
0.94
0.34
g2
b2
0.92
TABLE I: A comparison between the actual and the estimated
time constants in the 9-bus system
0.32
0.9
0.88
0.3
0.86
0.28
τg1
τg2
τg3
τb1
τb2
τb3
0.84
0.26
0.82
0.8
0
200
400
600
time(s)
800
0.24
0
1000
200
400
600
time(s)
800
1000
actual value (s)
1.0000
3.0000
0.2000
5.0000
7.0000
0.8000
estimated value (s)
0.9145
2.9867
0.2122
4.7974
6.9777
0.7462
error
8.55%
0.44%
6.1%
4.05%
0.32%
6.72%
(a) Trajectory of g2 on [0s,1000s] (b) Trajectory of b2 on [0s,1000s]
1.4
g3
1.3
0.5
b3
0.45
1.2
0.4
1.1
1
0.35
0.9
0.3
0.8
0.25
0.7
0
200
400
600
time(s)
800
1000
0.2
0
200
400
600
time(s)
800
1000
(c) Trajectory of g3 on [0s,1000s] (d) Trajectory of b3 on [0s,1000s]
1.025
V2
1.06
Section IV-A according to the IEEE Standards [28] [29].
A comparison between the actual and the estimated time
constants are presented in Table. II. It is observed that the
proposed method provides similar accuracy to the case without
the measurement noises, which indicates that the method is
relatively robust under measurement noise.
TABLE II: A comparison between the actual and the estimated
time constants in the 9-bus system with the measurement
noises
V3
1.02
τg1
τg2
τg3
τb1
τb2
τb3
1.04
1.015
1.01
1.02
1.005
1
1
0.995
0
200
400
600
time(s)
800
1000
0.98
0
200
400
600
time(s)
800
actual value (s)
1.0000
3.0000
0.2000
5.0000
7.0000
0.8000
estimated value (s)
0.9144
2.9819
0.2121
4.7752
6.9426
0.7443
error
8.56%
0.60%
6.06%
4.50%
0.82%
6.97%
1000
(e) Trajectory of V2 on [0s,1000s](f) Trajectory of V3 on [0s,1000s]
C. Further Validation
Fig. 1: Trajectories of some dynamic conductances, susceptances, and voltages in the 9-bus system.
For further validation, we apply the method to
a larger system—the IEEE 39-bus 10-generator test
system, the parameters of which are available online:
https://github.com/xiaozhew/PES-load-parameter-estimation.
In particular, 10 dynamic loads have been added to buses
1-10, and their corresponding time constants range from 0.1s
to 5s. A comparison between the actual and the estimated
time constants are presented in Table. III. The simulation
results further demonstrate that the proposed method is able
to provide good estimations for time constants of the dynamic
loads.
dynamic conductance, susceptances and their corresponding
sample covariance matrixes:
1.41 × 10−3 9.84 × 10−5 2.24 × 10−4
Qgg = 9.84 × 10−5 4.16 × 10−4 6.57 × 10−5 (24)
2.24 × 10−4 6.57 × 10−5 5.75 × 10−3
Qbb
2.63 × 10−4
= −1.47 × 10−6
6.46 × 10−6
−1.47 × 10−6
1.75 × 10−4
−7.57 × 10−6
6.46 × 10−6
−7.57 × 10−6 (25)
1.62 × 10−3
It is expected that both Qgg and Qbb are nearly diagonal as
the stochastic perturbations are independent.
Since each entry of (P s )2 (Σp )2 and (Qs )2 (Σq )2 is set to
be 0.0025, Tg and Tb can be readily estimated from (22)(23). A comparison between the estimated τg , τb and their
actual values are shown in Table. I. It’s observed that the
proposed algorithm provides fairly accurate estimation for time
constants of each load.
B. Impact of Measurement Noise
Like other measurement-based methods, the performance of
the proposed algorithm may be affected by PMU measurement noise. In order to investigate the potential influence,
measurement noises with standard deviation of 10−3 have
been added to g, b and V in the 9-bus example shown in
V.
CONCLUSIONS AND PERSPECTIVES
In this paper, we have proposed a novel method to estimate parameter values of dynamic load from ambient PMU
measurements. The accuracy and robustness of the method
have been demonstrated through numerical studies. Unlike
conventional methods, the proposed technique does not require
the existence of large disturbance to systems, and thus can be
implemented continuously in daily operation to provide up-todate dynamic load parameter values.
In the future, we plan to further validate the method by using
real PMU data and extend the method to estimate dynamic
load parameters without knowing their static characteristics.
R EFERENCES
[1] D. Karlsson, and D. J. Hill, Modelling and identification of nonlinear
dynamic loads in power systems. IEEE Transactions on Power Systems,
vol. 9, no. 1, pp. 157-166.
5
TABLE III: A comparison between the actual and the estimated time constants in the 39-bus system
τg1
τg2
τg3
τg4
τg5
τg6
τg7
τg8
τg9
τg10
τb1
τb2
τb3
τb4
τb5
τb6
τb7
τb8
τb9
τb10
actual value (s)
0.1000
0.6000
1.1000
1.6000
2.1000
2.6000
3.1000
3.6000
4.1000
4.6000
0.5000
1.0000
1.5000
2.0000
2.5000
3.0000
3.5000
4.0000
4.5000
5.0000
estimated value (s)
0.1225
0.5860
1.1086
1.6924
2.1323
2.6130
3.0715
3.3004
4.5080
4.6636
0.5277
0.9831
1.5325
2.0559
2.5784
3.3522
3.1273
4.1105
4.3084
5.1723
error
22.55%
2.33%
0.78%
5.78%
1.54%
0.50%
0.92%
8.32%
9.95%
1.38%
5.53%
1.69%
2.17%
2.79%
3.14%
11.74%
10.65%
2.76%
4.26%
3.45%
[2] H. D. Nguyen, and K. Turitsyn, Robust Stability Assessment in the
Presence of Load Dynamics Uncertainty. IEEE Transactions on Power
Systems, vol. 31, no. 2, pp. 1579-1594.
[3] I. A. Hiskens, and J. Alseddiqui, Sensitivity, approximation, and uncertainty in power system dynamic simulation. IEEE Transactions on Power
Systems, vol. 21, no. 4, pp. 1808-1820.
[4] A. M. Najafabadi, and A. T. Alouani, Real time estimation of sensitive
parameters of composite power system load model. In Transmission and
Distribution Conference and Exposition (T&D), 2012 IEEE PES (pp. 18).
[5] A. Gebreselassie, and J. H. Chow, Investigation of the effects of load
models and generator voltage regulators on voltage stability. International
Journal of Electrical Power & Energy Systems, vol. 16, no. 2, pp. 83-89,
1994.
[6] W. S. Kao, C. J. Lin, C. T. Huang, Y. T. Chen, and C. Y.
Chiou,Comparison of simulated power system dynamics applying various
load models with actual recorded data. IEEE Transactions on Power
Systems, vol. 9, no. 1, pp. 248-254.
[7] J. V. Milanovic, and I. A. Hiskens, Effects of load dynamics on power
system damping. IEEE Transactions on Power Systems, vol. 10, no. 2,
pp. 1022-1028, 1995.
[8] I. A. Hiskens, and J. V. Milanovic, Load modelling in studies of power
system damping. IEEE Transactions on Power Systems, vol. 10, no. 4,
pp. 1781-1788, 1995.
[9] W. W. Price, K. A. Wirgau, A. Murdoch, J. V. Mitsche, E. Vaahedi,
and M. El-Kady, Load modeling for power flow and transient stability
computer studies. IEEE Transactions on Power Systems, vol. 3, no. 1,
pp. 180-187, 1988.
[10] B. K. Choi, H. D. Chiang, Y. Li, H. Li, Y. T. Chen, D. H. Huang,
and M. G. Lauby, Measurement-based dynamic load models: derivation,
comparison, and validation. IEEE Transactions on Power Systems, vol.
21, no. 3, pp. 1276-1283, 2006.
[11] Y. Li, H. D. Chiang, B. K. Choi, Y. T. Chen, D. H. Huang, and M. G.
Lauby,Load models for modeling dynamic behaviors of reactive loads:
Evaluation and comparison. International Journal of Electrical Power &
Energy Systems, vol. 30, no. 9, pp. 497-503, 2008.
[12] J. Ma, D. Han, R. M. He, Z. Y. Dong, and, D. J. Hill, Reducing
identified parameters of measurement-based composite load model. IEEE
Transactions on Power Systems, vol. 23, no. 1, pp. 76-83, 2008.
[13] P. Ju, E. Handschin, and D. Karlsson, Nonlinear dynamic load modelling: model and parameter estimation. IEEE Transactions on Power
Systems, vol. 11, no. 4, pp. 1689-1697, 1996.
[14] H. Bai, P. Zhang, and V. Ajjarapu, V, A novel parameter identification
approach via hybrid learning for aggregate load modeling. IEEE
Transactions on Power Systems, vo. 24, no. 3, pp. 1145-1154, 2009.
[15] T. Hiyama, M. Tokieda, W. Hubbi, and H. Andou, Artificial neural
network based dynamic load modeling. IEEE transactions on Power
Systems, vol. 12, no. 4, pp. 1576-1583, 1997.
[16] J. Y. Wen, L. Jiang, Q. H. Wu, and S. J. Cheng, Power system load
modeling by learning based on system measurements. IEEE Transactions
on Power Delivery, vol. 18, no. 2, pp. 364-371, 2003.
[17] D. Han, J. Ma, R. M. He, and Z. Y. Dong, A real application of
measurement-based load modeling in large-scale power grids and its
validation. IEEE Transactions on Power Systems, vol. 24, no. 4, pp.
1756-1764.
[18] H. D. Nguyen, and K. Turitsyn, Voltage multistability and pulse emergency control for distribution system with power flow reversal. IEEE
Transactions on Smart Grid, vol. 6, no. 6, pp. 2985-2996.
[19] H. D. Chiang, Direct Methods for Stability Analysis of Electric Power
Systems-Theoretical Foundation, BCU Methodologies, and Applications.
New Jersey: John Wiley & Sons, Inc, 2011.
[20] A. Pai, Energy function analysis for power system stability. Springer
Science & Business Media, 2012.
[21] H. Mohammed, and C. O. Nwankpa, Stochastic analysis and simulation
of grid-connected wind energy conversion system. IEEE Transactions on
Energy Conversion, vol. 15, no. 1, pp. 85-90.
[22] C. O. Nwankpa, S. M. Shahidehpour, and Z. Schuss, A stochastic
approach to small disturbance stability analysis. IEEE Transactions on
Power systems, vol. 7, no. 4, pp. 1519-1528.
[23] T. Van Cutsem, C. Vournas. Voltage Stability of Electric Power Systems.
Springer Science & Business Media, 1998.
[24] C. Gardiner, Stochastic Methods: A Handbook for the Natural and Social
Sciences. Springer Series in Synergetics. Springer, Berlin, Germany, 2009.
[25] X. Wang, K. Turitsyn, Data-driven diagnostics of mechanism and source
of sustained oscillations. IEEE Transactions on Power Systems, to appear.
[26] G. Ghanavati, P. D. H. Hines, and T. I. Lakoba, Identifying useful statistical indicators of proximity to instability in stochastic power systems.
IEEE Transactions on Power Systems, in press, 2015. arXiv preprint
arXiv:1410.1208 (2014).
[27] F. Milano, An open source power system analysis toolbox. IEEE
Transactions on Power Systems, vol. 20, no. 3, pp. 1199-1206.
[28] IEEE Standard for Synchrophasor Measurements for Power Systems.
IEEE Std C37.118.1-2001 (Revision of IEEE Std C37.118-2005), pp. 161, Dec. 2011.
[29] IEEE Standard for Synchrophasor Measurements for Power SystemsAmendment 1: Modification of Selected Performance Requirements. IEEE
Std C37.118.1a-2014 (Amendment to IEEE Std C37.118.1-2011), pp. 125, April 2014.
| 3cs.SY
|
Fast and Compact Exact Distance Oracle for Planar Graphs
Vincent Cohen-Addad, Søren Dahlgaard∗, and Christian Wulff-Nilsen
University of Copenhagen
[vincent.v,soerend,koolooz]@di.ku.dk
arXiv:1702.03259v3 [cs.DS] 2 May 2017
Abstract
For a given a graph, a distance oracle is a data structure that answers distance queries
between pairs of vertices. We introduce an Opn5{3 q-space distance oracle which answers exact
distance queries in Oplog nq time for n-vertex planar edge-weighted digraphs. All previous
distance oracles for planar graphs with truly subquadratic space (i.e., space Opn2´ q for
some constant ą 0) either required query time polynomial in n or could only answer
approximate distance queries.
Furthermore, we show how to trade-off time and space: for any S ě n3{2 , we show how
5{2
to obtain an S-space distance oracle that answers queries in time Op Sn3{2 log nq. This is a
polynomial improvement over the previous planar distance oracles with opn1{4 q query time.
∗
Research partly supported by Mikkel Thorup’s Advanced Grant DFF-0602-02499B from the Danish Council
for Independent Research under the Sapere Aude research career programme.
1
Introduction
Efficiently storing distances between the pairs of vertices of a graph is a fundamental problem
that has receive a lot of attention over the years. Many graph algorithms and real-world problems
require that the distances between pairs of vertices of a graph can be accessed efficiently. Given
an edge-weighted digraph G “ pV, Eq with n vertices, a distance oracle is a data structure that
can efficiently answer distance queries between pairs of vertices u, v P V .
A naive approach consists in storing an nˆn distance matrix, giving a distance query time of
Op1q by a simple table lookup. The obvious downside is the huge Θpn2 q space requirement which
is in many cases impractical. For example, several popular routing heuristics (e.g.: for the travelling salesman problem) require fast access to distances between pairs of vertices. Unfortunately
the inputs are usually too big to allow to store an n ˆ n distance matrix (see e.g.: [1])1 .
Fast and compact data structures for distances are also critical in many routing problems.
One important challenge in these applications is to process a large number of online queries
while keeping the space usage low, which is important for systems with limited memory or
memory hierarchies. Therefore, the alternative naive approach consisting in simply storing the
graph G and answering a query by running a shortest path algorithm on the entire graph is also
prohibitive for many applications.
Since road networks and planar graphs share many properties, planar graphs are often
used for modeling various transportation networks (see e.g.: [26]). Therefore obtaining good
space/query-time trade-offs for planar distance oracles has been studied thoroughly over the
past decades [10, 3, 8, 12, 30, 4, 26].
If S represents the space usage and Q represents the query time, the trivial solutions described
above would suggest a trade-off of Q “ n2 {S 2 . Up to logarithmic factors, this trade-off is achieved
by the oracles of Djidjev [10] and Arikati, et al.
? [3]. The oracle of Djidjev further improves on
this trade-off obtaining an oracle with Q “ n{ S for the range S P rn4{3 , n3{2 s suggesting that
this trade-off might instead be the correct one. Extending this trade-off to the full range of S was
the subject of several subsequent papers by Chen and Xu [8], Cabello [4], Fakcharoenphol and
Rao [12], and finally Mozes
of Nussbaum [27]) obtaining
? and Sommer [26] (see also the result
2
a query time of Q “ n{ S for the entire range of S P rn, n s (again ignoring constant and
logarithmic factors).
It is worth noting that the above mentioned trade-off between space usage and query
time is no better than the trivial solution of simply storing the n ˆ n distance matrix when
constant (or even polylogarithmic) query time is needed. In fact the best known result in
this case due to Wulff-Nilsen [30] who manages to obtain very slightly subquadratic space of
Opn2 polyloglogpnq{ logpnqq and constant query time. It has been a major open question whether
an exact oracle with truly subquadratic (that is, Opn2´ε q for any constant ε ą 0) space usage
and constant or even polylogarithmic query time exists. Furthermore, the trade-offs obtained in
the literature suggest that this might not be the case.
In this paper we break this quadratic barrier:
Theorem 1. Let G “ pV, Eq be a weighted planar digraph with n vertices. Then there exists
a data structure with Opn2 q preprocessing time and Opn5{3 q space and a data structure with
Opn11{6 q space and Opn11{6 q expected preprocessing time. Given any two query vertices u, v P V ,
both oracles report the shortest path distance from u to v in G in Oplog nq time.
In addition to Theorem 1 we also obtain a distance oracle with a trade-off between space
and query time.
1
In these cases, the inputs are then embedded into the 2-dimensional plane so that the distances can be
computed in Op1q time at the expense of working with incorrect distances.
2
Using the Opnq shortest path algorithm for planar graphs of Henzinger et al. [16].
1
Theorem 2. Let G “ pV, Eq be a weighted planar digraph with n vertices. Let S denote the
space, P denote the preprocessing time, and Q denote the query time. Then there exists planar
distance oracles with the following properties:
5{2
• P “ Opn2 q, S ě n3{2 , and Q “ Op Sn3{2 log nq.
11{5
• P “ S, S ě n16{11 , and Q “ Op nS 6{5 log nq.
In particular, this result improves on the current state-of-the-art [26] trade-off between space
and query time for S ě n3{2 . The main idea is to use two r-divisions, where we apply our
structure from Theorem 1 to one and do a brute-force search over the boundary nodes of the
other.
Recent developments We note that the main focus of this paper is on space usage and query
time, and the the preprocessing time follows directly from our proofs (and by applying the result
of [5] for subquadratic time).
After posting a preliminary version of this paper on arXiv [9], the algorithm of Cabello [5]
was improved by Gawrychowski et al. [15] to run in Õpn5{3 q deterministically. As noted in [15]
this also improves the preprocessing time of our Theorem 1 to Õpn5{3 q while keeping the space
usage at Opn5{3 q. It is also possible to use Gawrychowski et al. to speed-up the pre-processing
time of our distance oracle described in Theorem 2. This yields a distance oracle (in the notation
5{2
of Theorem 2) with P “ ÕpSq, S ě n3{2 , and Q “ Op Sn3{2 log nq, thus eliminating entirely the
need of the second bullet point of Theorem 2.
Techniques
We derive structural results on Voronoi diagrams for planar graphs when the centers of the
Voronoi cells lie on the same face. The key ingredients in our algorithm are a novel and technical
separator decomposition and point location structure for the regions in an r-division allowing us
to perform binary search to find a boundary vertex w lying on a shortest-path between a query
pair u, v. These structures are applied on top of weighted Voronoi diagrams, and our point
location structure relies heavily on partitioning each region into small “easy-to-handle” wedges
which are shared by many such Voronoi diagrams. More high-level ideas are given in Section 3.
Our approach bears some similarities with the recent breakthrough of Cabello [5]. Cabello
showed that abstract Voronoi diagrams [24, 25] studied in computational geometry combined
with planar r-division can be used to obtain fast planar graphs algorithms for computing the diameter and wiener index. We start from Cabello’s approach of using abstract Voronoi diagrams.
While Cabello focuses on developing fast algorithms for computing abstract Voronoi diagrams of
a planar graph, we introduce a decomposition theorem for abstract Voronoi diagrams of planar
graphs and a new data structure for point location in planar graphs.
1.1
Related work
In this paper, we focus on distance oracles that report shortest path distances exactly. A closely
related area is approximate distance oracles. In this case, one can obtain near-linear space and
constant or near-constant query time at the cost of a small p1 ` q-approximation factor in the
distances reported [28, 21, 19, 20, 32].
One can also study the problem in a dynamic setting, where the graph undergoes edge
insertions and deletions. Here the goal is to obtain the best trade-off between update and query
time. Fakcharoenphol and Rao [12] showed how to obtain Õpn2{3 q for both updates and queries
2
?
and a trade-off of Oprq and Opn{ rq in general. Several follow up works have improved this
result to negative edges and shaving further logarithmic factors [22, 17, 18, 14]. Furthermore,
Abboud and Dahlgaard [2] have showed that improving this bound to Opn1{2´ε q for any constant
ε ą 0 would imply a truly subcubic algorithm for the All Pairs Shortest Paths (APSP) problem
in general graphs.
In the seminal paper of Thorup and Zwick [29], a p2k ´ 1q-approximate distance oracle
is presented for undirected edge-weighted n-vertex general graphs using Opkn1`1{k q space and
Opkq query time for any integer k ě 1. Both query time and space has subsequently been
improved to Opn1`1{k q space and Op1q query time while keeping an approximation factor of
2k ´ 1 [31, 6, 7]. This is near-optimal, assuming the widely believed and partially proven girth
conjecture of Erdős [11].
2
Preliminaries and Notations
Throughout this paper we denote the input graph by G and we assume that it is a directed
planar graph with a fixed embedding. We assume that G is connected (when ignoring edge
orientations) as otherwise each connected component can be treated separately.
Section 4 will make use of the geometry of the plane and associate Jordan curves to cycle
separators. Let H be a planar embedded edge-weighted digraph. We use V pHq to denote the
set of vertices of H and we denote by H ˚ the dual of H (with parallel edges and loops) and
view it as an undirected graph. We assume a natural embedding of H ˚ i.e., each dual vertex
is in the interior of its corresponding primal face and each dual edge crosses its corresponding
primal edge of H exactly once and intersects no other edges of G. We let dH pu, vq denote the
shortest path distance from vertex u to vertex v in H.
r-division We will rely on the notion of r-division introduced by Frederickson [13] and further
developed by Klein et al. [23]. For a subgraph H of G, a vertex v of H is a boundary vertex if G
contains an edge not in H that is incident to v. We let δH denote the set of boundary vertices
of H. Vertices of V pHqzδH are called internal vertices of H. A hole of a subgraph H of G is a
face of H that is not a face of G.
Let c1 and c2 be constants. For a number r, an r-division with few holes of (connected) graph
G (with respect to c1 , c2 ) is a collection R of subgraphs of G, called regions, with the following
properties.
1. Each edge of G is in exactly one region.
2. The number of regions is at most c1 |V pGq|{r.
3. Each region contains at most r vertices.
?
4. Each region has at most c2 r boundary vertices.
5. Each region contains only Op1q holes.
We make the simplifying assumption that each hole H of each region R is a simple cycle and
that all its vertices belong to δR. We can always reduce to this case as follows. First, turn H
into a simple cycle by duplicating vertices that are visited more than once in a walk of the hole.
Then for each pair of consecutive boundary vertices in this walk, add a bidirected edge between
them unless they are already connected by an edge of R; the new edges are embedded such that
they respect the given embedding of R. We refer to the new simple cycle obtained as a hole and
it replaces the old hole H.
3
We also make the simplifying assumption that each face of a region R is either a hole or a
triangle and that each edge of R is bidirected. This can always be achieved by adding suitable
infinite-weight edges that respect the current embedding of R.
Non-negative weights and unique shortest paths As mentioned earlier, we may assume
w.l.o.g. that G has non-negative edge weights. Furthermore, we assume uniqueness of shortest
paths i.e., for any two vertices x, y of a graph G, there is a unique path from x to y that minimizes
the sum of the weights of its edges. This can be achieved either with random perturbations of
edge weights or deterministically with a slight overhead as described in [5]; we need the shortest
paths uniqueness assumption only for the preprocessing step and thus the overhead only affects
the preprocessing time and not the query time of our distance oracle.
Voronoi diagrams We now define the key notion of Voronoi diagrams. Let G be a graph and
r ą 0. Consider an r-division with few holes R of G and a region R P R and let H be a hole of
R. Let u be a vertex of G not in R.
Let RH be the graph obtained from R by adding inside each hole H 1 ‰ H of R, a new vertex
in its interior and infinite-weight bidirected edges between this vertex and the vertices of H 1
(which by the above simplifying assumption all belong to δR), embedding the edges such that
they are pairwise non-crossing and contained in H 1 .
Some of the following definitions are illustrated in Figure 1. Consider the shortest path
tree Tu in G rooted at u. For any vertex x P V pTu q, define Tu pxq to be the subtree of Tu
rooted at x. For each vertex x P V pHq we define the Voronoi cell of x (w.r.t. u, R, and H)
as the set of vertices of RH that belong to the subtree Tu pxq and not to any subtree Tu pyq
for y P pV pHq ´ txuq X V pTu pxqq. The weighted Voronoi diagram of u w.r.t. R and H is the
collection of all the Voronoi cells of the vertices in H. For each vertex x P H we say that its
weight is the shortest path distance from u to x in G. Note that since we assume unique shortest
paths and bidirectional edges, the weighted Voronoi diagram of u w.r.t. R and H is a partition
of the vertices of RH . Furthermore, each Voronoi cell contains exactly one vertex of H. For
any Voronoi cell C, we define its boundary edges to be the edges of RH that have exactly one
˚ be the subgraph of R˚ consisting of the (dual) boundary edges over
endpoint in C. Let BH
H
˚ is an
all Voronoi cells w.r.t. u and RH ; we ignore edge orientations and weights so that BH
˚
unweighted undirected graph. We define VorH pR, uq to be the multigraph obtained from BH
by replacing each maximal path whose interior vertices have degree two by a single edge whose
embedding coincides with the path it replaces. When H is clear from context, we simply write
VorpR, uq.
3
High-level description
We now give a high-level description of our distance oracle where we omit the details needed to
get our preprocessing time bounds. Our data structure is constructed on top of an r-division
of the graph. For each region R of the r-division we store a look-up table of the distance in G
between each ordered pair of vertices u, v P V pRq. We also store a look-up table of distances in G
?
from each vertex u P V to the boundary vertices of R. In total this part requires Opnr ` n2 { rq
space.
The difficult case is when two vertices u and v from different regions are queried. To do
this we will use weighted Voronoi diagrams. More specifically, for every vertex u, every region
R, and every hole H of R, we construct a recursive separator decomposition of the weighted
Voronoi diagram of u w.r.t. R and H. The goal is to determine the boundary vertex w such
that v is contained in the Voronoi cell of u. If we can do this, we know that dG pu, vq “
4
1
4
2
5
3
2
1
Figure 1: Illustration of the weighted Voronoi diagram of a vertex u (not shown) w.r.t. a region
R (grey) and a hole H (dotted edges and seven grey vertices). Edges of R are bidirected and
have weight 1. The number next to a vertex of H is the weight of that vertex. Vertices of the
Voronoi diagram (which are dual vertices and hence faces of the primal) and its boundary are
shown in black except the one embedded inside H. The graph VorH pR, uq has seven edges. Note
that this illustration does not have unique shortest paths as assumed in this paper. Note also
that the shortest path from u to the node with weight 5 goes through the nearby node with
weight 2. Thus the rest of the shortest-path tree from u has been assigned to the node with
weight 5. Note also that the illustration is not triangulated.
dG pu, wq ` dRH pw, vq for one of the holes H of R. To determine this we use a carefully selected
recursive decomposition. This decomposition is stored in a compact way and we show how it
enables binary search to find w in Oplog rq time.
In order to store all of the above mentioned parts efficiently we will employ the compact
representation of the abstract Voronoi diagram (namely VorpR, uq as defined in the previous
?
?
section). This requires only Op rq space for each choice of u, R, and H for a total of Opn2 { rq
space. This also dominates the space for storing the recursive decompositions.
Finally, we store for each graph RH and each possible separator of RH the set of vertices on
one side of the separator, since this is needed to perform the binary search. This is done in a
compact way requiring only Opr2 q space per region. Thus, the total space requirement of our
?
distance oracle is Opnr ` n2 { rq. Picking r “ n2{3 gives the desired Opn5{3 q space bound.
4
Recursive Decomposition of Regions
In this section, we consider a region R in an r-division of a planar embedded graph G “ pV, Eq, a
vertex u P V ´V pRq, and a hole H of R which we may assume is the outer face of R. To simplify
notation, we identify R with RH and let δR denote the boundary vertices of R belonging to H,
i.e., δR “ V pHq (by our simplifying assumption regarding holes in the preliminaries). The dual
vertex corresponding to H is denoted v8 pR, uq or just v8 .
We assume that R contains at least three boundary vertices. Recall that each face of R other
than the outer face is a triangle and so, each vertex of VorpR, uq other than v8 has degree 3.
Moreover, every cell of VorpR, uq contains exactly one boundary vertex, therefore the boundary
of each cell of VorpR, uq contains exactly one occurrence of v8 and contains at least one other
vertex. Also note that the cyclic ordering of cells of VorpR, uq around v8 is the same as the
cyclic ordering δR of boundary vertices of R.
5
Construct a plane multigraph R∆ puq from VorpR, uq as follows. First, for every Voronoi cell
C, add an edge from v8 to each vertex of C other than v8 itself; these edges are embedded such
that they are fully contained in C and such that they are pairwise non-crossing. For each such
edge e, denote by Cpeq the cell it is embedded in. To complete the construction of R∆ puq, remove
every edge incident to v8 belonging to VorpR, uq. The construction of R∆ puq is illustrated in
Figure 2.
v∞
Figure 2: Illustration of R∆ puq highlighted in bold black edges. To avoid clutter, most edges
incident to v8 are only sketched.
Recursive decomposition using a Voronoi diagram In this section, we show how to
obtain a recursive decomposition of R∆ puq into subgraphs called pieces. A piece Q is decomposed
into two smaller pieces by a cycle separator S of size 2 containing v8 . Each of the two subgraphs
of Q is obtained by replacing the faces of Q on one side of S by a single face bounded by S. The
separator S is balanced w.r.t. the number of faces of Q on each side of S. The recursion stops
when a piece with at most six faces is obtained. It will be clear from our construction below
that the collection of cycle separators over all recursive calls form a laminar family, i.e., they are
pairwise non-crossing.
We assume a linked list representation of each piece Q where edges are ordered clockwise
around each vertex.
Lemma 2 below shows how to find the cycle separators needed to obtain the recursive decomposition into pieces. Before we can prove it, we need the following result.
Lemma 1. Let Q be one of the pieces obtained in the above recursive decomposition. Then for
every vertex v of Q other than v8 ,
1. v has at least two edges incident to v8 ,
2. for each edge pv, wq of Q where w ‰ v8 , the edge preceding and the edge following pv, wq
in the clockwise ordering around v are both incident to v8 , and
3. for every pair of edges e1 “ pv, v8 q and e2 “ pv, v8 q where e2 immediately follows e1 in
the clockwise ordering around v, if both edges are directed from v to v8 then the subset of
the plane to the right of e1 and to the left of e2 is a single face of Q.
Proof. The proof is by induction on the depth i ě 0 in the recursion tree of the node corresponding to piece Q. Assume that i “ 0 and let v ‰ v8 be given. The first and third part of the
6
lemma follow immediately from the construction of R∆ puq and the assumption that |δR| ě 3.
To show the second part, it suffices by symmetry to consider the edge e following pv, wq in the
cyclic ordering of edges around v in Q “ R∆ puq. Since e and pv, wq belong to the same face of
Q and since each face of Q contains v8 and at most three edges, the second part follows.
Now assume that i ą 0 and that the claim holds for smaller values. Let v ‰ v8 be given.
Consider the parent piece Q1 of Q in the recursive decomposition tree and let S be the cycle
separator that was used to decompose Q1 . Then S contains v8 and one additional vertex v 1 . To
show the inductive step, we claim that we only need to consider the case when v 1 “ v. This is
clear for the first and second part since if v 1 ‰ v then v has the same set of incident edges in Q1
and in Q. It is also clear for the third part since Q is a subgraph of Q1 .
It remains to show the induction step when v 1 “ v. The first part follows since the two edges
of S are incident to v and to v8 and belong to Q. The second part follows by observing that
the clockwise ordering of edges around v in Q is obtained from the clockwise ordering around v
in Q1 by removing an interval of consecutive edges in this ordering; furthermore, the first and
last edge in the remaining interval are both incident to v8 . Applying the induction hypothesis
shows the second part.
For the third part, if e2 immediately follows e1 in the clockwise ordering around v in Q1 then
the induction hypothesis gives the desired. Otherwise, e1 and e2 must be the two edges of S and
Q is obtained from Q1 by removing the faces to the right of e1 and to the left of e2 and replacing
them by a single face bounded by S.
In the following, let Q be a piece with more than six faces. The following lemma shows that
Q has a balanced cycle separator of size 2 which can be found in Op|Q|q time.
Lemma 2. Q as defined above contains a 2-cycle S containing v8 such that the number of faces
of Q on each side of S is a fraction between 1{3 and 2{3 of the total number of faces of Q.
Furthermore, S can be found in Op|Q|q time.
Proof. We construct S iteratively. In the first iteration, pick an arbitrary vertex v1 ‰ v8 of Q
and let S1 consist of two distinct arbitrary edges, both incident to v1 and v8 . This is possible
by the first part of Lemma 1.
Now, consider the ith iteration for i ą 1 and let vi´1 and v8 be the two vertices of the
2-cycle Si´1 obtained in the previous iteration. If Si´1 satisfies the condition of the lemma, we
let S “ Si´1 and the iterative procedure terminates.
Otherwise, one side of Si´1 contains more than 2{3 of the faces of Q. Denote this set of
faces by Fi´1 and let Ei´1 be the set of edges of Q incident to vi´1 , contained in faces of Fi´1 ,
and not belonging to Si´1 . We must have Ei´1 ‰ H; otherwise, it follows from the third part
of Lemma 1 that Fi´1 contains only a single face of Q (bounded by Si´1 ), contradicting our
assumption that Q contains more than six faces and that Fi´1 contains more than 2{3 of the
faces of Q.
If Ei´1 contains an edge incident to v8 , pick an arbitrary such edge ei´1 . This edge partitions
1
Fi´1 into two non-empty subsets; let Fi´1
be the larger subset. We let vi “ vi´1 and let Si be
the 2-cycle consisting of ei´1 and the edge of Si´1 such that one side of Si contains exactly the
1 .
faces of Fi´1
Now, assume that none of the edges of Ei´1 are incident to v8 . Then by the second part
of Lemma 1, Ei´1 contains exactly one edge ei´1 . We let vi ‰ vi´1 be the other endpoint of
ei´1 and we let Si consist of the two edges incident to vi which belong to the two faces of Fi´1
incident to ei´1 .
To show the first part of the lemma, it suffices to prove that the above iterative procedure
terminates. Consider two consecutive iterations i ą 1 and i ` 1 and assume that the procedure
7
does not terminate in either of these. We claim that then Fi Ă Fi´1 . If we can show this, it
follows that |F1 | ą |F2 | ą . . . which implies termination.
1
If Ei´1 contains an edge incident to v8 then Fi´1
contains more than 1{3 of the faces of Q.
1
Since the procedure does not terminate in iteration i ` 1, Fi´1
must in fact contain more than
1
2{3 of the faces so Fi “ Fi´1 Ă Fi´1 , as desired.
Now, assume that none of the edges of Ei´1 are incident to v8 . Then one side of Si contains
exactly the faces of Fi´1 excluding two. Since we assumed that Q contains more than six faces,
this side of Si contains more than 2{3 of these faces. Hence, Fi Ă Fi´1 , again showing the
desired.
For the second part of the lemma, note that counting the number of faces of Q contained
in one side of a 2-cycle containing v8 can be done in the same amount of time as counting the
number of edges incident to v8 from one edge of the cycle to the other in either clockwise or
counter-clockwise order around v8 . This holds since every face of Q contains v8 . It now follows
easily from our linked list representation of Q with clockwise orderings of edges around vertices
that the ith iteration can be executed in Op|Fi´1 | ´ |Fi |q time for each i ą 1. This shows the
second part of the lemma.
Corollary 1. Given VorpR, uq and R∆ puq, its recursive decomposition can be computed in
?
Op r log rq time.
?
Proof. VorpR, uq has complexity |VorpR, uq| “ Op rq and R∆ puq can be found in time linear in
this complexity. Since the recursive decomposition of R∆ puq has Oplog rq levels and since the
?
total size of pieces on any single level is Op|R∆ puq|q “ Op|VorpR, uq|q “ Op rq, the corollary
follows from Lemma 2.
?
Lemma 3. The recursive decomposition of R∆ puq can be stored using Op rq space.
?
Proof. Observe that the number of nodes of the tree decomposition is Op rq and each separator
consists of two edges and so takes Op1q space.
Embedding of R∆ puq: We now provide a more precise definition of the embedding of R∆ puq.
Let f8 be the face of R corresponding to v8 in R˚ , i.e., f8 is the hole H. Consider the graph
R̃ that consists of R plus a vertex ṽ8 located in f8 and an edge between each vertex of f8 and
ṽ8 . The rest of R̃ is embedded consistently with respect to the embedding of R.
Now, consider the following embedding of R∆ puq. First, embed v8 to ṽ8 . We now specify
the embedding of each edge adjacent to v8 . Recall that each edge e that is adjacent to v8 lies
in a single cell Cpeq of VorpR, uq. For each such edge e going from v8 to a vertex w˚ of R∆ puq,
we embed it so that it follows the edge from ṽ8 to the boundary vertex be of Cpeq, then the
shortest path in R̃ from be to the vertex of Cpeq on the face corresponding to w˚ in R̃. Note
that by definition of R∆ puq such a vertex exists. We also remark that since the edges follow
shortest paths and because of the uniqueness of the shortest paths they may intersect but not
cross (and hence do not contradict the definition of R∆ puq).
It follows that there exists a 1-to-1 correspondence between 2-cycle separators going through
v8 of R∆ puq and cycle separators of R̃ consisting of an edge pu, vq, the shortest paths between
u and a boundary vertex b1 and v and a boundary vertex b2 and pb1 , ṽ8 q and pb2 , ṽ8 q. We call
the set tb1 , u, v, b2 u the representation of this separator. This is illustrated in Figure 3. Thus,
for any 2-cycle separator S going through v8 , we say that the set of vertices of R that is in
the interior (resp. exterior) of S is the set of vertices of R that lie in the bounded region of
the place defined by the Jordan curve corresponding to the cycle separator in R̃ that is in 1-1
correspondence with S.
We can now state the main lemma of this section.
8
b2
v
u
b1
~
v∞
Figure 3: Example of a 2-cycle separator of R∆ puq and its corresponding embedding into the
actual graph.
Lemma 4. Let w be a vertex of R. Assume there exists a data structure that takes as input
a the representation tb1 , x, y, b2 u of a 2-cycle separator S of VorpR, uq going through v8 and
answers in t time queries of the following form: Is w in the bounded closed subset of the plane
with boundary S? Then there exists an algorithm running in time Opt log rq that returns a set
of at most 6 Voronoi cells of VorpR, uq such that one of them contains w.
Proof. The algorithm uses the recursive decomposition of R∆ puq described in this section. Note
that the decomposition consists of 2-cycle separators going through v8 . Thus, using the above
embedding, each of the 2-cycle of the decomposition corresponds to a separator consisting of an
edge px, yq and the shortest paths PR px, b1 q and PR py, b2 q where b1 , b2 are boundary vertices of
R. Additionally, y belongs to the Voronoi cell of b2 in VorpR, uq and x belongs to the Voronoi
cell of b1 in VorpR, uq. Therefore, PR px, b1 q and PR py, b2 q are vertex disjoint and so the data
structure can be used to decide on which side of such a separator w is.
The algorithm is the following: proceed recursively along the recursive decomposition of
R∆ puq and for each 2-cycle separator of the decomposition use the data structure to decide in
t time in which side of the 2-cycle w is located and then recurse on this side. If w belongs to
both sides, i.e., if w is on the 2-cycle separator, recurse on an arbitrary side. The algorithm
stops when there are at most 6 faces of R∆ puq and then it returns the Voronoi cells of VorpR, uq
intersecting those 6 faces.
Observe that the separators do not cross. Thus, when the algorithm obtains at a given
recursive call that w is in the interior (resp. exterior) of a 2-cycle S and in the exterior (resp.
interior) of the 2-cycle separator S 1 corresponding to the next recursive call, we can deduce that
w lies in the intersection of the interior of S and the exterior of S 1 and hence deduce that it
belongs to a Voronoi cell that lies in this area of the plane.
Note that by Lemma 2 the number of faces of R∆ puq in a piece decreases by a constant
?
factor at each step. Thus, since the number of boundary vertices is Op rq, the procedure takes
at most Opt log rq time.
Finally, observe that each face of R∆ puq that is adjacent to v8 lies in a single Voronoi cell
of VorpR, uq. Thus, since at the end of the recursion there are at most 6 faces in the piece, they
correspond to at most 6 different Voronoi cells of VorpR, uq. Hence the algorithm returns at
most 6 different Voronoi cells of VorpR, uq.
9
5
Preprocessing a Region
Given a query separator S in a graph RH “ R and given a query vertex w in R, our data
structure needs to determine in Op1q time the side of S that w belongs to. In this section, we
describe the preprocessing needed for this.
In the following, fix R as well as an ordered pair pu, vq of vertices of R such that either pu, vq
or pv, uq is an edge of R. The preprocessing described in the following is done over all such
choices of R and pu, vq (and all holes H).
The vertices of δR are on a simple cycle and we identify δR with this cycle which we orient
clockwise (ignoring the edge orientations of R). We let b0 , . . . , bk denote this clockwise ordering
where bk “ b0 . It will be convenient to calculate indices modulo k so that, e.g., bk`1 “ b1 .
Given vertices w and w1 in R, let P pw, w1 q denote the shortest path in R from w to w1 .
Given two vertices bi , bj P δR, we let δpbi , bj q denote the subpath of cycle δR consisting of the
vertices from bi to bj in clockwise order, where δpbi , bj q is the single vertex bi if i “ j and
δpbi , bj q “ δR if j “ i ` k. We let ∆pw, bi , bj q denote the subgraph of R contained in the closed
and bounded region of the plane with boundary defined by P pbi , wq, P pbj , wq, and δpbi , bj q.
We refer to ∆pw, bi , bj q as a wedge and call it a basic wedge if bi and bj are consecutive in the
clockwise order, i.e., if j “ i ` 1 pmod kq. We need the following lemma.
Lemma 5. Let w be a given vertex of R. Then there is a data structure with Oprq preprocessing
time and size which answers in Op1q time queries of the following form: given a vertex x P V pRq
and two distinct vertices bi1 , bi2 P δR, does x belong to ∆pw, bi1 , bi2 q?
Proof. Below we present a data structure with the bounds in the lemma which only answers
restricted queries of the form “does x belong to ∆pw, b0 , bi q?” for query vertices x P V pRq
and bi P δR. In a completely symmetric manner, we obtain a data structure for restricted
queries of the form “does x belong to ∆pw, bi , bk q?” for query vertices x P V pRq and bi P δR.
We claim that this suffices to show the lemma. For consider a query consisting of x P V pRq
and bi , bj P δR. If b0 P δpbi , bj q then ∆pw, bi , bj q “ ∆pw, bi , bk q Y ∆pw, b0 , bj q and otherwise,
δpw, bi , bj q “ ∆pw, b0 , bj q X ∆pw, bi , bk q. Hence, answering a general query can be done using
two restricted queries and checking if b0 P δpb1 , b2 q can be done in constant time by comparing
indices of the query vertices.
It remains to present the data structure for restricted queries of the form “does x belong
to ∆pw, b0 , bi q?”. In the preprocessing step, each v P V pRq is assigned the smallest index iv P
t0, . . . , ku for which v P ∆pw, b0 , biv q. Clearly, this requires only Oprq space and below we show
how to compute these indices in Oprq time.
Consider a restricted query specified by a vertex x of R and a boundary vertex bi P δR where
0 ď i ď k. Since x P ∆pw, b0 , bi q iff ix ď i, this query can clearly be answered in Op1q time.
It remains to show how the indices iv can be computed in a total of Oprq time. Let R1 be
R with all its edge directions reversed. In Oprq time, a SSSP tree T 1 from w in R1 is computed.
Let T be the tree in R obtained from T 1 by reversing all its edge directions; note that all edges
of T are directed towards w and for each v P V pRq, the path from v to w in T is a shortest path
from v to w in R.
Next, ∆pw, b0 , b0 q “ P pb0 , wq is computed and for each vertex v P ∆pw, b0 , b0 q, set iv “ 0.
The rest of the preprocessing algorithm consists of iterations i “ 1, . . . , k where iteration i assigns
each vertex v P V p∆pw, b0 , bi qqzV p∆pw, b0 , bi´1 qq the index iv “ i. This correctly computes
indices for all vertices of R. In the following, we describe how iteration i is implemented.
First, the path P pbi , wq is traversed in T until a vertex vi is encountered which previously
received an index. In other words, vi is the first vertex on P pbi , wq belonging to ∆pw, b0 , bi´1 q.
Note that vi is well-defined since w P ∆pw, b0 , bi´1 q. Vertices that are in V pP pbi , vi qqztvi u or in
10
a subtree of T rooted in a vertex of V pP pbi , vi qqztvi u and extending to the right of this path
are assigned the index value i. Furthermore, vertices belonging to a subtree of T rooted in a
vertex of V pP pbi´1 , vi qqztvi u and extending to the left of this path are assigned the index value
i, except those on P pbi´1 , vi q (as they belong to ∆pw, b0 , bi´1 q).
Since R is connected, it follows that the vertices assigned an index of i are exactly those
belonging to V p∆pw, b0 , bi qqzV p∆pw, b0 , bi´1 qq and that the running time for making these assignments is Op|V p∆pw, b0 , bi qqzV p∆pw, b0 , bi´1 qq| ` |P pbi´1 , vi q ´ vi | ` 1q. Over all i, total
running time is Oprq; this follows by a telescoping sums argument and by observing that vertex
sets V pP pbi´1 , vi qqztvi u are pairwise disjoint.
Given distinct vertices bi1 , bi2 P δR, if P pbi1 , uq and P pbi2 , vq do not cross (but may touch and
then split), let lpbi1 , bi2 , u, vq denote the subgraph of R contained in the closed and bounded
region of the plane with boundary defined by P pbi1 , uq, P pbi2 , vq, δpbi1 , bi2 q, and an edge of R
between vertex pair pu, vq. In order to simplify notation, we shall omit u and v and simply write
lpbi1 , bi2 q.
It follows from planarity that there is at most one bi P δR such that pu, vq belongs to
Ep∆pu, bi , bi`1 qqzEpP pbi`1 , uqq when ignoring edge orientations. If bi exists, we refer to it as
buv ; otherwise buv denotes some dummy vertex not belonging to R.
The goal in this section is to determine whether a given query vertex belongs to a given
query subgraph lpbi1 , bi2 q. The following lemma allows us to decompose this subgraph into
three simpler parts as illustrated in Figure 4. We will show how to answer containment queries
for each of these simple parts.
v
u
bi
bi
2
1
bi+1
bi
Figure 4: Example of decomposing the region into three parts using bi : The dashed wedges
represent shortest paths to u, the dotted edges represent shortest paths to v, and the dasheddotted box lpbi , bi`1 q.
Lemma 6. Let bi1 and bi2 be distinct vertices of δR and assume that P pbi1 , uq and P pbi2 , vq are
vertex-disjoint. Then lpbi1 , bi2 q “ ∆pu, bi1 , bi q Y lpbi , bi`1 q Y ∆pv, bi`1 , bi2 q where bi “ buv if
buv P δpbi1 , bi2 ´1 q and bi “ bi2 ´1 otherwise.
Proof. Figure 5 gives an illustration of the proof. We first show the following result: given a
vertex bj P δpbi1 `1 , bi2 q such that buv R δpbi1 , bj´1 q, P pbj 1 , uq is contained in lpbi1 , bi2 q for each
11
bj 1 P δpbi1 , bj q. The proof is by induction on the number i of edges in δpbi1 , bj 1 q. The base case
i “ 0 is trivial since then bj 1 “ bi1 so assume that i ą 0 and that the claim holds for i ´ 1.
If pv, uq is the last edge on P pbj 1 , uq, the induction step follows from uniqueness of shortest
paths. Otherwise, neither pu, vq nor pv, uq belong to ∆pu, bj 1 ´1 , bj 1 q (since bj 1 ´1 ‰ buv ). By
the induction hypothesis, P pbj 1 ´1 , uq is contained in lpbi1 , bi2 q and since P pbj 1 , uq cannot cross
P pbj 1 ´1 , uq, P pbj 1 , uq cannot cross P pbi1 , uq. Also, P pbj 1 , uq cannot cross P pbi2 , vq since then
either pu, vq or pv, uq would belong to ∆pu, bj 1 ´1 , bj 1 q. Since bj 1 R tbi1 `1 , bi2 ´1 u, it follows that
P pbj 1 , uq is contained in lpbi1 , bi2 q which completes the proof by induction.
Next, assume that buv R δpbi1 , bi2 ´1 q so that bi “ bi2 ´1 . Note that ∆pv, bi`1 , bi2 q “ P pbi2 , vq.
Picking bj “ bi2 above implies that ∆pu, bi1 , bi q is contained in lpbi1 , bi2 q and hence lpbi1 , bi2 q “
∆pu, bi1 , bi q Y lpbi , bi`1 q Y ∆pv, bi`1 , bi2 q.
Now consider the other case of the lemma where bi “ buv P δpbi1 , bi2 ´1 q. Picking bj “ bi
above, it follows that ∆pu, bi1 , bi q is contained in lpbi1 , bi2 q. It suffices to show that P pbi`1 , vq is
contained in lpbi1 , bi2 q since this will imply that lpbi , bi`1 q is well-defined and that lpbi , bi`1 q Y
∆pv, bi`1 , bi2 q is contained in lpbi1 , bi2 q and hence that ∆pu, bi1 , bi qYlpbi , bi`1 qY∆pv, bi`1 , bi2 q “
lpbi1 , bi2 q.
Assume for contradiction that P pbi`1 , vq is not contained in lpbi1 , bi2 q.
By uniqueness of shortest paths, P pbi`1 , uq does not cross P pbi , uq. Since bi “ buv , we
have that when ignoring edge orientations, pu, vq belongs to Ep∆pu, bi , bi`1 qqzEpP pbi`1 , uqq Ď
Ep∆pu, bi1 , bi`1 qqzEpP pbi`1 , uqq. Hence P pbi`1 , uq is not contained in lpbi1 , bi2 q so it crosses
P pbi2 , vq. Let x be a vertex on P pbi`1 , uq X P pbi2 , vq such that the successor of x on P pbi ` 1, uq
is not contained in lpbi1 , bi2 q. By our assumption above that P pbi`1 , vq is not contained in
lpbi1 , bi2 q, there is a first vertex y on P pbi`1 , vq such that its successor y 1 does not belong to
lpbi1 , bi2 q. By uniqueness of shortest paths, y cannot belong to P pbi2 , vq so it must belong to
P pbi1 , uq. This also implies that y ‰ x since x P P pbi2 , vq and P pbi1 , uq and P pbi2 , vq are vertexdisjoint. Since P py, vq is a subpath of P pbi`1 , vq and y ‰ x, shortest path uniqueness implies
that P py, vq and P pbi`1 , uq are vertex-disjoint.
Since P pbi`1 , yq is contained in lpbi1 , bi2 q, v belongs to the subgraph of ∆pu, bi1 , bi`1 q contained in the closed region of the plane bounded by P pbi`1 , yq, P py, uq, and P pbi`1 , uq. Since
P py, vq does not intersect P pbi`1 , uq, P py 1 , vq thus intersects either P pbi`1 , yq or P py, uq. However, it cannot intersect P pbi`1 , yq since then P pbi`1 , vq would be non-simple. By uniqueness of
shortest paths, P py 1 , vq also cannot intersect P py, uq since P py 1 , vq is a subpath of P py, vq and
y 1 R P py, uq. This gives the desired contradiction, concluding the proof.
Let P be a collection of subpaths such that for each path δpbi1 , bi2 q in P, buv R δpbi1 , . . . , bi2 ´1 q
and bvu R δpbi1 `1 , . . . , bi2 ´1 q. We may choose the paths such that |P| “ Op1q and such that
all edges of δR except pbuv , bvu q (if it exists) belongs to a path of P. It is easy to see that
this is possible by considering a greedy algorithm which in each step picks a maximum-length
path which is edge-disjoint from previously picked paths and which satisfies the two stated
requirements.
The next lemma allows us to obtain a compact data structure to answer queries of the form
“does face f belong to lpbi , bi`1 q” for given query face f and query index i.
Lemma 7. Let P “ δpbi1 , bi2 q P P be given. Then
1. an index jpP q exists with i1 ď jpP q ď i2 such that lpbi , bi`1 q is undefined for i1 ď i ă jpP q
and well-defined for jpP q ď i ă i2 ,
2. for each face f ‰ δR of R, there is at most one index jf pP q with jpP q ď jf pP q ď i2 ´ 2
such that f Ď lpbjf pP q , bjf pP q`1 q and f Ę lpbjf pP q`1 , bjf pP q`2 q, and
12
v
u
y'
x
y
bi
bi
1
2
bi
Figure 5: Illustration of the proof of Lemma 6. The figure shows how unique shortest paths
imply a contradiction (highlighted with grey) if the path from bi`1 to v is not contained in
lpbi1 , bi2 q.
3. for each face f ‰ δR of R, there is at most one index jf1 pP q with jpP q ď jf1 pP q ď i2 ´ 2
such that f Ę lpbjf1 pP q , bjf1 pP q`1 q and f Ď lpbjf1 pP q`1 , bjf1 pP q`2 q.
Furthermore, there is an algorithm which computes the index jpP q and for each face f ‰ δR of
R the indices jf pP q and jf1 pP q if they exist. The total running time of this algorithm is Opr log rq
and its space requirement is Oprq.
Proof. Let path P “ δpbi1 , bi2 q P P and face f ‰ δR of R be given. To simplify notation in the
proof, we shall omit reference to P and write, e.g., j instead of jpP q.
Let j be the smallest index such that lpbj , bj`1 q is well-defined; if j does not exist, pick
instead j “ i2 . We will show that j satisfies the first part of the lemma. This is clear if j “ i2
so assume therefore in the following that j ă i2 .
We prove by induction on i that lpbi , bi`1 q is well-defined for j ď i ă i2 . By definition of j,
this holds when i “ j. Now, consider a well-defined subgraph lpbi , bi`1 q where j ď i ď i2 ´ 2.
We need to show that lpbi`1 , bi`2 q is well-defined. Since bi ‰ buv , P pbi`1 , uq is contained in
lpbi , bi`1 q and since bi`1 ‰ bvu , P pbi`2 , vq is contained in the closed region of the plane bounded
by the boundary of lpbi , bi`1 q and not containing lpbi , bi`1 q. In particular, lpbi`1 , bi`2 q is welldefined. This shows the first part of the lemma.
Next, we show that j can be computed in Opr log rq time. Checking that lpbi , bi`1 q is welldefined (i.e., that paths P pbi , uq and P pbi`1 , vq do not cross) for a given index i can be done in
Oprq time. Because of the first part of the lemma, a binary search algorithm can be applied to
identify j in Oplog rq steps where each step checks if lpbi , bi`1 q is well-defined for some index i.
This gives a total running time of Opr log rq, as desired. Space is clearly Oprq.
To show the second part of the lemma, assume that there is an index jf with i1 ď jf ď
i2 ´ 2 such that f Ď lpbjf , bjf `1 q and f Ę lpbjf `1 , bjf `2 q. It follows from the observations
in the inductive step above that ∆pu, bjf , bjf `1 q contains exactly the faces of R contained in
lpbjf , bjf `1 q and not in lpbjf `1 , bjf `2 q which implies that f Ď ∆pu, bjw , bjw `1 q. Since no face
of R belongs to more than one graph of the form ∆pu, bi , bi`1 q, jf must be unique, showing the
second part of the lemma.
13
Next, we give an Oprq time and space algorithm that computes indices jf . Let T be constructed as in the proof of Lemma 5. Initially, vertices of P pbi2 ´1 , uq are marked and all other
vertices of R are unmarked. The remaining part of the algorithm consists of iterations i2 ´2, . . . , j
in that order. In iteration i, P pbi , uq is traversed until a marked vertex vi is visited and then
the vertices of P pbi , vi q are marked. The faces of R contained in the bounded region of the
plane defined by P pbi , vi q, P pbi`1 , vi q, and δpbi , bi`1 q are exactly those that should be given an
index value of i. The algorithm performs this task by traversing each subtree of T emanating to
the right of P pbi , vi q and each subtree of T emanating to the left of P pbi`1 , vi q; for each vertex
visited, the algorithm assigns the index value i to its incident faces.
We now show that the algorithm for computing indices jf has Oprq running time. Using the
same arguments as in the proof of Lemma 5, the total time to traverse and mark paths P pbi , vi q
is Oprq. The total time to assign indices to faces is Oprq; this follows by observing that the time
spent on assigning indices to faces incident to a vertex of T is bounded by its degree and this
vertex is not visited in other iterations.
The third part of the lemma follows with essentially the same proof as for the second part.
We can now combine the results of this section to obtain the data structure described in the
following lemma.
Lemma 8. Let pu, vq be a vertex pair connected by an edge in R. Then there is a data structure
with Opr log rq preprocessing time and Oprq space which answers in Op1q time queries of the
following form: given a vertex w P R and two distinct vertices bi , bj P δR such that P pbi , uq and
P pbj , vq are vertex-disjoint, does w belong to lpbi , bj , u, vq?
Proof. We present a data structure Dpu, vq satisfying the lemma. First we focus on the preprocessing. Boundary vertices buv and bvu and set P as defined above are precomputed and stored.
Vertices of δR are labeled with indices b0 , . . . , b|V pδRq|´1 according to a clockwise walk of δR.
Each path of the form δpbi1 , bi2 q (including each path in P) is represented by the ordered index
pair pi1 , i2 q. Checking if a given boundary vertex belongs to such a given path can then be done
in Op1q time.
Next, two instances of the data structure in Lemma 5 are set up, one for u denoted Du ,
and one for v denoted Dv . Then the following is done for each path P “ δpbi1 , bi2 q P P. First,
the algorithm in Lemma 7 is applied. Then if lpbjpP q , bjpP q`1 , u, vq is well-defined, its set of
faces FjpP q is computed and stored; otherwise, FjpP q “ H. Similarly, if lpbi2 ´1 , bi2 , u, vq is welldefined, its set of faces Fi2 ´1 is computed and stored, and otherwise Fi2 ´1 “ H. If pbuv , bvu q P
δR, Dpu, vq computes and stores the set Fuv of faces of R contained in lpbuv , bvu , u, vq. This
completes the description of the preprocessing for Dpu, vq. It is clear that preprocessing time is
Opr log rq and that space is Oprq.
Now, consider a query specified by a vertex w P R and two distinct vertices bi1 , bi2 P δR
such that P pbi1 , uq and P pbi2 , vq are pairwise vertex-disjoint. First, Dpu, vq identifies a boundary
vertex bi such that lpbi1 , bi2 , u, vq “ ∆pu, bi1 , bi q Y lpbi , bi`1 q Y ∆pv, bi`1 , bi2 q; this is possible
by Lemma 6. Then Du and Dv are queried to determine if w P ∆pu, bi1 , bi q Y ∆pv, bi`1 , bi2 q; if
this is the case then w P lpbi1 , bi2 , u, vq and Dpu, vq answers “yes”. Otherwise, Dpu, vq identifies
an arbitrary face f ‰ δR of R incident to w. At this point, the only way that w can belong to
lpbi1 , bi2 , u, vq is if w belongs to the interior of lpbi , bi`1 , u, vq which happens iff f is contained
in lpbi , bi`1 , u, vq. If pbi , bi`1 q “ pbuv , bvu q, Dpu, vq checks if f P Fuv and if so outputs “yes”.
Otherwise, there exists a path P P P containing pbi , bi`1 q and Dpu, vq identifies this path. It
follows from Lemma 7 and from the definition of P that f is contained in lpbi , bi`1 , u, vq iff at
least one of the following conditions hold:
1. jf pP q and jf1 pP q are well-defined and jf1 pP q ă i ď jf pP q.
14
2. f P Fi2 ´1 , jf1 pP q is well-defined, and i ą jf1 pP q,
3. f P FjpP q , jf pP q is well-defined, and i ď jf pP q,
4. f P FjpP q and jf pP q is undefined,
Data structure Dpu, vq checks if any one these conditions hold and if so outputs “yes”; otherwise
it outputs “no”. Two of the cases are illustrated in Figure 6.
bi
2
bjf(P)
bi+1
bi
bi
bj'f(P)
bi+1
bi
1
bi
1
bj(P)
2
bj'f(P)
bi
bj(P)
Figure 6: Illustration of how to determine if the face f belongs to lpbi , bi`1 , u, vq. The illustration includes cases 1 and 4 in the proof of Lemma 8. The large gray subpath indicates the part,
where f is contained in each box defined by consecutive boundary nodes.
It remains to show that Dpu, vq has query time Op1q. By Lemma 6, identifying bk can be
done in Op1q time. Querying Du and Dv takes Op1q time by Lemma 5. Checking whether
f P Fuv can clearly be done in Op1q time since this set of faces is stored explicitly. With our
representation of paths by the indices of their endpoints, identifying P takes Op1q time. Finally,
since sets FjpP q and FjpP q are explicitly stored, the four conditions above can be checked in Op1q
time.
6
The Distance Oracle
In this section we give a detailed presentation of both our algorithm for answering distance
queries and our distance oracle data structure.
Combining Lemmas 9, 10 and 11 with r “ n2{3 directly implies Theorem 1.
6.1
The Data Structure
We present the algorithm for building our data structure.
Preprocessing G
1. Compute an r-division R of G. Let δ be the set of all boundary vertices.
2. Store for each internal vertex the region to which it belongs.
3. Compute and store the distances from each vertex to each boundary vertex.
4. For each region R P R, compute and store the distances between any pair of internal
vertices of R.
15
5. For each region R, for each vertex u R R, for each hole H, compute VorH pR, uq and store
a separator decomposition as described in Section 4.
6. For each region R, for each edge px, yq P R, for each hole H, compute and store the data
structure described in Section 5.
?
Lemma 9. The total size of the data structure computed by Preprocessing is Opn2 { r`n¨rq.
?
Proof. Recall that by definition of the r-division, there are Opn{ rq boundary vertices and
Opn{rq regions. Thus, the number of distances stored at step 3 of the algorithm is at most
?
Opn2 { rq.
For a given region, storing the pairwise distances between all its internal vertices takes Opr2 q
space. Since there are Opn{rq regions in total, Step 4 takes memory Opn ¨ rq.
We now bound the space taken by Step 5. There are n{r choices for R and n choices for
?
u. By Lemma 3, each decomposition can be stored using Op rq space. Thus, this step takes
?
Opn2 { rq total space.
We finally bound the space taken by Step 6. There n{r choices for R and r choices for an
edge px, yq.By Lemma 8, for a given edge px, yq, the data structure takes Oprq space. Hence,
the total space taken by this step is Opn ¨ rq and the lemma follows.
?
Theorem 3. The execution of Preprocessing takes Opn2 q time and Opn ¨ r ` n2 { rq space.
Proof. We analyze the procedure step by step. Computing an r-division with Op1q holes can be
done in linear time and space using the algorithm of Klein et al., see [23].
?
We now analyze Step 3. There are Opn{ rq boundary vertices. Computing single-source
shortest paths can be done in linear time using the algorithm of Henzinger et al. [16]. Hence,
?
Step 3 takes at most Opn2 { rq time and space.
Step 4 takes Opn ¨ rq time and space using the following algorithm. The algorithm proceed
region by region and hole by hole. For a given region and hole, the algorithm adds an edge
between each pair of boundary vertices that are on the hole of length equal to the distance
between these vertices in the whole graph. Note that this is already in memory and was computed
at Step 3. Now, for each vertex of the region, the algorithm runs a shortest path algorithm. Since
?
there are Op rq boundary vertices, the number of edges added is Oprq. Thus, the algorithm is
run on a graph that has at most Oprq edges and vertices. The algorithm spends at most Oprq
time per vertex of the region. Since there are Opn{rq regions and Oprq vertices per region, both
the running time and the space are Opn ¨ rq.
Step 5 takes Opn2 q using the following algorithm. The algorithm proceeds vertex by vertex,
region by region, hole by hole. For a given vertex u, a given region R, and a given hole H the
algorithm computes VorH pR, uq. This can be done by adding a “dummy” vertex reprensenting
u and connecting it to each boundary vertex x of the hole by an edge of length distpu, xq. Thus,
this takes time Oprq using the single-source shortest path algorithm of Henzinger et al [16].
Furthermore, by Lemma 3 and Corollary 1, computing the separator decomposition of Section 4
?
?
given VorH pR, uq takes Õp rq time and Op rq space. Thus, over all vertices, regions and holes,
?
this step takes Opn2 q time and Opn2 { rq memory.
Finally, we show that Step 6 takes Õpn ¨ rq time and Opn ¨ rq space. The algorithm proceeds
region by region, hole by hole, and edge by edge. By Lemma 8, for a given edge of the region,
computing the data structure of Section 5 takes Opr ¨ log rq time and space. Since the total
number of region is Opn{rq and the total number of edges per region is Oprq, the proof is
complete.
Corollary 2. There exists a distance oracle with total space Opn11{6 q and expected preprocessing
time Opn11{6 q.
16
Proof. We apply Procedure Preprocessing with r “ n1{3 . By Lemma 9, the total size of the
data structure output is Opn11{6 q.
We now analyse the total preprocessing time. For Steps 3, 4, and 6, we mimicate the analysis
?
of the proof of Theorem 3 and obtain a total preprocessing time of Opn ¨ r ` n2 { rq.
We now explain how to speed-up Step 5. We show that Step 5 can be done in time Opn ¨ r5{2 q
using Cabello’s data structure [5] for computing weighted Voronoi diagrams of a given region.
More formally, Cabello introduces a data structure that allows to compute weighted Voronoi
?
diagrams of a given region in expected time Õp rq. This data structure has preprocessing time
Opr7{2 q. Hence the total preprocessing time for computing the data structure for all the regions
is Õpn ¨ r5{2 q.
Then, for each vertex u, each region R, each hole H, the algorithm
1. uses the data structure to compute the weighted Voronoi diagram VorH pR, uq in expected
?
time Õp rq and
?
2. computes the separator decomposition of Section 4 in time Õp rq (by Lemma 3 and
Corollary 1).
?
This results in an expected preprocessing time of Opn ¨ r5{2 ` n2 { rq. Choosing r “ n1{3 yields
a bound of Opn11{6 q.
6.2
Algorithm for Distance Queries
This section is devoted to the presentation of our algorithms for answering distance queries
between pairs of vertices.
We show that any distance query between two vertices u, v can be performed in Oplog rq
time. In the following, let u, v be two vertices of the graph. The algorithm is the following.
Distance Query u, v
1. If u, v belong to the same region or if either u or v is a boundary vertex, the query can be
answered in Op1q time since the distances between vertices of the same region and between
boundary vertices and the other vertices of the graph are stored explicitly.
2. If u and v are internal vertices that belong to two different regions we proceed as follows.
Let R be the region containing v and δR be the set of boundary vertices ofŤregion R. The
boundary vertices are partitioned into holes H “ tH0 , . . . , Hk u, such that HPH H “ δR.
For each H P H, we apply the following procedure. Let V be the weighted Voronoi diagram
where the sites are the vertices of H and the weight of x P H is the distance from u to x.
We now aim at determining to which cell of V, v belongs. We use the binary search
procedure of Lemma 4 on the decomposition of R induced by the separators of the weighted
Voronoi diagram. More precisely, we use the algorithm described in Section 4, Lemma 4,
and the query algorithm described in Section 5, Lemma 8 to identify a set of at most six
Voronoi cells so that one of them contains v. This induces a set of at most six boundary
vertices X “ tx0 , . . . , xk u that represent the centers of the cells.
Finally, we have the distances from both u and v to all the boundary vertices in X. Let
vpHq “ minxPX distpu, xq ` distpx, vq. The algorithm returns minHPH vpHq.
Lemma 10 (Running time). The Distance Query takes Oplog rq time.
17
Proof. Consider a distance query from a vertex u to a vertex v and assume that those vertices
are internal vertices of two different regions as otherwise the query takes Op1q time. Observe
that we can determine in Op1q time to which region v belongs. Fix a hole H. Let V be the
weighted Voronoi diagram where the sites are the vertices of H and the weight of x P δR is the
distance from u to x. We consider the decomposition of the region of v of VorpR, uq.
Lemma 8 shows that the query time for the data structure defined in Section 5 is t “ Op1q.
Applying Lemma 4 with t “ Op1q implies that the total time to determine in which Voronoi cell
v belongs is at most Oplog rq.
Finally, computing minxPX distpu, xq ` distpx, vq takes Op1q time. By definition of the rdivision there are Op1q holes.
Therefore, we conclude that the running time of the Distance Query algorithm is Oplog rq.
We now prove that the algorithm indeed returns the correct distance between u and v.
Lemma 11 (Correctness). The Distance Query on input u, v returns the length of the shortest
path between vertices u and v in the graph.
Proof. We remark that the distance from any vertex to a boundary vertex is stored explicitly
and thus correct. Hence, we consider the case where u and v are internal vertices of different
regions. Let P be the shortest path from u to v in G. Let x P P be the last boundary vertex of
R on the path from u to v and let Hx be the hole containing x. Let V be the weighted Voronoi
diagram where the sites are the boundary vertices of Hx and the weight of y P Hx is the distance
from u to y.
We need to argue that the data structure of Section 5, Lemma 8 satisfies the conditions
of Lemma 4. Observe that the separators defined in Section 4 consist of two shortest paths
PR pb1 , xq and PR pb2 , yq where b1 , b2 P δR and py, zqis an edge of R. Hence, the set of vertices
of the subgraph lpb1 , b2 , y, zq correspond to the set of all the vertices that are one of the two
sides of the separator. Thus, by Lemma 8 the data structure described in Section 5 satisfies the
condition of Lemma 4, with query time t “ Op1q.
We argue that v belongs to the Voronoi cell of x. Assume towards contradiction that v is
in the Voronoi cell of y ‰ x, we would have distpy, vq ` wpyq ď distpx, vq ` wpxq, where w is
the weight function associated with the Voronoi diagram. Thus, this implies that distpy, vq `
distpy, uq ď distpx, vq ` distpx, uq. Therefore, there exists a shortest path from u to v that goes
through y. Now observe that if v belongs to the Voronoi cell of y, the shortest path from v to
y does not go through x. Hence, assuming unique shortest paths between pairs of vertices, we
conclude that the last boundary vertex on the path from u to v is y and not x, a contradiction.
Thus, v belongs to the Voronoi cell of x.
Combining with Lemma 8, it follows that the Voronoi cell of x is in the set of Voronoi
cells X obtained at the end of the recursive procedure. Observe that for any x1 P X there
exists a path (possibly with repetition of vertices) of length distpx1 , uq ` distpx1 , vq. Therefore,
since we assume unique shortest paths between pairs of vertices, we conclude that distpu, vq “
distpx, uq ` distpx, vq “ minx1 PX distpx1 , vq ` distpx1 , uq “ vpHx q “ minH vpHq.
7
Trade-off
We now prove Theorem 2. We first consider the case with P “ Opn2 q and then extend it to the
case with efficient preprocessing time.
Let r1 ď r2 be positive integers to be defined later. The case r1 ď r2 will correspond exactly
?
to S ě n n. The data structure works as follows.
18
1. We compute an r1 -division and an r2 -division of G named R1 and R2 respectively. Let δ1
respectively δ2 denote all the boundary vertices of R1 resp. R2 .
2. For each region R of R1 and R2 , compute and store the pairwise distances of the nodes
of R.
3. For each u P δ1 and v P δ2 , compute and store the distance between u and v in G.
4. For each u P δ1 , region R P R2 and hole H P R, compute VorH pR, uq and store a separator
decomposition as described in Section 4.
5. For each Region R P R2 , for each edge px, yq P R, for each hole H, compute and store the
data structure described in Section 5.
We start by bounding the space.
Lemma 12. The total size of the data structure described above is
ˆ
˙
n2
O nr2 ` ?
.
r1 r2
Proof. Consider the steps above. By definition of R1 and R2 step 2 uses Opnr2 q space (since we
?
assumed r1 ď r2 ) and step 4 uses ?nr1 ¨ ?nr2 space. By Lemma 3 step 5 takes Opn{ r2 q for each
?
node of δ1 giving Opn2 { r1 r2 q in total. By Lemma 8 the total space of step 6 is Opnr2 q.
Now consider a query pair u, v. If u and v belong to the same region in R1 or R2 we return
the stored distance. Otherwise we iterate over each boundary node w in the region of u in R1 .
For each such boundary node we compute the distance to v using the data structures of steps
5 and 6 above similar to the query algorithm from Section 6.2. This is possible since we have
stored the distances between all the needed boundary nodes in step 4. The minimum distance
returned over all such w is the answer to the query.
?
From the description above it is clear that we get a query time of Op r1 logpr2 qq. The
correctness follows immediately from the discussion in the proof of Theorem 1. What is left now
is to balance the space to obtain Theorem 2. The expression of Lemma 12 is balanced when
r2 “
n2{3
1{3
.
r1
?
?
Now, since we assumed that S ě n n we can focus on the case when r2 ě n and thus we get
r1 ď r2 as we required. Plugging into the definition of Q we get exactly
?
n5{2
Q “ Op r1 log nq “ 3{2 log n ,
S
which gives us Theorem 2.
For pre-processing time, we consider two cases similar to Theorem 3 and Corollary 2. It
follows directly from the discussion above and Theorem 3 that the preprocessing can be performed in Opn2 q. We may, however, also consider pre-processing time as a parameter similar
to space and query time. This gives a 3-way trade-off. In Corollary 2 we showed how to lower
pre-processing time by increasing the space. Here we discuss the case of lowering pre-processing
time further by increasing the query time. It follows from the discussion above and Corollary 2
?
5{2
that we can perform pre-processing of the above structure in Opnr1 ` nr2 ` n2 { r1 r2 q time
5{2
and get the same space bound. If we assume that r1 ď r2 we get a data structure with query
19
?
?
5{2
time Q “ Op r1 log nq and space and pre-processing time S “ Opnr2 ` n2 { r1 r2 q. Up to
logarithimic and constant factors, this gives us Q “ n11{5 {S 6{5 . For any S ě n16{11 this sat5{2
isfies the requirement that r1 ď r2 . As an example, we get a data structure with space and
pre-processing time Opn16{11 q and a query time of Opn5{11 log nq.
20
References
[1] The TSPLIB. http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/. Accessed:
2010-09-30.
[2] Amir Abboud and Søren Dahlgaard. Popular conjectures as a barrier for dynamic planar
graph algorithms. In Proc. 57th IEEE Symposium on Foundations of Computer Science
(FOCS), pages 477–486, 2016.
[3] Srinivasa Rao Arikati, Danny Z. Chen, L. Paul Chew, Gautam Das, Michiel H. M. Smid,
and Christos D. Zaroliagis. Planar spanners and approximate shortest path queries among
obstacles in the plane. In Proc. 4th European Symposium on Algorithms (ESA), pages
514–528, 1996.
[4] Sergio Cabello. Many distances in planar graphs. Algorithmica, 62(1-2):361–381, 2012. See
also SODA’06.
[5] Sergio Cabello. Subquadratic algorithms for the diameter and the sum of pairwise distances
in planar graphs. In Proc. 28th ACM/SIAM Symposium on Discrete Algorithms (SODA),
pages 2143–2152, 2017.
[6] Shiri Chechik. Approximate distance oracles with constant query time. In Proc. 46th ACM
Symposium on Theory of Computing (STOC), pages 654–663, 2014.
[7] Shiri Chechik. Approximate distance oracles with improved bounds. In Proc. 47th ACM
Symposium on Theory of Computing (STOC), pages 1–10, 2015.
[8] Danny Z. Chen and Jinhui Xu. Shortest path queries in planar graphs. In Proc. 22nd ACM
Symposium on Theory of Computing (STOC), pages 469–478, 2000.
[9] Vincent Cohen-Addad, Søren Dahlgaard, and Christian Wulff-Nilsen. Fast and compact
exact distance oracle for planar graphs. CoRR, abs/1702.03259, 2017.
[10] Hristo Djidjev. On-line algorithms for shortest path problems on planar digraphs. In
Graph-Theoretic Concepts in Computer Science, 22nd International Workshop, WG ’96,
Cadenabbia (Como), Italy, June 12-14, 1996, Proceedings, pages 151–165, 1996.
[11] Paul Erdős. Extremal problems in graph theory. In IN “THEORY OF GRAPHS AND ITS
APPLICATIONS,” PROC. SYMPOS. SMOLENICE. Citeseer, 1964.
[12] Jittat Fakcharoenphol and Satish Rao. Planar graphs, negative weight edges, shortest paths,
and near linear time. Journal of Computer and System Sciences, 72(5):868–889, 2006. See
also FOCS’01.
[13] Greg N. Frederickson. Fast algorithms for shortest paths in planar graphs, with applications.
SIAM Journal on Computing, 16(6):1004–1022, 1987.
[14] Pawel Gawrychowski and Adam Karczmarz. Improved bounds for shortest paths in dense
distance graphs. CoRR, abs/1602.07013, 2016.
[15] Paweł Gawrychowski, Haim Kaplan, Shay Mozes, Micha Sharir, and Oren Weimann.
Voronoi diagrams on planar graphs, and computing the diameter in deterministic o (n5/3)
time, 2017. arXiv preprint.
21
[16] Monika R Henzinger, Philip Klein, Satish Rao, and Sairam Subramanian. Faster shortestpath algorithms for planar graphs. Journal of Computer and System Sciences, 55(1):3–23,
1997.
[17] Giuseppe F Italiano, Yahav Nussbaum, Piotr Sankowski, and Christian Wulff-Nilsen. Improved algorithms for min cut and max flow in undirected planar graphs. In Proc. 43rd
ACM Symposium on Theory of Computing (STOC), pages 313–322, 2011.
[18] Haim Kaplan, Shay Mozes, Yahav Nussbaum, and Micha Sharir. Submatrix maximum
queries in monge matrices and monge partial matrices, and their applications. In Proc.
23rd ACM/SIAM Symposium on Discrete Algorithms (SODA), pages 338–355, 2012.
[19] Ken-ichi Kawarabayashi, Philip N Klein, and Christian Sommer. Linear-space approximate distance oracles for planar, bounded-genus and minor-free graphs. In International
Colloquium on Automata, Languages, and Programming, pages 135–146. Springer, 2011.
[20] Ken-ichi Kawarabayashi, Christian Sommer, and Mikkel Thorup. More compact oracles for
approximate distances in undirected planar graphs. In Proceedings of the Twenty-Fourth
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 550–563. SIAM, 2013.
[21] Philip N. Klein. Preprocessing an undirected planar network to enable fast approximate
distance queries. In Proc. 13th ACM/SIAM Symposium on Discrete Algorithms (SODA),
pages 820–827, 2002.
[22] Philip N Klein. Multiple-source shortest paths in planar graphs. In Proc. 16th ACM/SIAM
Symposium on Discrete Algorithms (SODA), volume 5, pages 146–155, 2005.
[23] Philip N Klein, Shay Mozes, and Christian Sommer. Structured recursive separator decompositions for planar graphs in linear time. In Proc. 45th ACM Symposium on Theory of
Computing (STOC), pages 505–514. ACM, 2013.
[24] Rolf Klein. Concrete and Abstract Voronoi Diagrams, volume 400 of Lecture Notes in
Computer Science. Springer, 1989.
[25] Rolf Klein, Elmar Langetepe, and Zahra Nilforoushan. Abstract voronoi diagrams revisited.
Comput. Geom., 42(9):885–902, 2009.
[26] Shay Mozes and Christian Sommer. Exact distance oracles for planar graphs. In Proc. 23rd
ACM/SIAM Symposium on Discrete Algorithms (SODA), pages 209–222, 2012.
[27] Yahav Nussbaum. Improved distance queries in planar graphs. In Proc. 12th Workshop on
Algorithms and Data Structures (WADS), pages 642–653, 2011.
[28] Mikkel Thorup. Compact oracles for reachability and approximate distances in planar
digraphs. Journal of the ACM, 51(6):993–1024, 2004. See also FOCS’01.
[29] Mikkel Thorup and Uri Zwick. Approximate distance oracles. Journal of the ACM, 52(1):1–
24, 2005. See also STOC’01.
[30] Christian Wulff-Nilsen. Algorithms for Planar Graphs and Graphs in Metric Spaces. PhD
thesis, University of Copenhagen, 2010.
[31] Christian Wulff-Nilsen. Approximate distance oracles with improved preprocessing time.
In Proceedings of the twenty-third annual ACM-SIAM symposium on Discrete Algorithms,
pages 202–208. Society for Industrial and Applied Mathematics, 2012.
22
[32] Christian Wulff-Nilsen. Approximate distance oracles for planar graphs with improved
query time-space tradeoff. In Proc. 27th ACM/SIAM Symposium on Discrete Algorithms
(SODA), pages 351–362, 2016.
23
| 8cs.DS
|
Optimal Belief Approximation
Reimar H. Leike, Torsten A. Enßlin
arXiv:1610.09018v6 [math.ST] 3 Aug 2017
Max-Planck-Institut für Astrophysik, Karl-Schwarzschildstr. 1, 85748 Garching, Germany
Ludwig-Maximilians-Universität München, Geschwister-Scholl-Platz 1, 80539 Munich, Germany
In Bayesian statistics probability distributions express beliefs. However, for many problems the
beliefs cannot be computed analytically and approximations of beliefs are needed. We seek a loss
function that quantifies how “embarrassing” it is to communicate a given approximation. We reproduce and discuss an old proof showing that there is only one ranking under the requirements that
(1) the best ranked approximation is the non-approximated belief and (2) that the ranking judges
approximations only by their predictions for actual outcomes. The loss function that is obtained in
the derivation is equal to the Kullback-Leibler divergence when normalized. This loss function is
frequently used in the literature. However, there seems to be confusion about the correct order in
which its functional arguments—the approximated and non-approximated beliefs—should be used.
The correct order ensures that the recipient of a communication is only deprived of the minimal
amount of information. We hope that the elementary derivation settles the apparent confusion.
For example when approximating beliefs with Gaussian distributions the optimal approximation is
given by moment matching. This is in contrast to many suggested computational schemes.
Keywords: information theory, Bayesian inference, loss function, axiomatic derivation, machine learning
I.
INTRODUCTION
In Bayesian statistics, probabilities are interpreted
as degrees of belief. For any set of mutually exclusive and exhaustive events, one expresses the state of
knowledge as a probability distribution over that set.
The probability of an event then describes the personal confidence that this event will happen or has
happened. As a consequence, probabilities are subjective properties reflecting the amount of knowledge
an observer has about the events; a different observer
might know which event happened and assign different probabilities. If an observer gains information, she
updates the probabilities she had assigned before.
If the set of possible mutually exclusive and exhaustive events is infinite, it is generally impossible to
store all entries of the corresponding probability distribution on a computer or communicate it through a
channel with finite bandwidth. One therefore needs
to approximate the probability distribution which describes one’s belief. Given a limited set X of approximative beliefs q(s) on a quantity s, what is the best
belief to approximate the actual belief as expressed by
the probability p(s)?
In the literature, it is sometimes claimed that the
best approximation is given by the q ∈ X that minimizes the Kullback–Leibler divergence (“approximation” KL) [1]
X
p(s)
KL(p, q) =
p(s) ln
(1)
q(s)
s
where q is the approximation and p is the real belief.
We refer to this functional as “approximation” KL to
emphasize its role in approximation, which will be derived in the course of this paper and to distinguish
it from the same functional, with q being a prior belief and p being the posterior belief to which this KL
is minimized in inference. We will refer to the functional with q being the input and p obtained through
minimization as “inference KL”. In Equation (1), minimization is done with respect to its second argument.
The derivation of this particular functional form varies
from field to field.
For example, in coding theory, one tries to minimize the amount of bandwidth needed to transmit a
message. Given a prior q over the symbols that the
message consists of, an optimal scheme can be derived.
The approximation KL gives the expected amount of
extra bits needed to transmit such a message if the
symbols are actually drawn from the probability distribution p instead of q [2]. If we know that p is the
real probability distribution, the best approximative
probability distribution q ∈ X to base a coding on is
therefore the one minimizing the approximation KL.
However, it is not clear that minimizing the amount
of bits transferred is the best or even only measure
expressing how good such an approximation is in general.
In machine learning and deep learning, neural networks are trained to understand abstract data d; for
example, to assign a label s to it. This task can be
viewed as fitting an approximative probability distribution q(s|d) to a true generating probability distribution p(s|d). For this, the approximative probability distribution is parametrized (to a neural network)
and then matched to the true, generating probability
distribution using a loss function and samples. The
most frequently used loss function is the cross entropy,
which is equivalent to the approximation KL. The reason to use this form is often either inspired from coding theory, or by experimental experience [3].
Another argument for minimizing the approximation KL is given in Chapter 13 of Reference [4], where
it is claimed that this yields the maximum likelihood
estimation to p(s) among the probability distributions in X and that it gives an unbiased and unique
approximation. Interchanging the arguments of the
Kullback–Leibler divergence (the inference KL used
in variational Bayes) generally leads to a biased estimate and does not necessarily yield a unique result.
These arguments undoubtedly give evidence for why
minimizing the approximation KL gives a good estimate. However, this does not exclude all other methods. Having an unbiased estimate refers to getting
the right mean. In our picture, this is a result of
2
optimal approximation and not a requirement for optimality. Additionally, this result was derived with
the help of information geometry, whose applicability to non-local problems is criticized, for example, in
References [5, 6].
Contrary to the evidence for minimizing the approximation KL, we find many examples where an approximation is made by minimizing other functionals; for
example, minimizing the inference KL (e.g., [7–13]).
For many but not all of them, this is because minimizing the approximation KL is not feasible in practice in their case due to the real distribution p not
being accessible.
In this paper, we seek to bring together the different motivations and give a full and consistent picture.
The proof we present here is not new; it goes back to
[14], where there is an exact mathematical derivation
for probability densities analogously to our derivation.
However, there are earlier publications dealing with
the discrete case [15–17]. Although this proof dates
back at least 40 years, its implication on approximating beliefs seem to be quite unknown—especially in
the community of physicists applying Bayesian methods. In this paper, we reproduce a slightly modified
version of this proof, give the result a new interpretation and add further justification for the prerequisites
used, laying emphasis on why one has to accept the axioms necessary for the derivation if one is a Bayesian.
We provide argumentation for why believe approximation is an important and omnipresent topic.
We lay the emphasis of this paper more on interpretation of results and justification of prerequisites,
and thus present an easy version of the proof where
the loss function is assumed to be differentiable. The
proof can however be extended to the general case of
non-differentiable loss [18]. The argument we reproduce gives evidence that minimizing the approximation KL is the best approximation in theory. This
argument does not rest on information geometry nor
is it restricted to coding theory. By imposing two consistency requirements, one is able to exclude all ranking functions with the exception of one for ranking
the approximative probability distributions q ∈ X.
For this, one employs the principle of loss functions
[19], also called cost functions, regret functions, (or
with flipped sign, utility functions or score functions)
and shows that the unique loss function for ranking
approximated probability distributions is the approximation KL. For us, a ranking is a total order indicating preference, whereas a loss is a map to R,
which induces a ranking but additionally gives an absolute scale to compare preferences. The presented
axiomatic derivation does not give rise to any new
method, but it enables a simple checking for whether a
certain approximation is most optimally done through
the approximation KL. There are many other examples of axiomatic derivations seeking to support information theory on a fundamental level. Some notable
examples are Cox derivation [20] of Bayesian probability theory as a unique extension of Boolean algebra
as well as a scientific discussion on the maximum entropy principle [21–23], establishing the inference KL
as unique inference tool (which gave rise to the naming
convention in this paper). Most of these arguments
rely on page-long proofs to arrive at the Kullback–
Leibler divergence. The proof that is sketched in this
paper is only a few lines long, but nonetheless standard literature for axiomatic derivation in Bayesianism does not cite this “easy” derivation (e.g., the influential Reference [21]). As already discussed, approximation is an important and unavoidable part of information theory, and with the axiomatic derivation
presented here we seek to provide orientation to scientists searching for a way to approximate probability distributions.
In Section II, we introduce the concept of loss functions, which is used in Section III to define an optimal scheme for approximating probability distributions that express beliefs. We briefly discuss the relevance of our derivations for the scientific community
in Section IV. We conclude in Section V.
II.
LOSS FUNCTIONS
The idea to evaluate predictions based on loss functions dates back 70 years, and was first introduced by
Brier [24]. We explain loss functions by the means of
parameter estimation. Imagine that one would like to
give an estimate of a parameter s that is not known,
which value of s should be taken as an estimate? One
way to answer this question is by using loss functions.
For this note that p(s) is now formally
´ a probability
measure,
however we choose to write dsp(s) instead
´
of dp(s) as if p(s) would be a probability density. A
loss function in the setting of parameter estimation is
a function that takes an estimate σ for s and quantifies how “embarrassing” this estimate is if s = s0
turns out to be the case:
L (σ, s0 )
The expected embarrassment can be computed by
using the knowledge p(s) about s:
ˆ
hL (σ, s0 )ip(s0 ) = ds0 L (σ, s0 ) p(s0 )
The next step is to take the estimate σ that minimizes the expected embarrassment; that is, the expectation value of the loss function. For different loss
functions, one arrives at different recipes for how to
extract an estimate σ from the belief p(s); for example, for s ∈ R:
L (σ, s0 ) =
Take σ such that
−δ(σ − s0 ) ⇒ p(s)|
s=σ is maximal
=
(2)
|σ − s0 |
⇒ Take σ to be the median
(σ − s0 )2
⇒ Take σ to be the mean
In the context of parameter estimation, there is no
general loss function that one should take. In many
scientific applications, the third option is favored, but
different situations might enforce different loss functions. In the context of probability distributions, one
3
has a mathematical structure available to guide the
choice. In this context, one can restrict the possibilities by requiring consistent loss functions.
III.
THE UNIQUE LOSS FUNCTION
How embarrassing is it to approximate a probability
distribution by q(s) even though it is actually p(s)?
We quantify the embarrassment in a loss function
q
L
, s0
(3)
m
which says how embarrassing it is to tell someone q(s)
is one’s belief about s in the event that later s is measured to be s0 . Here m is introduced as reference
measure to make L coordinate independent. For a
finite set coordinate independence is trivially fulfilled
and it might seem that having a reference measure
m is superficial. Note however, that it is a sensible
additional requirement to have the quantification be
invariant under splitting of events, i.e. mapping to a
bigger set where two now distinguishable events repq
is inresent one former large event. The quotient m
variant under such splitting of events, whereas q itself
is not. The reference measure m can be any measure
such that q is absolutely continuous with respect to
m.
Note further that we restrict ourselves to the case
that we get to know the exact value of s. This does
not make our approach less general; imagining that
we would instead take a more general loss L(q, q̃(s))
where q̃ is the knowledge about
s at some later point,
q
then we may define L m
, s0 = L(q, δss0 ) with δ denoting the Kronecker or Dirac delta function, and thus
restrict ourselves again to the case of exact knowledge.
This line of reasoning was spelled out in detail by John
Skilling [25]:
“If there are general theories, then they
must apply to special cases”.
To decide which belief to tell someone, we look at
the expected loss
ˆ
D q
E
q
L
, s0
= ds0 L
, s0 p(s0 ) (4)
m
m
p(s0 )
and try to find a q ∈ X that minimizes this expected
loss. To sum up, if we are given a loss function, we
have a recipe for how to optimally approximate the
belief. Which loss functions are sensible, though? We
enforce two criteria that a good loss function should
satisfy.
Criterion 1 (Locality) If s = s0 turned out to be
the case, L only depends on the prediction q actually
makes about s0 :
q
q(s0 )
L
, s0 = L
(5)
m
m(s0 )
Note that we make an abuse of notation here, denoting the function on both sides of the equation by the
same symbol. The criterion is called locality because
q
should be evalit demands that the functional of m
uated locally for every s0 . It also forbids a direct
dependence of the loss L on s0 which excludes losses
that are a priori biased towards certain outcomes s0 .
This form of locality is an intrinsically Bayesian
property. Consider a situation where one wants to
decide which of two rivaling hypotheses to believe. In
order to distinguish them, some data d are measured.
To update the prior using Bayes theorem, one only
needs to know how probable the measured data d are
given each hypothesis, not how probable other possible data d˜ 6= d that were not measured are. This
might seem intuitive, but there exist hypothesis decision methods (not necessarily based on loss functions) that do not fulfill this property. For example,
the non-Bayesian p-value depends mostly on data that
were not measured (all the data that are at least as
“extreme” as the measured data). Thus, it is a property of Bayesian reasoning to judge predictions only
by using what was predicted about things that were
measured.
The second criterion is even more natural. If one is
not restricted in what can be told to others, then the
best thing should be to tell them the actual belief p.
Criterion 2 (Optimality of the actual belief, properness.) Let X be the set of all probability distributions
over s. For all p and all m, the probability distribution
q ∈ X with minimal expected loss is q = p:
0=
E
∂ D q
L
, s0
∂q(s)
m
p q=p
(6)
The last criterion is also referred to as a proper
loss (score) function in the literature; see Reference
[26] for a mathematical overview of different proper
scoring rules. Our version of this property is slightly
modified to the version that is found in the literature
as we demand this optimum to be obtained independently of a reference measure m. There is a fundamental Bayesian desiderata stating that “If there are
multiple ways to arrive at a solutions, then they must
agree.” We’d like to justify why this is a property
that is absolutely important. If one uses statistics
as a tool to answer some question, then if that answer would depend on how statistics is applied, then
this statistic itself is inconsistent. In our case, where
the defined loss function is dependent on an arbitrary
reference measure m, the result is thus forced to be
independent of that m.
Note furthermore that although intuitively we want
the global optimum to be at the actual belief p (referred to as strictly proper in the literature), mathematically we only need it to be an extreme value for
our derivation.
Having these two consistency requirements fixed, we
derive which kind of consistent loss functions are possible. We insert Equation (5) into Equation (6), expand the domain of the loss function to not necessarily
normalized positive vectors q(s) but introduce λ as a
Lagrange multiplier to account for the fact that we
minimize under the constraint of normalization. We
4
compute
´
q(s0 )
∂
ds
L
p(s
)
+
λ
q(s
)
0 =
0
0
0
∂q(s)
m(s0 )
q=p
´
p(s0 )
0)
=
ds0 ∂L m(s0 ) δ(s−s
p(s
)
+
λ
δ(s
−
s
)
0
0
m(s0 )
p(s)
p(s)
= ∂L m(s) m(s) + λ
p(s)
⇒ ∂L m(s)
= − λm(s)
p(s)
(7)
Here ∂L denotes the derivative of L . In the next
p(s)
for the quotient. Note
step we substitute x := m(s)
that Equation (7) holds for all positive real values of
x ∈ R+ since the requested measure independence of
the resulting ranking permits to insert any measure
m. We then obtain
∂L (x) = − λx
⇒ L (x) = −C ln (x) + D
(8)
where C > 0 and D are constants with respect to q.
Note that through the two consistency requirements
one is able to completely fix what the loss function is.
In the original proof of [14], there arises an additional
possibility for L if the sample space consists of 2 elements. In that case, the locality axiom, as it is used
in the literature, does not constrain L at all. In our
case, where we introduced m as a reference measure,
we are able to exclude that possibility. Note that the
constants C and D are irrelevant for determining the
optimal approximation as they do not affect where the
minimum of the loss function is.
To sum up our result if one is restricted to the closed
set X of probability distributions, one should take
that q ∈ X that minimizes
ˆ
D q
E
q(s0 )
L
, s0
(9)
= − ds0 p(s0 ) ln
m
m(s0 )
p(s0 )
in order to obtain the optimal approximation, where
it is not important what m is used.
If one takes m = 1, this loss is the cross entropy
h−ln (q(s0 ))ip(s0 ) .
If one desires a rating of how good an approximation is, and not only a ranking which approximation
is best, one could go one step further and enforce a
third criterion:
Criterion 3 (zero loss of the actual belief) For all p,
the expected loss of the probability distribution p is 0:
D p
E
0= L
, s0
(10)
m
p
This criterion trivially forces m = p and makes the
quantification unique while inducing the same ranking. Thus we arrive at the Kullback-Leibler divergence
ˆ
p(s0 )
KL(p, q) = ds0 p(s0 )ln
q(s0 )
as the optimal rating and ranking function.
To phrase the result in words, the optimal way to
approximate the belief p is such that given the approximated belief q, the amount of information KL(p, q)
that has to be obtained for someone who believes q
to arrive back at the actual belief p is minimal. We
should make it as easy as possible for someone who
got an approximation q to get to the correct belief p.
This sounds like a trivial statement, explaining why
the approximation KL is already widely used for exactly that task.
IV.
DISCUSSION
We briefly discuss the implications of these results.
In comparison to Reference [2, 4], we presented another more elementary line of argumentation for the
claim that the approximation KL is the correct ranking function for approximating which holds in a more
general setting.
Other works that base their results on minimizing
the inference KL (KL(q, p)) for belief approximation
are not optimal with respect to the ranking function
we derived in Section III. One reason for preferring the
for this purpose non-optimal inference KL is that it
is computationally feasible for many applications, in
contrast to the optimal approximation. As long as the
optimal scheme is not computationally accessible, this
argument has its merits.
Another reason for minimizing the inference KL for
approximation that is often cited (e.g., [27]) is that it
gives a lower bound to the log-likelihood
p(d, s)
ln(p(d|s)) = ln
+ KL(q, p) (11)
q(s)
q(s)
which for example gives rise to the expectation maximization (EM- ) algorithm [28]. However, the method
only gives rise to maximum a posteriori or maximum
likelihood solutions, which corresponds to optimizing
the δ-loss of Equation (2).
In Reference [11], it is claimed that minimizing the
inference KL yields more desirable results since for
multi-modal distributions, individual modes can be
fitted with a mono-modal distribution such as a Gaussian distribution, whereas the resulting distribution
has a very large variance when minimizing the approximation KL to account for all modes. In Figure 1
there is an example of this behavior. The true distribution of the quantity s is taken to be a mixture of
two standard Gaussians with means ±3. It is approximated with one Gaussian distribution by using the
approximation KL and the inference KL. When using
the approximation KL, the resulting distribution has
a large variance to cover both peaks. Minimizing the
inference KL leads to a sharply peaked approximation around one peak. A user of this method might
be very confident that the value of s must be near
3, even though the result is heavily dependent on the
initial condition of the minimization and could have
become peaked around −3 just as well.
We find that fitting a multi-modal distribution with
a mono-modal one will yield suboptimal results irrespective of the fitting scheme. An approximation
should always have the goal to be close to the target
that is being approximated. If it is already apparent
that this goal cannot be achieved, it is recommended
5
to rethink the set of approximative distributions and
not dwell on the algorithm used for approximation.
In Reference [12] an approximative simulation
scheme called information field dynamics is described.
There, a Gaussian distribution q is matched to a
time-evolved version U (p) of a Gaussian distribution
p. This matching is done by minimizing the inference KL. In this particular case (at least for information preserving dynamics), the matching can be
made optimal without making the algorithm more
complicated. Since for information preserving dynamics time evolution is just a change of coordinates
and the Kullback–Leibler divergence is invariant under such transformations, one can instead match the
Gaussian distribution p and U −1 (q) by minimizing
KL(p, U −1 (q)) = KL(U (p), q), which is just as difficult in terms of computation.
accordance to the behavior of other approximations.
For example, when step-wise rounding the real number 1.49, one gets 2 if it is first rounded to one decimal and then to integer precision compared to being
rounded to integer precision directly where one gets
1. If information is lost due to approximation, it is
natural for further approximations to be less precise
than if one were to approximate in one go.
There also exist cases where we could not find
any comments explaining why the arguments of the
Kullback–Leibler divergence are in that particular order. In general, it would be desirable that authors
provide a short argumentation for why they choose
a particular order of the arguments of the KL divergence.
V.
probability density
0.35
0.30
target
approximation KL
inference KL
0.25
0.20
0.15
0.10
0.05
0.00
−10
−5
0
s
5
10
Figure 1: Results of approximating a target
distribution in s with a Gaussian distribution. KL:
Kullback–Leibler.
In Reference [13] it is claimed that the inference KL
yields an optimal approximation scheme fulfilling certain axioms. This result is the exact opposite of our
result. This disagreement is due to an assumed consistency of approximations. In Reference [13], further
approximations are forced to be consistent with earlier
approximations; i.e., if one does two approximations,
one gets the same result as with one joint approximation. Due to this requirement, the derived functional
cannot satisfy some of the axioms that we used. In
our picture, it is better to do one large approximation instead of many small approximations. This is in
[1] S. Kullback and R. Leibler, Annals of Mathematical
Statistics 22 (1), 79 (1951).
[2] T. M. Cover and J. A. Thomas, (2006).
[3] I. Goodfellow, Y. Bengio,
and A. Courville,
Deep Learning (MIT Press, 2016) http://www.
deeplearningbook.org.
[4] M. Opper and D. Saad, Advanced mean field methods:
Theory and practice (MIT press, 2001).
[5] J. Skilling, in BAYESIAN INFERENCE AND MAXIMUM ENTROPY METHODS IN SCIENCE AND
ENGINEERING: Proceedings of the 33rd Interna-
CONCLUSIONS
Using the two elementary consistency requirements
on locality and optimality, as expressed by Equations
(5) and (6), respectively, we have shown that there
is only one ranking function that ranks how good
an approximation of a belief is, analogously to Reference [14]. By minimizing KL(p, q) with respect to
its second argument q ∈ X, one gets the best approximation to p. This is claimed at several points in the
literature. Nevertheless, we found sources where other
functionals were minimized in order to obtain an approximation. This confusion is probably due to the
fact that for the slightly different task of updating a
belief q under new constraints, KL(p, q) has to be minimized with respect to p, its first argument [29, 30].
We do not claim that any of the direction of Kullback–
Leibler divergence are wrong by themselves, but one
should be careful of when to use which.
We hope that for the case of approximating a probability distribution p by another q we have given convincing and conclusive arguments for why this should
be done by minimizing KL(p, q) with respect to q, if
this is feasible.
VI.
ACKNOWLEDGEMENTS
We would like to thank A. Caticha, J. Skilling, V.
Böhm, J. Knollmüller, N. Porqueres, and M. Greiner
and six anonymous referees for the discussions and
their valuable comments on the manuscript.
tional Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Engineering
(MaxEnt 2013), Vol. 1636 (AIP Publishing, 2014) pp.
24–29.
[6] J. Skilling, in Bayesian inference and maximum entropy methods in science and engineering (MAXENT
2014), eds. A. Mohammad-Djafari and F. Barbaresco,
AIP Conference Proceedings, Vol. 1641 (2015) pp. 27–
42.
[7] C. W. Fox and S. J. Roberts, Artificial intelligence
review 38, 85 (2012).
6
[8] T. A. Enßlin and C. Weig, Phys. Rev. E 82, 051112
(2010), arXiv:1004.2868 [astro-ph.IM].
[9] D. M. Blei, A. Kucukelbir, and J. D. McAuliffe,
ArXiv e-prints (2016), arXiv:1601.00670 [stat.CO].
[10] F. J. Pinski, G. Simpson, A. M. Stuart,
and
H. Weber, ArXiv e-prints (2014), arXiv:1408.1920
[math.NA].
[11] F. Pinski, G. Simpson, A. Stuart, and H. Weber,
ArXiv e-prints (2013), arXiv:1310.7845 [math.PR].
[12] T. A. Enßlin, Phys. Rev. E 87, 013308 (2013),
arXiv:1206.4229 [physics.comp-ph].
[13] C.-Y. Tseng and A. Caticha, Physica A Statistical
Mechanics and its Applications 387, 6759 (2008),
arXiv:0808.4160 [cond-mat.stat-mech].
[14] J. M. Bernardo, The Annals of Statistics , 686 (1979).
[15] J. Aczél and J. Pfanzagl, Metrika 11, 91 (1967).
[16] J. McCarthy, Proceedings of the National Academy
of Sciences 42, 654 (1956).
[17] I. J. Good, Journal of the Royal Statistical Society.
Series B (Methodological) , 107 (1952).
[18] P. Harremoës, arXiv preprint arXiv:1701.01010
(2017).
[19] H. Cramér, On the mathematical theory of risk (Centraltryckeriet, 1930).
[20] R. T. Cox, American Journal of Physics 14, 1 (1946).
[21] E. T. Jaynes, Probability Theory, by E. T. Jaynes
and Edited by G. Larry Bretthorst, pp. 758. ISBN
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
0521592712. Cambridge, UK: Cambridge University
Press, June 2003., edited by Bretthorst, G. L. (2003).
J. Skilling, in Maximum-Entropy and Bayesian Methods in Science and Engineering (Springer, 1988) pp.
173–187.
A. Caticha, in Bayesian Inference and Maximum Entropy Methods in Science and Engineering, American Institute of Physics Conference Series, Vol. 707,
edited by G. J. Erickson and Y. Zhai (2004) pp. 75–96,
physics/0311093.
G. W. Brier, Monthly weather review 78, 1 (1950).
J. Skilling, in Maximum entropy and Bayesian methods (Springer, 1989) pp. 45–52.
T. Gneiting and A. E. Raftery, Journal of the American Statistical Association 102, 359 (2007).
C. Bishop, Pattern Recognition and Machine Learning
(Information Science and Statistics), 1st edn. 2006.
corr. 2nd printing edn (2007).
A. P. Dempster, N. M. Laird, and D. B. Rubin, Journal of the royal statistical society. Series B (methodological) , 1 (1977).
I. Csiszar, The annals of statistics , 2032 (1991).
A. Caticha, ArXiv e-prints (2014), arXiv:1412.5644.
| 10math.ST
|
arXiv:1508.03536v1 [cs.PL] 14 Aug 2015
Proceedings of the 3rd Workshop on Domain-Specific
Language Design and Implementation
DSLDI’15
Tijs van der Storm
Centrum Wiskunde & Informatica (CWI)
[email protected]
Sebastian Erdweg
TU Darmstadt
[email protected]
July 7th, 2015
2
3
DSLDI’15
Introduction
DSLDI’15 is the 3rd workshop on Domain-Specific Language Design and Implementation, which
was held at ECOOP 2015, on Tuesday, July 7th, 2015.
DSLDI’15 was organized by:
• Tijs van der Storm
Centrum Wiskunde & Informatica (CWI)
[email protected]
• Sebastian Erdweg
TU Darmstadt
[email protected]
The submitted talk proposal were reviewed by the following program committee:
• Emilie Balland
• Martin Bravenboer (LogicBlox)
• Hassan Chafi (Oracle Labs)
• William Cook (UT Austin)
• Shriram Krishnamurthi (Brown University)
• Heather Miller (EPFL)
• Bruno Oliveira (University of Hong Kong)
• Cyrus Omar (CMU)
• Richard Paige (University of York)
• Tony Sloane (Macquarie University)
• Emma Söderberg (Google)
• Emma Tosch (University of Massachusetts, Amherst)
• Jurgen Vinju (CWI)
The website of DSLDI’15 is: http://2015.ecoop.org/track/dsldi-2015-papers.
Informal Post-Proceedings DSLDI’15
This document contains informal post-proceedings of DSLDI’15. It contains:
• A snapshot of the home page of DSLDI’15
• The detailed program of the workshop
• The accepted talk proposals.
• A summary of the panel discussion on Language Composition.
4
5
Program
6
SCROLL - A Scala-based library for Roles at Runtime
Max Leuthäuser
Technische Universität Dresden
Software Technology Group
[email protected]
1
2
See: github.com/max-leuthaeuser/SCROLL
See: scala-lang.org/api/current/#scala.Dynamic
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
SCROLL
2015
JavaStage
2012 [3]
NextEJ
2009 [8]
ScalaRoles
2008 [10]
Rumer
2007 [2]
powerJava
2006 [1]
Rava
2006 [5]
OT/J
2005 [6]
Chameleon
2003 [4]
way, e.g. adding, removing and transferring roles or accessing role functions and attributes with the +-Operator
(e.g. in line 14, right listing).
The following limitation (apart from some role features
not implemented yet, see table 1) is the major subject of
the future work on SCROLL. The underlying technique
(compiler rewrite rules) hides important typing information to the tooling typically used by most developers, i.e.
IDEs with debugger and link tracers. Writing plugins for
those (e.g. Eclipse, Intellij) overcoming this issue would be
one solution and is currently under development.
Finally, it is necessary to investigate how well this implementation blends into coeval approaches. We use a classification scheme established in two successive surveys on
role-based modeling and programming languages, namely
[9,11]. This revolves 26 classifying features of roles incorporating both the relational and the context-dependent
nature of roles (see table 1). The most sophisticated, competing approach so far is ObjectTeams/Java [6]. It allows
to override methods of their player by aspect weaving. Besides that, it introduces Teams to represent compartments
whose inner classes automatically become roles. Supporting both the inheritance of roles and teams leads to family
polymorphism [7]. On the downside, it does not support
multiple unrelated player types for a role type. In sum
SCROLL provides a simple and clean testing playground
in an unmodified Scala for using roles at runtime.
Feature [9]
Today’s software systems always need to anticipate changing context. New business rules and functions should be
implemented and adapted. The concept of role modeling
and programming is frequently discussed for decades across
many scientific areas. It allows the modeling and implementation of context dependent information w.r.t. dynamically changing context. Hence future software infrastructures have the intrinsic need to introduce such a role concept. Until now the implementation with existing object
oriented languages always requires the generation of a specific runtime environment and management code. The expressiveness of these languages is not able to cope with
essential role-specific features, such as true delegation or
binding roles dynamically. In this work we present how a
relatively simple implementation with Scala based on its
Dynamic trait allows to augment an object’s type at runtime implementing dynamic (compound-) role types. It enables role-based implementations that lead to more reuse
and better separation of concerns.
Currently, only a handful, mostly unusable (e.g. because
they are not providing a running compiler or have been
abandoned by the developer) role-based programming languages exists. The field of research is highly fragmented,
due to the fact that every research area relies on a different set of role-related features [9]. Therefore it is necessary to establish a basic role concept at runtime and
build an appropriate tooling around it to make it more
useful for developers. A prototypic Scala implementation
for roles (SCROLL - SCala ROLes Language1 ) was developed as library approach, enabling the user to specify roles
and context dependent behavior. A basic example of smart
(autonomous) cars driving around demonstrates part of its
capabilities (see page 2). Two persons (class Person) and
two cars (Car) are bound to the roles Driver, Passenger
and NormalCar, SmartCar respectively. Each role modifies
the default behavior implemented in the players classes, as
demonstrated at runtime in the console output listing. Dynamic role selection (like selecting an object playing the
Source location role) supports filtering for attributes and
evaluating function call results (e.g. line 17 and 26, right
listing).
Internally the following two technical aspects are the
most considerable ones. First, making use of the Dynamic
Trait. All calls to role functions (i.e. functions that are not
natively available on the player object) are translated by
the compiler using certain rules2 . These are adjustable resulting in customizable, dynamic role dispatch. Second, applying implicit conversions. Scala’s implicit classes allow for packing in player and role objects to compound dynamic types. All important role features are exposed this
Table 1: Comparison of coeval approaches for etablishing roles
at runtime based on 26 classifying features extracted from the
literature [9,11]. It differentiates between fully (), partly ()
and not supported () features.
1
2
3
class Person(val name: String)
class Car(val licenseID: String)
class Location(val name: String)
4
5
6
7
8
9
10
11
12
13
14
15
16
17
1
2
3
4
class Transportation() extends Compartment {
object AutonomousTransport extends Compartment {
class SmartCar() {
def drive() {
info("I am driving autonomously!")
}
}
class Passenger() {
def brake() {
info(s"I can’t reach the brake. I am ${+this name}
,→ and just a passenger!")
}
}
}
18
5
val transportation = new Transportation {
val peter = new Person("Peter")
val harry = new Person("Harry")
val googleCar = new Car("A-B-C-001")
val toyota = new Car("A-B-C-002")
6
new Location("Munich") play new Source()
new Location("Berlin") play new Source()
new Location("Dresden") play new Target()
7
8
9
10
harry play new ManualTransport.Driver()
toyota play new ManualTransport.NormalCar()
11
12
13
+toyota drive()
ManualTransport play
new TransportationRole(
one[Source]("name" ==# "Berlin"),
one[Target]()) travel()
14
15
16
17
18
19
object ManualTransport extends Compartment {
class NormalCar() {
def drive() {
info(s"I am driving with a driver called
,→ ${+one[Driver]() name}.")
}
}
class Driver() {
def brake() {
info(s"I am ${+this name} and I am hitting the
,→ brakes now!")
}
}
}
19
20
21
22
23
24
25
26
27
28
29
30
31
class TransportationRole(source: Source, target: Target) {
def travel() {
val kindOfTransport = this player match {
case ManualTransport => "manual"
case AutonomousTransport => "autonomous"
}
info(s"Doing a $kindOfTransport transportation with
,→ the car ${one[Car]().licenseID} from
,→ ${+source name} to ${+target name}.")
}
}
32
33
34
35
36
37
38
39
40
peter play new AutonomousTransport.Passenger()
googleCar play new AutonomousTransport.SmartCar()
20
21
22
+googleCar drive()
AutonomousTransport play
new TransportationRole(
one[Source]("name" ==# "Munich"),
one[Target]()) travel()
23
24
25
26
27
28
+peter brake(); +harry brake()
29
30
}
Fig. 1: The SmartCar example (instance code, top) and
the corresponding model code (left).
1
2
3
4
5
6
I am driving with a driver called Harry.
Doing a manual transportation with the car A-B-C-002 from
,→ Berlin to Dresden.
I am driving autonomously!
Doing a autonomous transportation with the car A-B-C-001
,→ from Munich to Dresden.
I can’t reach the brake. I am Peter and just a passenger!
I am Harry and I am hitting the brakes now!
41
class Target()
class Source()
42
43
44
}
Fig. 2: Running the SmartCar example generates the
console output shown above.
Acknowledgements
This work is funded by the German Research Foundation (DFG) within the Research Training Group “Role-based Software
Infrastructures for continuous-context-sensitive Systems” (GRK 1907).
References
1. Baldoni, M., Boella, G., van der Torre, L.: Roles as a coordination construct: Introducing powerjava. Electr. Notes Theor.
Comput. Sci 150(1), 9–29 (2006)
2. Balzer, S., Gross, T., Eugster, P.: A relational model of object collaborations and its use in reasoning about relationships.
In: Ernst, E. (ed.) ECOOP. Lecture Notes in Computer Science, vol. 4609, pp. 323–346. Springer (2007)
3. Barbosa, F.S., Aguiar, A.: Modeling and programming with roles: introducing javastage. Tech. rep., Instituto Politécnico
de Castelo Branco (2012)
4. Graversen, K.B., Østerbye, K.: Implementation of a role language for object-specific dynamic separation of concerns. In:
AOSD03 Workshop on Software-engineering Properties of Languages for Aspect Technologies (2003)
5. He, C., Nie, Z., Li, B., Cao, L., He, K.: Rava: Designing a java extension with dynamic object roles. In: Engineering of
Computer Based Systems, 2006. ECBS 2006. 13th Annual IEEE International Symposium and Workshop on. pp. 7–pp.
IEEE (2006)
6. Herrmann, S.: Programming with roles in ObjectTeams/Java. Tech. rep., AAAI Fall Symposium (2005)
7. Herrmann, S., Hundt, C., Mehner, K.: Translation polymorphism in object teams. Tech. rep., TU Berlin (2004)
8. Kamina, T., Tamai, T.: Towards safe and flexible object adaptation. In: International Workshop on Context-Oriented
Programming. p. 4. ACM (2009)
9. Kühn, T., Leuthäuser, M., Götz, S., Seidl, C., Aßmann, U.: A metamodel family for role-based modeling and programming languages. In: Combemale, B., Pearce, D., Barais, O., Vinju, J. (eds.) Software Language Engineering, Lecture
Notes in Computer Science, vol. 8706, pp. 141–160. Springer International Publishing (2014), http://dx.doi.org/10.
1007/978-3-319-11245-9_8
10. Pradel, M., Odersky, M.: Scala Roles - A lightweight approach towards reusable collaborations. In: International Conference
on Software and Data Technologies (ICSOFT ’08) (2008)
11. Steimann, F.: On the representation of roles in object-oriented and conceptual modelling. Data & Knowledge Engineering
35(1), 83–106 (2000)
A case for Rebel
A DSL for Product Specifications
Jouke Stoel
CWI, Amsterdam, The Netherlands
[email protected]
1.
Introduction
Large service organisations like banks have a hard time
keeping grips on their software landscape. This is not only
visible while performing maintenance on existing applications but also when developing new applications.
One of the problems these organisations face is that they
often do not have a clear and uniform descriptions of their
products like savings- and current account, loans and mortgages. This makes it hard to reason about changes to existing products and hampers the introduction of new ones. The
specifications that are written down often contain ambiguities or are out-of-date. Next to this, specification are almost
always written down using natural language which is known
to lead to numerous deficiencies [1].
To counter these problems we introduce Rebel, a DSL for
product specifications. Rebel lets users specify their product
in a high-level, unambiguous manner. These specification
can then be simulated which enables users to explore their
products before they are build.
We have created Rebel for a large Dutch bank and are
currently in the process of specifying existing banking products.
Since Rebel is in the early stages of development we
would like to use DSLDI to gather feedback on its current
design and proposed future directions.
2.
Rebel
Rebel is a domain specific language for product specifications. It is inspired on formal methods like Z [2], B [3] and
Alloy [4]. It is aimed at helping a large Dutch bank in bridging the gap between informal specifications written down in
natural language or passed on mouth-to-mouth towards unambiguous, machine interpretable specifications. The main
idea behind Rebel is to present the user with a easy to understand syntax and interface while it exploits powerful tooling
like verification to check whether the specifications hold under the hood.
Rebel is implemented in R ASCAL [5] as a stand-alone
DSL.
2.1
Requirements
The language needed to fit the following requirements:
• Flexibility - it should be possible to tune it to the problem
of the bank we were working with.
• Integration - it should be possible to integrate existing
tools like model checkers and connect to existing systems
in the banks application landscape.
• Adaptation - it should be easy to learn and the tooling like
an IDE should be similar to the tooling currently used.
Considering these requirements we decided to create a new
language. This new language needed to be a linguistic hybrid
to be able to support both the definition of single products as
well as the overlying process.
2.2
Design
Rebel is a declarative language and centres around specifications. Figure 1 shows an example of such a specification.
A specification describes one product. Specifications contain fields, events, invariants and life cycle. Fields declare
the data used in the specification. Events describe the possible mutations on the data under certain conditions. Invariants
describe global rules which should always hold and life cycle constrains the order of events.
The definition of events and invariants is separated from
usage in specifications. This is to promote reuse and to
separate the responsibility of implementing an event from
using an event in a specification.
Defined fields can only be of built-in types. Events can
only reference fields declared in the specification, not fields
of other specifications. We made this choice so that the
potential state space is smaller when applying verification
techniques like model checking.
Events are described using pre- and postconditions. An
example event is shown in Figure 2. The semantics are
straightforward; if the precondition holds then the postcondition will hold after the event is raised. Events contain runtime instance variables as well as configuration variables.
Configuration variables are keyword parameters that can
have a default value and can be set when the event is referenced by a particular specification. For instance, the us-
specification SavingsAccount {
fields {
balance: Time -> Integer
}
events {
openAccount[minimumDeposit=50]
withdraw[]
deposit[]
close[]
}
invariants {
positiveBalance
}
lifeCycle {
initial new -> opened: openAccount
opened -> opened: withdraw, deposit
-> closed: close
final closed
}
}
Figure 1. Example Rebel specification
initial event openAccount
[minimumDeposit : Integer = 0]
(accountNumber: String, initialDeposit : Integer) {
preconditions {
initialDeposit >= minimumDeposit;
}
postconditions {
new this.balance(now) == initialDeposit;
}
}
Figure 2. Example of an event definition
age declaration of openAccount (Figure 1) sets the event
configuration parameter minimumDeposit meaning that the
SavingsAccount uses 50 as a minimumDeposit when an account is opened.
Invariants are global rules. They use quantifiers over data
to express certain constrains that should always hold. Figure 3 shows an example that states that at all time, saving
accounts should have a balance equal to or above zero.
invariant positiveBalance {
all sa:SavingsAccount | all t:Time {
sa.balance(t) >= 0
}
}
Figure 3. Example of an invariant
2.3
Simulating specifications
The simulation is aimed at helping product owners and developers gain insight into their specified product. It can be
used to check if the specification meets the expectations of
the user. Figure 4 shows a screenshot of the simulation of
Figure 4. Screenshot of simulating the SavingsAccount
specification
a SavingsAccount. The simulation is implemented with the
use of the Z3 SMT solver [6].
3.
Future work
The current version of Rebel supports the definition of single products. Next to this it is also needed to define composition of these products. In other words, the process. Since
the specifications only contain fields of built-in types and
can only reference themselves it is not possible to compose
specifications. To overcome this we propose the use of process algebra [7] for specifying how the individual specification events should be composed. This will give us the ability
to specify choices, sequencing, concurrency and communications between specifications. The question will be if we
will still be able to reason about (certain parts of) the specifications since composing the specifications will have a large
impact on the state space.
An orthogonal aspect is the tooling for Rebel specifications. Next to the simulation we will explore the possibility
of model checking. The model checker could be used to find
event traces that lead to violations of the invariants. Earlier
work has shown that it is possible to translate Rebel specification to Alloy. Alloys analyser was used to find traces
which would break the specification. The problem with this
approach was scalability. An alternative would be to exploit
an SMT solver for the same purpose [8]. One of the challenges here will be how we can bound the data in a smart
way to limit the state space.
Ultimately, running systems should be generated from
Rebel specifications. Since Rebel is a declarative language
it will not always be straightforward to generate a correct
system from this. Again SMT solvers might hold the key as
shown in other work like [9].
References
[1] B. Meyer. On Formalism in Specifications. IEEE Software,
2(1):6–26, 1985.
[2] J.M. Spivey and J.R. Abrial. The Z notation: A Reference
Manual. Prentice Hall Hemel Hempstead, second edition,
1992.
[3] J.R. Abrail. The B-Book: Assinging programs to meaning.
Cambridge University Press, 1996.
[4] D. Jackson. Alloy: a lightweight object modelling notation.
ACM Transactions on Software Engineering and Methodology,
11(2):256–290, 2002.
[5] P. Klint, T. van der Storm, and J. Vinju. RASCAL: A Domain
Specific Language for Source Code Analysis and Manipulation. In 2009 Ninth IEEE International Working Conference
on Source Code Analysis and Manipulation, pages 168–177.
IEEE, 2009.
[6] L. De Moura and N. Bjorner. Z3: An efficient SMT solver.
Tools and Algorithms for the Construction and Analysis of
Systems, pages 337–340, 2008.
[7] J. Baeten, T. Basten, and M.A. Reniers. Process algebra:
equational theories of communicating processes. Cambridge
university press, 2010.
[8] A. Milicevic and H. Kugler. Model checking using SMT and
theory of lists. In NASA Formal Methods, pages 282–297.
Springer, 2011.
[9] R. Alur, R. Bodik, G. Juniwal, M.M.K. Martin,
M. Raghothaman, S. Seshia, R. Singh, A. Solar-Lezama,
E. Torlak, and A. Udupa. Syntax-guided synthesis. 2013
Formal Methods in Computer-Aided Design, pages 1–17,
October 2013.
Flick: A DSL for middleboxes
Network-as-a-Service (NaaS) project∗
We argue the need for specialised languages to program application-level
middleboxes, and describe our design for such a language, through which we
seek to make available suitable abstractions for middlebox programming, and
constrain what kinds of programs can be expressed.
A middlebox is a non-standard router; it carries out a computation on network traffic (beyond decrementing a TTL and recomputing a checksum) and
the routing decision on that traffic may be influenced by the computation’s
result. Examples of middleboxes include firewalls, protocol accelerators, VPN
gateways, transcoders, load-balancers and proxies. Middleboxes perform a vital
function in today’s Internet, in datacentres, and even in home and corporate
networks.
An application-level middlebox is one that does not deal solely in networklevel primitives or protocols, but also (and perhaps exclusively) in applicationlevel protocols and data. Examples include a memcached caching reverse-proxy,
and a load-balancer that unsheaths an HTTP/SSL connection before passing
the cleartext HTTP data to a backend server.
Middleboxes are usually written in general-purpose languages—usually C.
Such languages are sufficiently expressive, enjoy a broad developer population,
and a compiler is likely to exist for the intended architecture. It has been observed that even for systems programming, the disadvantages of using a generalpurpose language sometimes outweigh the benefits [3, 5]. We believe that middlebox programming is an example of this. By their nature, general-purpose
languages do not provide suitable abstractions for middlebox programming. Another disadvantage is that general-purpose languages are much too expressive
for writing middleboxes, which often do not implement complex behaviour. Indeed middleboxes cannot implement complex behaviour if they are to operate
at line rate at high bandwidth.
This suggests the need for a framework in which to write middleboxes. This
could be implemented in two ways: as a DSL, or as a library. Either approach
could provide more suitable abstractions for middleboxes than is usually provided by a general-purpose language alone. Because of this, either approach
could lead to shorter, more readable, code, without significant regression in
performance. Work on both approaches has been described in the literature.
DSLs for middlebox programming include Click [4], POF [6], and P4 [2].
All of these are designed for processing packets. (Click can also be regarded
∗ http://www.naas-project.org/
1
as a library, but even there it is oriented towards the implementation of packet
processors.) These languages seek to support the use of arbitrary protocols;
part of the programmer’s task is to encode the packet format.
But not all middleboxes are most naturally defined as packet processors—
this is particularly the case for application-level middleboxes which we seek to
support. We can think of middleboxes more generally as processors of arbitrary
data extracted from byte streams. xOMB [1] is a library for programming
middleboxes in C++. Unlike packet processors, xOMB elements can operate on
higher-layer data. But a library does not allow us to impose suitable constraints
on programs: using a middlebox library leaves you at liberty to write arbitrary
functions in the host language.
We therefore opted to design a DSL for implementing application-level middleboxes. Our language, called Flick, has the following features. First, it is
statically typed, and features algebraic types. In addition to being used for
type checking and inference, types can be used to synthesise serialisation and
deserialisation code for values of a type. This is achieved by decorating a type’s
declaration with serialisation annotations. These specify the precision of integers, for example, and byte ordering for multi-byte values. Second, Flick only
allows bounded recursion. It is a Turing-incomplete language; only terminating
computations can be expressed in Flick. This is an important constraint that
middlebox DSLs enforce, but that libraries cannot, as mentioned earlier. Third,
channels and processes are language primitives; such concepts seem native to
middleboxes. Channels are typed, and connect processes to other processes, or
to external (network) sources or destinations of data. Processes can run concurrently, and contain the middlebox’s logic. Fourth, processes encapsulate their
own state, but middleboxes may also share state. That is, message-passing is the
method used to both notify a process of new data, and to provide that data; but
shared memory can be used when notification is not necessary. We found this
useful for describing shared caches, such as in the encoding of a load-balancer
for the memcached key-value database, shown below.
proc Memcached: (cmd/cmd client, [cmd/cmd] backends)
global cache := empty_dict
client => test_cache(client, backends, cache)
backends => update_cache(cache) => client
Here the process Memcached has a channel client, and an array of channels
backends with which it can communicate with other processes. All channels
in this process can yield and accept values of type cmd, values of which are
memcached commands. update_cache and test_cache are functions. The
body of the process simply forwards requests from client to a backend unless
the reply has been cached; and it forwards replies from backends to the client,
after caching them locally.
So far we have a formal semantics for the core expression language, and a
partial compiler to a runtime of our devising. Future work involves language
features such as exception handling and resource estimation, as well as extending
the compiler to include more targets, including reconfigurable hardware.
2
References
[1] James W. Anderson, Ryan Braud, Rishi Kapoor, George Porter, and Amin
Vahdat. xOMB: Extensible Open Middleboxes with Commodity Servers.
In Proceedings of the Eighth ACM/IEEE Symposium on Architectures for
Networking and Communications Systems, ANCS ’12, pages 49–60, New
York, NY, USA, 2012. ACM.
[2] Pat Bosshart, Dan Daly, Glen Gibb, Martin Izzard, Nick McKeown, Jennifer
Rexford, Cole Schlesinger, Dan Talayco, Amin Vahdat, George Varghese,
and David Walker. P4: Programming Protocol-independent Packet Processors. SIGCOMM Comput. Commun. Rev., 44(3):87–95, July 2014.
[3] Pierre-Evariste Dagand, Andrew Baumann, and Timothy Roscoe. Filet-ofish: Practical and Dependable Domain-specific Languages for OS Development. SIGOPS Oper. Syst. Rev., 43(4):35–39, January 2010.
[4] Eddie Kohler, Robert Morris, Benjie Chen, John Jannotti, and M. Frans
Kaashoek. The Click Modular Router. ACM Trans. Comput. Syst.,
18(3):263–297, August 2000.
[5] Adrian Schüpbach, Andrew Baumann, Timothy Roscoe, and Simon Peter.
A Declarative Language Approach to Device Configuration. ACM Trans.
Comput. Syst., 30(1):5:1–5:35, February 2012.
[6] Haoyu Song. Protocol-oblivious Forwarding: Unleash the Power of SDN
Through a Future-proof Forwarding Plane. In Proceedings of the Second
ACM SIGCOMM Workshop on Hot Topics in Software Defined Networking,
HotSDN ’13, pages 127–132, New York, NY, USA, 2013. ACM.
3
Towards a Next-Generation Parallel Particle-Mesh
Language∗
Sven Karol1 , Pietro Incardona2,3 , Yaser Afshar2,3 , Ivo F.
Sbalzarini2,3 , and Jeronimo Castrillon1
1
Chair for Compiler Construction, Center for Advancing Electronics Dresden,
TU Dresden, Dresden, Germany
[sven.karol|jeronimo.castrillon]@tu-dresden.de
MOSAIC Group, Chair of Scientific Computing for Systems Biology, Faculty
of Computer Science, TU Dresden, Dresden, Germany
Center for Systems Biology Dresden, Max Planck Institute of Molecular Cell
Biology and Genetics, Dresden, Germany
[afshar|incardon|ivos]@mpi-cbg.de
2
3
Abstract
We present our previous and current work on the parallel particle-mesh language PPML—a DSL
for parallel numerical simulations using particle methods and hybrid particle-mesh methods in
scientific computing.
1
Introduction
During the past years, domain-specific languages (DSLs) gained central importance in
scientific high-performance computing (HPC). This is due to the trend towards HPC clusters
with heterogeneous hardware—today, mainly using multi-core CPUs as well as streaming
processors such as GPUs—in the future, using many-core CPUs, and potentially also
reconfigurable processors or data-flow processing units. Writing programs for these machines
is a challenging and time-consuming task for scientific programers, who do not only need
to develop efficient parallel algorithms for the specific problem at hand, but also need to
tune their implementations in order to take advantage of the cluster’s hardware performance.
This does not only require experience in parallel programming, e.g. using OpenMP, OpenCL,
or MPI, but also in computer architectures and numerical simulation methods, leading to the
so-called “knowledge gap” in program efficiency [12]. Besides, it renders the simulation codes
hardly portable. DSLs can help two-fold: First, they allow scientific programmers to write
programs using abstractions closer to the original mathematical representation, e.g., partial
differential equations. Second, they transparently encapsulate hardware-specific knowledge.
In the proposed talk, we focus on the parallel particle-mesh language (PPML) [3]. This
language provides a macro-based frontend to the underlying PPM library [13, 2] as a
parallel run-time system. We analyze PPML’s implementation as well as its advantages
and disadvantages w.r.t. state-of-the-art DSL implementation techniques. Based on this
analysis, we discuss our early efforts in realizing the next version of PPML (Next-PPML) in
conjunction with a redesign of the PPM library in C++.
∗
This work is partially supported by the German Research Foundation (DFG) within the Cluster of
Excellence “Center for Advancing Electronics Dresden”.
2
Towards a Next-Generation Parallel Particle-Mesh Language
2
Particle and Mesh Abstractions
In scientific computing, discrete models are naturally simulated using particles that directly
represent the discrete entities of the model, such as atoms in a molecular-dynamics simulation
or cars in a traffic simulation. These particles carry properties and interact with each other
in order to determine the evolution of these properties and of their spatial location. But also
continuous models, written as partial differential equations, can be simulated using particles.
In this case, the particle interactions discretize differential operators, such as in the DC-PSE
method [14]. This is often combined and complemented with mesh-based discretizations,
such as the finite-difference method. Mesh and particle discretizations are equivalent in that
they approximate the simulated system by a finite number of discrete degrees of freedom that
are the particles or the mesh cells. When using particles together with meshes, it is sufficient
to consider regular Cartesian (i.e., checkerboard) meshes, since all irregular and sub-grid
phenomena are represented on the particles, which can arbitrarily distribute in the domain.
Particles and meshes hence define data abstractions. A particle is a point abstraction
that associates a location in space with arbitrary properties, like color, age, or the value of
a continuous field at that location. These properties, as well as the particle locations, are
updated at discrete time steps over the simulation period by computing interactions with
surrounding particles within a given cut-off radius. Meshes are topological abstractions with
defined neighborhood relations between cells. The properties are stored either on the mesh
nodes or the mesh cells. The PPM library supports both types of abstractions, and also
provides conversion operators between them (i.e., particle-mesh interpolation).
3
Current Status of PPM(L)
Currently, PPM is implemented in object-oriented Fortran2003 [13, 2] and PPML is a macro
system embedded in Fortran2003 [3]. PPML and PPM support transparently distributed
mesh and particle abstractions, as well as parallel operations over them. This also includes
properties and iterators. Different domain-decomposition algorithms allow for the automatic
distribution of data over the nodes of an HPC cluster. Assigning data and work to processing
elements is automatically done by a graph partitioning algorithm, and communication between
processing elements in transparently handled by PPM “mappings”. The mathematical
equations of the model to be simulated are written in LaTeX-like math notation with
additional support for differential operators and dedicated integration loops [3].
Syntactically, PPML is an extension of Fortran2003 providing the aforementioned abstractions as domain-specific language concepts. Technically, the language is implemented as
a source-to-source transformation relying on a mixture of macro preprocessing steps where
macro calls are interspersed with standard Fortran2003 code. Besides C-style preprocessor
directives, PPML also supports non-local macros. These are implemented in Ruby using
eRuby as a macro language and the ANTLR parser generator for recognizing macro output
locations, such as integration loops. Hence, PPML is partially realized using an island
grammar [11].
In this preliminary form, PPML has already nicely demonstrated the benefits of embedded
DSLs for scientific HPC. It has reduced both the size and the development time of scientific
simulation codes by orders of magnitude [3]. It hides much of the parallelization intricacies
(PPML automatically generates MPI) from scientific programmers without preventing them
from using all features of the underlying programming language. The latter is essential since
a DSL may not cover all potential corner cases, and may not always deliver top performance.
However, the current light-weight implementation of PPML has severe disadvantages when
S. Karol and Y. Afshar and P. Incardona and I. F. Sbalzarini and J. Castrillon
Figure 1 Compiler and grammarware-based language processing chain of Next-PPML.
it comes to code analysis algorithms targeting the whole program and domain-specific
optimizations based thereon. Moreover, PPML programs are difficult to debug due to a lack
of semantic error messages. We hence present our intended improvements addressing these
issues in Next-PPML.
4
Approach to Next-PPML
Next-PPML is a language extension using grammarware and compilerware. This allows us to
analyze larger portions of the program code. Examples such as the universal form language
(UFL) [1] for finite-element meshes, the Liszt language for mesh-based solvers [7], and the
Blitz++ [15] stencil template library have shown that domain-specific analyses and built-in
abstractions are beneficial for scientific computing DSLs. Hence, similar concepts will be
considered in the Next-PPML language.
Figure 1 conceptually illustrates the planned tool chain. First, the embedded DSL
program is parsed to an AST-based intermediate representation. This representation already
contains control-flow edges. After computing domain-specific static optimizations on this
intermediate representation, including optimizations to the communication pattern of the
parallel program, the Next-PPML compiler generates an executable (or source code) which
is then used to run the simulation on a parallel HPC cluster. During the simulation run,
the application continuously self-optimizes, e.g., for dynamic load balancing. While static
optimizations are handled by the DSL compiler, dynamic runtime optimization are handled
by the PPM library, which may rely on information provided by the DSL program.
Ideally, the new language uses a declarative approach that bases on an existing programminglanguage grammar and extends it with new productions. Some well-known candidates for
this are Stratego/XT [4], TXL [6], JastAdd [8] or EMFText [9]. However, the target language
is C++11 which has no simple declarative specification. Hence, it is difficult to estimate if
the above-mentioned tools would scale, and implementing a C++ frontend is a huge project
on its own. Therefore, we prefer Clang [10] as an implementation framework, which already
provides built-in analyses that can be adopted and extended.
5
Conclusions
Hybrid particle-mesh simulations are the only scientific computing framework that is able
to simulate models of all four kingdoms: continuous/deterministic, continuous/stochastic,
discrete/deterministic, and discrete/stochastic. This versatility makes the hybrid particlemesh paradigm a prime target for a generic parallel HPC DSL for scientific computing. Prior
work has shown the power of parallelization middleware libraries like PPM, and embedded
DSLs like PPML. Over the past 10 years, they have reduced code development times for
parallel scientific simulations from years to hours, and enabled unprecedented scalability
3
4
REFERENCES
on large HPC machines [5]. The envisioned Next-PPML will address current shortcomings
in code generation, static and dynamic optimization, and semantic error checking and
reporting. It is co-developed with the next generation of the PPM library in C++ using a
semi-declarative language design.
References
[1] Martin Sandve Alnæs et al. “Unified Form Language: A domain-specific language for
weak formulations of partial differential equations.” In: CoRR abs/1211.4047 (2012).
[2] Omar Awile, Ömer Demirel, and Ivo F. Sbalzarini. “Toward an Object-Oriented Core of
the PPM Library.” In: Proc. ICNAAM, Numerical Analysis and Applied Mathematics,
International Conference. AIP, 2010, pp. 1313–1316.
[3] Omar Awile et al. “A domainspecific programming language for particle simulations
on distributed-memory parallel computers.” In: Proceedings of the 3rd International
Conference on Particle-based Methods. 2013.
[4] Martin Bravenboer et al. “Stratego/XT 0.17. A Language and Toolset for Program
Transformation.” In: Science of Computer Programming 72.1-2 (2008): Second Issue of
Experimental Software and Toolkits (EST), pp. 52–70. issn: 0167-6423.
[5] Philippe Chatelain et al. “Billion Vortex Particle Direct Numerical Simulations of
Aircraft Wakes.” In: Comput. Method. Appl. Mech. Engrg. 197 (2008), pp. 1296–1304.
[6] James R. Cordy. “The TXL Source Transformation Language.” In: Science of Computer Programming 61.3 (2006): Special Issue on The Fourth Workshop on Language
Descriptions, Tools, and Applications (LDTA ’04), pp. 190–210. issn: 0167-6423.
[7] Zachary DeVito et al. “Liszt: a domain specific language for building portable meshbased PDE solvers.” In: Proceedings of 2011 International Conference for High Performance Computing, Networking, Storage and Analysis. ACM. 2011, p. 9.
[8] Torbjörn Ekman and Görel Hedin. “The JastAdd System—Modular Extensible Compiler Construction.” In: Science of Computer Programming 69.1-3 (2007): Special Issue
on Experimental Software and Toolkits, pp. 14–26. issn: 0167-6423.
[9] Florian Heidenreich et al. “Model-Based Language Engineering with EMFText.” In:
GTTSE IV. Vol. 7680. LNCS. Springer, 2013, pp. 322–345. isbn: 978-3-642-35991-0.
[10] Chris Lattner. “LLVM and Clang: Next generation compiler technology.” The BSD
Conference. 2008.
[11] Leon Moonen. “Generating Robust Parsers Using Island Grammars.” In: Proceedings
of the Eighth Working Conference on Reverse Engineering 2001. Los Alamitos, CA,
USA: IEEE Computer Society, 2001, pp. 13–22. isbn: 0-7695-1303-4.
[12] I. F. Sbalzarini. “Abstractions and middleware for petascale computing and beyond.”
In: Intl. J. Distr. Systems & Technol. 1(2) (2010), pp. 40–56.
[13] I.F. Sbalzarini et al. “PPM – A highly efficient parallel particle–mesh library for the
simulation of continuum systems.” In: Journal of Computational Physics 215.2 (2006),
pp. 566–588. issn: 00219991.
[14] Birte Schrader, Sylvain Reboux, and Ivo F. Sbalzarini. “Discretization Correction of
General Integral PSE Operators in Particle Methods.” In: J. Comput. Phys. 229 (2010),
pp. 4159–4182.
[15] Todd L. Veldhuizen. “Blitz++: The Library that Thinks it is a Compiler.” In: Advances
in Software Tools for Scientific Computing. Ed. by Hans Petter Langtangen, Are Magnus
Bruaset, and Ewald Quak. Lecture Notes in Computational Science and Engineering
10. Berlin/Heidelberg: Springer, 2000, pp. 57–87. isbn: 978-3-642-57172-5.
DSLs for Graph Algorithms and Graph Pattern Matching
Proposal for DSLDI 2015 - submitted 8 April 2015
Oracle Labs
Oskar van Rest
Sungpack Hong
Hassan Chafi
While SQL and procedural languages like PL/SQL have proven very successful for relational-data-based
applications, the increasing importance of graph-data-based applications is fueling the need for graph
processing DSLs. On the one hand is the need for a procedural-like language that allows you to implement
graph algorithms, like Dijkstra’s shortest path algorithm and Brandes’ Betweenness Centrality algorithm,
while on the other hand is the need for a declarative graph pattern-matching DSL, much like SQL for
graphs. We believe that our DSLs Green-Marl and GMQL are tailored for these two types of graph
workloads and we aim at making the languages standards in the field of graph processing.
Green-Marl is a procedural DSL for implementing graph algorithms for use cases such as product
recommendation, influencer identification and community detection. The language allows you to
intuitively define a wide set of graph algorithms by providing 1) graph-specific primitives like nodes and
edges, 2) built-in graph traversals like BFS and DFS, and 3) built-in iterations over neighbors and incoming
neighbors of nodes in the graph. Furthermore, Green-Marl algorithms can be automatically translated
into parallel implementations in C++, Java or other general purpose languages. Thus, Green-Marl users
can intuitively define their graph algorithms using high-level graph constructs, and then run their
algorithms efficiently on large graphs.
Our other language, GMQL, is a declarative graph pattern-matching query language that borrows syntax
from Neo4j’s Cypher and from the SPARQL query language. It is like SQL for graphs, but instead allows you
to query using concepts like nodes, edges and paths, instead of tables, rows and columns. A graph pattern
can be defined in the form of a set of nodes that are connected via edges or arbitrary-length paths
together with a set of constraints on (properties) of the nodes, edges and paths. More complicated
patterns that are typical for query languages, such as negation and optional matching, are also supported.
The execution of a GMQL query comes down to finding all the instances of the specified pattern inside
the graph. Again, execution can be done in parallel to support the efficient processing of large graphs.
The Green-Marl and GMQL compilers that we created, target our efficient, parallel, and in-memory graph
analytic framework PGX. PGX supports loading graphs from various popular flat file graph formats and can
also load graphs from an Oracle database. Graphs are efficiently stored in memory using the Compressed
Sparse Row (CSR) format. This format allows for huge graphs to fit into the memory of a single machine,
in such a way that they can be processed efficiently. Our framework is also very portable: The Green-Marl
and GMQL compilers target both our Java-based and C++-based PGX backbends. Furthermore, we are
working on a distributed backend that allows for the processing of even bigger graphs that do not fit into
the memory of a single machine. In such distributed mode, graphs are partitioned across multiple
machines and data is exchanged between machines using high-speed Infiniband or 10G-E.
So far, we have seen great performance improvements with PGX, when comparing it to other graph
frameworks: Green-Marl algorithms are processed one to two orders of magnitude faster than
corresponding implementations with the popular machine learning framework GraphLab. Furthermore,
GMQL queries are processed two to four orders of magnitude faster than corresponding Cypher queries
with Neo4j.
Because our PGX framework can process large graphs very efficiently, we decided to also create support
for the popular RDF graph data model. This means we import RDF graphs into PGX and query them using
either GMQL or the standard RDF query language SPARQL. In order to import RDF graphs, we created a
translation from RDF graphs into PGX’ property graphs. Furthermore, in order to query using SPARQL, we
have created a translation from SPARQL to GMQL. This translation currently supports a subset of W3C’s
SPARQL 1.1.
The Green-Marl, GMQL and SPARQL compilers were implemented using the Spoofax language workbench.
Spoofax provides high-level DSLs for specifying grammars, name binding, type systems and
transformations. By using these DSLs, we can keep our code base small and maintainable, while we obtain
much compiler functionality, such Eclipse IDE integration, from Spoofax for free.
Our Spoofax-based SPARQL implementation initially functioned merely as a building block for the SPARQLto-GMQL translation. However, the Eclipse editor that we obtained from Spoofax has become a product
on its own and is useful for anyone who wants to write SPARQL queries. The editor has full support for
W3C’s SPARQL 1.1 and provides editor features like formatting, code completion, syntax checking, namebased checks for variables and prefixes, and editor navigation from name uses to their definitions. Many
of the features came with little implementation effort. For example, our SPARQL grammar definition in
Spoofax’ grammar definition formalism SDF3, gave us a parser, a formatter, syntax-checks and error
recovery rules, while our SPARQL name-binding definition in Spoofax’ name binding language NaBL gave
us name-based code completion, name checks and editor navigation. We were even able to encode the
more complicated name-binding rules needed for SPARQL constructs like NOT EXISTS, MINUS and for
Subqueries in NaBL, in very few lines of code.
Finally, we also allow our Green-Marl, GMQL and SPARQL compilers to be linked against a particular graph
such that we can perform additional compile-time error checking. For example, our GMQL compiler is
able to warn a user when they have misspelled a property name, based on the node and edge properties
that are available in a particular graph. We also provide such kinds of checks when querying schema-less
RDF graphs, by extracting schema information from an RDF graphs when it is loaded into PGX. In our
Eclipse editors, such error checking even happens in real-time.
DSLs of Mathematics, Theorems and
Translations
Cezar Ionescu∗
Patrik Jansson∗
In this talk, we present some of the ideas behind the course on DSLs of Mathematics (DSLM), currently in preparation in Chalmers.
We view mathematics as a rich source of examples of DSLs. For example, the
language of group theory, or the language of probability theory, embedded in
that of measure theory. The idea that the various branches of mathematics are
in fact DSLs embedded in the “general purpose language” of set theory was
(even if not expressed in these words) the driving idea of the Bourbaki project,
which exerted an enormous influence on present day mathematics.
In DSLM, we consequently develop this point of view, aiming to show computer
science students that they can use the tools from software engineering and functional programming in order to deal with the classical continuous mathematics
they encounter later in their studies.
In this talk, we’ll start with the simple example of the standard development of
a calculus of derivatives. This can be seen as a DSL whose semantics are given in
terms of limits of real sequences. We can try to give alternative semantics to this
language, in terms of complex numbers. This leads to the notion of holomorphic
function, and to an essentially different calculus than in the real case.
Our second example is that of extending the language of polynomials to power
series. This DSL can also be interpreted in various domains: real numbers,
complex numbers, or intervals.
In the case of complex numbers, a fundamental theorem creates a bridge between the DSL of derivatives and that of power series, through the identity of
holomorphic and (regular) analytic functions. This leads to the discussion of
translation between DSLs, an aspect which is fundamental in mathematics, but
has been somewhat neglected by computer science. Thus, we believe that a
closer examination of the DSLs of mathematics can also be relevant for practical
software engineering.
∗ Chalmers
University of Technology
Check Syntax: An Out-of-the-Box Tool for Macro-Based DSLs
Spencer Florence
Northwestern University
Ryan Culpepper
Northeastern University
Abstract
Racket supports DSL construction through an open compiler,
where the compiler is made open through its macro system. Programmers define new syntactic constructs by macro-expansion to
existing constructs, using the Racket module system to hide the
originals and rename the new ones as needed. Taken together, these
facilities permit the definition of new languages, even enabling new
languages to give new semantics to familiar syntax.
In this world, all programs are compiled to a well-known intermediate language, and tools can operate on that structure to compute information about programs—including programs written in
a DSL that the tool knows nothing about. Crucially, binding information and original source locations are available in the intermediate language. This information allows tools to provide keybindings
to hop between bound and binding occurrences of identifiers and
rename them. Since Racket’s documentation system is based on
binding, tools can also conveniently access API documentation.
1.
Building a Macro-based Language in Racket
Programming tools typically support multiple languages by compiling all programs into a shared intermediate language. In Racket,
the representation of the intermediate language is the same as for
macro transformations: an enriched form of S-expressions that embeds source-location and binding information. This representation
offers an especially convenient way to make new languages via
macros, where the new languages inherit tools that operate on
macro-expanded programs.
For example, to build a DSL that uses call-by-need evaluation
instead of Racket’s usual call-by-value convention, we start with a
wait construct for delaying the evaluation of an expression. The
wait form is defined using define-syntax-rule, which directs
the compiler to replace expressions matching one subexpression
with another.
(define-syntax-rule
(wait e)
(Wait #t (λ () e)))
The definition of wait specifies that the body e in any (wait e)
is wrapped in a thunk (thus delaying the evaluation of e) and the
thunk is packaged into a Wait structure:
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
, .
Copyright © ACM [to be supplied]. . . $15.00.
http://dx.doi.org/10.1145/
Matthew Flatt
Robert Bruce Findler
University of Utah
Northwestern University
(struct Wait (waiting? TorV) #:mutable)
This struct declaration indicates that the Wait structure has two
mutable fields, waiting? and TorV, and it defines supporting
functions: Wait-waiting?, which accepts a Wait structure and
returns the value in the first field; set-Wait-waiting?!, which
accepts a Wait structure and a value and a new waiting? field
value and mutates the structure; Wait-TorV, which returns the
value of a Wait structure’s second field; and set-Wait-TorV!,
which mutates the value of a Wait structure’s second field.
To force a delayed evaluation, we define an act function. It
operates on a Wait structure, returning the result of the thunk
encapsulated in the second field and caching the thunk’s result.
(define (act w)
(cond
[(Wait? w)
(when (Wait-waiting? w)
(set-Wait-TorV! w ((Wait-TorV w)))
(set-Wait-waiting?! w #f))
(act (Wait-TorV w))]
[else w]))
Next, we define macros that compile the various constructs of
our new language into constructs that already exist in Racket. In
this case, we’ll use the wait functionality and exploit Racket’s λ,
a conditional form, and some simple arithmetic operations, lifting
them into our lazy language. The following 8 lines provide the rest
of a lazy language implementation, using a mixture of function
definitions and macros.
(define-syntax-rule
(app f x ...)
((act f) (wait x) ...))
(define (multiply a b) (* (act a) (act b)))
(define (subtract a b) (- (act a) (act b)))
(define-syntax-rule
(if0 e1 e2 e3)
(if (= (act e1) 0) e2 e3))
Using these constructs directly would be awkward. They do not
have the right names, and it is easy to accidentally step outside of
the language and use, for example, Racket’s * operation instead
of the multiply. To avoid those problems, we put the definitions
into a module, we use provide’s rename-on-export capability to
provide those operations with their expected names, and we hide
everything else (by simply not providing anything else). The only
subtle point here is that, when the Racket macro expander sees an
open parenthesis with no macro following it, it inserts a reference
to the #%app macro to make function application explicit, so we
must export app as #%app to cooperate with this feature of macro
expansion. Then, we can install the language as a Racket package,
so that if the first line of a file is #lang mini-lazy, the rest of the
file sees our bindings. For example, running the following program
(using the call-by-name Y combinator) produces 3628800.
#lang mini-lazy
(((λ (f) ((λ (x)
(f (x x)))
(λ (x)
(f (x x)))))
(λ (fac)
(λ (n)
(if0 n 1 (* n (fac (- n 1)))))))
10)
Even though this language has a radically different order of
evaluation from Racket, it still compiles into Racket and, even
better, scope is preserved by this compilation.
2.
Check Syntax
Check Syntax runs continuously as part of the DrRacket IDE. Each
time the user edits a program, Check Syntax takes the content as a
string, parses it using the current language’s parser, and hands it off
to the macro system, resulting in a program where the macros have
all been expanded away. In this state, programs are in a well-known
language that contains functions, conditionals, variable binding,
variable reference, and other core forms of Racket.1
Check Syntax traverses a program’s expansion, searching for
bindings and references. The expanded form of a program is an
enriched form of S-expressions known as syntax objects. A syntax
object includes information about whether or not a form appeared
in the original program and its location in the source, and identifiers
extracted from an expanded program can be compared to determine
whether they refer to the same binding. Check Syntax collects this
information into pairs of source locations, which DrRacket uses to
draw binding information in the editor window.
For the example at the end of the previous section, these are the
set of arrows that Check Syntax collects for the lexical variables:2
#lang mini-lazy
(((λ (f) ((λ (x)
(f (x x)))
(λ (x)
(f (x x)))))
(λ (fac)
(λ (n)
(if0 n 1 (* n (fac (- n 1)))))))
10)
Check Syntax uses this information in several ways. First, when
the mouse is over a variable occurrence, Check Syntax draws arrows to the other occurrences. Second, Check Syntax provides keybindings to jump between the occurrences of a variable in a file.
Finally, Check Syntax offers a bound-variable rename facility. The
user selects a variable and supplies a new name, and Check Syntax
follows the arrows to rename the variables.
The technique of traversing expanded programs can support
a myriad of different languages, even parenthesis-challenged languages like Datalog:
#lang datalog
ancestor(A, B) :parent(A, B).
ancestor(A, B) :parent(A, C), ancestor(C, B).
1 http://docs.racket-lang.org/reference/
syntax-model.html#%28part._fully-expanded%29
2 Naturally, the figures in this paper are generated by running Check Syntax
on the shown program text as the paper is typeset.
and Algol 60:
#lang algol60
begin
integer procedure SIGMA(x, i, n);
value n; integer x, i, n;
begin
integer sum;
sum:=0;
for i:=1 step 1 until n do
sum:=sum+x;
SIGMA:=sum;
end;
integer q;
printnln(SIGMA(q*2-1, q, 7));
end
3.
Scaling Up
For many constructs, the process described in the previous section
is enough to reconstruct the binding structure of a program. Other
constructs appear to the programmer as binding forms and variable
references, but they are compiled into data-structure accesses. For
example, one of the modularity constructs in Racket, unit, turns
variable definitions into the creation of a pointer and variable references into pointer dereferences. Similarly, the pattern-matching
part of define-syntax-rule turns variable references into function calls that destructure syntax objects.3
To handle such forms, Check Syntax needs a little cooperation
from the implementing macro. When producing an expanded form,
a macro can add properties to the result syntax objects. These
properties have no effect on how the code runs, but they are used by
Check Syntax to draw additional arrows. More precisely, the macro
can add a property indicating that some syntax object (i.e., a piece
of the input to the transformation) was a conceptually a binding or
a reference; Check Syntax uses those syntax objects for renaming
and navigation, just like the program’s other syntax objects.
Check Syntax also recognizes properties for tool-tip information. A macro can put an arbitrary string as a property on the result
of a macro, and Check Syntax displays the string in a tooltip. For
example, Typed Racket adds properties so that Check Syntax displays the types of expressions.
In Racket, the build process for documentation creates a database
that maps module names and exports to the file containing the documentation and to a function’s contract or a syntactic form’s grammar. Check Syntax can examine a variable in the fully expanded
program and discover which module exported the variable. It uses
that information to consult the database and can render the contract/grammar directly in the DrRacket window, including a link to
the full documentation.
In the following Typed Racket program, highlighting shows the
locations where Check Syntax finds documentation in the database:
#lang typed/racket
(: fib (-> Integer Integer))
(define (fib n)
(cond
[(= n 0) 0]
[(= n 1) 1]
[else (+ (fib (- n 1))
(fib (- n 2)))]))
Acknowledgments. Thanks to Matthias Felleisen for comments on
earlier drafts and feedback on Check Syntax over the years.
3 The define-syntax-rule form is itself is a macro that expands into
the primitive macro-building form with a compile-time syntax transformer
that is synthesized from the specified pattern and template.
Dynamic Compilation of DSLs
Vojin Jovanovic and Martin Odersky
EPFL, Switzerland
{firstname}.{lastname}@epfl.ch
Domain-specific language (DSL) compilers use domain knowledge to perform domain-specific optimizations that can yield several orders of magnitude
speedups [4]. These optimizations, however, often require knowledge of values known only at program runtime. For example, in matrix-chain multiplication, knowing matrix sizes allows choosing the optimal multiplication order [2,
Ch. 15.2] and in relational algebra knowing relation sizes is necessary for choosing the right join order [6]. Consider the example of matrix-chain multiplication:
v a l (m1, m2, m3) = ... // matrices of unknown size
m1 * m2 * m3
In this program, without knowing the matrix sizes, the DSL compiler can not
determine the optimal order of multiplications. There are two possible orders
(m1*m2)*m3 with an estimated cost c1 and m1*(m2*m3) with an estimated
cost c2 where:
c1 = m1.rows*m1.columns*m2.columns+m1.rows*m2.columns*m3.rows
c2 = m2.rows*m2.columns*m3.columns+m1.rows*m2.rows*m3.columns
Ideally we would change the multiplication order at runtime only when the
condition c1 > c2 changes. For this task dynamic compilation [1] seems ideal.
Yet, dynamic compilation systems—such as DyC [3] and JIT compilers—
have shortcomings. They use runtime information primarily for specialization.
In these systems profiling tracks stability of values in the user program. Then,
recompilation guards and code caches are based on checking equality of current
values and previously stable values.
To perform domain-specific optimizations we must check stability, introduce guards, and code caches, based on the computation specified in the DSL
optimizer—outside the user program. Ideally, the DSL optimizer should be agnostic of the fact that input values are collected at runtime. In the example
stability is only required for the condition c1 > c2, while the values c1 and c2
themselves are allowed to be unstable. Finally, recompilation guards and code
caches would recompile and reclaim code based on the same condition.
An exception to existing dynamic compilation systems are Truffle [7] and
Lancet [5]. They allow creation of user defined recompilation guards. However,
with Truffle, language designers do not have the full view of the program, and
thus, can not perform global optimizations (e.g., matrix-chain multiplication
optimization). Further, recompilation guards must be manually introduced and
the code in the DSL optimizer must be modified to specially handle decisions
based on runtime values.
2
We propose a dynamic compilation system aimed for domain specific languages where:
– DSL authors declaratively, at the definition site, state the values that are
of interest for dynamic compilation (e.g., array and matrix sizes, vector and
matrix sparsity). These values can regularly be used for making compilation
decisions throughout the DSL compilation pipeline.
– The instrumented DSL compiler transparently reifies all computations on
the runtime values that will affect compilation decisions. In our example,
the compiler reifies and stores all computations on runtime values in the
unmodified dynamic programming algorithm [2] for determining the optimal
multiplication order (i.e., c1 > c2).
– Recompilation guards are introduced automatically based on the stored
DSL compilation process. In the example the recompilation guard would
be c1 > c2.
– Code caches are automatically managed and addressed with outcomes of the
DSL compilation decisions instead of stable values from user programs. In
the example the code cache would have two entries addressed with a single
boolean value computed with c1 > c2.
The goal of this talk is to foster discussion on the new approach to dynamic
compilation with focus on different policies for automatic introduction of recompilation guards: i) heuristic, ii) DSL author specified, and iii) based on domain
knowledge.
References
1. Joel Auslander, Matthai Philipose, Craig Chambers, Susan J Eggers, and Brian N
Bershad. Fast, effective dynamic compilation. In International Conference on Programming Language Design and Implementation (PLDI), 1996.
2. Thomas H Cormen, Charles E Leiserson, Ronald L Rivest, Clifford Stein, et al.
Introduction to algorithms, volume 2. MIT press Cambridge, 2001.
3. Brian Grant, Markus Mock, Matthai Philipose, Craig Chambers, and Susan J Eggers. DyC: an expressive annotation-directed dynamic compiler for C. Theoretical
Computer Science, 248(1), 2000.
4. Tiark Rompf, Arvind K. Sujeeth, Nada Amin, Kevin J. Brown, Vojin Jovanović,
HyoukJoong Lee, Manohar Jonnalagedda, Kunle Olukotun, and Martin Odersky.
Optimizing data structures in high-level programs: New directions for extensible
compilers based on staging. In Symposium on Principles of Programming Languages
(POPL), 2013.
5. Tiark Rompf, Arvind K. Sujeeth, Kevin J. Brown, HyoukJoong Lee, Hassan Chafi,
Kunle Olukotun, and Martin Odersky. Project Lancet: Surgical precision JIT compilers. In International Conference on Programming Language Design and Implementation (PLDI), 2013.
6. P Griffiths Selinger, Morton M Astrahan, Donald D Chamberlin, Raymond A Lorie,
and Thomas G Price. Access path selection in a relational database management
system. In International conference on Management of data (SIGMOD), pages 23–
34, 1979.
7. Thomas Würthinger, Christian Wimmer, Andreas Wöß, Lukas Stadler, Gilles Duboscq, Christian Humer, Gregor Richards, Doug Simon, and Mario Wolczko. One
VM to rule them all. In Symposium on New Ideas, New Paradigms, and Reflections
on Programming & Software (Onward!), 2013.
A practical theory of language-integrated query
—and—
Everything old is new again
Philip Wadler
The University of Edinburgh
[email protected]
15 April 2015
The proposed talk would consist of summaries of two recent pieces of work,
both concerned with applying quotation and the subformula property to domainspecific languages. Both will also be the subject of an invited talk by the author
at Curry On.
A practical theory of language-integrated query
James Cheney, Sam Lindley, Philip Wadler
submitted to ICFP 2015
http://homepages.inf.ed.ac.uk/wadler/topics/recent.html#essence-of-linq
Language-integrated query is receiving renewed attention, in part
because of its support through Microsoft’s LINQ framework. We
present a theory of language-integrated query based on quotation
and normalisation of quoted terms. Our technique supports abstraction over values and predicates, composition of queries, dynamic
generation of queries, and queries with nested intermediate data.
Higher-order features prove useful even for constructing first-order
queries. We prove that normalisation always succeeds in translating any query of flat relation type to SQL. We present experimental
results confirming our technique works, even in situations where Microsoft’s LINQ framework either fails to produce an SQL query or,
in one case, produces an avalanche of SQL queries.
Everything old is new again: Quoted Domain Specific Languages.
Shayan Najd, Sam Lindley, Josef Svenningsson, Philip Wadler
ICFP 2013
http://homepages.inf.ed.ac.uk/wadler/topics/recent.html#qdsl
We describe a new approach to domain specific languages (DSLs),
called Quoted DSLs (QDSLs), that resurrects two old ideas: quotation, from McCarthy’s Lisp of 1960, and the subformula property,
from Gentzen’s natural deduction of 1935. Quoted terms allow the
DSL to share the syntax and type system of the host language.
Normalising quoted terms ensures the subformula property, which
1
guarantees that one can use higher-order types in the source while
guaranteeing first-order types in the target, and enables using types
to guide fusion. We test our ideas by re-implementing Feldspar,
which was originally implemented as an Embedded DSL (EDSL), as
a QDSL; and we compare the QDSL and EDSL variants.
2
28
Panel Discussion
Figure 1: Discussion Panel at DSLDI’15. From left to right: Matthew Flatt, Jonathan Aldrich,
Andrzej Wąsowski Laurence Tratt and Sebastian Erdweg
Position Statements of the Panelists
Jonathan Aldrich Position: DSL frameworks should guarantee the absence of syntactic conflict, and support unanticipated interoperation between DSLs in the code and during typechecking,
execution, and debugging, without losing aspects that are special to each DSL.
Rationale: Programming today is all about composition; developers gain enormous leverage
from libraries, and expect them to work together even if they were designed separately. Conflicts
that prevent compilation when you merely import two different DSLs are completely unacceptable
in this world. Furthermore, most of the value from DSLs comes when they work like ”real”
languages, with checking, execution, and debugging facilities that are natural; an 80% solution is
not going to convince most real-world developers to adopt a DSL. In a composition-based world,
therefore, all these facilities must work even when multiple DSLs are used together.
Concrete Illustration: Here’s a multi-part challenge problem for language composition in DSL
frameworks:
• (A) Have different developers independently design and build DSLs for state machines and
structured synchronous programming
29
30
• (B) The DSL framework should guarantee that these DSLs can be used together without
having to resolve any syntactic conflicts
• (C) Write a state machine and a structured synchronous program that drives its state transitions, ideally with no visible role played by the DSL framework.
• (D) Statically verify that the structured synchronous program does not misuse the state
machine (e.g. by generating transitions that aren’t appropriate for the machine’s state)
• (E) With respect to task D, report any errors in a way that is consistent with both the
structured synchronous program and the state machine (rather than some translation of
each).
The two developers in A are not allowed to communicate or to anticipate tasks B-E. Tasks
B–E must be done without changing the DSLs developed in A.
Sebastian Erdweg The same language features reoccur in the design of many DSLs: operations
on primitive data, operations on structured data, conditionals and backtracking, error handling,
and many more. Yet, we have no principled way of composing basic language blocks into working
DSLs and we have no way of detecting and eliminating interactions between language features.
This is one of the big open challenges in the area of DSLs.
Matthew Flatt How can language-composition tools mediate extensions that depend on (or,
alternatively, adapt to) different semantics of shared constructs, such as function application?
For example, what happens when a form whose implementation depends on eager evaluation
is used in an otherwise lazy context? Or what happens when a from that implies a function
application is used in a language where function application is meant to be syntactically restricted
to first-order functions?
Racket’s hygienic-macro approach reflects core constructs like function application through
macros, such as #%app, whose use is typically implicit. A macro by default adopts its definitionsite implementation of such macros, which is usually the right approach. That means, however,
that a macro that uses eager function application has questionable behavior in a lazy use context.
Similarly, macros tend not to respect the function-application constraints of a context like Beginner
Student Language. A macro can adapt to a use-site notion of #%app by non-hygienically referencing
#%app from the use context, but that approach is relatively tedious not not commonly followed.
Laurence Tratt Language composition challenge: Integrating existing languages into a language composition framework. Controversial statement: People have not shown themselves hugely
interested in the forms of language composition we’ve given them thus far.
Andrzej Wąsowski
• I believe that language composition is not a language problem, but a software engineering
problem.
• If you are lucky then the DSLs are composed by framework designers, which are usually
means very good programmers (or at least above-average).
• More often languages are composed by framework *users*, who design systems (often average
programmers or worse). You rarely find serious project using less than 5 languages, and 20
is a norm. Many of them DSLs.
• The challenge is how to allow language composition (or integration) for non-language designers, but for system designers (the language users), so that they still get static checking,
meaningful messages, across language testing, etc.
31
Summary of the Discussion
• The practical value of a DSL does not grow linearly with the quality of the implementation.
You need a to have really polished framework/ecosystem to deliver value. (80/20 or 20/80
(todo)).
• The software engineering perspective on language composition is different from the programming language perspective. The software engineering perspective emphasizes language
interopability, cross-language IDE support, and DSLs that are not necessarily like programming languages (e.g., config files, build files, deployment descriptors, data mapping files,
etc.).
• Is using different languages in the same file an essential aspect of language composition? Is
language interopability an instance of language composition, and what are the consequences
for performance?
• Translating all language features down to a single virtual machine for execution, is possible,
but there can be costs in terms of performance. Example: the JVM is a state-of-the-art
virtual machine, but is not suitable for executing Prolog.
• Modular language components seems to be an extremely hard to achieve goal, but it is
necessary at the same time. Without it, the number of feature interactions quickly explodes.
• There is a need for language interfaces: what features, services or constructs are exported
from a language component?
• Integrating syntax and semantics are only two language aspects that need to be composed
for a realistic programming experience of a composed language. Examples include: name
binding, type checking, IDE features, etc.
• Cross language name analysis seems feasible and would solve real problems of programmers, now. A successful example is Jetbrains’ IntelliJ which integrates all references in Web
framework configuration files with Java.
| 6cs.PL
|
Accelerated Methods for Deep Reinforcement Learning
Adam Stooke 1 Pieter Abbeel 1
arXiv:1803.02811v1 [cs.LG] 7 Mar 2018
Abstract
Deep reinforcement learning (RL) has achieved
many recent successes, yet experiment turnaround time remains a key bottleneck in research
and in practice. We investigate how to optimize
existing deep RL algorithms for modern computers, specifically for a combination of CPUs and
GPUs. We confirm that both policy gradient and
Q-value learning algorithms can be adapted to
learn using many parallel simulator instances. We
further find it possible to train using batch sizes
considerably larger than are standard, without
negatively affecting sample complexity or final
performance. We leverage these facts to build
a unified framework for parallelization that dramatically hastens experiments in both classes of
algorithm. All neural network computations use
GPUs, accelerating both data collection and training. Our results include using an entire NVIDIA
DGX-1 to learn successful strategies in Atari
games in single-digit minutes, using both synchronous and asynchronous algorithms.
1. Introduction
Research in deep reinforcement learning (RL) has relied heavily on empirical evaluation, making experiment
turnaround time a key limiting factor. Despite this critical
bottleneck, many reference implementations do not fulfill
the potential of modern computers for throughput, unlike in
supervised learning (see e.g. (Goyal et al., 2017)). In this
work, we study how to adapt deep RL algorithms–without
changing their underlying formulations–to better leverage
multiple CPUs and GPUs. The result is a significant gain
in efficiency and scale of hardware utilization and hence in
learning speed.
Today’s leading deep RL algorithms have roughly clustered
into two families: (i) Policy gradient methods, which rely on
on-policy samples for training. A representative example is
Asynchronous Advantage Actor-Critic (A3C) (Mnih et al.,
1
Department of Computer Science, University of California,
Berkeley, California, USA. Correspondence to: Adam Stooke
<[email protected]>.
2016), in which sampler-learner processes train a centralized
neural network in parallel, using only a CPU. (ii) Q-value
learning methods, which permit off-policy training. A representative example is Deep Q-Networks (DQN) (Mnih et al.,
2015), in which a single simulator feeds a replay buffer of
transitions, and this buffer is randomly sampled for training
(with a GPU). Traditionally, implementations of policy gradient methods and Q-learning methods have differed from
each other; in this paper we unify them under the same
framework for scaling.
Modern computers host many CPU cores and multiple
GPUs on one motherboard, and we target this hardware
mix. In deep RL the neural net participates both in learning
and in data collection (to execute the agent’s current policy). We use GPUs for hardware acceleration of all neural
network computations; to do so efficiently requires batched
processing. During data collection, we run many simulator
instances and batch calls to the policy or Q-function for
inference. During learning, we use large batch sizes and/or
asynchronous methods to expedite computation.
Our contribution is a framework for parallelized deep RL
including novel techniques for GPU acceleration. Within
this framework, we demonstrate multi-GPU versions of
the following algorithms: Advantage Actor-Critic (Mnih
et al., 2016), Proximal Policy Optimization (PPO) (Schulman et al., 2017), DQN (Mnih et al., 2015), Categorical
DQN (Bellemare et al., 2017), and Rainbow (Hessel et al.,
2017). Our target hardware is an NVIDIA DGX-1, which
contains 40 CPU cores and 8 P100 GPUs. To provide calibrated results, we test our implementations in the heavily
benchmarked Atari-2600 domain via the Arcade Learning
Environment (ALE) (Bellemare et al., 2013).
We found that highly parallel sampling using batched inferences can accelerate experiment turn-around time of all
algorithms without hindering training. We further found that
neural networks can learn using batch sizes considerably
larger than are standard, without harming sample complexity or final game score, and this dramatically speeds up
learning.
The primary benefit of our implementation is improved execution speed. Our policy gradient algorithms can leverage
the entire DGX-1 to learn successful game strategies in under 10 minutes, rather than hours. We reduced the duration
Accelerated Methods for Deep Reinforcement Learning
of some standard Q-value-learning runs from 10 days to
under 2 hours. We believe that these results promise to
accelerate research in deep RL, and we suggest directions
for further investigation and development.
following the policy after first using action at to advance.
With these definitions in hand, we briefly review the deep
RL algorithms used in our experiments.
3.1. Policy Gradient Methods
2. Related Work
Efforts to parallelize and accelerate deep RL algorithms
have been underway for several years. Gorila (Nair et al.,
2015) parallelized DQN using distributed computing. It
achieved significant although sub-linear speedups using
hundreds of computing units as samplers or learners, with
central parameter servers for managing parameter updates.
This effort suffered in sample complexity relative to singlethreaded DQN. More recently, (Horgan et al., 2018) showed
that a distributed, prioritized replay buffer can support faster
learning while using hundreds of CPU cores for simulation and a single GPU for training. The same work used
increased batch sizes, with a brief study of the effect of
learning rate.
The policy gradient method A3C is itself a parallelized algorithm. In GA3C (Babaeizadeh et al., 2016), a speedup
over CPU-only A3C was achieved by using a GPU. It was
employed asynchronously, with “predictor” and “trainer”
threads queuing observations and rewards for batched inferences and training updates. GA3C introduced a “policy
lag” between generation and consumption of training data,
worsening sample complexity. In independent work simultaneous to ours, (Espeholt et al., 2018) extended policy
gradient methods to a distributed setting, enabling an alternative approach to multi-GPU training called IMPALA.
They introduced a more heavily modified algorithm, V-trace,
to mitigate policy lag–which we avoid–and did not employ
GPU inference. In PAAC (Clemente et al., 2017), the authors explored the use of many simulators and increased
batch sizes learning rates in (single-GPU) batched A2C–
ideas central to our studies. Our contributions to actor-critic
methods exceed this work in a number of ways, chiefly:
improved sampling organization, tremendously enhanced
scale and speed using multiple GPUs, and inclusion of asynchronous optimization.
3. RL Algorithm Background
In reinforcement learning, an agent interacts with an environment and uses the experience to optimize a decision-making
policy. In a standard RL formulation, the agent
to maxPaims
∞
imize the sum of discounted rewards, Rt = k=0 γ k rt+k ,
in expectation, i.e. E [Rt ]. γ is the discount factor (e.g.
0.99), and rt is the reward received at time t. The value
of a state, V (st ) = E [Rt |st ], is defined as the expected
discounted return from that state under a given policy, π.
The Q-value, Q(st , at ) = E [Rt |st , at ] is the same but for
In policy gradient methods, the policy is directly parameterized in the form π(a|s; θ), where π is a probability distribution over actions a when observing state s, as parameterized
by θ, a neural network. The agent exercises the policy in the
environment, recording experiences. Periodically, it uses
the samples to update θ by estimating the gradient ∇θ E [Rt ].
Typically, the agent then discards these samples and repeats,
optimizing the policy iteratively.
Advantage Actor-Critic:
In advantage actorcritic,
the policy gradient is computed as
E [∇θ log π(at |st ; θ)(Rt − V (st ))].
The agent estimates V (st ) from the data, for instance using a separate
output from the same network used for π. (Rt − Vt )
estimates the advantage: A(s, a) = Q(s, a) − V (s). Rt
is computed using the discounted sum of as many future
returns as are observed in a given batch, up to rtmax , and is
bootstrapped with V (stmax +1 ), appropriately discounted.
The estimator V (s; θ) is trained using e.g. a squared-error
loss, simultaneously to π. Lastly, an entropy bonus is added
to the gradient: ∇θ H(π(·|st ; θ)), to promote exploration
and discourage premature convergence.
In A3C (Mnih et al., 2016), separate actor-learner threads
sample environment steps and update a centralized copy of
the parameters asynchronously to each other. In (batched)
A2C, which performs similarly to A3C (see, e.g. (Schulman
et al., 2017)), separate environment instances are sampled,
but the learner is single-threaded. It gathers all data into one
minibatch to compute the gradient.
Proximal Policy Optimization: Proximal Policy Optimization (PPO) (Schulman et al., 2017) is a closely
related algorithm that improves sample complexity
by increasing the training use of each sample. It
maximizes a “surrogate” objective E [ρt (θ)At ], where
ρt (θ) = π(at |st ; θ)/π(at |st ; θold ) is the likelihood
ratio of the recorded action between the updated
and original policies. We use the clipping-objective
version of PPO, which optimizes the expression
E [min (ρt (θ)At , clip(ρt (θ), 1 − , 1 + )At )],
noting
that ρt (θold ) = 1. Unlike A3C, PPO performs multiple
parameter updates using (minibatches from) each set of
samples; the clipping objective prevents excessive policy
updates from destabilizing learning.
3.2. Q-Value-Learning Methods
Instead of directly parameterizing a policy, Q-value learning
methods estimate the Q-function as Q(s, a; θ). The greedy
Accelerated Methods for Deep Reinforcement Learning
policy selects the (discrete) action maximizing value: a∗ =
argmaxa Q(s, a; θ). Exploration can be performed using
an -greedy policy, which chooses a uniform random action
with probability and otherwise uses the greedy action. By
its off-policy nature, Q-learning permits repeated training
use of samples.
Deep Q-Networks: Deep Q-Networks (DQN) (Mnih et al.,
2015) optimize the Q-value estimator
by repeated grah
i
2
dient descent on the objective: E (yi − Q(si , ai ; θ)) ,
where yi is the data-estimated Q-value given by yi =
ri + γ maxa Q(si+1 , a; θ− ) (although in practice the Huber
loss is used). DQN gains stability by 1) using a separate
“target” network θ− , periodically copied from θ, and 2) training from a replay buffer, to which it continually adds new
experiences. Several enhancements to DQN have improved
performance.
Categorical DQN: Categorical DQN (Bellemare et al.,
2017) discretizes the possible Q-values into a fixed range
and number of “atoms” and learns a distribution for each
Q(s, a; θ), so as to represent stochastic features of the environment. The squared-error loss function is replaced by
the Kullback-Leibler divergence against a target distribution.
The target is constructed by contracting the “future” atoms
according to γ and shifting them by Rt . Action-selection is
still performed greedily according to expected value over
the distribution.
-Rainbow: Recently, the use of distributional learning
was combined with five other enhancements to DQN under
a single name: Rainbow (Hessel et al., 2017). Without
describing them here, the enhancements include: 1) DoubleDQN (Van Hasselt et al., 2016), 2) Dueling Networks (Wang
et al., 2016), 3) Prioritized Replay (Schaul et al., 2016), 4)
n-step learning (Peng & Williams, 1994), and 5) NoisyNets
(Fortunato et al., 2018). In our experiments, we use the
-greedy version of Rainbow, without parameter noise.
4. Parallel, Accelerated RL Framework
We consider learning scenarios where the environment is
a CPU-based simulator and the policy uses a deep neural
network. We describe here a complete set of parallelization
techniques that achieve high throughput during both sampling and optimization. It is applicable to many deep RL
algorithms, including those outlined above. We treat GPUs
homogeneously; each performs the same sampling-learning
procedure. This strategy scales straightforwardly to various
numbers of GPUs.
4.1. Synchronized Sampling
We begin by associating multiple CPU cores with a single
GPU. We run multiple simulators in parallel processes on
the CPU cores, and these processes perform environment
steps in a synchronized fashion. At each step, all individual
observations are gathered into a batch for inference, which
is called on the GPU once the last observation is submitted.
The simulators step again once the actions are returned,
and so on, as in (Clemente et al., 2017). System shared
memory arrays provide fast communication between the
action-server and simulator processes.
The number of CPU cores might limit the inference batch
size, so we use multiple simulator instances per process.
These are stepped sequentially (within each process) and
always all together in every inference batch. This arrangement adds the benefit that it mitigates synchronization losses.
Variance in simulator stepping time arises from differing
simulator states and other random fluctuations. The straggler
effect worsens with increased number of parallel processes
but diminishes with increased number of sequential simulators. A schematic is shown in Figure 1. Slowdowns caused
by long environment resets can be avoided by resetting only
during optimization pauses.1
Figure 1. Synchronized sampling schematic, showing multiple simulator processes, each with multiple simulator instances, interacting synchronously with a GPU-based action-server through time
(alternating scheme not shown).
When running this sampling engine, hardware component
utilization reflects the relative load of each side of computation. If simulation and inference loads are balanced, this
would cause each component to sit idle half of the time. We
avoid this idleness by forming two alternating groups of
simulator processes. While one group awaits an inference,
the other steps, and the GPU alternates between servicing
each group. Alternation keeps utilization high; moreover, it
hides the execution time of whichever computation is the
quicker of the two.
We found it beneficial to fix the CPU assignment of each
simulator process. To use multiple GPUs, we simply repeat
the template, allocating available CPU cores evenly. We
have found it sufficient to reserve one CPU core to run each
1
For example, one may either 1) ignore a simulator in need of
reset or 2) immediately swap in a fresh instance held in reserve.
Accelerated Methods for Deep Reinforcement Learning
GPU. The experiments section contains measurements of
sampling speed, which indeed increases with the number of
environment instances.
4.2. Synchronous Multi-GPU Optimization
In our synchronous algorithms, all GPUs maintain identical parameter values. We leverage the data-parallelism
of stochastic gradient estimation and use the well-known
update procedure, on every GPU: 1) compute a gradient
using locally-collected samples, 2) all-reduce the gradient
across GPUs, 3) use the combined gradient to update local
parameters. We use the NVIDIA Collective Communication
Library for fast communication among GPUs.
4.3. Asynchronous Multi-GPU Optimization
In asynchronous optimization, each GPU acts as its own
sampler-learner unit and applies updates to a central parameter store held in CPU memory. The training batch size does
not aggregate across GPUs, thus avoiding a potential limit
to scaling.
Use of accelerators compels a choice of where to perform
the parameter update. In our experience, applying common
update rules to the network is faster on the GPU. Our general update procedure includes three steps: 1) compute the
gradient locally and store it on the GPU, 2) pull current central parameters onto the GPU and apply the update rule to
them using the pre-computed gradient, 3) write the updated
parameters back to the central store. After this sequence, the
local GPU parameters are in sync with the central values,
and sampling proceeds again. Following (Mnih et al., 2016),
we centralize the update rule parameters in addition to the
network weights.
Rather than add update increments to the central parameters,
which requires CPU computation, we overwrite the values.
Therefore, we employ a lock around steps (2) and (3) above,
preventing other processes from reading or writing parameter values concurrently. Typically, we divide the parameters
into a small number of disjoint chunks which are updated
separately, each with its own lock (steps 2-3 become a loop
over chunks). This balances update call efficiency against
lock contention and can provide good performance. When
large intermediate updates occur, it can be beneficial to pull
the central network weights more frequently (e.g. during
sampling and/or before gradient computation). In the surrogate loss of PPO, we use the action probabilities experienced
during sampling.
5. Experiments
We used the Atari-2600 domain to study the scaling characteristics of highly parallelized RL, addressing the following:
1) How efficient is our synchronized sampling engine, and
what speeds can it achieve? 2) Can policy gradient and Qlearning algorithms be adapted to learn using many parallel
simulator instances without diminishing learning performance? 3) Can large-batch training and/or asynchronous
methods speed up optimization without worsening sample
complexity?
In all learning experiments, we maintained the original training intensity–meaning average number of training uses of
each sampled data point. For A3C, PPO, and DQN+variants,
the reference training intensities are 1, 4, and 8, respectively.
All learning curves shown here are averages over at least
two random seeds. For policy gradient methods, we tracked
online scores, averaging over the most recent 100 completed trajectories. For DQN and variants, we paused every
1-million steps to evaluate for up to 125,000 steps, with
maximum path length of 27,000 steps, as is standard. The
appendices contain learning curves beyond those we highlight here.
5.1. Atari Frame Processing
Our frame pre-processing closely resembles that originally
described in (Mnih et al., 2015). The sole difference is that
we abandon the square frame dimensions in favor of simply
downsizing by a factor of 2, which provides crisp images
at minimal computational cost. Before downsizing, we
crop two rows, making the final image size 104 × 80. This
simplifies selection of convolution size, stride, and padding.
Otherwise, we kept all standard settings. For Q-learning
experiments, we used the 3-convolutional-layer network as
in (Mnih et al., 2015) (DQN-Net) or its algorithm-specific
variants, and for policy gradients the 2-convolutional-layer
feed-forward network of (Mnih et al., 2016) (A3C-Net).
The second (and third) convolution layers have padding 1,
so the convolution output is always 25 × 19.
5.2. Sampling
A series of sampling-only measurements demonstrated that
despite the potential for stragglers, the synchronized sampling scheme can achieve good hardware utilization.
First, we studied the capacity of a single GPU at serving
inferences for multiple environments. Figure 2 shows measurements running a trained A3C-Net policy on a P100
GPU while playing the game B REAKOUT. Aggregate sampling speed, normalized by CPU core count, is plotted as
a function of the number of (sequential) Atari simulators
running on each core.2 The minimum was 2 simulators per
core by the alternating scheme. Different curves represent
different numbers of CPU cores running simulations. For
reference, we include the sampling speed of a single core
2
Intel Turboboost was disabled for this test only, keeping the
clock speed of every core at 2.2 GHz.
Accelerated Methods for Deep Reinforcement Learning
Table 1. Sampling speeds on the DGX-1 with A3C-Net, by total
simulator count, in thousands of samples per second.
# S IMS ( PER C ORE )
S YNC
A SYNC
64 (2)
128 (4)
256 (8)
512 (16)
29.6
33.0
35.7
35.8
31.9
34.7
36.7
38.4
running without inference–the dashed line for a single process, and the dotted line one process on each of the two
Hyperthreads. Running with inferences and with a single
core, the sampling speed increased with simulator count
until the inference time was completely hidden. Synchronization losses appeared for higher core count. But at as
little as 8 environments per core, the GPU supported even
16 CPU cores running at roughly 80% of the inference-free
speed.
Figure 2. Sampling speed vs number of simulators per core, playing B REAKOUT using 1 GPU. Curves are by number of cores
performing simulations. Running multiple simulators per core
mitigates synchronization losses and hides NN inference time,
resulting in high throughput.
Next, we measured the sampling-only speed of the same
A3C-Net playing B REAKOUT parallelized across an entire
NVIDIA DGX-1. Table 1 shows results for various total
simulator counts. In the synchronized setting, a barrier was
placed across all GPU processes every five steps (mimicking
the optimization in A2C). Otherwise, each GPU and its
associated cores ran independently. At simulator counts of
256 (8 per core) and above, our sampling engine achieved
greater than 35,000 samples per second. This equates to
500 million emulator frames per hour, possibly the fastest
reported Atari sampling rate using a single computer.
5.3. Learning with Many Simulator Instances
To leverage the high throughput of parallel sampling, we
attempted to adapt existing deep RL algorithms to learn with
many simulator instances. The following results show that
indeed all algorithms we considered can be adapted without
losing performance, and with only minor changes. We
experimented with different techniques for each algorithm,
which we describe and evaluate here. Interestingly, scaling
affects synchronous and asynchronous learning somewhat
differently, looking across games.
A2C: We experimented with A2C by letting the optimization batch size grow with the number of simulators while
keeping the sampling horizon fixed (at 5 steps as in A3C).
Correspondingly fewer parameter update steps were made
per sample gathered. Unlike in (Clemente et al., 2017), we
empirically determined that increasing the learning rate with
the square root of the batch size worked best. A gradual
decay in efficiency remained. The top panel of Figure 3
shows learning curves vs total sample count for a test set of
games, with simulator count ranging from 16 to 512 (batch
size 80 to 2,560). We used RMSProp with a base learning
rate of 7 × 10−4 (scaled up to 3 × 10−3 here).
A3C: An asynchronous adaptation we experimented with
used a 16-environment A2C agent as the base samplerlearner unit. Figure 3 shows learning curves vs aggregate
sample count for numbers of learners ranging from 1 to 32,3
corresponding to 16 to 512 total simulators. We found no
hyperparameter adjustments to be necessary. The resulting
learning curves were nearly indistinguishable in most cases,
although some did degrade at the largest scales.
PPO: The large batch size already used to benchmark PPO
(8-simulator x 256-horizon = 2,048) provided a different
route to using many simulators. We adapted PPO by decreasing the sampling horizon such that the total batch size
remained fixed. Figure 3 shows learning curves vs sample
count for simulator counts ranging from 8 to 512, with corresponding sampling horizons ranging from 256 down to
4 steps. We changed no optimization settings. Learning
improved or worsened slightly on some test games.
APPO: We also experimented with an asynchronous version of PPO. We used an 8-simulator PPO agent as the base
learner unit, using the same sampling horizon and update
sequence as original PPO. The bottom panel in Figure 3
shows learning curves from a study of 8 learners running
on 8 GPUs, with varying communication frequency. Standard PPO uses 4 gradient updates per epoch, and 4 epochs
per optimization; we experimented with 1-4 gradient updates between synchronizations (update rule provided in
supplementary material). Relative to PPO, we introduced
gradient-norm-clipping, reduced the learning rate by a factor
of four, and removed the learning rate schedule, all of which
benefited learning. We found it helpful to periodically pull
the central policy parameters during sampling, and did this
with a horizon of 64 steps in all cases. In several games, the
learning remained fairly consistent; it may be possible to
reduce communication in asynchronous settings.
3
Learner counts in excess of 8 were run with multiple separate
learners sharing GPUs.
Accelerated Methods for Deep Reinforcement Learning
Figure 3. Scaling investigations for policy gradient algorithms: game scores vs aggregate sample count. Top) A2C with various batch
sizes (proportional to environment count), Upper) A3C with various numbers of 16-environment learner processes, Lower) PPO with
varied number of simulators, Bottom) Asynchronous PPO, 8 learners with varied communication period. In most but not all cases, the
scaled/adapted versions match the baseline performance.
DQN + Variants: To adapt DQN and variants to run with
many simulator instances, we performed learning experiments using an experience replay buffer organized by simulator. The total buffer size remained at 1 million transitions,
so a correspondingly shorter history was held for each simulator. We observed learning performance to be largely
independent of simulator count up to over 200, provided
the number of update steps per optimization cycle is not too
high (large batch size ameliorates this).
Start State Decorrelation: Learning failed very early in
some many-simulator experiments. We attributed this to
correlation in starting game states, which resulted in large
but poorly informed learning signals. To correct this, all
our experiments start by stepping every simulator through
a random number of uniform-random actions. Once this
measure is taken, we found learning rate warmup (Goyal
et al., 2017) to have no effect. On commencing sampling
for training, game resets proceed as usual.
5.4. Q-Value Learning with Large Training Batches
We performed experiments to explore the effects of training
batch size in Q-value learning. They confirm that sampleefficient, large-batch training is possible with DQN. Further,
we begin to shed light on the optimization effects of batch
size.
5.4.1. A LGORITHM P ERFORMANCE
DQN: When growing the simulator count and batch size
in DQN, we maintained training intensity by adjusting the
sampling horizon and the number of update steps per optimization phase. We experimented with batch sizes ranging
from the standard 32 up to 2,048, and found consistent
learning performance up to 512. Beyond this batch size,
it became increasingly difficult to find a single learning
rate which performed well in all tested games, and some
maximum scores suffered. Figure 4 shows DQN learning
performance for the batch sizes 32, 512, and 1024, using
learning rates 2.5, 7.5, 15 × 10−4 , respectively. Other hyperparameters were as in original DQN (Mnih et al., 2015).
In several games, a larger batch size improved learning. We
also found asynchronous DQN to learn well using up to 4
GPU learners, each using batch size 512, with no further
adjustments.
Categorical DQN: We found Categorical DQN to scale further than DQN. The lower panel of Figure 4 shows learning
curves for batch sizes up to 2,048, with no diminution in
maximum scores. This is possibly due to richer content
of the gradient signal. We employed the published setting
for epsilon in the Adam optimizer: 0.01/L, where L is the
batch size. Notably, learning was delayed for the largest
batch sizes in the game S EAQUEST, but greater maximum
Accelerated Methods for Deep Reinforcement Learning
scores were eventually reached. In contrast to previous algorithms, we found a single learning rate to apply broadly,
using 4.2 × 10−4 at all batch sizes of 256 and above. We
return to study this surprising result shortly.
-Rainbow: Despite its use of distributional learning, we
found -Rainbow to lose performance above batch size 512
in some games. Scores for batch size 512 appear in Figure 12 in an appendix. They roughly match (in some cases
exceeding) those reported in the literature for batch size 32
(Hessel et al., 2017). We used the published hyperparameters without scaling the learning rate. We scaled epsilon in
the Adam update rule as 0.005/L, with L the batch size.
5.4.2. E FFECTS OF BATCH S IZE ON O PTIMIZATION
Some possible factors limiting training batch sizes include:
1) reduced exploration, since fewer different networks are
exercised in the environment, and 2) difficulties in numerical optimization of network weights. We conducted experiments to begin to identify these factors.
Secondary-Learner Experiment: We configured a secondary DQN learner to train using only the replay buffer
of a normal DQN agent. The secondary learner was initialized with the same parameter values as the primary,
“sampler-learner”, and the two networks trained simultaneously, at the same rate of data consumption. Each sampled
its own training batches. In the game of B REAKOUT, 64and 2048-sampler-learners achieved the same score, but
the 2048-learner required more samples, despite using the
fastest stable learning rate (the number refers to training
batch size). When training a 64-secondary-learner using a
2048-sampler-learner, the secondary learner’s score tracked
that of the primary. In the reverse scenario, however, the
2048-secondary-learner failed to learn. We posit this was
due to the slower optimization of the decreased number of
parameter updates–it was unable to track the rapid changes
to the Q-value estimates near initialization and became too
off-policy to learn. In the same test using two 256-learners,
their scores matched. Had the 2048-secondary-learner outpaced the 2048-sampler-learner, it would have suggested
exploration to be a more important factor than optimization.
See the supplementary materials for figures.
Update Rule: We conducted another experiment to isolate the effect of update rule on optimization. We found
the Adam (Kingma & Ba, 2014) formula to be superior to
RMSProp (Tieleman & Hinton) in providing large-batch
learners with capability to traverse parameter-space during
learning. When comparing learners achieving the same
learning curves, those using smaller batch sizes (and hence
performing more update steps) tended to have larger parameter vector-norms at all points in training. Unlike RMSProp,
the Adam rule resulted in a fairly tight spread in parameter
norms between batch sizes without changing the learning
Table 2. Hours to complete 50 million steps in policy gradient
learning on the DGX-1, by GPU count. A2C/A3C used 16 environments per GPU, PPO/APPO used 8.
A LGORITHM
1 (5)
A2C
A3C
PPO
APPO
3.8
–
4.4
–
# GPU (# CPU)
2 (10) 4 (20) 8 (40)
2.2
2.4
2.6
2.8
1.2
1.3
1.5
1.5
0.59
0.65
1.1
0.71
rate. This explains the lack of need to scale the learning rate
in Categorical DQN and -Rainbow, and indicates that the
update rule plays an important role in scaling. Further details, including differences in trends between convolutional
and fully connected layers, appear in an appendix.
5.5. Learning Speed
Finally, we demonstrated the learning speeds our framework
obtains when running an entire DGX-1 to learn a single
game. Figure 5 shows results for well-performing configurations of the policy gradient methods A2C, A3C, PPO, and
APPO. Several games exhibit a steep initial learning phase;
all algorithms here completed that phase in under 10 minutes
in all such games shown. Notably, PPO mastered Pong in 4
minutes. A2C with 256 environments reached a processing
speed in excess of 25,000 samples per second, equating to
over 90 million steps per hour (360 million frames). Scaling
measurements are listed in Table 2, showing in excess of 6x
speedup using 8 GPUs relative to 1.
We ran synchronous versions of DQN and its variants, with
training times shown in Table 3. Using 1 GPU and 5 CPU
cores, DQN and -Rainbow completed 50 million steps (200
million frames) in 8 and 14 hours, respectively–a significant
gain over the reference times of 10 days. We achieved comparable learning speeds to those in (Horgan et al., 2018),
which used 1 GPU and 376 CPU cores (see e.g. Figure 2
therein for 10-hour learning curves). Using multiple GPUs
and more cores sped up our implementations. By virtue of
a larger batch size, Categorical-DQN scaled best and completed training in under 2 hours using the entire DGX-1, a
speedup of over 6x relative to 1 GPU. DQN and -Rainbow,
however, experienced diminishing returns beyond 2 GPUs.
We were unable to find asynchronous configurations that
further boosted learning speed without curbing performance
in some games (we only tested fully-communicating algorithms). Opportunities may exist to improve on our scaling.
6. Conclusions and Discussion
We have introduced a unified framework for parallelizing
deep RL that uses hardware accelerators to achieve fast
learning. The framework is applicable to a range of al-
Accelerated Methods for Deep Reinforcement Learning
Figure 4. Scaling investigations for DQN (top) and Categorical-DQN (bottom): game scores vs sample count. Both learn well using
training batch sizes as large as 512; Categorical-DQN succeeds using up to 2,048.
Figure 5. Policy gradient algorithms using an entire DGX-1 for a single learning run: game scores vs aggregate sample count and vs time,
in minutes. Asynchronous and synchronous versions of A2C and PPO learn successful game strategies in under 10 minutes.
Table 3. Hours to complete 50 million steps (200 million frames)
in Q-learning on the DGX-1, by GPU count (synchronous).
A LGORITHM -B.S.
DQN-512
-R AINBOW-512
C AT-DQN-2048
*
1 (5)
8.3
14.1
10.7
# GPU (# CPU)
2 (10)
4 (20)
4.8
8.6
6.0
3.1/3.9
6.6
2.8
*
8 (40)
2.6
6.4
1.8
Asynchronous
gorithms, including policy-gradient and Q-value learning
methods. Our experiments show that several leading algorithms can learn a variety of Atari games in highly parallel
fashion, without loss of sample complexity and at unprecedented wall-clock times. This result indicates a promising
direction to significantly boost experiment scale. We will
release the code-base.
We note several directions for extension of this framework. First is to apply it to domains other than Atari, especially ones involving perception. Second, our framework is
likely to scale favorably to more sophisticated neural network agents, due to GPU acceleration of both inference
and training. Moreover, as network complexity increases,
scaling could become easier, as GPUs may run efficiently
with smaller batch sizes, although communication overhead
would worsen. Reduced-precision arithmetic could hasten
learning–a topic yet to be explored in deep RL due to use
of CPU-based inference. Larger RL problems may use the
current, single-node implementation as a building block for
distributed algorithms.
Questions remain as to the extent of parallelization possible
in deep RL. We have not conclusively identified the limiting
factor to scaling, nor if it is the same in every game and
algorithm. Although we have seen optimization effects in
large-batch learning, we do not know their full nature, and
other factors remain possible. Limits to asynchronous scaling remain unexplored; we did not definitively determine the
best configurations of these algorithms, but only presented
some successful versions. Better understanding may enable
further gains in scaling, which is a promising direction for
the advancement of deep RL.
Accelerated Methods for Deep Reinforcement Learning
Acknowledgements
Adam Stooke gratefully acknowledges the support of the
Fannie & John Hertz Foundation. The DGX-1 used for
this research was donated by the NVIDIA Corporation. We
thank Frédéric Bastien and the Theano development team
(Theano Development Team, 2016) for their framework
and helpful discussions during development of GPU-related
methods. Thanks to Rocky Duan et al for the rllab codebase (Duan et al., 2016) out of which this work evolved.
References
Babaeizadeh, Mohammad, Frosio, Iuri, Tyree, Stephen,
Clemons, Jason, and Kautz, Jan. GA3C: gpu-based A3C
for deep reinforcement learning. arXiv preprint arXiv:
1611.06256, 2016.
Bellemare, Marc G, Naddaf, Yavar, Veness, Joel, and Bowling, Michael. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial
Intelligence Res., 47:253–279, 2013.
Bellemare, Marc G, Dabney, Will, and Munos, Rémi. A distributional perspective on reinforcement learning. arXiv
preprint arXiv: 1707.06887, 2017.
Clemente,
Alfredo
V.,
Martı́nez,
Humberto
Nicolás Castejón, and Chandra, Arjun.
Efficient parallel methods for deep reinforcement
learning.
CoRR, abs/1705.04862, 2017.
URL
http://arxiv.org/abs/1705.04862.
Duan, Yan, Chen, Xi, Houthooft, Rein, Schulman, John, and
Abbeel, Pieter. Benchmarking deep reinforcement learning for continuous control. CoRR, abs/1604.06778, 2016.
URL http://arxiv.org/abs/1604.06778.
Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih,
V., Ward, T., Doron, Y., Firoiu, V., Harley, T., Dunning,
I., Legg, S., and Kavukcuoglu, K. IMPALA: Scalable
Distributed Deep-RL with Importance Weighted ActorLearner Architectures. ArXiv e-prints, February 2018.
Fortunato, Meire, Azar, Mohammad Gheshlaghi, Piot, Bilal, Menick, Jacob, Hessel, Matteo, Osband, Ian, Graves,
Alex, Mnih, Volodymyr, Munos, Remi, Hassabis, Demis,
Pietquin, Olivier, Blundell, Charles, and Legg, Shane.
Noisy networks for exploration. In International Conference on Learning Representations, 2018.
Goyal, Priya, Dollr, Piotr, Girshick, Ross, Noordhuis, Pieter,
Wesolowski, Lukasz, Kyrola, Aapo, Tulloch, Andrew, Jia,
Yangqing, and He, Kaiming. Accurate, large minibatch
sgd: Training imagenet in 1 hour, 2017.
Hessel, Matteo, Modayil, Joseph, van Hasselt, Hado, Schaul,
Tom, Ostrovski, Georg, Dabney, Will, Horgan, Dan, Piot,
Bilal, Azar, Mohammad, and Silver, David. Rainbow:
Combining improvements in deep reinforcement learning.
arXiv preprint arXiv: 1710.02298, 2017.
Horgan, D., Quan, J., Budden, D., Barth-Maron, G., Hessel,
M., van Hasselt, H., and Silver, D. Distributed Prioritized
Experience Replay. ArXiv e-prints, March 2018.
Kingma, Diederik P. and Ba, Jimmy. Adam: A method
for stochastic optimization. CoRR, abs/1412.6980, 2014.
URL http://arxiv.org/abs/1412.6980.
Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David,
Rusu, Andrei A, Veness, Joel, Bellemare, Marc G, Graves,
Alex, Riedmiller, Martin, Fidjeland, Andreas K, Ostrovski, Georg, Petersen, Stig, Beattie, Charles, Sadik,
Amir, Antonoglou, Ioannis, King, Helen, Kumaran, Dharshan, Wierstra, Daan, Legg, Shane, and Hassabis, Demis.
Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
Mnih, Volodymyr, Badia, Adrià Puigdomènech, Mirza,
Mehdi, Graves, Alex, Lillicrap, Timothy P, Harley, Tim,
Silver, David, and Kavukcuoglu, Koray. Asynchronous
methods for deep reinforcement learning. In International
Conference in Machine Learning, 2016.
Nair, Arun, Srinivasan, Praveen, Blackwell, Sam, Alcicek, Cagdas, Fearon, Rory, De Maria, Alessandro, Panneershelvam, Vedavyas, Suleyman, Mustafa, Beattie,
Charles, Petersen, Stig, Legg, Shane, Mnih, Volodymyr,
Kavukcuoglu, Koray, and Silver, David. Massively parallel methods for deep reinforcement learning. arXiv
preprint arXiv: 1507.04296, 2015.
Peng, Jing and Williams, Ronald J. Incremental multi-step
q-learning. In Machine Learning Proceedings 1994, pp.
226–232. Elsevier, 1994.
Schaul, Tom, Quan, John, Antonoglou, Ioannis, and Silver,
David. Prioritized experience replay. In International
Conference on Learning Representations, 2016.
Schulman, John, Wolski, Filip, Dhariwal, Prafulla, Radford,
Alec, and Klimov, Oleg. Proximal policy optimization
algorithms. arXiv preprint arXiv: 1707.06347, 2017.
Theano Development Team. Theano: A Python framework
for fast computation of mathematical expressions. arXiv
e-prints, abs/1605.02688, May 2016. URL http://
arxiv.org/abs/1605.02688.
Tieleman, T. and Hinton, G. RMSprop Gradient Optimization.
URL http://www.cs.toronto.
edu/˜{}tijmen/csc321/slides/lecture_
slides_lec6.pdf.
Accelerated Methods for Deep Reinforcement Learning
Van Hasselt, Hado, Guez, Arthur, and Silver, David. Deep
reinforcement learning with double q-learning. In AAAI,
volume 16, pp. 2094–2100, 2016.
Wang, Ziyu, Schaul, Tom, Hessel, Matteo, van Hasselt,
Hado, Lanctot, Marc, and de Freitas, Nando. Dueling
network architectures for deep reinforcement learning. In
International Conference on Machine Learning, 2016.
Accelerated Methods for Deep Reinforcement Learning
Supplementary Materials
A. Update Rule for Multi-Step Asynchronous Adam
Our asynchronous PPO experiments used the update rule described here, which permits multiple local gradient steps per
synchronization with the central parameters. The usual Adam update rule (Kingma & Ba, 2014) is the following. It has
fixed hyperparameters r, β1 , β2 , and ; g stands for the gradient; and all other values except the network parameters θ are
initialized at 0:
t←t+1
p
1 − β2t
a←r
1 − β1t
m ← β1 m + (1 − β1 )g
v ← β2 v + (1 − β2 )g 2
am
s← √
v
θ ←θ−s .
We kept these rules for making local updates and introduced the additional, local accumulation variables, also zero-initialized:
ag ← β1 ag + g
ag2 ← β2 ag2 + g 2
as ← as + s
.
When applying an update to the central parameters, denoted with a tilde, we used the following assignments:
θ, θ̃ ← θ̃ − as
m, m̃ ← β1n m̃ + (1 − β1 )ag
v, ṽ ← β2n ṽ + (1 − β2 )ag2
ag , ag2 , as ← 0
where n is the number of local gradient steps taken between synchronizations. This rule reduces to the usual Adam update
rule in the case of a single learner thread.
Accelerated Methods for Deep Reinforcement Learning
B. Figures for Secondary-Learner Experiment (DQN)
(a)
(b)
Figure 6. Learning the game B REAKOUT with a secondary-learner using only the replay buffer of the normal, sampler-learner, both using
DQN. a) The 64-batch-size secondary-learner kept pace with its 2048-batch-size sampler-learner, but b) the 2048-batch-size secondary
learner failed to track its 64-batch-size sampler-learner or even learn at all. (Curves averaged over two random trials.)
(a)
(b)
Figure 7. Neural network parameter vector-norms (l-2) during training. In both cases, the large-batch learner lagged behind the small
batch learner. In b) the parameters of large-batch secondary-learner continued to grow while its game score remained nil.
Figure 8. Learning the game B REAKOUT, where a secondary-learner using the same batch-size as the sampler-learner tracked (albeit
imperfectly) the game score, learning successfully. (Curves averaged over two random trials.)
Accelerated Methods for Deep Reinforcement Learning
C. Observations on Update Rules and Batch Size Scaling
We present observations of the effects of scaling the training batch size on neural net optimization under two different
parameter update rules: Adam and RMSProp (RMSProp without momentum and only direct accumulation of the squared
gradients, see e.g. https://github.com/Lasagne/Lasagne/blob/master/lasagne/updates.py). We
trained agents on the game Q*B ERT with learning rates adjusted to yield very similar performance curves for all settings,
and we tracked the L-2 vector-norms of several quantities during learning. These included the gradients, parameter update
steps, and the parameter values themselves. As in all DQN experiments in this paper, the training intensity was fixed at 8, so
that the number of parameter update steps during learning scaled inversely with the batch size. Two random seeds were run
for each setting.
Although the game scores roughly matched throughout training, the exact solutions found at any point did not, as evidenced
by the differing parameter norms. No regularization was used. The following paragraphs follow the panels in Figure 9,
where curves are labeled by batch-size and learning rate. When viewing the network as a whole (i.e. norms of all weights
and biases as a single vector), trends reflected those seen in FC-0, where most of the weights are.
i) Learning Curves: We controlled for game score, adjusting the learning rate as needed. For the case of RMSProp with
batch-size 64, we included a learning rate that was slightly too low (1 × 10−4 ), yielding slow learning and lower final score,
and a learning rate that was slightly too high (5 × 10−4 ), yielding lower final score due to instability–these are the dashed
lines in all panels.
ii) Fully-Connected-0 Weights-Norm: The Adam optimizer yielded fairly tight grouping despite using the same learning
rate for all settings. The RMSProp learner, on the other hand, needed to scale the learning rate by 20x between batch sizes
64 and 1,024, which then produced very similar norms. At batch-size 64, slow / unstable learning was characterized by
small / large norms, respectively. The large norm of the batch-size 256 runs suggests this learning rate was likely near the
upper limit of stability.
iii) Fully-Connected-0 Gradients-Norm: Under both update rules, large batch sizes always produced smaller gradient
vectors–reduced variance led to reduced magnitudes. We also observed this pattern in policy gradient methods, when
looking at the total gradient norm. Here, the magnitude of the gradients depended inversely on the parameter norm; see the
RMSProp 64-batch-size curves. This effect was opposed and outweighed by the effect of batch size.
iv) Fully-Connected-0 Step-Norm: The Adam optimizer yielded significantly bigger step sizes for the bigger batch
learners, despite the smaller gradients. RMSProp required an adjusted learning rate to produce the same effect. Under
both update rules, the amount of step size increase did not fully compensate for the reduction in step count, indicating
that the larger batch learners followed straighter trajectories through parameter space. RMSProp led to significantly larger
steps overall, but despite this ended learning at smaller weights–its learning trajectories were apparently less direct, more
meandering.
v) Convolution-0 Weights-Norm: The Adam optimizer gave much greater spread in norms here than in the FC-0 layer; as
batch size increased, the learning emphasis shifted away from Conv-0. But in RMSProp the increased learning rate led the
first convolution layer to grow larger for larger batch sizes, placing more emphasis on this layer.
vi) Convolution-0 Gradients-Norm: The Adam update rule produced an intriguing cross-over in gradient norm; the large
batch learner actually started higher, bucking the trend seen in other cases. The pattern under RMSProp matched that for
FC-0.
vii) Convolution-0 Step-Norm: Unlike for FC-0, the step norm did not change significantly with batch size under Adam.
RMSProp yielded a similar pattern as in FC-0.
Overall, the Adam optimizer appeared to compensate for batch size in the FC-0 layer, but less so in the Conv-0 layer, leading
to de-emphasized learning in Conv-0 for large batches. The increased learning rate in RMSProp compensated for batch
size in the FC-0 layer and increased the emphasis on learning in Conv-0. This sort of pattern could have implications
for learning representations vs game strategies. Further study of these clear trends could yield insights into the causes of
learning degradation and possible solutions for large batch RL.
Accelerated Methods for Deep Reinforcement Learning
Figure 9. L-2 vector-norms of the parameters, gradients (average), and parameter steps (average) for the first convolution layer and the
first fully connected layer while learning to play Q*B ERT with Categorical DQN at various batch sizes: Adam vs RMSProp.
Accelerated Methods for Deep Reinforcement Learning
D. Additional Learning Curves
Figure 10. Learning curves for Advantage Actor-Critic: baseline (A2C-16env) and scaled configurations, including synchronous and
asynchronous. Only in ATLANTIS, G OPHER, and possibly K RULL does the baseline stand out above both scaled versions.
Accelerated Methods for Deep Reinforcement Learning
Figure 11. Learning curves for Proximal Policy Optimization: baseline (PPO-8env) and scaled configurations, including synchronous and
asynchronous. Only in A STEROIDS and B OWLING does the baseline stand out above both scaled versions.
Accelerated Methods for Deep Reinforcement Learning
Figure 12. Learning curves for scaled versions of DQN (synchronous only): DQN-512, Categorical-DQN-2048, and -Rainbow-512,
where the number refers to training batch size. The anomalously low scores for -Rainbow in B REAKOUT also appeared for smaller
batch sizes, but was remedied when setting the reward horizon to 1 or with asynchronous optimization (cause unknown; reward horizon 3
usually helped).
| 2cs.AI
|
Formal Concept Analysis for Knowledge
Discovery from Biological Data
arXiv:1506.00366v1 [cs.AI] 1 Jun 2015
Khalid Raza
Abstract Due to rapid advancement in high-throughput techniques, such as microarrays and next generation sequencing technologies, biological data are increasing
exponentially. The current challenge in computational biology and bioinformatics
research is how to analyze these huge raw biological data to extract biologically
meaningful knowledge. This chapter presents the applications of formal concept
analysis for the analysis and knowledge discovery from biological data, including
gene expression discretization, gene co-expression mining, gene expression clustering, finding genes in gene regulatory networks, enzyme/protein classifications,
binding site classifications, and so on. It also presents a list of FCA-based software
tools applied in biological domain and covers the challenges faced so far.
1 Introduction
After Human Genome Project, there is unprecedented growth in biological data. Due
to technological advancement in high throughput technologies, such as Microarray
and Next Generation Sequencing, it is possible to produce high quality biological
data with rapid speed. The biological data can be broadly classified as Genomics,
Transcriptomic and Proteomics. For example, gene expressions are transcriptomic
data that quantify the state of genes in a cell. When these gene expression data
are analyzed properly, it may reveal many hidden cellular processes and biological
knowledge. These knowledge discoveries from biological data may lead to better
understanding of disease mechanism and further it guide for better diagnosis and
therapy of the disease.
Formal Concept Analysis (FCA), introduced by R. Wille in early 1980s [1], is a
method based on lattice theory for the analysis of binary relational data. Since its
Khalid Raza
Department of Computer Science, Jamia Millia Islamia (Central University), New Delhi, India
e-mail: [email protected]
1
2
Khalid Raza
inception, FCA has been found to have potential applications in many areas including data mining, knowledge discovery and machine learning. Like other computational technique, FCA has also been applied in microarray analysis, gene expression
mining, gene expression clustering, finding genes in gene regulatory networks, enzyme/protein classifications, binding site classifications, and so on. In this chapter,
we will present the current status of FCA for the analysis and knowledge discovery
from biological data and also cover challenges faced so far.
2 Biological Databases
Due to availability of high-throughput techniques, biological database are being
generated exponentially and the modern biology has turned into a data-rich science.
Some of the important biological data are nucleotide and protein sequences, protein
3D structure produced by X-ray crystallography and NMR, metabolic pathways,
complete genomes and maps, gene expression and protein-protein interaction, and
so on.
Biological databases are broadly divided into sequences databases and structure
databases. Sequence data are applicable for both DNA and protein, but structural
databases are applicable for proteins only. Today, most of the biological databases
are freely available to the researchers. In general, biological databases can be classified as primary, secondary and composite databases. A primary databases stores
information of either sequence or structure. For example, Uni-PROT and PIR for
protein sequence, GenBank and DDBJ for Genome sequence and the Protein Databank for protein structure. Secondary database stores information which are derived
from the primary database source, such as conserved sequence information, active
site residues of the protein families arrived by multiple sequence alignment of a
set of related proteins, etc. The SCOP, CATCH, PROSITE are few examples of
secondary databases. A composite database is a collection of variety of different
primary database sources that avoid the need for searching into multiple database
sources. The National Centre for Biotechnology Information (NCBI) is the main
central host that links multiple database sources and makes these resources freely
available to us. For more information about biological databases, refer tutorials in
[2] and [3].
3 Microarray Analysis
3.1 Mining gene expression data
Due to rapid advancement in high throughput technology such as Microarray and
Next Generation Sequencing, transcriptomic data has been produced in unprece-
Formal Concept Analysis for Knowledge Discovery from Biological Data
3
dented way. But analysis and interpretation of these data remains a challenge for
the researchers due to complexity of the biological systems. The motivation and
biological background that need to be considered for gene expression mining are:
i) mostly single gene participates in many biological processes, i.e., it has several
functions, ii) a biological process implies a small subset of genes, iii) a biological
process of interest may be active in many, all or none situation for a given dataset,
and iv) differentially expressed genes over different samples are not frequent.
In transcriptomic, researchers routinely analyze expression level of genes in different situations such as in tumor samples versus normal samples. Formal Concept
Analysis (FCA) has been successfully applied in the field of transciptomics. Some
of the studies identified set of genes that are sharing same transcriptional behavior using FCA [4, 5, 6]. Due to availability of large gene expression datasets, it is
possible to apply data mining tools to identify patterns of interest in the gene expression data. One of the most widely used data mining technique is association rules
which can be applied for the analysis of gene expression data. Association rules
may uncover biologically relevant associations between genes, or between different
environmental conditions and gene expression. An association rules can be written
in the form S1 → S2 , where S1 and S2 are disjoint sets of data items. The set S2 is
likely to occur whenever the set S1 occurs. Here, the data items may include highly
expressed or repressed genes, or other relevant facts stating the cellular environment
of genes such as diagnosis of a disease samples. Association rules mining has been
applied for gene expression data mining by many researchers including [8].
In this section, we have discussed only the application of different variants of
FCA for gene expression data mining, especially extracting co-expressed groups of
genes sharing similar expression. Most of the methods for co-expressed genes mining are based on binary biclustering methods. Here, scaling of data is done using a
single threshold and one expression value. The expression values above this threshold are considered as over-expressed and represented by 1; otherwise it is considered as under-expressed and represented by 0. Once, the gene expression values are
discretized to binary table then strong relationships can be extracted having biologically meaningful information. Kaytoue-Uberall et al. (2008) [9] proposed intervalbased FCA to extract groups of co-expressed genes. Given a set of genes G, a set of
relationships S, and set of ordered intervals T, (g, (s,t)) ∈ I, where g ∈ G, s ∈ S,t ∈ T
and I is binary relation means gene expression value of gene g is interval of index
t for situation S. Hence, formal concept of the context (G, S × T, I) shows groups
of genes having G ∈ V in same interval. Although a priori determination of these
intervals are difficult.
Messai et al. (2008) [14] proposed interval-free FCA based method to cluster
gene expression values. However, this algorithm does not deal with large data set
and also no link to interordinal scaling was done. To overcome these problems,
Kaytoue-Uberall et al. (2009) [6] introduced two FCA-based methods for clustering
gene expression data. The first method is based on interordinal scaling and second
one is based on pattern structures that require adaptation of algorithm computed with
interval algebra. Between these two algorithms by Kaytoue-Uberall et al. (2009) [6],
second method has been proved to be more computationally efficient and provide
4
Khalid Raza
more readable results. These algorithms have been tested on microarray gene expression data of fungus Laccaria biocolor taken from Gene Expression Omnibus
databases (GSE9784) composed of 22,294 genes and five different conditions. For
the dimension reduction, cyber-T [7] tool was used that filter dataset and returned
10,225 genes.
DNA methylate affects the expression of genes and their regulation may cause
several cancer-specific diseases. It is observed in many investigations that hypomethylation of DNA have been associated with many cancers including breast
cancer. Amin et al., (2012) [10] applied FCA for mining the hypomethylated genes
among breast cancer tumors. They constructed formal concepts lattices with significant hypomethylated genes for every breast cancer subtypes. The constructed
lattice reflects the biological relationships among breast cancer tumor subtypes. The
proposed filter method has two stages: non-specific filter and specific filters. The
non-specific filtering step determines the hypomethlated CPGs by computing the
difference between the mean of methylation level for the corresponding adjacent
normal tissue. The second stage (specific filtering) receives the output of the first
stage as input and performs one-sample Kolmogorov Smirnov test to check the normality of each breast cancer subtypes. If the given dataset follows normal distribution then paired t-test is applied, otherwise Wilcoxon signed ranked is applied.
Once, the filtering of hypomethylated genes is done then FCA has been applied
to determine breast cancer subtypes. Here, Java-based FCA analysis software tool,
called ConExp [11] was used to generate the lattice diagram.
3.2 Clustering gene expression data
For grouping set of genes and/or grouping experimental conditions having similar gene expression pattern, clustering algorithms are the most popularly applied
method. Some of the most widely used clustering algorithms are hierarchical, kmeans, self organizing maps, fuzzy c-means, and so on [12]. However, FCA has
also been used for grouping genes, as an alternative approach to clustering. Choi
et al. (2008) [13] proposed FCA-based approach for grouping genes based on their
gene expression pattern. FCA builds a lattice from the gene expression data together with some additional biological information, where each vertex corresponds
to a subset of genes which are clubbed together based on their expression values and
some other functional information. The lattice structures of gene sets are assumed
to show biological relationship in the gene expression dataset. Here, similarities and
dissimilarities between different experiments are determined by corresponding lattices. This approach consists of three main steps: i) building a binary relation, ii)
construction of concept lattice, and iii) defining a distance measure and comparing
the lattices. In the first step, the objects are genes, their discretized gene expression
attributes and biological attributes. In the second step, for each experiment a binary
relationship is constructed using concept lattice algorithm. Finally, third step calculates distance and compares the lattices. The work of Choi et al. (2008) [13] is an
Formal Concept Analysis for Knowledge Discovery from Biological Data
5
attempt to apply FCA for gene clustering but the distance measure employed was
quite fundamental and it did not properly exploit the properties of the lattice structure. Hence, other possible distance measures such as spectral distance, maximal
common sublattice based distance, etc. can also be investigated [13]. In addition to
global lattice comparison, local structure (sublattice) can also be investigated that
may assist in identification of particular biological pathways.
Melo and collaborators [15] proposed an FCA-based approach combined with
association rule and visual analytics to find out overlapping groups of genes in gene
expression and analyzed it in an analytical tool called CUBIST. The workflow of
CUBIST involves querying a semantic databases and transforming the result into
formal context and then it is visualized as a concept lattice and associated charts.
The CUBIST tool address the challenges of gene expression analysis by filtering
and grouping large amount of datasets, interactive exploration of data and presents
various relevant statistics.
3.3 Clustering multi-experiment expression data
Due to availability of high-throughput techniques, presently we have large number of gene expression datasets. Combining datasets taken from multiple microarray experiments is research question. It has been proved and suggested by many
recent studies that the analysis and integration of multi-experiment datasets are expected to give more accurate, reliable and robust results. The reason is that integrated datasets would be based on large number of gene expression samples and
the effects of individual study-specific biases are reduced. For the consensus integration of multi-experiment expression data, FCA has been successfully applied
by Hristoskova and collaborators [16]. They proposed a generic consensus clustering which applied FCA for consolidation and analysis of clustering solutions taken
from multiple microarray experiments. Initially, the datasets are broken into multiple groups of related experiments based on some predefined criteria. In the next
step, a consensus clustering technique is deployed to each group that results on
clustering solution per group. Further, these solutions are pooled together and analyzed by FCA that enables extracting valuable insights from the data and generate
a gene partition over all the experiments. The FCA-enhanced consensus clustering
algorithm proposed by Hristoskova and collaborators [16] is depicted in Fig. 1. The
algorithm is divided into three steps: initialization, clustering and FCA-based analysis. In the initialization step, multi-experiment data are divided into r groups of
related datasets. Clustering step applies consensus clustering that generates r different solutions. FCA-based analysis step construct concept lattice that partitions the
genes into a set of disjoint clusters, as shown in Fig. 1. The advantages of FCAenhanced clustering approach proposed in [16] are as follows: i) Uses all data that
allow each group of related experiments to have a different set of genes, i.e., total set of studies genes is not limited to those present in all the datasets, ii) it can
be better tuned for each samples by identifying initial number of clusters for each
6
Khalid Raza
Fig. 1 Schematic representation of the FCA-enhanced consensus clustering algorithm [16]
group of related experiment, depending upon the number, composition and quality
of expression profiles, and iii) the problem with ties is avoided by applying FCA
to analyze together all partitioned results and find out the final clustering solution
representation as the entire experiment collection.
One another attempt for the application of FCA for knowledge discovery and
knowledge integration from gene expression data has been done by Benabderrahmane (2014) [17]. Benabderrahmane [17] introduced a symbolic data mining approach based on FCA involving bi-clustering of genes, for knowledge discovery
and knowledge integration. Firstly, datasets are represented as a formal context (objects attributes), where objects are genes and attributes are their expression profiles
plus additional information was used such as GO terms that they annotate, the list of
pathways they are involved and their genetic interactions. The algorithm has eight
steps, the outline of the algorithm is depicted in Fig. 2. This algorithm integrates different kinds of datasets such as genes having similar expression profiles and share
similar biological function (GO ontology), knowledge-base of pathways and interactors (KEGG, BioGrid, STRING, etc.)
3.4 Gene expression data comparison
Finding and understanding the similarities among various diseases is an import research problem in translational bioinformatics. Understanding disease similarities
Formal Concept Analysis for Knowledge Discovery from Biological Data
7
Fig. 2 An overview of the proposed framework proposed by Benabderrahmane (2014)[17]
may help us in refining disease classification, identifying common etiology of comorbidities in genetic studies and finding analogies between closely related diseases
and finally identify common treatments [18]. Bhavnani and collaborators [19] applied network analysis approach to find similarities among renal disease using gene
expression data.
In addition to many computational techniques, FCA has also been applied for
finding disease similarities. The work of Keller et al., (2012) [18] shows the application of FCA for identification of disease similarity. They identified formal concepts
using gene disease associations that indicate hidden relationship among diseases
having same set of associated genes, and gene that are associated with same set of
disease. The FCA approach has advantages over network analysis approach, such as
i) FCA allows representation of relationships among several diseases, ii) it provide
results in algebraic form allowing to consider relationship among concepts, and iii)
additional gene annotation can be added to refine concepts that assist for the identification of functional gene relationships within disease groups. FCA has been applied
on renal disease dataset that finds unexpected relationships among disease which
are promising but it suffers from few disadvantages. The difficulty with FCA is that
many of the formal concepts may not be useful because only a few formal concepts
indicate relationships.
8
Khalid Raza
3.5 Identifying genes of gene regulatory networks
Gene regulatory networks (GRNs) are the systematic biological networks describing
interaction among set of genes in the form of a graph, where node represents genes
and edges defines their regulatory interactions. Understanding the GRNs helps in
understanding interactions among genes, biological and environmental effects and
to identify the target genes for drug against the diseases. GRNs have been proved to
be a very useful tool used to describe and explain complex dependencies between
key developmental transcription factors (TFs), their target genes and regulators [20,
21]. For the better understanding of a gene regulatory network (GRN), it is necessary
to know set of genes belonging to it. Identification of these set of genes correctly
is a challenging task, even for a small subnetworks. In fact, only few genes of a
GRN are known and rest of the genes are guessed based on experience or informed
speculation [22]. Hence, it is better to rely on experimental data to support these
guesses.
Gebert and collaborators [22] presented a new FCA based method to detect
unknown members of GRN using time-series gene expression data. Suppose that
G = {g1 , g2 , , gn } is the set of all genes in an organism and S ⊆ G is set of seed
genes. The goal is to find subset S0 ⊆ G\S of genes which interact strongly with
GRN defined by set S. Let R ⊆ G × G be a relation having interactions and M an
n × l matrix that consists of time-series gene expression profiles having length l.
If pair (gi , g j ) ∈ R then it is known that gi and g j interaction to each other. The
FCA-based approach proposed by Gebert et al., (2008) [22] has three main steps
described as follows. First step is preprocessing step that uses the relation R to get
an initial list of interesting genes. If interaction data are not available, this step is
skipped and entire gene set G is taken as the initial list of genes. In the second step,
concept lattice is constructed using gene expression data that reduces the number
of genes on the initial list. The last step computes probabilities for the correlation
coefficient between genes that result from the second step and genes of S in order to
get list of significant interactions.
4 Classification and prediction of enzymes, ligand and
domain-domain identification
The classification and study of relations in FCA is focused on the basis of the objects and various types of related attributes (binary, nominal, ordinal etc.), therefore
it is quiet helpful for computational scientist working on Biological data, who may
wish to skip the inside details. With several advantages, including strong mathematical basis, FCA serves in several applications to explore biological data, enzyme
classifications, identification of important protein domains (including protein binding sites) and related drug molecules. FCA is also reported useful in the integration
of Biological activity with chemical spaces. This list is not exhaustive; FCA has
Formal Concept Analysis for Knowledge Discovery from Biological Data
9
also been used to understand the structural classification of glasses [23] and several
other studies. In this section, we discussed some important application of FCA for
the classification of enyzmes, binding site identification and discovery of ligand as
drug molecules and so on.
4.1 Enzyme classification
Enzymes are proteins which catalyses biological reactions and they are named and
classified according to the reaction they catalyse. For example, hydrolases are those
types of enzymes which are involved in the reactions by addition or deletion of water
molecules. Though the sequences of most of the enzymes are available in numerous
biological databases, it is tedious task to predict the function of the enzymes from
their respective sequences due to varied activity from small sequence combinations.
Considering that the new enzyme family may emerge, an effort was made for enzyme classifications using FCA which classifies the enzymes that does not belong
to known family [24]. They comment: it is easier to predict the super-families of
the proteins as compared to the families of the proteins. In this study, the labelled
and unlabelled enzyme sequences were objects whereas attributes represent the enzyme blocks. Enzyme blocks are formed by sequential arrangements of the amino
acids, which correspond to specific functions like catalytic site, lining residues of
important pockets or binding sites. In this method of classification, more than half
unlabelled sequences were found to be correctly classified.
Another attempt for the classification of protein using FCA has been done by
Han and collaborators [25]. They proposed FCA-based approach for protein classification that uses protein domain and Gene-Ontology annotation information. Protein domains represent the evolutionary information forming a protein, while GeneOntology describes other properties of proteins that includes structure of protein,
molecular interactions, etc. Han and collaborators [25] applied tripartite lattice for
interpenetrations among protein, domain and GO terms. With the help of tripartite
lattice, they classified protein from domain composite and their corresponding GO
term description. They extracted concrete information using tripartite lattice in the
corresponding domain that co-occur in proteins because they are more likely to exhibit common functions, as annotated in GO terms.
4.2 Binding site identification
Protein binding sites (PBS) and ligand binding sites identification are vital to
protein- protein and protein-ligand interactions, respectively. This eventually helps
the medical science in identification of better drug or therapy for several important
diseases. There are several ways to identify the binding sites. Most commonly, the
10
Khalid Raza
protein docking protocol helps in identifying the binding site by forming complex
with one protein to other protein or a ligand (which is a drug in most of the cases).
Bresso et al., (2012) [26] in their report highlight: Majority of the reported methods utilising the structure based prediction methods for protein-protein interactions
consider the attributes, which are physico-chemical properties like hydrophobicity,
residue constituents but lack the representation of properties (e.g. accessible surface of a particular residue) of binding components or spatial relation between two
components (residues). Considering these limitations and knowing the flexibility of
FCA, Bresso et al., (2012) [26] utilised available protein 3D structures for characterizing PBS. In this concept, Inductive Logic Programming (ILP) was linked with
FCA, which enabled identifications and discovery of distinct binding pockets of
protein-protein interactions.
4.3 Discovering Ligand from database as a drug molecule
Using FCA, several attempts have been made to identify suitable ligands from number of chemical database like IUPHAR, ZINC and many more. The reports suggest
that FCA helps in the identification of drug molecules. Drug molecules can be either agonist (activators) or antagonist (inhibitors). For a given protein, these drug
molecules, would likely act as agonist or antagonist. The one which do not binds
and do not show the changes, are not considered as drug molecules. In addition
to the ADMETox properties, the chemical molecules, which follow the Lipinski’s
Rule, are considered as suitable drugs.
Actually, when we talk about drugs, a chemical compound has number of physical properties: Hydrogen bond donors, acceptors; rotatable bonds; topological surface area; molecular weight; XlogP and chemical properties: absorption; digestion;
metabolism; excretion; toxicity. Using these properties as attributes for the object
ligand which could be possibly a drug molecule, one can identify and differentiate them from a bulk of chemical molecules in the database using FCA. To take
an example, similar attempt was made by Sugiyama et al., in 2012 [27] . They
considered the physical features discussed above, including the number of Lipinskis rule broken to set as attribute in order to identify the ligands from IUPHAR
database. They designed an algorithm, LIFT (LIgand Finding via Formal ConcepT
Analysis) for semi-supervised multi-labelled classification from mixed type data.
Results of the algorithm were effective and proved to be efficient system of classification to identify the ligands from the training data. Fragment Formal Concept
Analysis (FragFCA) introduced by Lounkine et al., (2008) [29] has the ability to
identify the selective hits in high-throughput screening data sets. In the concept design of FragFCA, combinations of molecular fragments are the ’objects’ and their
’attributes’ includes the compound activity and potency information.
The effectiveness of better drug identification can be improved, when the attributes classifying the ligands could be slightly updated, so as to filter non-peptide
molecules from the bulk of drug molecules. It has been identified that the peptide
Formal Concept Analysis for Knowledge Discovery from Biological Data
11
molecules have limited in vivo efficacy due to pharmacological constraints: solubility, stability and selectivity. Hence, for reliable and safer drug therapy, discovery
and optimisation of non-peptide inhibitors/drugs is necessary [28]. Moreover, in a
recent in silico identifications of the drug molecules for Cathepsin L (SmCL1) of the
organism, Schistosoma mansoni responsible for the disease schistosomiasis, it was
revealed that the non-peptide molecules could be better drug molecules as compared
to peptide drugs molecules [30]. The list of popularly used software tool based on
FCA is shown in Table 1.
So, to conclude, FCA can set an excellent framework to deal with variety of
problems. Before application of the concept on to the biological data minor optimisations and through understanding of the domain is the need in current study for
better research.
Table 1 List of FCA based software tools applied in biological domain
S.No.Tool Name
1.
ConExp
Descriptions
Java-based FCA analysis software tool
used to generate the lattice diagram.
2. FcaStone
Tool for format conversion and
command-line lattice generation.
3. Contextual Role FCA tool that work with Eclipse modelEditor
ing tool.
4. FcaBedrock
A tool for creating context files for Formal Concept Analysis. It can convert
existing data sets in flat-file CSV or
3-column CSV, to Burmeister (.cxt) or
FIMI (.dat) context files.
5. Lattice Miner
FCA tool for the construction, manipulation and visualization of concept lattices.
6. CUBIST
Gene expression analysis tool that combines FCA with association rule and visual analytics. It provides filtering and
grouping large data sets, its interactive
exploration and provides various relevant statistics.
7. Galicia
Galois lattice integrative constructor
8. OpenFCA
This project comprises of a set of tools
for performing FCA activities, including
creation of context, visualization of lattice and attribute exploration
9. LIFT
LIgand Finding via Formal ConcepT
Analysis (LIFT) for semi-supervised
multi-labelled classification from mixed
type data.
10. FragFCA
Fragment Formal Concept Analysis
(FragFCA) identifies the selective hits in
high-throughput screening data sets.
References
Serhiy (2000) [11]
Priss (2008) [31]
Mhle & Wende
(2010) [32]
Andrews & Orphanides
(2010)
[35]
Lahcen & Kwuida
(2010) ) [34]
Melo et al., (2013)
[15]
Galicia [33]
Borza et al., (2010)
[36]
Sugiyama et
(2012) [27]
al.,
Lounkine et
(2008) [29]
al.,
12
Khalid Raza
5 Conclusions and Discussions
Biological data are growing with unprecedented rate. High throughput technologies
fuelled in the production of high quality biological data. These data when analyzed
properly then one can discover several fruitful knowledge hidden inside biological
data. Formal Concept Analysis (FCA) is a method based on lattice theory for the
analysis of binary relational data and has been found to have potential applications
in many areas of bioinformatics and computational biology, beside other applications. In this chapter, we presented the current status of FCA for the analysis and
knowledge discovery from biological data including gene expression discretization,
gene co-expression mining, gene clustering, finding genes in gene regulatory networks, enzyme/protein classifications, binding site classifications, and so on. It also
presented a brief list of FCA-based software tools applied in biological domain and
covered some challenges faced so far.
References
1. Wille R (1982). Restructuring lattice theory: an approach based on hierarchies of concepts.
In I. Rival (Ed.). Ordered sets. Reidel. Dordrecht-Boston, 445–470
2. Babu MM (1997) Biological databases and protein sequence analysis. Center for Biotechnology, Anna University, Chennai
3. Kraulis P (2001) Databases in bioinformatics. Available on:
http://www.avatar.se/lectures/strbio2001/databases/index.html Accessed on May 20, 2015
4. Rioult F, Boulicaut J-F, Cremilleux B, Bessoner J (2003) Using transposition for pattern discovery from microarray data. In: Proceedings of DMKD, 73–79.
5. Rioult F, Robardet C, Blachon S, Cremilleux B, Olivier G, and Boulicaut J-F (2003) Mining
concepts from large SAGE gene expression matrices. In: Proceedings of KDID03 co-located
with ECML-PKDD 2003, Catvat- Dubrovnik (Croatia), 107–118.
6. Kaytoue-Uberall M, Duplessis S, Kuznetsov SO, Napoli A (2009) Two FCA-based methods
for mining gene expression data. In Sebastien Ferre and Sebastian Rudolph, editors, ICFCA,
Lecture Notes in Computer Science, Springer 5548:251–266.
7. Kayala MA, Baldi P (2012) Cyber-T web server: differential analysis of high-throughput data.
Nucleic acids research, 40(W1):W553–W559.
8. Creighton C, Hanash S (2003) Mining gene expression databases for association rules. Bioinformatics 19(1):79-86.
9. Kaytoue-Uberall M, Duplessis S, Napoli A (2008) Using formal concept analysis for the
extraction of groups of co-expressed genes. In Modelling, Computation and Optimization in
Information Systems and Management Sciences 439–449, Springer Berlin Heidelberg.
10. Amin I, Kassim SK, Hassanien A, Hefny HA (2012) Using formal concept analysis for mining hyomethylated genes among breast cancer tumors subtypes. In IEEE 12th International
Conference on Intelligent Systems Design and Applications (ISDA) 764–769.
11. Serhiy A. Yevtushenko (2000) System of data analysis Concept Explorer (In Russian). In
Proceedings of the 7th national conference on Artificial Intelligence KII-2000 Russia, 127–
134.
12. Raza K (2014) Clustering analysis of cancerous microarray data. Journal of Chemical and
Pharmaceutical Research, 6(9):488–493.
13. Choi V, Huang Y, Lam V, Potter D, Laubenbacher R, Duca K (2008) Using formal concept
analysis for microarray data comparison. Journal of bioinformatics and computational biology, 6(01):65–75
Formal Concept Analysis for Knowledge Discovery from Biological Data
13
14. Messai N, Devignes MD, Napoli A, Smail-Tabbone M (2008) Many-valued concept lattices
for conceptual clustering and information retrieval. In ECAI 178:127–131
15. Melo C, Aufaure MA, Orphanides C, Andrews S, McLeod K, Burger A (2013) A conceptual
approach to gene expression analysis enhanced by visual analytics. In Proceedings of the 28th
Annual ACM Symposium on Applied Computing, ACM, 1314–1319
16. Hristoskova A, Boeva V, Tsiporkova E (2014) A formal concept analysis approach to consensus clustering of multi-experiment expression data. BMC Bioinformatics 15(151):1–16
17. Benabderrahmane S (2014) Formal concept analysis and knowledge integration for highlighting statistically enriched functions from microarrays data. International Work-Conference on
Bioinformatics and Biomedical Engineering, IWBBIO 2014, Granda, Spain, 1.
18. Keller BJ, Eichinger F, Kretzler M (2012) Formal concept analysis of disease similarity.
AMIA Summits on Translational Science Proceedings, 42–51
19. Bhavnani SK, Eichinger F, Martini S, Saxman P, Jagadish HV, Kretzler M (2009) Network
analysis of genes regulated in renal diseases: implications for a molecular-based classification.
BMC Bioinformatics 10(Suppl 9):S3
20. Raza K, Parveen R (2012) Soft computing approach for modeling genetic regulatory networks. Advances in Computing and Information Technology, Springer-Verlag, 178:1–12, doi:
10.1007/978-3-642-31600-5 1
21. Raza K, Jaiswal R (2013) Reconstruction and analysis of cancer-specific gene regulatory networks from gene expression profiles. International Journal on Bioinformatics & Biosciences
3(2):25–34, doi: 10.5121/ijbb.2013.3203
22. Gebert J, Motameny S, Faigle U, Forst CV, Schrader R (2008) Identifying genes of gene regulatory networks using formal concept analysis. Journal of Computational Biology 15(2):185–
194
23. Bartel HG, Nofz M (1997) Exploration of NMR data of glasses by means of formal concept
analysis. Chemometrics and Intelligent Laboratory Systems, 36(1):53–63.
24. Coste F, Garet G, Groisillier A, Nicolas J, Tonon T (2014) Automated enzyme classification
by formal concept analysis. In Formal Concept Analysis, Springer International Publishing,
235–250.
25. Han1 M-R, Chung H-J , Kim Jihun, Noh D-Y, Kim JH (2007) Protein classification from
protein-domain and gene-ontology annotation information using formal concept analysis. Y.
Shi et al. (Eds.): ICCS 2007, Part II, Springer-Verlag Berlin Heidelberg, LNCS 4488:347–
354.
26. Bresso E, Grisoni R, Devignes MD, Napoli A, Tabbone M (2012) Formal concept analysis
for the interpretation of relational learning applied on 3D protein-binding sites. In Proceeding
of 4th International Conference on Knowledge Discovery and Information Retrieval - KDIR
2012, Barcelona, Spain, 12.
27. Sugiyama M, Imajo K, Otaki K, Yamamoto A (2012) Semi-supervised ligand finding using
formal concept analysis. IPSJ Transactions on Mathematical Modeling and Its Applications
5(2):39–48, doi: 10.2197/ipsjtrans.5.114
28. Mugumbate G, Newton AS, Rosenthal PJ, Gut J, Moreira R, Chibale K, et al. (2013) Novel
anti-plasmodial hits identified by virtual screening of the ZINC database. J Comput Aided
Mol Des. 27:859–71, doi: 10.1007/s10822-013-9685-zPMID:24158745
29. Lounkine E, Auer J, and Bajorath J (2008) Formal concept analysis for the identification of
molecular fragment combinations specific for active and highly potent compounds. J. Med.
Chem. 51:5342–5348, doi: 10.1021/jm800515r
30. Zafar A, Ahmad S, Rizvi A, Ahmad M (2015) Novel non-peptide inhibitors against SmCL1 of
schistosoma mansoni: In silico elucidation, implications and evaluation via knowledge based
drug discovery. PLoS ONE 10(5):e0123996, doi:10.1371/journal.pone.0123996
31. Priss U (2008) FcaStone - FCA file format conversion and interoperability software. Conceptual Structures Tool Interoperability Workshop (CS-TIW 2008).
32. Mhle H, Wende C (2010) Describing role-oriented software models in terms of formal concept analysis. In Proceedings of the 8th International Conference on Formal Concept Analysis, 241–254. Springer.
14
Khalid Raza
33. Galicia: Galois lattice interactive constructor. http://www.iro.umontreal.ca/galicia/features.html,
Accessed on May 15, 2015.
34. Lahcen B, Kwuida L (2010) Lattice Miner: A tool for concept lattice construction and exploration. In Suplementary Proceeding of International Conference on Formal concept analysis
(ICFCA’10).
35. Andrews S, Orphanides C (2010) FcaBedrock, a formal context creator. In Conceptual Structures: From Information to Intelligence, Springer Berlin Heidelberg, 181–184.
36. Borza PV, Sabou O, Sacarea C (2010) OpenFCA, an open source formal concept analysis
toolbox. In Automation Quality and Testing Robotics (AQTR), 2010 IEEE International Conference on, 3:1–5.
| 5cs.CE
|
arXiv:1104.2835v3 [math.AC] 13 Oct 2013
Combinatorial properties and characterization of
glued semigroups
J. I. Garcı́a-Garcı́a∗
M.A. Moreno-Frı́as†
A. Vigneron-Tenorio‡
Abstract
This work focuses on the combinatorial properties of glued semigroups
and provides its combinatorial characterization. Some classical results for
affine glued semigroups are generalized and some methods to obtain glued
semigroups are developed.
Keywords: Gluing of semigroups, semigroup, semigroup ideal, simplicial complex, toric ideal.
MSC-class: 20M14 (Primary), 20M05 (Secondary).
Introduction
Let S = hn1 , . . . , nl i be a finitely generated commutative semigroup with zero
element which is reduced (i.e. S ∩ (−S) = {0}) and cancellative (if m, n, n0 ∈ S
and m+n = m+n0 then n = n0 ). Under these settings if S is torsion-free then it
is isomorphic to a subsemigroup of Np which means it is an affine semigroup (see
[11]). From now on assume that all the semigroups appearing in this work are
finitely generated, commutative, reduced and cancellative, but not necessarily
torsion-free.
Let k be a field and k[X1 , . . . , Xl ] the polynomial ring in l indeterminates.
This polynomial ring is obviously an S−graded ring (by assigning the S-degree
Pl
ni to the indeterminate Xi , the S-degree of X α = X1α1 · · · Xlαl is i=1 αi ni ∈
S). It is well known that the ideal IS generated by
(
)
l
l
X
X
α
β
X −X |
α i ni =
βi ni ⊂ k[X1 , . . . , Xl ]
i=1
i=1
∗ Departamento
de Matemáticas, Universidad de Cádiz, E-11510 Puerto Real (Cádiz,
Spain). E-mail: [email protected]. Partially supported by MTM2010-15595 and Junta de
Andalucı́a group FQM-366.
† Departamento de Matemáticas, Universidad de Cádiz, E-11510 Puerto Real (Cádiz,
Spain). E-mail: [email protected]. Partially supported by MTM2008-06201-C02-02
and Junta de Andalucı́a group FQM-298.
‡ Departamento de Matemáticas, Universidad de Cádiz, E-11405 Jerez de la Frontera (Cádiz, Spain). E-mail: [email protected]. Partially supported by the grant
MTM2007-64704 (with the help of FEDER Program), MTM2012-36917-C03-01 and Junta
de Andalucı́a group FQM-366.
1
is an S−homogeneous binomial ideal called semigroup ideal (see [6] for details).
If S is torsion-free, the ideal obtained defines a toric variety (see [12] and the
references therein). By Nakayama’s lemma, all the minimal generating sets of
IS have the same cardinality and the S−degrees of its elements can be determinated.
The main goal of this work is to study the semigroups which result from
the gluing of others two. This concept was introduced by Rosales in [10]
and it is closely related with complete intersection ideals (see [13] and the
references therein). A semigroup S minimally generated by A1 t A2 (with
A1 = {n1 , . . . , nr } and A2 = {nr+1 , . . . , nl }) is the gluing of S1 = hA1 i and
S2 = hA2 i if there exists a set of generators ρ of IS of the form ρ = ρ1 ∪
0
ρ2 ∪ {X γ − X γ }, where ρ1 , ρ2 are generating sets of IS1 and IS2 respectively,
γ
γ0
X − X ∈ IS and the supports of γ and γ 0 verify supp (γ) ⊂ {1, . . . , r}
and supp (γ 0 ) ⊂ {r + 1, . . . , l}. Equivalently, S is the gluing of S1 and S2 if
0
IS = IS1 + IS2 + hX γ − X γ i. A semigroup is a glued semigroup when it is the
gluing of others two.
As seen, glued semigroups can be determinated by the minimal generating
sets of IS which can be studied by using combinatorial methods from certain
simplicial complexes (see [1], [4] and [7]). In this work the simplicial complexes
used are defined as follows: for any m ∈ S, set
Cm = {X α = X1α1 · · · Xlαl |
l
X
αi ni = m}
(1)
i=1
and the simplicial complex
∇m = {F ⊆ Cm | gcd(F ) 6= 1},
(2)
with gcd(F ) the greatest common divisor of the monomials in F.
Furthermore, some methods which require linear algebra and integer programming are given to obtain examples of glued semigroups.
The content of this work is organized as follows. Section 1 presents the tools
to generalize to non torsion-free semigroups a classical characterization of affine
gluing semigroups (Proposition 2). In Section 2, the non-connected simplicial
complexes ∇m associated to glued semigroups are studied. By using the vertices of the connected components of these complexes we give a combinatorial
characterization of glued semigroups as well as their glued degrees (Theorem
6). Besides, in Corollary 7 we deduce the conditions in order the ideal of a
glued semigroup to be uniquely generated. Finally, Section 3 is devoted to the
construction of glued semigroups (Corollary 10) and affine glued semigroups
(Subsection 3.1).
1
Preliminaries and generalizations about glued
semigroups
A binomial of IS is called indispensable if it is an element of all system of generators of IS (up to a scalar multiple). This kind of binomials were introduced in
[9] and they have an important role in Algebraic Statistics. In [8] the authors
characterize indispensable binomials by using simplicial complexes ∇m . Note
2
that if IS is generated by its indispensable binomials, it is uniquely generated
up to a scalar multiples.
With the above notation, the semigroup S is associated
Pl to the lattice ker S
formed by the elements α = (α1 , . . . , αl ) ∈ Zl such that i=1 αi ni = 0. Given
G a system of generators of IS , the lattice ker S is generated by the elements
α − β with X α − X β ∈ G and ker S also verifies that ker S ∩ Nl = {0} if
and only if S is reduced. If M(IS ) is a minimal generating set of IS , denote
by M(IS )m ⊂ M(IS ) the set of elements whose S−degree is equal to m ∈ S
and by Betti(S) the set of the S−degrees of the elements of M(IS ). When IS
is minimally generated by rank(ker S) elements, the semigroup S is called a
complete intersection semigroup.
Let C(∇m ) be the number of connected components of ∇m . The cardinality
of M(IS )m is equal to C(∇m ) − 1 (see Remark 2.6 in [1] and Theorem 3 and
Corollary 4 in [7]) and the complexes associated to the elements in Betti(S) are
non-connected.
Construction 1. ([4, Proposition 1]). For each m ∈ Betti(S) the set M(IS )m
is obtained by taking C(∇m ) − 1 binomials with monomials in different connected components of ∇m satisfying that two different binomials have not their
corresponding monomials in the same components and fulfilling that there is at
least a monomial of every connected component of ∇m . This let us construct a
minimal generating set of IS in a combinatorial way.
Let S be minimally1 generated by A1 t A2 with A1 = {a1 , . . . , ar } and
A2 = {b1 , . . . , bt }. From now on, identify the sets A1 and A2 with the matrices
(a1 | · · · |ar ) and (b1 | · · · |bt ). Denote by k[A1 ] and k[A2 ] the polynomial rings
k[X1 , . . . , Xr ] and k[Y1 , . . . , Yt ], respectively. A monomial is a pure monomial
if it has indeterminates only in X1 , . . . , Xr or only in Y1 , . . . , Yt , otherwise it is
a mixed monomial. If S is the gluing of S1 = hA1 i and S2 = hA2 i, then the
binomial X γX −Y γY ∈ IS is a glued binomial if M(IS1 )∪M(IS2 )∪{X γX −Y γY }
is a generating set of IS and in this case the element d = S-degree(X γX ) ∈ S is
called a glued degree.
It is clear that if S is a glued semigroup, the lattice ker S has a basis of the
form
{L1 , L2 , (γX , −γY )} ⊂ Zr+t ,
(3)
where the supports of the elements in L1 are in {1, . . . , r}, the supports of the
elements in L2 are in {r + 1, . . . , r + t}, ker Si = hLi i (i = 1, 2) by considering
only the coordinates in {1, . . . , r} or {r+1, . . . , r+t} of Li , and (γX , γY ) ∈ Nr+t .
Moreover, since S is reduced, one has that hL1 i ∩ Nr+t = hL2 i ∩ Nr+t = {0}.
Denote by {ρ1i }i the elements in L1 and by {ρ2i }i the elements in L2 .
The following Proposition generalizes [10, Theorem 1.4] to non-torsion free
semigroups.
Proposition 2. The semigroup S is the gluing of S1 and S2 if and only if there
exists d ∈ (S1 ∩ S2 ) \ {0} such that G(S1 ) ∩ G(S2 ) = dZ, where G(S1 ), G(S2 )
and dZ are the associated commutative groups of S1 , S2 and {d}, respectively.
1 We consider a minimal generator set of S, in the other case S is trivially the gluing of the
semigroup generated by one of its non minimal generators and the semigroup generated by
the others.
3
Proof. Assume that S is the gluing of S1 and S2 . In this case, ker S is generated
by the set (3). Since (γX , −γY ) ∈ ker S, the element d is equal to A1 γX =
A2 γY ∈ S and d ∈ S1 ∩ S2 ⊂ G(S1 ) ∩ G(S2 ). Let d0 be in G(S1 ) ∩ G(S2 ),
then there exists (δ1 , δ2 ) ∈ Zr × Zt such that d0 = A1 δ1 = A2 δ2 . Therefore
(δ1 , −δ2 ) ∈ ker S because (A1 |A2 )(δ1 , −δ2 ) = 0 and so there exist λ, λρi 1 , λρi 2 ∈ Z
satisfying
P ρ1
(δ1 , 0) =
i λi ρ1i + λ(γX , 0)
(0, δ2 )
−
=
P
i
λρi 2 ρ2i + λ(0, γY ),
P
and d0 = A1 δ1 = i λρi 1 (A1 |0)ρ1i + λA1 γX = λd. We conclude that G(S1 ) ∩
G(S2 ) = dZ with d ∈ S1 ∩ S2 .
Conversely, suppose there exists d ∈ (S1 ∩S2 )\{0} such that G(S1 )∩G(S2 ) =
dZ. We see IS = IS1 +IS2 +hX γX −Y γY i. Trivially, IS1 +IS2 +hX γX −Y γY i ⊂ IS .
Let X α Y β −X γ Y δ be a binomial in IS . Its S−degree is A1 α+A2 β = A1 γ +A2 δ.
Using A1 (α − γ) = A2 (β − δ) ∈ G(S1 ) ∩ G(S2 ) = dZ, there exists λ ∈ Z such
that A1 α = A1 γ + λd and A2 δ = A2 β + λd. We have the following cases:
• If λ = 0,
X αY β − X γ Y δ = X αY β − X γ Y β + X γ Y β − X γ Y δ =
= Y β (X α − X γ ) + X γ (Y β − Y δ ) ∈ IS1 + IS2 .
• If λ > 0,
X αY β − X γ Y δ =
= X α Y β −X γ X λγX Y β +X γ X λγX Y β −X γ X λγY Y β +X γ X λγY Y β −X γ Y δ =
= Y β (X α − X γ X λγX ) + X γ Y β (X λγX − Y λγY ) + X γ (Y λγY Y β − Y δ ).
Using that
X
λγX
−Y
λγY
= (X
γX
−Y
γY
)
λ−1
X
!
X
(λ−1−i)γX
Y
iγY
,
i=0
the binomial X α Y β − X γ Y δ belongs to IS1 + IS2 + hX γX − Y γY i.
• The case λ < 0 is solved similarly.
We conclude that IS = IS1 + IS2 + hX γX − Y γY i.
From above proof it is deduced that given the partition of the system of
generators of S the glued degree is unique.
2
Glued semigroups and combinatorics
Glued semigroups by means of non-connected simplicial complexes are characterized. For any m ∈ S, redefine Cm from (1), as
Cm = {X α Y β = X1α1 · · · Xrαr Y1β1 · · · Ytβt |
r
X
i=1
4
αi ai +
t
X
i=1
βi bi = m}
and consider the sets of vertices and the simplicial complexes
A1
Cm
= {X1α1 · · · Xrαr |
r
X
A1
1
αi ai = m}, ∇A
m = {F ⊆ Cm | gcd(F ) 6= 1},
i=1
A2
Cm
= {Y1β1 · · · Ytβt |
t
X
A2
2
βi bi = m}, ∇A
m = {F ⊆ Cm | gcd(F ) 6= 1},
i=1
where A1 = {a1 , . . . , ar } and A2 = {b1 , . . . , bt } as in Section 1. Trivially, the
A2
1
relations between ∇A
m , ∇m and ∇m are
A1
A2
A2
1
∇A
m = {F ∈ ∇m |F ⊂ Cm }, ∇m = {F ∈ ∇m |F ⊂ Cm }.
(4)
The following result shows an important property of the simplicial complexes
associated to glued semigroups.
Lemma 3. Let S be the gluing of S1 and S2 , and m ∈ Betti(S). Then all the
connected components of ∇m have at least a pure monomial. In addition, all
mixed monomials of ∇m are in the same connected component.
Proof. Suppose that there exists C, a connected component of ∇m only with
mixed monomials. By Construction 1 in all generating sets of IS there is at
least a binomial with a mixed monomial, but this does not occur in M(IS1 ) ∪
M(IS2 ) ∪ {X γX − Y γY } with X γX − Y γY a glued binomial.
Since S is a glued semigroup, ker S has a system of generators as (3). Let
X α Y β , X γ Y δ ∈ Cm be two monomials such that gcd(X α Y β , X γ Y δ ) = 1. In
this case, (α, β) − (γ, δ) ∈ ker S and there exist λ, λρi 1 , λρi 2 ∈ Z satisfying:
P ρ1
(α − γ, 0) =
i λi ρ1i + λ(γX , 0)
(0, β − δ)
=
P
i
λρi 2 ρ2i − λ(0, γY )
• If λ = 0, α − γ ∈ ker S1 and β − δ ∈ ker S2 , then A1 α = A1 γ, A2 β = A2 δ
and X α Y δ ∈ Cm .
P
• If λ > 0, (α, 0) = i λρi 1 ρ1i + λ(γX , 0) + (γ, 0) and
X ρ
A1 α =
λi 1 (A1 |0)ρ1i + λA1 γX + A1 γ = λd + A1 γ,
i
then X λγX X γ Y β ∈ Cm .
• The case λ < 0 is solved likewise.
In any case, X α Y β and X γ Y δ are in the same connected component of ∇m .
We now describe the simplicial complexes that correspond to the S−degrees
which are multiples of the glued degree.
Lemma 4. Let S be the gluing of S1 and S2 , d ∈ S the glued degree and
d0 ∈ S \ {d}. Then CdA0 1 6= ∅ =
6 CdA0 2 if and only if d0 ∈ (dN) \ {0}. Furthermore,
the simplicial complex ∇d0 has at least one connected component with elements
in CdA0 1 and CdA0 2 .
5
Pr
Pt
Proof. If there exist X α , Y β ∈ Cd0 , then d0 = i=1 αi ai = i=1 βi bi ∈ S1 ∩S2 ⊂
G(S1 ) ∩ G(S2 ) = dZ. Hence, d0 ∈ dN.
Conversely, let d0 = jd with j ∈ N and j > 1 and take X γX − Y γY ∈
IS be a glued binomial. It is easy to see that X jγX , Y jγY ∈ Cd0 and thus
{X jγX , X (j−1)γX Y γY } and {X (j−1)γX Y γY , Y jγY } belong to ∇d0 .
The following Lemma is a combinatorial version of [5, Lemma 9] and it is a
necessary condition of Theorem 6.
Lemma 5. Let S be the gluing of S1 and S2 , and d ∈ S the glued degree. Then
the elements of Cd are pure monomials and d ∈ Betti(S).
Proof. The order S defined by m0 S m if m − m0 ∈ S is a partial order on S.
Assume there exists a mixed monomial T ∈ Cd . By Lemma 3, there exists
a pure monomial Y b in Cd such that {T, Y b } ∈ ∇d (the proof is analogous
if we consider X a with {T, X a } ∈ ∇d ). Now take T1 = gcd(T, Y b )−1 T and
Y b1 = gcd(T, Y b )−1 Y b . Both monomials are in Cd0 , where d0 is equal to d
minus the S−degree of gcd(T, Y b ). By Lemma 4, if CdA0 1 6= ∅ then d0 ∈ dN, but
since d0 ≺S d this is not possible. So, if T1 is a mixed monomial and CdA0 1 = ∅,
then CdA0 2 6= ∅. If there exists a pure monomial in CdA0 2 connected to a mixed
monomial in Cd0 , we perform the same process obtaining T2 , Y b2 ∈ Cd00 with T2
a mixed monomial and d00 ≺S d0 . This process can be repeated if there exist a
pure monomial and a mixed monomial in the same connected component. By
degree reasons this cannot be performing indefinitely, an element d(i) ∈ Betti(S)
verifying that ∇d(i) is not connected having a connected component with only
mixed monomials is found. This contradicts Lemma 3.
After examining the structure of the simplicial complexes associated to glued
semigroups, we enunciate a combinatorial characterization by means of the nonconnected simplicial complexes ∇m .
Theorem 6. The semigroup S is the gluing of S1 and S2 if and only if the
following conditions are fulfilled:
1. For all d0 ∈ Betti(S), any connected component of ∇d0 has at least a pure
monomial.
2. There exists a unique d ∈ Betti(S) such that CdA1 6= ∅ 6= CdA2 and the
elements in Cd are pure monomials.
3. For all d0 ∈ Betti(S) \ {d} with CdA0 1 6= ∅ =
6 CdA0 2 , d0 ∈ dN.
Besides, the above d ∈ Betti(S) is the glued degree.
Proof. If S is the gluing of S1 and S2 , the result is obtained from Lemmas 3, 4
and 5.
Conversely, by hypothesis 1 and 3, given d0 ∈ Betti(S)\{d} the set M(IS1 )d0
is constructed from CdA0 1 and M(IS2 )d0 from CdA0 2 as in Construction 1. Analogously, if d ∈ Betti(S), the set M(IS )d is obtained from the union of M(IS1 )d ,
M(IS2 )d and the binomial X γX −Y γY with X γX ∈ CdA1 and Y γY ∈ CdA2 . Finally
G
M(IS1 )m t M(IS2 )m t {X γX − Y γY }
m∈Betti(S)
is a generating set of IS and S is the gluing of S1 and S2 .
6
From Theorem 6 we obtain an equivalent property to Theorem 12 in [5] by
using the language of monomials and binomials.
Corollary 7. Let S be the gluing of S1 and S2 , and X γX − Y γY ∈ IS a glued binomial with S−degree d. The ideal IS is minimally generated by its indispensable
binomials if and only if the following conditions are fulfilled:
• The ideals IS1 and IS2 are minimally generated by their indispensable binomials.
• The element X γX − Y γY is an indispensable binomial of IS .
• For all d0 ∈ Betti(S), the elements of Cd0 are pure monomials.
Proof. Suppose that IS is generated by its indispensable binomials. By [8,
Corollary 6], for all m ∈ Betti(S) the simplicial complex ∇m has only two
vertices. By Contruction 1 ∇d = {{X γX }, {Y γY }} and by Theorem 6 for all
A2
1
d0 ∈ Betti(S) \ {d} the simplicial ∇d0 is equal to ∇A
d0 or ∇d0 . In any case,
γX
γY
X −Y
∈ IS is an indispensable binomial, and IS1 , IS2 are generated by
their indispensable binomials.
Conversely, suppose that IS is not generated by its indispensable binomials.
Then, there exists d0 ∈ Betti(S) \ {d} such that ∇d0 has more than two vertices
in at least two different connected components. By hypothesis, there are not
mixed monomials in ∇d0 and thus:
A2
1
• If ∇d0 is equal to ∇A
d0 (or ∇d0 ), then IS1 (or IS2 ) is not generated by its
indispensable binomials.
• Otherwise, CdA0 1 6= ∅ =
6 CdA0 2 and by Lemma 4, d0 = jd with j ∈ N, therefore
(j−1)γX γY
X
Y
∈ Cd0 which contradicts the hypothesis.
We conclude IS is generated by its indispensable binomials.
The following example taken from [13] illustrates the above results.
Example 8. Let S ⊂ N2 be the semigroup generated by the set
{(13, 0), (5, 8), (2, 11), (0, 13), (4, 4), (6, 6), (7, 7), (9, 9)} .
In this case, Betti(S) is
{(15, 15), (14, 14), (12, 12), (18, 18), (10, 55), (15, 24), (13, 52), (13, 13)}.
Using the appropriated notation for the indeterminates in the polynomial
ring k[x1 , . . . , x4 , y1 , . . . , y4 ] (x1 , x2 , x3 and x4 for the first four generators of
S and y1 , y2 , y3 , y4 for the others), the simplicial complexes associated to the
elements in Betti(S) are those that appear in Table 1. From this table and by
using Theorem 6, the semigroup S is the gluing of h(13, 0), (5, 8), (2, 11), (0, 13)i
and h(4, 4), (6, 6), (7, 7), (9, 9)i and the glued degree is (13, 13). From Corollary
7, the ideal IS is not generated by its indispensable binomials (IS has only four
indispensable binomials).
7
C(15,15) = {y12 y3 , y2 y4 }
C(14,14) = {y12 y2 , y32 }
C(12,12) = {y13 , y22 }
C(10,55) = {x21 x4 , x53 }
∇(15,15)
∇(14,14)
∇(12,12)
∇(10,55)
Ys
Ys
Ys
Xs
C(18,18) = {y12 y2 , y23 ,
y1 y32 , y42 }
∇(18,18)
C(15,24) = {x1 x2 x3 , x32 ,
x3 y1 y4 , x3 y2 y3 }
∇(15,24)
C(13,52) = {x2 x43 , x1 x44 ,
x34 y1 y4 , x34 y2 y3 }
∇(13,52)
C(13,13) = {x1 x4 , y1 y4 ,
y2 y3 }
∇(13,13)
Ys
Xs
Ys
Xs
Ys
Ys
Table 1: Non-connected simplicial complexes associated to Betti(S).
3
Generating glued semigroups
In this section, an algorithm to obtain examples of glued semigroups is given.
Consider A1 = {a1 , . . . , ar } and A2 = {b1 , . . . , bt } two minimal generator sets of
the semigroups T1 and T2 and let Lj = {ρji }i be a basis of ker Tj with j = 1, 2.
Assume that IT1 and IT2 are nontrivial proper ideals of their corresponding
polynomial rings. Consider γX and γY be two nonzero elements in Nr and Nt
respectively2 , and the integer matrix
L1
0
L2 .
(5)
A= 0
γX −γY
Let S be a semigroup such that ker S is the lattice generated by the rows of
the matrix A. This semigroup can be computed by using the Smith Normal
Form (see [11, Chapter 3]). Denote by B1 , B2 two sets of cardinality r and t
respectively satisfying S = hB1 , B2 i and ker(hB1 , B2 i) is generated by the rows
of A.
The following Proposition shows that the semigroup S satisfies one of the
necessary conditions to be a glued semigroup.
Proposition 9. The semigroup S verifies G(hB1 i) ∩ G(hB2 i) = (B1 γX )Z =
(B2 γY )Z with d = B1 γX ∈ hB1 i ∩ hB2 i.
Proof. Use that ker S has a basis as (3) and proceed as in the proof of the
necessary condition of Proposition 2.
Due to B1 ∪ B2 may not be a minimal generating set, this condition does not
assure that S is a glued semigroup. For instance, taking the numerical semigroups T1 = h3, 5i, T2 = h2, 7i and (γX , γY ) = (1, 0, 2, 0), the matrix obtained
2 Note
that γX ∈
/ ker T1 and γY ∈
/ ker T2 because these semigroups are reduced.
8
Xs
from formula (5) is
5
0
1
−3 0
0
7
0 −2
0
−2
0
and B1 ∪ B2 = {12, 20, 6, 21} is not a minimal generating set. The following
result solve this issue.
Corollary 10. The semigroup S is a glued semigroup if
r
X
γXi > 1 and
i=1
t
X
γY i > 1.
(6)
i=1
Proof. Suppose that the set of generators B1 ∪ B2 of S is non-minimal, thus
one of its elements is a natural combination of the others. Assume that this
element is the first of B1 ∪ B2 , then there exist λ2 , . . . , λr+t ∈ N such that
B1 (1, −λ2 , . . . , −λr ) = B2 (λr+1 , . . . , λr+t ) ∈ G(hB1 i)∩G(hB2 i). By Proposition
9, there exists λ ∈ Z satisfying B1 (1, −λ2 , . . . , −λr ) = B2 (λr+1 , . . . , λr+t ) =
B1 (λγX ). Since B2 (λr+1 , . . . , λr+t ) ∈ S, λ ≥ 0 and thus
ν = (1 − λγX1 , −λ2 − λγX2 , . . . , −λr − λγXr ) ∈ ker(hB1 i) = ker T1 ,
{z
}
|
≤0
with the following cases:
• If λγX1 = 0, then T1 is not minimally generated which it is not possible
by hypothesis.
• If λγX1 > 1, then 0 > ν ∈ ker T1 , but this is not possible due to T1 is a
reduced semigroup.
• If λγX1 = 1, then λ = γX1 = 1 and
ν = (0, −λ2 − γX2 , . . . , −λr − γXr ) ∈ ker T1 .
{z
}
|
≤0
If λi + γXi 6= 0 for some i = 2, . . . , r, then T1 is not a reduced semigroup.
This implies λi = γXi = 0 for all i = 2, . . . , r.
We have just proved that γX = (1, 0, . . . , 0). In the general case, if S is not
minimally generated it is because either γX or γY are elements in the canonical
bases of Nr or Nt , respectively.
To avoidPthis situation, it is sufficient to take
Pr
t
γX and γY satisfying i=1 γXi > 1 and i=1 γY i > 1.
From the above result we obtain a characterization of glued semigroups: S
is a glued semigroup if and only if ker S has a basis as (3) satisfying Condition
(6).
Example 11. Let T1 = h(−7, 2), (11, 1), (5, 0), (0, 1)i ⊂ Z2 and T2 = h3, 5, 7i ⊂
N be two reduced affine semigroups. We compute their associated lattices
ker T1 = h(1, 2, −3, −4), (2, −1, 5, −3)i and ker T2 = h(−4, 1, 1), (−7, 0, 3)i.
9
If we take γX = (2, 0, 2, 0) and γY = (1, 2, 1), the matrix A is
1 2 −3 −4 0
0
0
2 −1 5 −3 0
0
0
0 0
0
0
−4
1
1
0 0
0
0 −7 0
3
2 0
2
0 −1 −2 −1
and the semigroup S ⊂ Z4 × Z2 is generated by
{(1, −5, 35), (3, 12, −55), (1, 5, −25), (0, 1, 0), (2, 0, 3), (2, 0, 5), (2, 0, 7)}.
|
{z
} |
{z
}
B1
B2
The semigroup S is the gluing of the semigroups hB1 i and hB2 i and ker S is generated by the rows of the above matrix. The ideal IS ⊂ C[x1 , . . . , x4 , y1 , . . . , y3 ]
is generated3 by
{x1 x83 x4 − x32 , x1 x22 − x33 x44 , x21 x53 − x2 x34 , x31 x2 x23 − x77 ,
y1 y3 − y22 , y13 y2 − y32 , y14 − y2 y3 , x21 x23 − y15 y2 },
{z
}
|
glued binomial
then S is really a glued semigroup.
3.1
Generating affine glued semigroups
From Example 11 it be can deduced that the semigroup S is not necessarily
torsion-free. In general, a semigroup T is affine (or equivalently it is torsionfree) if and only if the invariant factors4 of the matrix whose rows are a basis
of ker T are equal to one. Assume zero-columns of the Smith Normal Form of
a matrix are located on its right side. We now show conditions for S being
torsion-free.
Take L1 and L2 the matrices whose rows form a basis of ker T1 and ker T2 ,
respectively and let P1 , P2 , Q1 and Q2 be some matrices with determinant ±1
(i.e. unimodular matrices) such that D1 = P1 L1 Q1 and D2 = P2 L2 Q2 are the
Smith Normal Form of L1 and L2 , respectively. If T1 and T2 are two affine
semigroups, the invariant factors of L1 and L2 are equal to 1. Then
D1 0
P1 0 0
L1
0
Q1 0
0 D2 = 0 P2 0 0
L2
, (7)
0 Q2
0
γX
γY0
0
0 1
γX −γY
{z
}
|
=:A
0
where γX
= γX Q1 and γY0 = −γY Q2 . Let s1 and s2 be the numbers of zerocolumns of D1 and D2 (s1 , s2 > 0 because T1 and T2 are reduced, see [11,
Theorem 3.14]).
Lemma 12. The semigroup S is an affine semigroup if and only if
0
gcd {γXi
}ri=r−s1 ∪ {γY0 i }ti=t−s2 = 1.
3 See
[14] to compute IS when S has torsion.
invariant factors of a matrix are the diagonal elements of its Smith Normal Form (see
[2, Chapter 2] and [11, Chapter 2]).
4 The
10
Proof. With the conditions fulfilled by T1 , T2 and (γX , γY ), the necessary and
sufficient
condition for the invariant
factors of A to be all equal to one is
0
t
0
r
gcd {γXi }i=r−s1 ∪ {γY i }i=t−s2 = 1.
The following Corollary gives the explicit conditions that γX and γY must
satisfy to construct an affine semigroup.
Corollary 13. The semigroup S is an affine glued semigroup if and only if:
1. T1 and T2 are two affine semigroups.
2. (γX , γY ) ∈ Nr+t .
Pr
Pt
3.
i=1 γXi ,
i=1 γY i > 1.
4. There exist fr−s1 , . . . , fr , gt−s2 , . . . , gt ∈ Z such that
0
0
, . . . , γXr
(fr−s1 , . . . , fr )·(γX(r−s
)+(gt−s2 , . . . , gt )·(γY0 (t−s2 ) , . . . , γY0 t ) = 1.
1)
Proof. It is trivial by the given construction, Corollary 10 and Lemma 12.
Therefore, to obtain an affine glued semigroup it is enough to take two affine
semigroups and any solution (γX , γY ) of the equations of the above corollary.
Example 14. Let T1 and T2 be the semigroups of Example 11. We compute
two elements γX = (a1 , a2 , a3 , a4 ) and γY = (b1 , b2 , b3 ) in order to obtain an
affine semigroup. First of all, we perform a decomposition of the matrix as (7)
by computing the integer Smith Normal Form of L1 and L2 :
1
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
a1 a1 − 2a2 − a3 −7a1 + 11a2 + 5a3 2a1 + a2 + a4
0
0
0
0
0
0
=
1
0
0
0
1
0
−b1 b1 + 2b2 + 3b3 −3b1 − 5b2 − 7b3
1 0
2 −1
0 0
0 0
0 0
0 0
0 0
−2 1
7 −4
0 0
0
1 2 −3 −4
2 −1 5 −3
0
0
0 0 0 0
0 0 0 0 0
1
a1 a2 a3 a4
0
0
0
0
−4 1
−7 0
−b1 −b2
1 1 −7 2
0 −2 11 1
0
0
0 −1 5 0
1
0 0 0 1
3
0 0 0 0
−b3 0 0 0 0
0 0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
1 −1 3
0 −2 5
0 −3 7
Second, by Corollary 13, we must find a solution to the system:
a1 + a2 + a3 + a4 > 1
b1 + b2 + b3 > 1
f1 , f2 , g1 ∈ Z
f1 (−7a1 + 11a2 + 5a3 ) + f2 (2a1 + a2 + a4 ) + g1 (−3b1 − 5b2 − 7b3 ) = 1
with a1 , a2 , a3 , a4 , b1 , b2 , b3 ∈ N. Such solution is computed (in less than a second) using FindInstance of Wolfram Mathematica (see [15]):
FindInstance[(−7a1 +11a2 +5a3 )∗f1 +(2a1 +a2 +a4 )∗f2 +(−3b1 −5b2 −7b3 )∗g1 == 1
&& a1+a2+a3+a4 > 1&& b1 +b2 +b3 > 1&& a1 ≥ 0&& a2 ≥ 0&& a3 ≥ 0&& a4 ≥ 0
11
&& b1 ≥ 0&& b2 ≥ 0&& b3 ≥ 0, {a1 , a2 , a3 , a4 , b1 , b2 , b3 , f1 , f2 , g1 }, Integers ]
{{a1 → 0, a2 → 0, a3 → 3, a4 → 0, b1 → 1, b2 → 1, b3 → 0, f1 → 1, f2 → 0, g1 → 0}}
We now take γX = (0, 0, 3, 0) and γY = (1, 1, 0), and construct the matrix
1
2 −3 −4
0
0 0
2 −1
5 −3
0
0 0
0
0
0
0
−4
1
1
A=
.
0
0
0
0 −7
0 3
0
0
3
0 −1 −1 0
We have the affine semigroup S ⊂ Z2 which is minimally generated by
{(2, −56), (1, 88), (0, 40), (1, 0), (0, 45), (0, 75), (0, 105)}
|
{z
} |
{z
}
B1
B2
satisfies that ker S is generated by the rows of A and it is the result of gluing
the semigroups hB1 i and hB2 i. The ideal IS is generated by
{x1 x83 x4 − x32 , x1 x22 − x33 x44 , x21 x53 − x2 x34 , x31 x2 x23 − x74 ,
y1 y3 − y22 , y13 y2 − y32 , y14 − y2 y3 , x33 − y1 y2 },
| {z }
glued binomial
therefore, S is a glued semigroup.
All glued semigroups have been computed by using our programm ecuaciones
which is available in [3] (this programm requires Wolfram Mathematica 7 or
above to run).
References
[1] E. Briales, A. Campillo, C. Marijuán, P. Pisón. Minimal Systems
of Generetors for Ideals of Semigroups. J. Pure Appl. Algebra 124 (1998),
7–30.
[2] H. Cohen. A Course in Computational Algebraic Number Theory, Graduate Texts in Mathematics, 138, Springer-Verlag, 1996.
[3] Ecuaciones.
http://www.uca.es/dpto/C101/pags-personales/
alberto.vigneron/1ecuaciones.rar
[4] S. Eliahou. Courbes monomiales et algébre de Rees symbolique. PhD
Thesis. Université of Genève, 1983.
[5] P. A. Garcia-Sanchez, I. Ojeda. Uniquely presented finitely generated
commutative monoids. Pacific J. Math. 248(1) (2010), 91–105.
[6] J. Herzog. Generators and relations of abelian semigroups and semigroup
rings. Manuscripta Math. 3 (1970) 175–193.
12
[7] I. Ojeda, A. Vigneron-Tenorio. Simplicial complexes and minimal
free resolution of monomial algebras. J. Pure Appl. Algebra 214 (2010),
850–861.
[8] I. Ojeda, A. Vigneron-Tenorio. Indispensable binomials in semigroup
ideals. Proc. Amer. Math. Soc. 138 (2010), 4205–4216.
[9] H. Ohsugi, T. Hibi. Toric ideals arising from contingency tables. Commutative Algebra and Combinatorics. Ramanujan Mathematical Society
Lecture Notes Series, Vol. 4, Ramanujan Mathematical Society, Mysore,
India, 2007, pp. 91-115.
[10] J.C. Rosales. On presentations of subsemigroups of Nn . Semigroup Forum 55 (1997), no. 2, 152–159.
[11] J.C. Rosales, P.A. Garcı́a-Sánchez. Finitely generated commutative
monoids. Nova Science Publishers, Inc., New York, 1999.
[12] B. Sturmfels. Gröbner bases and convex polytopes, volume 8 of University Lecture Series. American Mathematical Society, Providence, RI,
1996.
[13] A. Thoma. Construction of Set Theoretic Complete Intersections via
Semigroup Gluing. Beiträge Algebra Geom. 41(1) (2000), 195–198.
[14] A. Vigneron-Tenorio. Semigroup Ideals and Linear Diophantine Equations. Linear Algebra and its Applications 295 (1999), 133–144.
[15] Wolfram Mathematica. http://www.wolfram.com/mathematica/
13
| 0math.AC
|
September 12, 2014
The Distributed Language
V1.0.3 ( alpha )
White Paper
Boris Burshteyn
[email protected]
Hello White Paper
V1.0.3 (alpha)
One Solution
Abstract
Hello is a general-purpose, object-oriented, protocol-agnostic
distributed programming language. This paper explains the ideas
that guided design of Hello. It shows the spirit of Hello using two
brief expressive programs and provides a summary of language
features. In addition, it explores historical parallels between the
binary programming of early computers and the distributed programming of modern networks.
Hello documentation, translator and runtime engine are freely
available at [14].
1 Preface
“Whether oratory is a creation of rules, or of training,
or of natural gifts, it is the most difficult of all things to
achieve.” Cicero, “The Brutus,” 46 B.C.E.
Cicero – the most gifted orator of all times – had realized the extreme difficulty of communication more than two
thousand years ago. Obviously, there were no computers
when he made his famous speeches in the Roman Courts,
Assemblies, and Senate, and wrote treatises about the oratory art. At that time, his talents flourished in the area of oneway human communication.
Today computers exchange code and data across the
networks with the help of sophisticated hardware devices,
intricate networking protocols and elaborate software systems. This immensely complex communication infrastructure makes creation of networking software an extremely
difficult and costly enterprise. Thus, the Cicero’s quote
remains true when applied to modern computer networking
– distributed software is one of the hardest fields of software engineering.
The distributed applications orchestrate computations
across the network while utilizing a multitude of collaborating computers. Their development demands significant
investments in labor, time, and finances. Therefore, a relentless quest for a silver bullet that would reduce the development efforts continues.
This paper presents Hello -- a general-purpose, objectoriented, imperative, protocol-agnostic programming language for writing distributed software. Its unique first-class
distributed language features put Hello in an excellent position to cut the costs and ease the pain of the distributed
software development.
2 Background
The idea to create a distributed programming language
emerged during the long experience of designing and implementing commercial-grade networking applications as
diverse as a Distributed Database, High Availability Cluster
Monitor, MAPI-based Email Server (via wire-level protocol
discovery), and a Video-Conferencing Server. During that
work, communication libraries implemented abstractions
that hid the underlying network protocol. These abstractions
represented interconnected servers communicating over the
network between themselves and with clients.
Page 2
September 12, 2014
3 Three Problems
Although the applications were quite different, they abstracted the distributed hardware and software in a similar
way. Still, the abstractions had to be implemented each time
anew as all systems had re-invented many of the same
building blocks: the re-programming involved remote objects, connections, messages, queues, jobs, replies, timeouts,
state changes, failures, and others.
Because the language in hand had no first-class distributed features, it needed help of extraneous distributed libraries. However, since the language knew nothing about the
libraries, which offered only low level networking interface,
the coding and debugging were difficult for programmers
while the optimization was impossible for a translator. At
the end, there were no time and project resources left to
improve efficiency and reliability of the communication
components. In summary, the following three problems
characterized many distributed systems:
1) Repeated re-coding of the same distributed primitives.
2) Challenging development process due to the combination of a programming language and extraneous library.
3) Inefficient and unreliable distributed code.
4 One Solution
Those repetitive efforts had built determination to develop distributed primitives for reuse in a diverse spectrum
of applications. To embed them in the language was a logical choice because it offered simultaneous simplicity of
thought and efficiency of code1. The distributed programming language has nicely solved all three problems mentioned above:
1) Since the first-class distributed language features are
always available, there is no need to re-implement them
for every new application.
2) These protocol-agnostic features, being higher level
than the library interfaces, cut the need for extraneous
tools while simplifying development and debugging.
3) The distributed object-oriented language architecture
and the optimizing translator raise the efficiency, reliability, and security of the distributed code.
The distributed elements, implemented in both the language translator and its runtime system, become as ubiquitous as are the local primitives like heap, stack, method,
class, variable, and object. Therefore, one does not need
extraneous libraries, but can rely only on the language translator, and runtime to achieve the productive software development combined with the efficient and reliable distributed
code. Thus, the distributed programming language Hello
was born to blend the protocol-agnostic network model into
the fabric of an object-oriented paradigm.
1
It was also natural because the author had extensive experience
with parser generators and a variety of language frontends.
V1.0.3 (alpha)
Hello, World!
5 Language Design Goals
One of the primary goals set during the Hello design
was its future widespread adoption by the software engineering community. Therefore, it was crucial to come up
with a general-purpose language, avoiding the danger of
falling into a niche of specialized languages [20]. Another
goal was to abstract in the language the components of the
entire network and its protocols. The abstraction level
should be high enough to conceal the mundane intricacies
of the rigid protocol interfaces, yet not too high to make the
language overly specialized. In addition, distributed primitives were meant to become a crucial part of the language,
but not its dominant part: the language core should remain
filled with general-purpose abstractions for sequential computations, which are running on a single host.
Towards these goals, the lower-level distributed abstractions embedded into the language’s type system, augmented by a few imperative operations, had been invented.
Any particular distributed semantics can now be derived
from a combination of the Hello distributed primitives -engines, hosts, queues, partitions, groups, messages, events,
and remote references -- and their operations.
6 Language Design Principle
Historically, single-host general-purpose languages abstract the computer memory, memory-stored programs, and
CPU from the computer hardware architecture first articulated in 1945 by J. von Neumann in [21]. However, distributed abstractions should differ from single-host abstractions
because of the following fundamental differences between
the single-host and distributed architectures.
As J. Backus explained in [1], imperative single-host
languages are designed so that their programs reflect the
linear structure of the von Neumann machine: variables and
statements from the source text of a program are translated
into data and CPU instructions stored in the contiguous
memory of a single host. Moreover, the linear textdescending control flow (with occasional gotos and loops)
in the source program corresponds to the linear addressascending control flow (with occasional jumps and branches) in its translated machine code, which is stored in a single
piece of memory at runtime.
However, distributed software does not run in a single
memory piece – its components execute on different computers with separate memory units. The computers connect,
communicate and compute in a network of any structure:
from linear – to hierarchical -- to the full graph. Network
topology and connections between application’s components often change with time. Thus, the linear text-tomemory mapping idea that guided design of the imperative
single-host languages fails for distributed languages. At the
same time, a distributed language should abstract the networked components such as computers, processes, threads,
Page 3
September 12, 2014
objects, messages, events, and code packages. It shall provide first-class language features to manage distributed
operations with inevitable timeouts and failures, and to hide
the underlying networking protocol.
Therefore, Hello design principle was to implement the
distributed abstractions inside the language’s type system:
1) Hello provides several built-in classes augmented with
a few operations and statements, organically interwoven into the language’s imperative syntax and objectoriented semantics. Instances of these classes represent
hosts, runtime engines, memory partitions, and execution queues spread out around the network; their purpose is to support first-class distributed language features transparently to Hello programs. At runtime, host
instances self-organize into distributed groups, which
graph structures reflect dynamic network connections.
2) The instances of user-defined classes – objects – reside
in memory partitions and engine heaps around the network referring to each other through local and remote
references. The runtime engines, with the help of the
translator and the instances of the built-in classes,
transparently navigate object references. The navigation allows Hello programs to manipulate and transfer
data, code and control flow over the network.
7 Hello, World!
The following example shows how Hello first-class distributed features and sequential core allow for a simple
design. In addition, it demonstrates fast coding of the short
and efficient distributed programs. This example highlights
Hello advantages over the single-host languages C [4], C++
[5] and Java [18] augmented with extraneous distributed
libraries.
The program “Hello, World!” shows just some firstclass distributed features of the Hello language, which set it
apart from any single-host programming language. The
whole program is a class named HelloWorld. It contains
a function main() that invokes a function print() on
the hosts referred to from the built-in group hosts:
package Hello_World;
class HelloWorld {
// Broadcast to all known hosts
public static void main() {
hosts.+print("Hello, World!\n" +
this_host.name() + ":-)\n");
}
};
The two Hello concepts illustrated by this example are a
group, which is a graph-ordered collection of objects, and
operator of bottom-up iteration .+, which invokes, in the
graph order, an iterator function on each of the group's
objects. In particular, the built-in group named hosts, has
its objects distributed across the network -- each object is a
separate instance of class host_group, residing within
the virtual or physical host, which this object describes. At
runtime, Hello engines maintain this group automatically:
two objects are connected in the group if their hosts can
V1.0.3 (alpha)
Distributed Architecture
communicate over the network. The definition of
host_group is located in the Hello package standard;
here is its relevant abbreviated fragment:
public external group class host_group {
public external
host
current_host;
public external copy
host_group[] children();
public external iterator void print(copy char[] str) {
current_host.print(str);
}
};
When main() is started on a runtime engine, it begins
iterating function print() on the objects from the group
hosts. Because each host object resides on a different
host, print() is invoked on different hosts -- once locally
on the current host referred to by current_host and
once remotely on each of the remote children hosts referred
to by references from the array returned by the function
with the reserved name children().
During the iteration, at each invocation print() writes
a greeting signed with the name of the host where main()
has been started and kept waiting for the iteration to finish;
this name is a character array returned by the method current_host.name(). At each invocation, print()
writes to the standard output of the host where print() is
invoked, local or remote. Its argument is a string combined,
with the operator concatenation +, from the three portions
of the greeting: two literal strings and one character array. It
is copied across the network during iteration and passed to
print() at each invocation.
Package standard is an internal part of the Hello
runtime – it contains types that support distributed architecture (e.g. classes host and host_group). A programmer
may want to learn a general functionality of some standard
types and operations described in Hello Programming
Guide [14]. However, there is no need to understand its
details or even import it into any package.
7.1 Hello, C++ & Java!
The simple imperative and object-oriented features
coupled with the distributed abstractions distinguish general-purpose Hello architecture from the specialized distributed architectures2. In addition, the following code snippets
present “Hello, World!” broadcasts in Java and C++ as seen
at [13] and [3].
Since neither Java nor C++ has distributed features,
both resort to extraneous libraries: Java to RMI, C++ to
BOOST.mpi. The difference with the “Hello, World!” written in Hello is apparent – the latter has no setup while both
C++ and Java use most of their code (in red) for setting up
translation and runtime contexts.
At the same time, Hello code is much shorter than either
C++ or Java. Moreover, the Java source shows only the
2
Like the object mobility in Emerald [11], actors in Salsa [25], or
process-orientation of Erlang [9].
Page 4
September 12, 2014
server part – the client portion from [13] adds even more
Java code while Hello broadcast does not distinguish between client and server.
package example.hello;
import java.rmi.registry.Registry;
import java.rmi.registry.LocateRegistry;
import java.rmi.RemoteException;
import java.rmi.server.UnicastRemoteObject;
public class Server implements Hello {
public Server() {}
public String sayHello() {
return "Hello, world!";
}
public static void main(String args[]) {
try {
Server obj = new Server();
Hello stub = (Hello)
UnicastRemoteObject.
exportObject(obj, 0);
// Bind the remote object stub in the registry
Registry registry = LocateRegistry.
getRegistry();
registry.bind("Hello", stub);
System.err.println("Server ready");
} catch (Exception e) {
System.err.println("Server exception: " +
e.toString());
e.printStackTrace();
}
}
}
Figure 1 Java broadcasting server program as seen at [13]
#include <boost/mpi.hpp>
#include <iostream>
#include <boost/serialization/string.hpp>
namespace mpi = boost::mpi;
int main(int argc, char* argv[]) {
mpi::environment env(argc, argv);
mpi::communicator world;
std::string value;
if (world.rank() == 0) {
value = "Hello, World!";
}
broadcast(world, value, 0);
std::cout << "Process #" << world.rank() <<
" says " << value
<< std::endl;
return 0;
}
Figure 2 C++ broadcasting program as seen at [3]
8 Distributed Architecture
In order to be useful, any distributed language should
provide first-class features that reflect the program’s
runtime distributed environment. They can manifest themselves through particular fundamental data types and their
operations, as well as through a runtime context. In order to
make software development productive and programs
runtime efficient, such architecture should reflect general
functionality of the major components of modern networks.
At the same time, it should be precise enough for the translator and runtime system to be able to map that architecture
on the runtime environment, and to optimize it.
To be accepted among software engineers, the distributed primitives should augment, but not overwhelm the sequential language elements. These elements must remain its
central part. Therefore, the Hello networking components
are designed in order to blend with the main sequential core
of the language, which resembles a simplified Java subset.
V1.0.3 (alpha)
Translation
The following list presents the essential elements of the
Hello distributed architecture:
1) Sourcepack is a set of all source files with Hello programs under a single directory that together constitute a
Hello package. Sources define hierarchies of related
classes and interfaces. Each class represents both data
that are instantiated as objects and functions which operate on objects at runtime. Interface defines shared data and function prototypes implemented by derived
classes. Each source file must have extension .hlo,
like HelloWorld.hlo.
2) Runpack is a shared library built from all sources of a
source package by Hello translator: a sourcepack abc
becomes runpack abc.so. Hello runtime engine loads
and executes multiple runpacks at runtime. Engines
from different computers exchange runpacks at runtime
on demand, following the program control flow. Runpack code accesses data from memory partitions located on the same or different hosts. Runpacks contain
machine instructions rather than intermediary code –
they are binary libraries, which execute their code directly on CPU.
3) Hello translator is an executable /usr/bin/het,
this translates Hello sourcepack into a runpack. The
translator also archives and installs Hello sourcepacks.
4) Computer is a physical or virtual computing machine
that may persistently store Hello packages in its directories, capable of running Hello packages, translator,
hosts and runtime engines. A computer may reside on a
network, which is used by runtime engines in order to
transfer the Hello data, control flow, and runpacks between different hosts.
5) Host is a uniquely named set of runtime engines and
partitions running on a computer. Each computer may
run a single primary host and a number of secondary
hosts. Engines from the same host directly access data
from all partitions of their host. Engines from different
hosts access each other’s data through the network even
if they run on the same computer.
6) Partition is a piece of virtual memory that holds shared
runtime data – objects, as well as arrays of references to
objects and primitive data. Each host may contain several partitions created by its engines. The partition data
is available for multiple programs executing concurrently on either local or remote hosts.
7) Runtime engine is an executable /usr/bin/hee, this
loads and executes runpacks. Runtime engine maintains
its data heap accessible only to programs executed on its
behalf. In addition, engines are responsible for automatic
and transparent transferring of code, data and control
flow between the hosts across the network. An engine
accesses partitions from its host by mapping its virtual
memory onto the partitions.
Page 5
September 12, 2014
8) Network is a physical or virtual medium, which connects hosts from the same or different computers. Hello
does not impose any particular operational characteristics on the network as long as the network is capable of
maintaining locally unique computer addresses and of
transferring code, data and control flow between the
engines from the connected hosts. Hello is a protocolagnostic language, as it imposes no requirements on the
underlying network protocol; neither has it exposed any
of the protocol elements to Hello programs 3.
9) Queue is a system object that queues up execution
requests and subsequently dispatches them onto an execution thread; each Hello program executes on behalf
of a thread designated to a particular queue.
10) Event is a call to a method with partially supplied arguments. Events accumulate on the queues, which execute events when subsequent calls complete the method’s argument list. Events serve for delayed method invocation as well as for synchronization between local
and remote queue threads.
11) SID and Privilege protect distributed data from unauthorized access. A Security Id (SID) is a 16-byte globally unique identifier that represents, via a bitmask, a
set of privileges at runtime. Every queue has a set of
pairs (SID, privilege) that determines operations allowed for the programs executed on behalf of the
queue. In addition, an individual remote object may also have a set of such pairs to control operations allowed
on the object. Finally, each host maintains a map of
such pairs in order to determine the kinds of permissible remote requests executed on the host.
12) Neighborhood and Path assist in reaching distributed
code and data across the network. A host neighborhood
contains connected hosts that communicate directly,
without any help from intermediary hosts. A host path
is a sequence of connected hosts between two disconnected hosts; disconnected hosts communicate through
the connected hosts from the path. The runtime engines
build neighborhoods, accumulate host paths, and use
them, transparently to Hello programs, while navigating
the network between the hosts. Hello programs can
control the neighborhood and paths formation as well as
the path navigation policies.
9 Translation
The following figure illustrates how Hello translator
het translates a source package into a runpack:
3
However, the current runtime engine implementation requires the
network to support the TCP protocol from IPV4.
V1.0.3 (alpha)
Hello sourcepack
(package in
source text)
Language Features
10.1
Hello
Translator
C++ sources
Hello runpack
(shared
binary library)
C++
Compiler
Figure 3 Hello Translation
Note that all source files from a given package participate in this operation. Although het can translate individual source files, it does not build the runpack unless it successfully translates all sources from the package. Another
important feature of the Hello translation process is that
het first translates Hello sources into C++ sources, and
then automatically invokes the C++ compiler that further
translates the C++ sources into a runpack – a shared binary
library.
10 Sample Code – Remote Shell
The Appendix at the end of this paper presents a sample Hello code -- Remote Shell -- a program that executes a
UNIX shell command on a remote host. Its code contains
extensive comments in order to assist in the detailed understanding. In addition, here is a brief high-level explanation
of its workings.
From the command line, the program accepts the name of
the remote host, the number of buffers to parallelize data
transfer, the UNIX command to execute on the remote host,
and the arguments required for the execution. The program
works in a remarkably simple way: it creates two instances
of class Shell – one on the local host and one on the remote host. After that, the local engine passes command
arguments to the remote instance, which executes the command. Then the local engine enters a loop that catches portions of the command’s stdout output one after another
and transfers it back to the local Shell instance. That instance dumps the data on the stdout of the local Hello
runtime engine.
All three (potentially massive) data operations of reading
data from the command stdout, transferring that data
back to the original host, and dumping it on the stdout of
the local host are performed in parallel. They involve Hello
built-in asynchronous queuing operators and passing method parameters by value. This program also utilizes other
distributed facilities such as creating instances of external
classes on remote hosts, and synchronizing local expressions via queues. Together, these first-class distributed
language features make Hello Shell simple to design and
easy to develop while the resulting binary ends up being as
efficient as its C counterpart rsh.
Page 6
September 12, 2014
Hello, C!
It is fascinating to compare the amount of code from
the Hello Shell program and the C code from the widely
known UNIX remote shell rsh [24]. After deleting empty
lines and all comments from the source code of both programs, running word count on both C rsh.c and Hello
Shell_World/Shell.hlo produces this output:
hellouser@think:~/hem$ wc rsh.c
390 1047 7606 rsh.c
hellouser@think:~/hem$ wc Shell.hlo
116 427 2929 Shell.hlo
hellouser@think:~/hem$
Clearly, Hello Shell code is more than two times smaller
than C rsh code. Obviously, some rsh code processes authentication and encryption while Shell code has no such
functionality (although one can still tunnel Hello traffic
through an SSL [22] proxy like stunnel [26]). However,
even after deleting encryption and authentication from
rsh.c, the word count still shows rsh code being significantly larger than Hello code:
hellouser@think:~/hem$ wc rsh.c.clear
243 680 4724 rsh.c.clear
hellouser@think:~/hem$
11 Language Features
A full description of the Hello programming language,
its translator, the runtime engine, and a number of working
code examples are free and available for download from
[14]. Here is the condensed digest of the language features.
The Hello distributed architecture had emerged from the
practical experience of programming diverse distributed
applications such as databases, cluster monitors, protocol
discovery, email and video servers. Because of that, Hello
design pursues the following two goals:
The general-purpose distributed language should provide superior productivity for developing a variety of
efficient and reliable distributed software.
A mainstream programmer should be able to adopt this
language in a short amount of time.
Achieving these goals could facilitate the widespread
transition from distributed tools, packages, and libraries to
this distributed language. Therefore, Hello combines the
following unique properties, which set it apart from other
distributed systems:
1) The first such property is the ability to create objects,
access their data, and invoke their methods anywhere
on the network, using simple Java-like syntax and semantics, which is also reminiscent of the RPC [23].
2) The second property is the language-embedded distributed architecture, whose elements directly represent the
graph-structured underlying networking resources.
These elements augment the single-host imperative
language constructs, which represent the linearly structured components of a von Neumann machine.
V1.0.3 (alpha)
Conclusion
3) In addition, Hello offers abstractions for distributed
operations such as queued asynchronous and synchronous requests for remote synchronization, first-class
bulk operations like group traversals and intelligent,
deep copy algorithms, as well as asynchronous queued
event processing.
Other Hello unique properties are safe data sharing in
memory partitions and automatic transfer of packages
across the network. Hello offers object-level security identifiers and privileges. It handles distributed failures via monitoring state label transitions, catching exceptions, and
timeouts. In addition, Hello translator parallelizes distributed operations for runtime efficiency. Finally, Hello packages do not run inside a virtual machine. Hello translator converts them into shared binary libraries through the intermediary C++ invocation; the C++ translator can harness the
power of its optimizer to optimize generated C++ code.
Hello translator allows for embedding into Hello programs
any C++ code and linking with any C++ library.
Hello programs are written inside object-oriented type
hierarchies organized in packages. Hello translator translates Hello programs into C++ programs and then calls a
C++ compiler to compile the generated C++ code into a 64bit binary dynamic shared library. Hello runtime engine
dynamically loads the libraries and executes them; it also
transfers the libraries on demand across the network. Engines manage memory in partitions where the programs
share runtime data. All this happens while executing concurrent threads under control of the queues. A runtime engine belongs to a host – a named collection of engines running on a single computer; any computer can run multiple
hosts.
Hello programs manipulate strongly typed primitive data,
arrays, strings, and class instances (objects). Local operations are performed on data from a thread stack, the engine
heap and host partitions; remote operations are performed
by navigating references to objects from other hosts. Method arguments and return values are transferred to any host
by reference or value following an intelligent, deep copy
algorithm. Remote and local objects assemble in groups for
automated traversals. Hosts connect and engines transfer
data, code, and control flow across the network transparently to the Hello programs. Hello programs can build neighbourhoods and paths of connected hosts and use them for
better reliability, efficiency and security.
Currently, Hello v1.0.3 (alpha) is available on 64-bit
Centos [6], Fedora [10] and Ubuntu [27] versions of the
Linux OS [19] on Intel x86_64 architecture [17].
12 Conclusion
In conclusion, we discover a historical precedent for the
emergence of the widely adopted distributed programming
languages by drawing an analogy with the development of
Page 7
September 12, 2014
FORTRAN [12], which was the first high-level programming language that had alleviated the difficulties of binary
programming. In addition, we rationalize the need for a
widely accepted distributed language by drawing parallels
between the difficulties of distributed and assembler software development [11]. Finally, we uncover the deeper
currents carrying on its waves the distributed programming
from the past -- through today -- into the future.
12.1
Historical Precedent
On the highest level, most general-purpose programming languages, in the context of a single computer, abstract the fundamental elements of software programs. Examples of these are code modules, data types, operations,
and control flow. However, these languages do not provide
first-class features for another significant programming
component – automatic transfer of and access to code modules, data types, runtime data, operations, and control flow
across the network between collaborating computers. Instead, they delegate such networking services to tools, libraries and packages outside the language proper.
In order to realize the importance of the first-class language features (i.e. those presented by language syntax and
semantics), one should recall that programming languages
serve as a bridge between a human understanding and the
hardware realization of the algorithm. This bridge is extremely valuable: it provides a much easier path for human
beings to control computers via high-level algorithmic abstractions rather than via low-level processor instructions.
Ultimately, it saves precious time and money while developing software systems.
For example, the first high-level programming language
FORTRAN developed in 1954--1957 at IBM [1], enjoyed
tremendous success because it was able to express algorithms with the imperative control flow primitives, and
centuries tested mathematical notation [15]. The language
had freed programmers from the daunting burden of programming machine codes and auto-codes. The FORTRAN
translator was able to translate program sources, check for
programming errors, and compile efficient binaries from the
program libraries without human intervention. At the same
time, programmers, empowered with FORTRAN, were able
to devote more time to solving application problems, instead of coding binary programs.
After FORTRAN invention, a plethora of programming
languages had evolved [20]. However, despite immense
richness of purposes and features, most of them share a
single property inherited from the first language – they
provide a bridge to only a single computer. Although many
traditional languages allow for developing distributed applications, they still delegate network operations to extraneous
tools, libraries and packages. This is simply because the
V1.0.3 (alpha)
Appendix – Remote Shell Source
languages lack the features to communicate over the network.
However, such extraneous software is complicated to
use since its interface, being rich in complex low-level
details, exposes the rigidity of the networking protocol to
the application level. As a result, the quality of networking
applications suffers because developers must spend significant efforts at a low-level coding in order to adhere to the
subtleties of the networking interfaces. While the language
itself has little or no knowledge of the networking operations, it is hard for a language translator to optimize network access; this, in turn, results in reduced efficiency and
higher traffic of the networking applications.
As in 1950-s, today software engineers still lack an essential tool to address an exceptionally hard problem. While
back then such problem was the low-level binary coding for
a single computer, the today’s problem is the low-level
networking coding for distributed applications. Therefore,
as the programming language FORTRAN solved the former
problem more than half a century ago by providing firstclass features to simplify single-host programming, another
programming language might also solve the latter problem
today. The solution could provide in the language the firstclass features to simplify multi-host networking operations.
12.2
Technical Parallels
Anyone who had experienced binary or assembler programming and, in addition, had been involved in a distributed software project, can think of the binary coding problems
facing programmers of early 1950’s being quite similar to
the networking coding problems of today. The following list
enumerates the exact parallels:
1) A well-defined central processing element: a CPU versus a server on the network;
2) A program that needs service from the central element:
an assembler, auto-code or binary program in memory,
executing CPU instructions versus a client program on
the network, submitting server requests;
3) A hard to use interface: assembler language, auto-code
or binary code versus networking interface;
4) A pattern of failure: crash of CPU or server usually
causes termination of assembler or client program;
5) Programming, debugging, and optimization: in both
cases, techniques are manual and complicated [11],
[28].
Obviously, it should be of no surprise that a similar solution could solve a similar problem. In other words, a distributed programming language might improve networking
programming today the same way FORTRAN had improved binary coding more than half a century ago.
Page 8
12.3
September 12, 2014
Connecting Past and Future
During millennia-long mathematical development [16]
and centuries-long technological progress [7], human beings
performed calculations individually. When needed, they
used to carry out computational and mathematical interactions through inter-personal collaboration, education, written correspondence, manuscripts, books and magazines. It is
only recently, after the invention of the computer networks
some 50 years ago [8], the automated distributed computations had become possible. Therefore, it is not surprising
that the area of distributed programming is less mature than
the field of single-host programming, which takes its roots
in the individual computational techniques evolved through
thousands of years.
One might speculate that it might take more centuries for
human civilization to develop (yet unknown) automated
paradigm for distributed computations. However, it appears
that the movement towards that goal can be started today.
Its first step should abandon a hodgepodge of single-host
languages and distributed tools in favour of widely accepted
general-purpose distributed programming languages.
Appendix – Remote Shell Source
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
// each source must begin with
// the package directive
package Shell_World;
// this class is declared external because
// its instances can be created on a remote host
// and also because once created they can
// be accessed from a remote host
external class Shell
{
// this is constructor of Shell class
external public Shell() {}
// entry point to the program -- must
// be called main; accepts command line
// arguments as an array of strings
static public int main(char [][]argv)
{
// built-in method sizear() returns
// the number of elements in an array
int argc = sizear(argv, 1);
// quit program if not enough arguments
if ( argc <= 1 )
return 0;
// first argument must be a host
// name were to execute command
// specified on the Shell command line;
// the built-in method hello() returns a
// reference (ref for short) to a host on
// the network with the given name
host hst = hello(argv[0]);
// if host not found then quit
if ( hst == null ) {
#C { cout << "host not found\n"; }
return -1;
}
// the 'create' expression is similar
// to 'new' expression, except it creates
// an instance not in the engine heap
// but in a partition from the specified
// host; if the host is remote, then an
// instance of class Shell is created
// on that host
Shell shl = create (hst) Shell();
// this line executes a method run() from
// class Shell on the just created
// remote object referred to by ref shl;
// after run() completes, this program exits
shl.run(argv, this_host);
V1.0.3 (alpha)
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
Appendix – Remote Shell Source
}
// this defines the size of a buffer for data
// to be transferred from the stdout of the
// command executed on the remote host to
// the local host -- this data will be printed
// on stdout of the program main() running on
// the local host
enum { BUFSIZE = 1024 * 1024 * 4 }
// this method executes on the remote host:
// it accepts command and its parameters
// in the character array [][]argv, and a ref to
// the remote host 'back' -- the host from where
// remote Shell command has been started
public external void run(copy char [][]argv,
host back) {
Page 9
// get the count of arguments
int argc = sizear(argv, 1);
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
if
-the first argument is the name
of this remote host;
-the second argument
specifies how many buffers to fill in
with the stdout data in parallel:
this argument is for experimentation,
just to see how the speed of data
transfer depends on the amount of
buffers being filled in parallel
-the third argument is the remote
command name
-the rest are the arguments for
the remote command
If no command is specified then quit
( argc <= 2 || back == null )
return;
// set the count of buffers from
// command line
int BUFCNT;
char []bc = argv[1];
#C { $BUFCNT = atoi($bc().__adc()); }
if ( BUFCNT < 0 )
BUFCNT = 0;
// create another Shell instance, this time
// on the source host from where the Shell
// has been launched
Shell rso
= create (back) Shell();
// on that same host create a queue
// for dumping command stdout output on
// the Shell command output in parallel
// with accepting the next portion of
// stdout from the remote host
queue rsq
= create (back) queue();
// create a queue on this host for
// sending portions of stdout back to the
// source host in parallel with accepting
// stdout from the command stdout
queue rtq
= create queue();
// create a character array and
// fill it with the command and its
// arguments using operations
// concatenation + and +=
char []line = create char[0];
for ( int i = 2; i < argc; i++ )
line += argv[i] + " ";
// also append redirection of stderr
// to make sure that any error messages
// are transferred back to the source host
line += "2>&1";
// create buffers in shared memory:
// the count of buffers 'buc' is set from
// command line, the size of each buffer
// is set from the enum BUFSIZE
int buc = BUFCNT?BUFCNT:1;
char [][]output = create char[buc][BUFSIZE];
// this is command exit code
int res
= 0;
// this holds the length of data
// in the buffer
int len
= 0;
// this indicates if all data has been
// received from the command stdout
bool done = false;
// this executes the command using C++ code:
// enclosed in the embedded block denoted
// by '#D' -- a UNIX system call popen()
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
//
//
//
//
#D
September 12, 2014
is called and the command with all
arguments is launched; a file descriptor
'pipe' is filled by this call: it is
used below to receive the command stdout
{
FILE *pipe = popen($line().__adc(),
"r");
if ( pipe == NULL ) $done = true;
}
// now the program enters a loop that reads
// command stdout; counter ‘i’ indicates
// which buffer to use; boolean 'done'
// is set inside the loop when all
// stdout data has been received
int i = 0;
while ( !done )
{
// check if all buffers have been used
if ( i >= BUFCNT )
{
// if all buffers are filled in
// by stdout data, then the
// constant expression '1' is
// evaluated on the local queue
// 'rtq' using the queued
// expression evaluation
// operator '<=>'; this operator
// places an expression at the
// end of the queue -- when
// that expression reaches the
// head of the queue then control
// returns back to the current
// thread
rtq <=> 1;
// because queue 'rtq' contained
// requests to send data to the
// remote source host, control
// reaches this point only after
// all such requests have
// finished; then the current
// buffer number is reset back
// to 0
i = 0;
}
// get next buffer and advance
// buffer count
char []outbuf = output[i++];
// remember the address of that buffer
#D { char *b = $outbuf().__adc(); }
// this is the count of bytes
// left to read into the buffer
// from command stdout
int toread = BUFSIZE;
// keep reading from command stdout
// until this buffer is full
#D {
char *o = b;
while ( $toread > 0 ) {
$len = fread(o, 1, $toread,
pipe);
$toread -= $len;
o += $len;
if ( feof(pipe) ) {
$done = true, $res = 0;
break;
}
else if ( ferror(pipe) ) {
$done = true, $res = -1;
break;
}
}
}
// this is the count of bytes read
// from the command stdout; it is
// supposed to be equal to BUFSIZE
// for all buffers except possibly
// the last one for which it can
// be less than BUFSIZE
len = BUFSIZE - toread;
//
//
//
//
//
//
//
//
//
//
if
if no buffers were ordered then
do not transfer any data back
to the source host; otherwise
send the method message named
'send()' to 'this' Shell object
on the queue 'rtq'; pass the data
buffer and the length, and the
refs to the remote Shell object
'rso' and to remote queue 'rsq',
and the completion indicator 'done'
( BUFCNT == 0 )
;
V1.0.3 (alpha)
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
Appendix – Remote Shell Source
else
rtq #> (this, send(rso, rsq, outbuf,
len, res, done));
// at this point the sent message 'send'
// begins the transfer of the
// accumulated data from 'outbuf'
// back to the source host via a
// invocation of external method
// 'rcv' which has a 'copy' parameter
// 'data' to which 'outbuf' buffer
// is passed as an argument; meanwhile,
// this loop continues to receive the
// rest of command stdout
}
// this queued constant expression
// evaluation is needed to wait until
// remote source Shell object completes
// its work, so that this target Shell
// object exits afterwards; premature exit
// of the target Shell will result in an
// exception on the source shell
rtq <=> 1;
}
// this method 'send' is declared as
// 'message' so that it can be placed on
// the queue for asynchronous execution
public message void send(Shell rso, queue rsq,
char []data, int len,
int res, bool done)
{
// after this message reaches the head
// of the queue it has been placed upon,
// it calls method 'rcv' on the remote
// source Shell instance 'rso'; the buffer
// 'data' is copied remotely by the
// runtime engine back onto the
// remote source host
rso.rcv(rsq, data, len, res, done);
}
// this method is declared 'external' because
// it is called remotely from the target host
// on a shell object from the source host;
// parameter 'data' is declared 'copy'
// for passing array of data by value across
// the host
public external void rcv(queue rsq,
copy char []data,
References
[1]
John Backus, “The history of FORTRAN I, II and III”, 1981,
http://www.softwarepreservation.org/projects/FORTRAN/paper/p25
-backus.pdf.
[2]
Andrew P. Black, Norman C. Hutchinson, Eric Jul, Henry M. Levy,
“The development of the Emerald programming language”, HOPL
III Proceedings of the third ACM SIGPLAN conference on History
of programming languages, 2007.
[3]
Boost Tutorial,
http://www.boost.org/doc/libs/1_38_0/doc/html/mpi/tutorial.html.
[4]
C (programming language),
http://en.wikipedia.org/wiki/C_(programming_language).
[5]
C++, http://en.wikipedia.org/wiki/C++.
[6]
Centos, http://www.centos.org/.
[7]
Computer History Museum, “Timeline of Computer History”,
http://www.computerhistory.org/timeline/.
[8]
Computer network (History),
http://en.wikipedia.org/wiki/Computer_network#History.
[9]
Erlang (programming language),
https://en.wikipedia.org/wiki/Erlang_(programming_language).
[10] Fedora, https://fedoraproject.org/.
[11] Anger Fog, “Optimizing subroutines in assembly language”, Technical University of Denmark, 1996-2014,
http://www.agner.org/optimize/optimizing_assembly.pdf.
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358 };
September 12, 2014
int len,
int res,
bool done)
{
// when the source host receives the
// command stdout in the 'copy' parameter
// 'data', it sends method 'dump' to
// 'this' object (which is actually 'rso'
// on the source host; method 'dump'
// actually writes command stdout from
// remote target host on the local
// source host)
rsq #> (this, dump(data, len, res, done));
//
//
//
//
if
when all data has been received, this
queued expression waits until
all output completes to avoid
abrupt interrupt of the output
( done )
rsq <=> 1;
}
// this method 'dump' just prints
// the data received from remote target host
// on stdout of the runtime engine from
// the local source host
public message void dump(char []data,
int len,
int res,
bool done) {
// output data using embedded C++ block
#C {
char *buf = $data().__adc();
int wrt = 0;
while ( $len > 0 ) {
wrt = write(1, buf, $len);
if ( wrt == -1 )
break;
len -= wrt, buf += wrt;
}
}
// if exit code indicated a failure (!=0)
// then dump that exit code
if ( done && res )
#C { cout << "exit code " << $res; }
}
[14] Hello Programming Guide for Orchestrating Local and Remote
Computations, http://www.amsdec.com/download.html.
[15] History of mathematical notation,
http://en.wikipedia.org/wiki/History_of_mathematical_notation.
[16] History of Mathematics,
http://en.wikipedia.org/wiki/History_of_mathematics.
[17] Intel 64 Architecture.
http://www.intel.com/content/www/us/en/architecture-andtechnology/microarchitecture/intel-64-architecture-general.html.
[18] Java, http://www.java.com/en/.
[19] Linux, http://www.linux.com/.
[20] Leo A. Meyerovich, Ariel S. Rabkin, “Empirical Analysis of Programming Language Adoption”, ACM SIGPLAN NOTICES –
OOPSLA 2013, V. 48-10, 2013.
[21] John von Neumann, “First Draft of a Report on the EDVAC”, 1945,
http://en.wikipedia.org/wiki/First_Draft_of_a_Report_on_the_EDV
AC.
[22] OpenSSL, https://www.openssl.org/.
[23] Remote Procedure Call,
https://en.wikipedia.org/wiki/Remote_procedure_call.
[24] rsh.c,
http://www.opensource.apple.com/source/network_cmds/network_c
mds-77/rsh.tproj/rsh.c.
[25] Salsa, http://wcl.cs.rpi.edu/salsa/.
[26] stunnel, https://www.stunnel.org/index.html.
[12] FORTRAN, http://en.wikipedia.org/wiki/FORTRAN.
[27] Ubuntu, http://www.ubuntu.com/.
[13] Getting Started Using Java RMI,
http://docs.oracle.com/javase/1.5.0/docs/guide/rmi/hello/helloworld.html.
[28] Carlos A. Varela, “Programming Distributed Computing Systems”,
MIT Press, 2013.
Page 10
| 6cs.PL
|
Cooperative Task Planning of Multi-Agent Systems Under Timed
Temporal Specifications
arXiv:1509.09137v1 [cs.SY] 30 Sep 2015
Alexandros Nikou, Jana Tumova and Dimos V. Dimarogonas
Abstract— In this paper the problem of cooperative task
planning of multi-agent systems when timed constraints are
imposed to the system is investigated. We consider timed
constraints given by Metric Interval Temporal Logic (MITL).
We propose a method for automatic control synthesis in a twostage systematic procedure. With this method we guarantee that
all the agents satisfy their own individual task specifications as
well as that the team satisfies a team global task specification.
I. I NTRODUCTION
Cooperative control of multi-agent systems has traditionally focused on designing local control laws in order to
achieve tasks such as consensus, formation, network connectivity, and collision avoidance ([1]–[7]). Over the last
decade or so, the field of control of multi-agent systems
with complicated behavior under complex high-level task
specifications has been gaining significant research attention.
Such high-level tasks may have the form of “Periodically
survey regions A, B, C while avoiding region D”, or “Visit
regions A, B, C, in this order”, and many others. Multiple
robotic vehicles then may perform these types of tasks faster
and more efficiently than a single robot. In this work, we aim
to introduce specific time bounds into the complex tasks,
such as “Periodically survey regions A, B, C, avoid region
D and always keep the longest time between two consecutive
visits to A below 5 time units”, or “Visit regions A, B, C,
in this order within 10 time units”.
The team of agents is usually associated with a set of tasks
that should be fulfilled by the group of agents as a whole
at the discrete level. A three-step hierarchical procedure
to address such a problem is described as follows ([8],
[9]): First the robot dynamics is abstracted into a finite or
countable, discrete transition system using sampling or cell
decomposition methods based on triangulations, rectangular
or other partitions. Second, invoking ideas from verification
methods, a discrete plan that meets the high-level task is
synthesized. Third, the discrete plan is translated into a
sequence of continuous controllers for the original system.
The specification language that has extensively been used
to express the tasks is the Linear Temporal Logic (LTL) (see,
e.g., [10]). LTL has proven a valuable tool for controller
synthesis, because it provides a compact mathematical formalism for specifying desired behaviors of a system. There is
a rich body of literature containing algorithms for verification
and synthesis of system obeying temporal logic specifications
([11], [12]). A common approach in multi-agent planning
under LTL specifications is the consideration of a centralized,
The authors are with the ACCESS Linnaeus Center, School of Electrical
Engineering, KTH Royal Institute of Technology, SE-100 44, Stockholm,
Sweden and with the KTH Centre for Autonomous Systems. Email:
{anikou,tumova,dimos}@kth.se. This work was supported by
the H2020 ERC Starting Grant BUCOPHSYS and the Swedish Research
Council (VR).
global task specification for the team of agents which is
then decomposed into local tasks to be accomplished by the
individual agents. For instance, the authors in [13] utilized
the parallel composition (synchronous products) of multirobot systems in order to decompose a global specification
that is given to a team of robots into individual specifications.
This method has been proven computationally expensive due
to state space explosion problem and in order to relax the
computational burden the authors in [14], [15] proposed
a method that does not require the computation of the
parallel composition. This method, however, is restrictive
to specifications that can be expressed in certain subclasses
of LTL. In [16], the specification formula was given in
LTL in parallel with the problem of minimum inter-robot
communication.
Explicit time constraints in the system modeling have been
included e.g., in [17], where a method of automated planning
of optimal paths of a group of agents satisfying a common
high-level mission specification was proposed. The mission
was given in LTL and the goal was the minimization of
a cost function that captures the maximum time between
successive satisfactions of the formula. Authors in [18],
[19] used a different approach, representing the motion of
each agent in the environment with a timed automaton. The
composition of the team automaton was achieved through
synchronization and the UPPAAL verification tool ([20]) was
utilized for specifications given in Computational Tree Logic
(CTL). In the same direction, authors in [21] modeled the
multi-robot framework with timed automata and weighted
transition systems considering LTL specifications and then,
an optimal motion of the robots satisfying instances of the
optimizing proposition was proposed.
Most of the previous works on multi-agent planning
consider temporal properties which essentially treat time in
a qualitative manner. For real applications, a multi-agent
team might be required to perform a specific task within a
certain time bound, rather than at some arbitrary time in the
future (quantitative manner). Timed specifications have been
considered in [22]–[25]. In [22], the authors addressed the
problem of designing high-level planners to achieve tasks
for switching dynamical systems under Metric Temporal
Logic (MTL) specification and in [23], the authors utilized a
counterexample-guided synthesis to cyber-physical systems
subject to Signal Temporal Logic (STL) specifications. In
[24], the MTL formula for a single agent was translated
into linear constraints and a Mixed Integer Linear Programming (MILP) problem was solved. However, these works
are restricted to single agent motion planning and are not
expendable to multi-agent systems in a straightforward way.
In [25], the vehicle routing problem was considered under the
presence of MTL specifications. The approach does not rely
on automata-based approach to verification, as it constructs
a set of linear inequalities from MTL specification formula
in order to solve an MILP problem.
In this work, we aim at designing automated planning
procedure for a team of agents that are given an individual,
independent timed temporal specification each and a single
global team specification. This constitutes the first step
towards including time constraints to temporal logic-based
multi-agent control synthesis. We consider a quantitative
logic called Metric Interval Temporal Logic (MITL) ([26])
in order to specify explicit time constraints. The proposed
solution is fully automated and completely desynchronized in
the sense that a faster agent is not required to stay in a region
and wait for the slower one. It is decentralized in handling
the individual specifications and centralized only in handling
the global team specification. To the best of the authors’
knowledge this is the first work that address the cooperative
task planning for multi-agent systems under individual and
global timed linear temporal logic specifications.
The remainder of the paper is structured as follows. In
Sec. II a description of the necessary mathematical tools,
the notations and the definitions are given. Sec. III provides
the model of the multi-agent system, the task specification,
several motivation examples as well as the formal problem
statement. Sec. IV discusses the technical details of the
solution. Sec. V is devoted to an illustrative example. Finally,
the conclusions and the future work directions are discussed
in Sec. VI.
II. N OTATION
AND
P RELIMINARIES
Given a set S, we denote by |S| its cardinality and by 2S
the set of all its subsets. An infinite sequence of elements of
S is called a infinite word over the set S and it is denoted by
w = w(0)w(1) . . . The i-th element of a sequence is denoted
with w(i). We denote by Q+ , N the set of positive rational
and natural numbers including 0, respectively. Let us also
define T∞ = T ∪ {∞} for a set of numbers T.
Definition 1. ([27]) A time sequence τ = τ (0)τ (1) · · · is a
infinite sequence of time values τ (j) ∈ T = Q+ , satisfying
the following constraints:
• Monotonicity: τ (j) < τ (j + 1) for all j ≥ 0.
• Progress: For every t ∈ T, ∃ j ≥ 1, such that τ (j) > t.
An atomic proposition p is a statement over the problem
variables and parameters that is either True (⊤) or False (⊥)
at a given time instance.
Definition 2. ([27]) Let AP be a finite set of atomic propositions. A timed word w over the set AP is an infinite sequence
w = (w(0), τ (0))(w(1), τ (1)) · · · where w(0)w(1) . . . is an
infinite word over the set 2AP and τ (0)τ (1) . . . is a time
sequence with τ (j) ∈ T, j ≥ 0. A timed language Lang T
over AP is a set of timed words over AP .
A. Weighted Transition System
Definition 3. A Weighted Transition System (WTS) is a tuple
(S, S0 , −→, d, AP, L) where S is a finite set of states; S0 ⊆ S
is a set of initial states; −→⊆ S × S is a transition relation;
d :−→−→ T is a map that assigns a positive weight to each
transition; AP is a finite set of atomic propositions; and
L : S −→ 2AP is a labeling function.
For simplicity, we use s → s′ to denote the fact that
(s, s′ ) ∈→.
Definition 4. A timed run of a WTS is an infinite sequence
rt = (r(0), τ (0))(r(1), τ (1)) . . ., such that r(0) ∈ S0 , and
for all j ≥ 1, r(j) ∈ S and r(j) → r(j + 1). The time
stamps τk (j), j ≥ 0 are inductively defined as
1) τ (0) = 0.
2) τ (j + 1) = τ (j) + d(r(j), r(j + 1)), ∀ j ≥ 1.
Every timed run rt generates a timed word w(rt ) =
(L(r(0)), τ (0)) (L(r(1)), τ (1)) . . . over the set 2AP where
w(j) = L(r(j)), ∀ j ≥ 0 is the subset of atomic propositions
that are true at state r(j) at time τ (j).
B. Metric Interval Temporal Logic and Timed Automata
The syntax of Metric Interval Temporal Logic (MITL) over
a set of atomic propositions AP is defined by the grammar
ϕ := p | ¬ϕ | ϕ1 ∧ ϕ2 |
I
ϕ | ♦I ϕ | I ϕ | ϕ1 UI ϕ2 (1)
where p ∈ AP , and
, ♦, and U is the next, future, always and until temporal operator, respectively. I ⊆
T is a non-empty time interval in one of the following
forms: [i1 , i2 ], [i1 , i2 ), (i1 , i2 ], (i1 , i2 ), [i1 , ∞], (i1 , ∞) where
i1 , i2 ∈ T with i1 < i2 . MITL can be interpreted either in
continuous or point-wise semantics. We utilize the latter one
and interpret MITL formulas over timed runs such as the
ones produced by a WTS (Def. 4).
Definition 5. ([28], [29]) Given a run rt
=
(r(0), τ (0))(r(1), τ (1)) . . . of a WTS and an MITL
formula ϕ, we define (rt , i) |= ϕ, for i ≥ 0 (read rt
satisfies ϕ at position i) as follows
(rt , i) |= p ⇔ p ∈ L(r(i))
(rt , i) |= ¬ϕ ⇔ (rt , i) 6|= ϕ
(rt , i) |= ϕ1 ∧ ϕ2 ⇔ (rt , i) |= ϕ1 and (rt , i) |= ϕ2
(rt , i) |= I ϕ ⇔ (rt , i + 1) |= ϕ and τ (i + 1) − τ (i) ∈ I
(rt , i) |= ♦I ϕ ⇔ ∃j, i ≤ j, s.t. (rt , j) |= ϕ, τ (j) − τ (i) ∈ I
(rt , i) |= I ϕ ⇔ ∀j, i ≤ j, τ (j) − τ (i) ∈ I ⇒ (rt , j) |= ϕ
(rt , i) |= ϕ1 UI ϕ2 ⇔ ∃j, i ≤ j, s.t. (rt , j) |= ϕ2 ,
τ (j) − τ (i) ∈ I and (rt , k) |= ϕ1 for every i ≤ k < j
Timed Büchi Automata (TBA) were introduced in [27] and
in this work, we also partially adopt the notation from [30],
[31]. Let X = {x1 , x2 , . . . , xM } be a finite set of clocks. The
set of clock constraints Φ(X) is defined by the grammar
φ := ⊤ | ¬φ | φ1 ∧ φ2 | x ⊲⊳ c
(2)
where x ∈ X is a clock, c ∈ T is a clock constant and
⊲⊳∈ {<, >, ≥, ≤, =}. A clock valuation is a function ν :
X → T that assigns a real value to each clock. A clock xi
has valuation νi for i ∈ {1, . . . , M }, and ν = (ν1 , . . . , νM ).
We denote by ν |= φ the fact that the valuation ν satisfies
the clock constraint φ.
Definition
6. A TBA is a tuple A
=
(S, S init , X, I, E, F, AP, L) where S is a finite set of
locations; S init ⊆ S is the set of initial locations; X is
a finite set of clocks; I : S → Φ(X) is the invariant;
E ⊆ S × Φ(X) × 2X × S gives the set of transitions; F ⊆ S
is a set of accepting locations; AP is a finite set of atomic
propositions; and L : S → 2AP labels every state with a
subset atomic propositions.
A state of A is a pair (s, ν) where s ∈ S and ν satisfies
the invariant I(s), i.e., ν |= I(s). The initial state of A is
(s(0), (0, . . . , 0)), where s(0) ∈ S0 . Given two states (s, ν)
and (s′ , ν ′ ) and an edge e = (s, γ, R, s′ ), there exists a
e
discrete transition (s, ν) −
→ (s′ , ν ′ ) iff ν satisfies the guard
of the transition γ, i.e., ν |= γ, ν ′ |= I(s′ ), and R is
the reset set, i.e., νi′ = 0 for xi ∈ R and νi′ = νi for
xi ∈
/ R. Given a δ ∈ T, there exists a time transition
δ
(s, ν) −
→ (s′ , ν ′ ) iff s = s′ , ν ′ = ν + δ and ν ′ |= I(s).
An infinite run of A starting at state (s(0), ν) is an infinite
δ0
sequence of time and discrete transitions (s(0), ν(0)) −→
δ1
e0
(s(1)′ , ν(1)′ ) . . ., where
(s(1), ν(1)) −→
(s(0)′ , ν(0)′ ) −→
(s(0), ν(0)) is an initial state. This run produces the timed
word w = (L(s(0)), τ (0))(L(s(1)), τ (1)) . . . with τ (0) = 0
and τ (i+1) = τ (i)+δi , ∀ i ≥ 1. The run is called accepting
if s(i) ∈ F for infinitely many times. A timed word is
accepted if there exists an accepting run that produces it.
The problem of deciding the emptiness of the language of a
given TBA A is PSPACE-complete [27]. In other words, we
can synthesize an accepting run of a given a TBA A, if one
exists.
Remark 1. Traditionally, the clock constraints and the TBAs
are defined with T = N, however, they can be extended to
accommodate T = Q+ ∪ {0}. By multiplying all the rational
numbers that are appearing in the state invariants and the
edge constraints with their least common multiple, we have
equivalently only natural numbers occurring to the TBA. For
the sake of physical understanding of the timed properties of
the under investigation framework, we will be working with
T = Q+ ∪ {0}.
Any MITL formula ϕ over AP can be algorithmically
translated to a TBA with the alphabet 2AP , such that the
language of timed words that satisfy ϕ is the language of
timed words produced by the TBA [26], [32], [33].
III. P ROBLEM F ORMULATION
A. System Model
Consider a multi-agent team composed by N agents
operating in a bounded workspace W0 ⊆ Rn . Let I =
{1, . . . , N } denote the index set of the agents. We assume
that the workspace W0 is partitioned into a finite number
(assume W ) of regions of interest π1 , . . . , πW where
[
W0 =
πi and πi ∩πj 6= ∅, ∀ i 6= j with i, j ∈ W (3)
i∈W
for the index set W = {1, . . . , W }. We denote by πik the
agent k being at region πi , where k ∈ I, i ∈ W. In this work,
we focus on interaction and high-level control strategies
rather than on nonlinear models, and we assume that the
dynamics of each agent is given by a single integrator
ẋi = ui , i ∈ I.
(4)
The partitioned environment (3) is a discretization that allows
us to control the agents with dynamics (4) using finite models
such as finite transition systems (e.g., [9], [34]–[36]). We
define a weighted transition system (see Def. 7) so that
• if there exists a controller ui , i ∈ I such that the agent
k can be driven from any point within the region π i to
a neighboring region π j , then we allow for a transition
πki →k πkj between the respective system states, and
• the weight of each transition estimates the time each
agent needs in order to move from one region to another.
In particular, the travel time is here determined as
the worst-case shortest time needed to travel from an
arbitrary point of the current region to the boundary of
the following region. This estimate is indeed conservative, however, it is sufficient for specifications that we
are generally interested in within multi-agent control.
Namely, it is suitable for scenarios where tasks are given
deadlines and upper rather than lower bound requirements are associated with events along the agents’ runs.
Definition 7. The motion of each agent k ∈ I in the
workspace is modeled by a WTS Tk = (Πk , Πinit
k , →k
, dk , APk , Lk ) where
k k
k
• Πk = π1 , π2 , . . . , πW
is the set of states of agent k.
Any state of an agent k can be denoted as πjk ∈ Πk for
k ∈ I, j ∈ W. The number of states for each agent is
|Πk | = W .
init
• Πk ⊆ Πk is the initial states of agent k, i.e. the set of
regions where agent k may start.
• →k ⊆ Πk × Πk is the transition relation. For example,
by π33 →3 π53 we mean that the agent 3 can move from
region π3 to region π5 .
• dk :→k → T is a map that assigns a positive weight
(duration) to each transition. For example, d2 (π22 , π52 ) =
0.7, where π22 →2 π52 , means that agent 2 needs at most
0.7 time units to move from any point of region π2 to
the boundary of the neighboring region π5 .
• AP k is a finite set of atomic propositions known to
agent k. Without loss of generality, we assume that
AP k ∩ AP k′ = ∅ for all k 6= k ′ ∈ I.
AP k
• Lk : Πk → 2
is a labeling function that assigns
to each state πjk ∈ Πk a subset of atomic propositions
APk that are satisfied when agent k is in region πj .
1) Individual Timed Runs and Words: The behaviors
of the individual agents can be captured through their
timed runs and timed words. The timed run rkt =
(rk (0), τk (0))(rk (1), τk (1)) · · · , k ∈ I of each WTS
Tk , k ∈ I and the corresponding timed words w(rkt ) =
(Lk (rk (0)), τk (0)) (Lk (rk (1)), τk (1)) · · · are defined by
using the terminology of Def. 4.
2) Collective Timed Run and Word: At the same time,
the agents form a team and we are interested in their
global, collective behaviors, which we formalize through the
following definition.
t
be individual timed runs of the
Definition 8. Let r1t , . . . , rN
agents 1, . . . , N , respectively, as defined above. Then, the
collective timed run rG = (rG (0), τG (0))(rG (1), τG (1)) . . .
of the team of agents is defined inductively as follows
1) (rG (0), τG (0)) = ((r1 (0), . . . , rN (0)), τG (0)).
2) Let (rG (i), τG (i)) = ((r1 (i1 ), . . . , rN (iN )), τG (i)),
where i ≥ 0 be the current state and time
stamp of the collective timed run. Then the next
state and time stamp (rG (i + 1), τG (i + 1)) =
((r1 (j1 ), . . . , rN (jN )), τG (i + 1)) are given by the
following
ℓ = argmin{τk (ik + 1)}.
•
τG (i + 1)(
= τℓ (iℓ + 1).
rℓ (iℓ + 1) if k = ℓ
rk (jk ) =
rk (iℓ )
if k 6= ℓ.
t
In what follows, rG
= (rG (0), τG (0))(rG (1), τG (1)) . . . ,
where rG (i) = (r1 (i1 ), . . . , rN (iN )), i, ik ≥ 0 and k ∈ I
denotes the collective timed run.
Definition 9. We define the global set of atomic propoN
[
AP k and for every state rG (i) =
sitions AP G =
k=1
(r1 (i1 ), . . . , rN (iN )) of a collective timed run, where i, ik ≥
0 and k ∈ I, we define the labeling
SN function LG :
Π1 . . . ΠN → AP G as LG (rG (i)) = k=1 Lk (rk (ik )).
t
A collective timed run rG
thus naturally produces a timed
t
word wG = (LG (rG (0)), τG (0))(LG (rG (1)), τG (1)) . . .
over AP G .
Example 1. Consider N = 2 robots operating in a
workspace with W = π1 ∪ π2 ∪ π3 , W0 = 3 and I = {1, 2}
modeled as the WTSs illustrated in Fig. 1. Let AP 1 =
{green}, and AP 2 = {red}. The labeling functions are
L1 (π11 ) = {green}, L1(π21 ) = L1 (π31 ) = ∅, and L2 (π12 ) =
L2 (π22 ) = ∅, L2 (π32 ) = {red}.
1.0
1.5
π21
2.0
π12
π31
0.5
2.0
WTS T2 :
0.5
π22
1.5
(r1 (4) = π11 , τ1 (4) = 5.0) . . .
r2t =(r2 (0) = π12 , τ2 (0) = 0.0)(r2 (1) = π22 , τ2 (1) = 2.0)
Given r1t and r2t the collective run rG is given according
to Def. 8 as follows:
Intuitively, given the current states r1 (i1 ), . . . , rN (iN ) and
the next states r1 (i1 + 1), . . . , rN (iN + 1) of the individual
agents at time τG (i), ℓ is the index of the agent k who
will finish its current transition from rℓ (iℓ ) to rℓ (iℓ + 1)
the soonest amongst all. The time of agent ℓ’s arrival to its
next state rℓ (iℓ + 1) becomes the new time stamp τG (i + 1)
of the collective timed run. The next state of the collective
timed run reflects that each agent k which cannot complete
its transition from rk (ik ) to rk (ik + 1) before τG (i + 1)
remains in rk (ik ).
π11
(r1 (2) = π31 , τ1 (2) = 2.5)(r1 (3) = π21 , τ1 (3) = 3.0)
(r2 (4) = π32 , τ2 (4) = 5.0) . . .
k∈I
WTS T1 :
r1t =(r1 (0) = π11 , τ1 (0) = 0.0)(r1 (1) = π21 , τ1 (1) = 1.0)
(r2 (2) = π32 , τ2 (2) = 2.5)(r2 (3) = π22 , τ2 (3) = 4.5)
•
•
Examples of the agents’ runs are:
π32
2.0
Fig. 1. WTSs T1 , T2 representing two agents in W. Π1 = {π11 , π21 , π31 },
Πinit
= {π11 }, Π2 = {π12 , π22 , π32 }, Πinit
= {π12 }, the transitions are
1
2
depicted as arrows which are annotated with the corresponding weights.
t
rG
=((π11 , π12 ), τG (0) = 0.0)((π21 , π12 ), τG (1) = 1.0)
| {z }
| {z }
rG (0)
((π21 , π22 ), τG (2)
| {z }
rG (2)
((π21 , π32 ), τG (4)
| {z }
rG (4)
=
rG (1)
2.0)((π31 , π32 ), τG (3)
= 2.5)
=
rG (3)
3.0)((π21 , π22 ), τG (5)
= 4.5)
| {z }
| {z }
rG (5)
((π11 , π32 ), τG (6) = 5.0) . . .
| {z }
rG (6)
The produced collective timed word is
t
wG
=({green}, 0.0)(∅, 1.0)(∅, 2.0)({red}, 2.5)
({red}, 3.0)(∅, 4.5)({green, red}, 5.0) . . . .
B. Specification
Several different logics have been designed to express
timed properties of real-time systems, such as MTL [37]
that extends the until operator of LTL with a time interval.
Here, we consider a fragment of MTL, called MITL (see
Sec. II for definition) which has been proposed in [26].
Namely, we utilize its point-wise semantics and interpret
its formulas over timed runs. Unlike MTL, MITL excludes
punctual constraints on the until operator. For instance, the
formula (a ⇒ ♦=1 b) saying that every a is followed by a b
precisely 1 time unit later, is not allowed in MITL, whereas
(a ⇒ ♦(0,1] b), saying that every a is followed by a b at
most after 1 time unit later, is. While MTL formulas cannot
be generally translated into TBAs, MITL formulas can [26].
1) Local Agent’s Specification: Each agent k, k ∈ I
is given an individual, local, independent specification in
the form of a MITL formula ϕk over the set of atomic
propositions AP k . The satisfaction of ϕk is decided from
the agent’s own perspective, i.e., on the timed run rkt .
2) Global Team Specification: In addition, the team of
agents is given a global team specification, which is a MITL
formula ϕG over the set of atomic propositions AP G . The
team specification satisfaction is decided on the collective
t
timed run rG
.
Example 1 (Continued). Recall the two agents from Example 1. Each of the agents is given a local, independent,
specification and at the same time, the team is given an
overall goal that may require collaboration or coordination. Examples of local specification formulas are ϕ1 =
♦≤10 (green) and ϕ2 = (red ⇒ ≤5 (¬red )) stating
that “The green region is periodically visited with at most 10
time units between two consecutive visits” and “Whenever a
red region is visited, it will not be visited for the following
5 time units again”, respectively. While ϕ1 is satisfied on
r1t , ϕ2 is not satisfied on r2t . An example of the global
specification is ϕG = ♦≤5 (green ∧ red ) that imposes
requirement on the agents’ collaboration; it states that agents
1 and 2 will periodically simultaneously visit the green and
the red region, respectively, with at most 5 time units between
two consecutive visits.
C. Problem Statement
Problem 1 (Run Synthesis). Given N agents governed by
dynamics as in (4), a task specification MITL formula ϕG
for the team of robots, over a set of atomic propositions
APG and N local task specifications ϕk over APk , k ∈ I,
t
synthesize a sequence of individual timed runs r1t , . . . , rN
such that the following hold
t
t
rG
|= ϕG ∧ r1t |= ϕ1 ∧ . . . ∧ rN
|= ϕN .
(5)
Though it might seem that the satisfaction of the individual
specificationsVϕ1 , . . . , ϕN can be treated as the satisfaction of
t
, this is
the formula k∈I ϕk on the collective timed run rG
generally not the case, as demonstrated through the following
example:
Example 1 (Continued). Recall the two agents
from
Example
1
and
a
local
specification
ϕ2
=
(red
⇒
≤2 (¬red )). While this
=
specification is satisfied on r2t since w(r2t )
(∅, 0.0)(∅, 2.0)({red}, 2.5)(∅, 4.5)({green, red}, 5.0) . . ., it
t
can be easily seen that it is not satisfied on rG
.
Formally, we have
^
t
t
rG
|=
ϕk < r1t |= ϕ1 ∧ . . . ∧ rN
|= ϕN .
4) We construct a global Büchi WTS TeG = TG ⊗AG . The
accepting timed runs of the TeG are the timed runs of
the TG that satisfy the team formula ϕG (Sect. IV-D).
t
5) We find an accepting timed run reG
of the global
f
Büchi WTS TG and project it onto timed runs of
the product Büchi WTS TG , then onto timed runs
of the local Büchi WTSs Te1 , . . . , TeN , and finally
t
onto individual timed runs r1t , . . . , rN
of the original
t
WTSs T1 , . . . , TN . By construction, r1t , . . . , rN
are
guaranteed to satisfy ϕ1 , . . . , ϕN , respectively, and
t
furthermore rG
satisfies ϕG (Sec. IV-E).
A. Construction of TBAs
As stated in Sec. II, every MITL formula ϕ can be translated into a language equivalent TBA. Several approaches
are proposed for that purpose, for instance [26], [32], [33],
[38]. Here, we translate each local specification ϕk , where
k ∈ I into a TBA Ak = (Sk , Skinit , Xk , Ik , Ek , Fk , APk , Lk ),
and the global specification ϕG into a TBA AG =
init
(SG , SG
, XG , IG , EG , FG , APG , LG ).
B. Construction of the local Büchi WTSs Te1 , . . . , TeN
Definition
10.
Given
a
WTS
Tk
=
,
→
,
AP
,
L
,
d
),
and
a
TBA
A
=
(Πk , Πinit
k
k
k
k
k
k
(Sk , Skinit , Xk , Ik , Ek , Fk , APk , Lk ) with Mk = |Xk |
and Ckmax being the largest constant appearing in Ak ,
we define their local Büchi WTS Tek = Tk ⊗ Ak =
e e
e
(Qk , Qinit
k , dk , Fk , APk , Lk ) as follows:
k ,
M
• Qk ⊆ {(rk , sk ) ∈ Πk × Sk : Lk (rk ) = Lk (sk )} × T∞k .
init
init
init
• Qk
= Πk × Sk × {0} × . . . × {0}.
{z
}
|
Mk products
(6)
•
k∈I
Hence, Problem 1 may not be treated in a straightforward,
fully centralized way. We propose a two-stage solution that
first pre-computes all timed runs of the individual agents in
a decentralized way and stores them efficiently in weighted
transition systems enhanced with a Büchi acceptance condition. Second, these are combined and inspected with respect
to guaranteeing the satisfaction of the team specification by
the collective timed run.
IV. P ROPOSED S OLUTION
In this section, we introduce a systematic solution to
Problem 1. Our overall approach builds on the following
steps:
1) We construct TBAs Ak , k ∈ I and AG that accept all
the timed words satisfying the specification formulas
ϕk , k ∈ I and ϕG , respectively (Sec. IV-A).
2) We construct a local Büchi WTS Tek = Tk ⊗ Ak , for
all k ∈ I. The accepting timed runs of Tek are the
timed runs of the Tk that satisfy the corresponding
local specification formula ϕk , k ∈ I (Sec. IV-B).
3) We construct a product Büchi WTS TG = Te1 ⊗· · ·⊗ TeN
such that its timed runs are collective timed runs of the
team and their projections onto the agents’ individual
timed runs are admissible by the local Büchi WTSs
Te1 , . . . TeN respectively (Sec. IV-C).
•
•
q
q ′ iff
◦ q = (r, s, ν1 , . . . , νMk ) ∈ Qk ,
′
q ′ = (r′ , s′ , ν1′ , . . . , νM
) ∈ Qk ,
k
◦ r →k r′ , and
◦ there exists γ, R, such that (s, γ, R, s′ ) ∈ Ek ,
′
ν1 , . . . , νMk |= γ, ν1′ , . . . , νM
|= Ik (s′ ), and for
k
all i ∈ {1, . . . , Mk }
0,
if xi ∈ R
ν + d (r, r′ ), if x 6∈ R and
i
k
i
νi′ =
νi + dk (r, r′ ) ≤ Ckmax
∞,
otherwise.
k
Then dek (q, q ′ ) = dk (r, r′ ).
Fek = {(rk , sk , ν1 , . . . , νMk ) ∈ Qk : sk ∈ Fk }.
e k (rk , sk , ν1 , . . . , νM ) = Lk (rk ).
L
k
Each local Büchi WTS Tek , k ∈ I is in fact a WTS with
a Büchi acceptance condition Fek . A timed run of Tek can
be written as rekt = (qk (0), τk (0))(qk (1), τk (1)) . . . using the
terminology of Def. 4. It is accepting if qk (i) ∈ Fek for
infinitely many i ≥ 0. An accepting timed run of Tek projects
onto a timed run of Tk that satisfies the local specification
formula ϕk by construction. Formally, the following lemma,
whose proof follows directly from the construction and and
the principles of automata-based LTL model checking (see,
e.g., [39]), holds:
Lemma 1. Consider an accepting timed run rekt =
(qk (0), τk (0))(qk (1), τk (1)) . . . of the local Büchi WTS Tek
defined above, where qk (i) = (rk (i), sk (i), νk,1 , . . . , νk,Mk )
denotes a state of Tek , for all i ≥ 1. The timed run rekt projects
onto the timed run rkt = (rk (0), τk (0))(rk (1), τk (1)) . . .
of the WTS Tk that produces the timed word w(rkt ) =
(Lk (rk (0)), τk (0))(Lk (rk (1)), τk (1)) . . . accepted by the
TBA Ak via its run ρk = sk (0)sk (1) . . . Vice versa, if
there exists a timed run rkt = (rk (0), τk (0))(rk (1), τk (1)) . . .
of the WTS Tk that produces a timed word w(rkt ) =
(Lk (rk (0)), τk (0))(Lk (rk (1)), τk (1)) . . . accepted by the
TBA Ak via its run ρk = sk (0)sk (1) . . . then there exist the
accepting timed run rekt = (qk (0), τk (0))(qk (1), τk (1)) . . . of
Tek , where qk (i) denotes (rk (i), sk (i), νk,1 (i), . . . , νk,Mk (i))
in Tek .
C. Construction of the product Büchi WTS TG
Now we aim to construct a finite product WTS TG whose
timed runs represent the collective behaviors of the team and
whose Büchi acceptance condition ensures that the accepting
timed runs account for the local specifications. In other
words, TG is a product of all the local WTS Tek built above.
In the construction of TG , we need to specifically handle the
cases when transitions of different agents are associated with
different time durations, i.e, different transition weights. To
this end, we introduce a vector b = (b1 , . . . , bN ) ∈ TN . Each
element of the vector is a rational number bk ∈ T, k ∈ I
which can be either 0, when the agent k has just completed
its transition, or the time elapsed from the beginning of the
agent’s current transition, if this transition is not completed,
yet. The state of the team of agents is then in the form
qG = (q1 , . . . , qN , b1 , . . . , bN , ℓ) where qk is a state of Tek ,
for all k ∈ I, and ℓ ∈ I has a special meaning in relation
to the acceptance condition of TG that will become clear
shortly. Taking the above into consideration we define the
global model TG as follows:
Definition 11. Given N local Büchi WTSs Te1 , . . . , TeN from
Def. 10, their product Büchi WTS TG = Te1 ⊗ . . . ⊗ TeN =
(QG , Qinit
G , →G , dG , FG , AP G , LG ) is defined as follows:
N
• QG ⊆ Q1 × · · · × QN × T × {1, . . . , N }.
init
init
init
• QG = Q1
× . . . × QN × {0} × . . . × {0} ×{1}.
{z
}
|
•
•
k=1
•
where dmin =
e k (qk ).
L
D. Construction of the global Büchi WTS TeG
Definition 12. Finally, given the product Büchi WTS
TG = (QG , Qinit
G , →G , dG , FG , APG , LG ), and a TBA AG =
init
(SG , SG
, XG , IG , EG , FG , APG , LG ) that corresponds to
the team specification formula ϕG with MG = |XG | and
max
CG
being the largest constant appearing in AG , we define
eG , Q
e init , G ,
their product WTS TeG = TG ⊗ AG = (Q
G
e
e
e
dG , FG , APG , LG ) as follows:
G
e G ⊆ {(q, s) ∈ QG × SG : LG (q) = LG (s)} × TM
• Q
∞ .
e init = Qinit × S init × {0} × . . . × {0} ×{1, 2}.
• Q
G
G
G
{z
}
|
MG −products
•
q
q ′ iff
◦ q = (r, s, ν1 , . . . , νMG , ℓ) ∈ QG ,
′
q ′ = (r′ , s′ , ν1′ , . . . , νM
, ℓ′ ) ∈ QG ,
G
′
◦ r →G r , and
◦ there exists γ, R, such that (s, γ, R, s′ ) ∈ EG ,
′
ν1 , . . . , νMG |= γ, ν1′ , . . . , νM
|= IG (s′ ), and for
G
all i ∈ {1, . . . , MG }
if xi ∈ R
0,
ν + d (r, r′ ), if x 6∈ R and
i
G
i
νi′ =
max
νi + dG (r, r′ ) ≤ CG
∞,
otherwise
G
◦
− bk ) is
(loosely speaking) the smallest time step that can
be applied, and
◦
(
ℓ,
if qℓ 6∈ Feℓ
′
ℓ =
((ℓ + 1) mod N ), otherwise
N
[
The product WTS TG is again a WTS with a Büchi
acceptance condition. Informally, the index ℓ in a state
qG = (q1 , . . . , qN , b1 , . . . , bN , ℓ) ∈ QG allows to project an
accepting timed run of TG onto an accepting run of every
one of the local Büchi WTS. The construction is based on
the standard definition of Büchi automata intersection (see,
e.g., [39]).
The following lemma follows directly from the construction and and the principles of automata-based LTL model
checking (see, e.g., [39]):
t
Lemma 2. For all k ∈ I, an accepting timed run rG
of the
product Büchi WTS TG projects onto an accepting timed run
rkt of the local Büchi WTS Tek that produces a timed word
w(rkt ) accepted by the corresponding TBA Ak . Vice versa,
if there exists a timed run rkt of the local Büchi WTS Tek that
produces a timed word w(rkt ) accepted by the TBA Ak for
t
each k ∈ I, then there exist an accepting timed run rG
of
TG .
′
qG →G qG
iff
◦ qG = (q1 , . . . , qN , b1 , . . . , bN , ℓ) ∈ QG ,
′
′
qG
= (q1′ , . . . , qN
, b′1 , . . . , b′N , ℓ′ ) ∈ QG ,
′′
◦ ∃ qk ∈ Qk : qk k qk′′ , for some k ∈ I,
◦
0,
if bk + dmin = dek (qk , qk′′ )
and qk′ = qk′′
b′k =
bk + dmin , if bk + dmin < dek (qk , qk′′ )
and qk′ = qk
min (dek (qk , qk′′ )
k∈{1,...,N }
LG ((q1 , . . . , qN , b1 , . . . , bN , ℓ) =
k=1
N products
•
′
Then dG (qG , qG
) = dmin .
FG = {(q1 , . . . , qN , b1 , . . . , bN , N ) ∈ QG : qN ∈ FeN }.
N
[
APk .
APG =
′
ℓ =
•
•
(
1 if ℓ = 1 and r 6∈ FG , or ℓ = 2 and s ∈ FG
2 otherwise
Then deG (q, q ′ ) = dG (r, r′ ).
FeG = {(r, s, ν1 , . . . , νMG , 1) ∈ QG : r ∈ FG }.
e G (rG , sG , ν1 , . . . , νMG ) = LG (rG ).
L
Analogously to above, the global Büchi WTS TeG is
a WTS with a Büchi acceptance condition. An accepting
timed run of TeG guarantees the satisfaction of the team
specification formula ϕG by construction. Furthermore, the
projected individual timed runs of the original T1 , . . . , TN
satisfy their respective local specifications. The following
lemma follows directly from the construction and and the
principles of automata-based LTL model checking (see, e.g.,
[39]):
t
Lemma 3. An accepting timed run reG
of the global Büchi
t
WTS TeG projects onto an accepting timed run rG
of the
t
product Büchi WTS TG that produces a timed word w(rG
)
accepted by the TBA AG . Vice versa, if there exists a timed
t
run rG
of the product Büchi WTS TG that produces a timed
t
word w(rG
) accepted by the TBA AG then there exist an
t
accepting timed run reG
of TeG .
E. Projection to the desired timed runs of T1 , . . . , TN
t
An accepting run reG
of the global Büchi WTS TeG can be
found efficiently leveraging ideas from automata-based LTL
model checking [39]. Namely, TeG is viewed as a graph that is
searched for a so-called accepting lasso; a cycle containing
an accepting state that is reachable from the initial state.
t
Once reG
is obtained, Lemmas 3, 2, and 1 directly provide
t
guidelines for projection of reG
onto the individual timed runs
t
of T1 , . . . , TN . In particular, reG
is projected onto a timed run
t
t
rG
of TG , which is projected onto timed runs re1t , . . . , reN
e
of T1 , . . . , TN , which are finally projected onto timed runs
t
r1t , . . . , rN
of T1 , . . . , TN , respectively. Such a projection
t
guarantees that r1t , . . . , rN
are a solution to Problem 1.
V. I LLUSTRATIVE E XAMPLE
For an illustrative example, consider 2 robots in the shared
workspace of Fig. 2. The workspace is partitioned into
W = 21 cells and a robot’s state is defined by the cell it is
currently present at. Agent 1 (R1) is depicted in green and
it is two times faster than Agent 2 (R2) which is depicted in
red. We assume that the environment imposes such moving
constraints that the traveling right and up is faster than left
and down. Let Agent 1 need 1 time unit for up and right
moves and 2 time units for down and left moves. Let also
Agent 2 need 2 time units for up and right moves and 4 time
units for down and left moves.
We consider a scenario where the robots have to eventually
meet at yellow regions (global team task), and at the same
time, they have to recharge within a certain time interval
in recharge locations (blue squares with the circles in the
respective color). The individual specifications are ϕ1 =
♦≤6 (recharge1 ) and ϕ2 = ♦≤12 (recharge2 ) stating that
agent 1 has to recharge within 5 time units and agent 2
within 10 units, respectively, and the team task is ϕG =
♦≤30 {(meet1A ∧meet2A )∨(meet1B ∧meet2B )} stating that the
agents have to meet either in yellow region A or B within
30 time units.
By following the process that was described in Section
t
=
IV step by step we have that an accepting timed run is reG
1
2
1
2
1
2
1
2
((π4 , π18 ), 0)((π11 , π18 ), 2) . . . ((π9 , π10 ), 6)((π16 , π3 ), 8) . . .
1
((π13
, π52 ), 13)((π61 , π62 ), 14) . . . with corresponding timed
t
2
word w(e
rG
) = (∅, 0)(∅, π18
), 2) . . . ({recharge1 }, 6)
2
A
({recharge2 }, 8) . . . (∅, π5 ), 13)(({meet A
1 , meet 2 }, 14) . . .
t
can be projected
which satisfies formula φG . The run reG
R1
π1
2
A
1 2
π14
π8
1
π21
B
R2
Fig. 2.
An illustrative example with 2 robots evolving in a common
workspace. Let W0 = π1 ∪ . . . ∪ π21 . We enumerate the regions starting
from the left region in every row and ending in the right. The initial positions
of robots R1 , R2 are depicted by a green and a red circle, respectively, the
desired meeting points in yellow and the recharging spots by the agents’
respective colors inside a blue box. The accepting runs for task specifications
φ1 , φ2 , φG are depicted with green and red arrows for agent 1 and agent
2 respectively.
1
1
onto individual the timed runs re1 t = (π41 , 0)(π11
, 2)(π10
, 4)
1
1
1
1
1
1
(π91 , 6)(π16
, 8)(π17
, 9)(π18
, 10)(π19
, 11)(π20
, 12)(π13
, 13)
2
2
2
(π61 , 14) . . . and re2 t = (π18
, 0)(π17
, 4)(π10
, 6)(π32 , 8)(π42 , 10)
2
2
(π5 , 12)(π6 , 14) . . .
(they
are
depicted
in
Fig.
2
with
green
and
red
arrows
respectively)
with
corresponding
timed
words
w(e
r1t )
=
(∅, 0)(∅, 2)(∅, 4)({recharge1 }, 6)(∅, 8)(∅, 9)(∅, 10)(∅, 11)
t
(∅, 12)(∅, 13)({meet A
1 }, 14) . . . and w(r2 ) = (∅, 0)(∅, 4)
(∅, 6)({recharge2 }, 8)(∅, 10)(∅, 12)({meet A
2 }, 14) . . .
which satisfy formulas φ1 and φ2 respectively. All
conditions from (5) are satisfied. The runs and the words of
the illustrative example are depicted in Fig. 3.
Consider now the alternative runs of the agents, where they
first meet in the meeting point A (after 8 time units) and then
recharge in the region π7 (after 9 and 10 time units, respectively). Regardless of how the agents continue, they have
accomplished the untimed formulas ϕ′1 = ♦(recharge1 ),
ϕ′2 = ♦(recharge2 ), and ϕ′G = ♦{(meet1A ∧ meet2A ) ∨
(meet1B ∧ meet2B )}. Although this is in fact a more efficient
way to satisfy the untimed formulas ϕ′1 , ϕ′2 , and ϕ′G than the
one described above, the formula ϕ1 is violated due to its
time constraint.
VI. C ONCLUSIONS AND F UTURE W ORK
We have proposed a systematic method for multi-agent
controller synthesis aiming cooperative planning under highlevel specifications given in MITL formulas. The solution
involves a sequence of algorithmic automata constructions
such that not only team specifications but also individual
specifications should be fulfilled. Future research directions
include the consideration of more complicated dynamics than
the fully actuated ones in (4), the decentralized solution such
that every agent has information only from his neighbors as
well as the modeling of the system with Markov Decision
Processes (MDPs) and probabilistic verification.
R EFERENCES
[1] W. Ren and R. Beard, “Consensus Seeking in Multi-agent Systems
under Dynamically Changing Interaction Topologies,” IEEE Transactions on automatic control, vol. 50, no. 5, pp. 655–661, 2005.
[2] R. Olfati-Saber and R. Murray, “Consensus Problems in Networks of
Agents with Switching Topology and Time-Delays,” IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1520–1533, 2004.
[3] A. Jadbabaie, J. Kin, and S. Morse, “Coordination of Groups of
Mobile Autonomous Agents Using Nearest Neighbor Rules,” IEEE
Transactions on Automatic Control, vol. 48, no. 6, pp. 988–1001, 2003.
1 , π 2 ), 0)
((π4
18
1 , π 2 ), 4)
((π10
17
1 , π 2 ), 2)
((π11
18
1 , π 2 ), 8)((π 1 , π 2 ), 10)((π 1 , π 2 ), 12)((π 1 , π 2 ), 14)
((π16
3
18
4
20
5
6
6
1 , π 2 ), 6)
((π9
10
1 , π 2 ), 9) ((π 1 , π 2 ), 11)((π 1 , π 2 ), 13)
((π17
3
9
4
13
5
Team
2
(π18
, 0)
2
(π10
, 6)
(π42 , 10)
2
(π17
, 4)
(π62 , 14)
(π32 , 8)
(π52 , 12)
Agent 2
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Agent 1
1
(π11
, 2)
(π41 , 0)
(π91 , 6)
1
(π10
, 4)
1
(π17
, 9)
1
(π16
, 8)
1
(π19
, 11)
1
(π18
, 10)
1
(π13
, 13)
1
(π20
, 12)
(π61 , 14)
t and the corresponding timed stamps. We denote with red dashed lines the times that both agents
Fig. 3. The accepting runs re1t , re2t , the collective run reG
have the same time stamps
[4] M. Zavlanos and G. Pappas, “Distributed Connectivity Control of
Mobile Networks,” IEEE Transactions on Robotics, vol. 24, no. 6,
pp. 1416–1428, 2008.
[5] M. Egerstedt and X. Hu, “Formation Constrained Multi-Agent Control,” IEEE Transactions on Robotics and Automation, vol. 17, no. 6,
pp. 947–951, 2001.
[6] H. Tanner, A. Jadbabaie, and G. Pappas, “Flocking in Fixed and
Switching Networks,” IEEE Transactions on Automatic Control,
vol. 52, no. 5, pp. 863–868, 2007.
[7] D. Dimarogonas and K. J. Kyriakopoulos, “On the Rendezvous
Problem for Multiple Nonholonomic Agents,” IEEE Transactions on
Automatic Control, vol. 52, no. 5, pp. 916–922, 2007.
[8] A. Bhatia, M. Maly, L. Kavraki, and M. Vardi, “Motion Planning with
Complex Goals,” IEEE Robotics and Automation Magazine,, vol. 18,
no. 3, pp. 55–64, 2011.
[9] H. Kress-Gazit, G. Fainekos, and G. Pappas, “Temporal-Logic-Based
Reactive Mission and Motion Planning,” Robotics, IEEE Transactions
on, vol. 25, no. 6, pp. 1370–1381, 2009.
[10] S. Loizou and K. Kyriakopoulos, “Automatic Synthesis of Multi-Agent
Motion Tasks Based on LTL Specifications,” IEEE Conference on
Decision and Control (CDC), vol. 1, pp. 153–158, 2004.
[11] M. Guo and D. Dimarogonas, “Multi-Agent Plan Reconfiguration Under Local LTL Specifications,” The International Journal of Robotics
Research, vol. 34, no. 2, pp. 218–235, 2015.
[12] S. Karaman and E. Frazzoli, “Linear Temporal Logic Vehicle Routing
with Applications to Multi-UAV Mission Planning,” International
Journal of Robust and Nonlinear Control, vol. 21, no. 12, pp. 1372–
1395, 2011.
[13] M. Kloetzer and C. Belta, “Automatic Deployment of Distributed
Teams of Robots From Temporal Logic Motion Specifications,” IEEE
Transactions on Robotics, vol. 26, no. 1, pp. 48–61, 2010.
[14] Y. Chen, X. Ding, A. Stefanescu, and C. Belta, “A Formal Approach
to Deployment of Robotic Teams in an Urban-Like Environment,”
Distributed Autonomous Robotic Systems, pp. 313–327, 2013.
[15] Y. Chen, Ding, X. Chu, A. Stefanescu, and C. Belta, “Formal Approach
to the Deployment of Distributed Robotic Teams,” IEEE Transactions
on Robotics, vol. 28, no. 1, pp. 158–171, 2012.
[16] M. Kloetzer, X. C. Ding, and C. Belta, “Multi-Robot Deployment from
LTL Specifications with Reduced Communication,” IEEE Conference
o Decision and Control (CDC), pp. 4867–4872, 2011.
[17] A. Ulusoy, S. Smith, X. Ding, C. Belta, and D. Rus, “Optimality
and Robustness in Multi-Robot Path Planning with Temporal Logic
Constraints,” The Int. Jour. of Rob. Res., vol. 32, no. 8, pp. 889–911.
[18] M. Quottrup, T. Bak, and R. Zamanabadi, “Multi-Robot Planning:
A Timed Automata Approach,” IEEE International Conference on
Robotics and Automation (ICRA), vol. 5, pp. 4417–4422, 2004.
[19] M. Andersen, R. Jensen, T. Bakand, and M. Quottrup, “Motion
Planning in Multi-Robot Systems Using Timed Automata,” 5th
IFAC/EURON, 2004.
[20] K. Larsen, P. Pettersson, and W. Yi, “UPPAAL in a Nutshell,” International Journal on Software Tools for Technology Transfer (STTT),
vol. 1, no. 1, pp. 134–152, 1997.
[21] A. Ulusoy, S. Smith, X. Ding, C. Belta, and D. Rus, “Optimal MultiRobot Path Planning with Temporal Logic Constraints,” 2011 IROS,
pp. 3087–3092, 2011.
[22] J. Liu and P. Prabhakar, “Switching Control of Dynamical Systems
from Metric Temporal Logic Specifications,” 2014 ICRA, pp. 5333–
5338, 2014.
[23] V. Raman, A. Donzé, D. Sadigh, R. Murray, and S. Seshia, “Reactive
Synthesis from Signal Temporal Logic Specifications,” pp. 239–248.
[24] Y. Zhou, D. Maity, and J. S. Baras, “Optimal Mission Planner with
Timed Temporal Logic Constraints,” European Control Conference
(ECC), 2015.
[25] S. Karaman and E. Frazzoli, “Vehicle Routing Problem with Metric
Temporal Logic Specifications,” IEEE Conference on Decision and
Control (CDC), pp. 3953–3958, Dec 2008.
[26] R. Alur, T. Feder, and T. A. Henzinger, “The Benefits of Relaxing
Punctuality,” Journal of the ACM (JACM), vol. 43, no. 1, pp. 116–
146, 1996.
[27] R. Alur and D. Dill, “A Theory of Timed Automata,” Theoretical
Computer Science, vol. 126, no. 2, pp. 183–235, 1994.
[28] D. D. Souza and P. Prabhakar, “On the Expressiveness of MTL in
the Pointwise and Continuous Semantics,” International Journal on
Software Tools for Technology Transfer, vol. 9, no. 1, pp. 1–4, 2007.
[29] J. Ouaknine and J. Worrell, “On the Decidability of Metric Temporal
Logic,” IEEE Symposium on Logic in CS, pp. 188–197, 2005.
[30] P. Bouyer, “From Qualitative to Quantitative Analysis of Timed
Systems,” Mémoire D ’ habilitation, Université Paris, vol. 7, pp. 135–
175, 2009.
[31] S. Tripakis, “Checking Timed Buchi Automata Emptiness on Simulation Graphs,” ACM TCL, vol. 10, no. 3, 2009.
[32] O. Maler, D. Nickovic, and A. Pnueli, “From MITL to Timed
Automata,” Formal Modeling and Analysis of Timed Systems, pp. 274–
289, 2006.
[33] T. Brihaye, M. Estivenart, and G. Geeraerts, “On MITL and Alternating Timed Automata,” Formal Modeling and Analysis of Timed
Systems, vol. 8053, pp. 47–61, 2013.
[34] P. Tabuada, Verification and Control of Hybrid Systems: a Symbolic
Approach. Springer Science and Business Media, 2009.
[35] A. Girard and G. J. Pappas, “Approximation Metrics for Discrete
and Continuous Systems,” IEEE Transactions on Automatic Control,
vol. 52, no. 5, pp. 782–798, 2007.
[36] R. Alur, T. Henzinger, G. Lafferriere, and G. Pappas, “Discrete
Abstractions of Hybrid Systems,” Proceedings of the IEEE, vol. 88,
no. 7, pp. 971–984, 2000.
[37] R. Koymans, “Specifying Real-Time Properties with Metric Temporal
Logic,” Real-time systems, vol. 2, no. 4, pp. 255–299, 1990.
[38] D. Ničković and N. Piterman, “From MTL to Deterministic Timed
Automata,” Formal Modeling and Analysis of Timed Systems, pp. 152–
167, 2010.
[39] C. Baier and J. Katoen, Principles of Model Checking. MIT Press,
2008.
| 3cs.SY
|
Splitting schemes with respect to physical processes for
double-porosity poroelasticity problemsI
A.E. Kolesova , P.N. Vabishchevichb,∗
arXiv:1601.06368v1 [cs.CE] 24 Jan 2016
b Nuclear
a North-Eastern Federal University, 58, Belinskogo, 677000 Yakutsk, Russia
Safety Institute, Russian Academy of Sciences, 52, B. Tulskaya, 115191 Moscow, Russia
Abstract
We consider unsteady poroelasticity problem in fractured porous medium within
the classical Barenblatt double-porosity model. For numerical solution of doubleporosity poroelasticity problems we construct splitting schemes with respect to physical processes, where transition to a new time level is associated with solving separate
problem for the displacements and fluid pressures in pores and fractures. The stability
of schemes is achieved by switching to three-level explicit-implicit difference scheme
with some of the terms in the system of equations taken from the lower time level and
by choosing a weight parameter used as a regularization parameter. The computational
algorithm is based on the finite element approximation in space. The investigation of
stability of splitting schemes is based on the general stability (well-posedness) theory
of operator-difference schemes. A priori estimates for proposed splitting schemes and
the standard two-level scheme are provided. The accuracy and stability of considered
schemes are demonstrated by numerical experiments.
Keywords: poroelasticity, double-porosity, operator-difference schemes, splitting
scheme, regularization
1. Introduction
Some aquifer and reservoir systems are formed by fractured rocks, which contain
both pores and interconnected fractures. The fluid flow in such porous medium is most
frequently modeled by the Barenblatt double-porosity model [4, 6], where we consider
two diffusion processes coupled by an exchange term. Also, the pressures of fluid in
both pores and fractures contribute to deformation of porous medium. Therefore, we
incorporate Biot poroelasticity model [8] into Barenblatt model.
The mathematical model of double-porosity poroelasticity problem includes the
Lame equation for the displacement and two nonstationary parabolic equations for the
I This
work was supported by the Russian Foundation for Basic Research (projects 14-01-00785).
author.
Email addresses: [email protected] (A.E. Kolesov), [email protected] (P.N.
Vabishchevich)
∗ Correspondibg
Preprint submitted to arXiv.org
March 15, 2018
fluid pressures in pores and fractures [7, 40]. Note that the displacement equation contains body forces, which are proportional to the gradient of pressures. In turn, the pressure equations includes dilation terms described by the divergence of the displacement
velocity.
The numerical solution of poroelasticity problems is usually based on a finite element approximation in space [33, 38]. It is well-known that in order to satisfy LBBcondition (inf-sup condition) [13] we must use mixed finite elements, where the order
of approximation for pressure is lower then the order of approximation for displacement. The finite element formulation of poroelasticity problem with double-porosity is
presented in [23].
For the time approximation the standard two-level finite difference schemes with
weights are used. In [11] the stability and convergence of two-level schemes for
double-porosity poroelasticity problems with finite-difference approximations in space
were analyzed using the general stability (correctness) theory of operator-difference
schemes [34, 35]. The numerical implementation of two-level schemes is associated
with solving coupled system of equations, which requires the use of special algorithms
[1, 19] or additive operator-difference schemes (splitting schemes) [30, 36].
Splitting schemes are designed for the efficient computational realization of various
unsteady problems by switching to a chain of simpler problems. The simplest example
of such schemes is splitting with respect to spatial variables (alternating direction methods). Regionally additive schemes or domain decomposition methods are widely used
for computations on parallel computers. For poroelasticity problems we use splitting
with respect to physical processes, when the transition to a new time level is performed
by sequentially solving separate problems for displacement and pressure.
Number of splitting schemes for poroelasticity problems are constructed and used
in [3, 14, 15, 22, 24, 25, 27, 31, 39]. The analysis of stability and performance of
these schemes is usually conducted using methodical computations without a theoretical study. Rigorous mathematical results concerning the stability of splitting schemes
for poroelasticity problems are described in [37].
In this work, using Samarskii’s regularization principle for operator-difference schemes we construct splitting schemes with respect to physical processes for doubleporosity poroelasticity problems. The paper is organized as follows. After a brief
description of the mathematical problem in Section 2, we introduce the variational formulation of problem in Section 3. Here, we obtain a priori estimate that ensures the
stability of the solution of problem. Discretization in space and time are performed in
Sections 4 and 5, respectively. Splitting schemes are constructed in Sections 6 and 7.
Numerical experiments that test the convergence and stability of considered schemes
are shown in Section 8. Conclusions follow.
2. Mathematical Model
We consider double-porosity poroelasticity problems arising when a fluid flows
through homogeneous and isotropic fractured medium where fractures and porous matrix represent two overlapping continua [4, 6]. In these continua two fluid flow processes with two different pressure fields occur. Therefore, the poroelasticity equations
2
can be expressed as
β1
∂p
∂t
β2
∂p
∂t
div σ(u) − α1 grad p1 − α2 grad p2 = 0,
!
∂ div u
k1
+ α1
− div
grad p1 + γ(p1 − p2 ) = f1 (x, t),
∂t
η
!
∂ div u
k2
+ α2
− div
grad p2 + γ(p2 − p1 ) = f2 (x, t).
∂t
η
(1)
(2)
(3)
Here, u is the displacement vector, p1 is the fluid pressure in pores, and p2 is fluid
pressure in fractures. The stress tensor σ is defined by the expression
σ = 2µε(u) + λ div uI,
where µ is the shear modulus, λ are the Lame coefficient, I is the unit tensor, and ε is
the strain tensor:
1
ε = (grad u + grad uT ).
2
The other notation is as follows: α is Biot coefficients, β = 1/M, M is the Biot modulus,
k1 , k2 are the permeability tensors, η is the viscosity of the fluid, γ is the exchange
parameter, and f1 , f2 are the functions describing given fluid sources (sinks). The
subscripts 1 and 2 represent notation associated with pores and fractures, respectively.
The system (1)–(3) is considered in a bounded domain Ω with a boundary Γ, on
which, for simplicity, we set the following homogeneous conditions for displacements
u = 0,
x ∈ ΓD ,
σn = 0,
x ∈ ΓN .
(4)
For pressures we set
p1 = 0,
x ∈ ΓN ,
−
k1 ∂p1
= 0,
η ∂n
x ∈ ΓD ,
(5)
k2 ∂p2
= 0, x ∈ ΓD .
(6)
η ∂n
Here, n is the unit normal to the boundary, Γ = ΓD + ΓN . In addition, we specify the
initial conditions for pressures as
p2 = 0,
x ∈ ΓN ,
p1 (x, 0) = s1 (x),
−
p2 (x, 0) = s2 (x),
x ∈ Ω.
(7)
The initial-boundary problem (1)–(7) for coupled parabolic and elliptic equations is the
basis for considering fluid flow in deformable fractured porous medium.
3. Variational Formulation
For the numerical solution, we use finite element approximation in space [12, 26],
so we need obtain a variational formulation of problem (1)–(7). For scalar quantities,
let us introduce the Hilbert space L2 (Ω) with the scalar product and norm given as
Z
(u, v) =
u(x)v(x), kuk = (u, u)1/2 .
Ω
3
For vector quantities, we use L2 (Ω) = [L2 (Ω)]m , where m = 2, 3 is the dimension of
domain Ω. Let H 1 (Ω) and H1 (Ω) be the Sobolev spaces. Next, we define the subspaces
of scalar and vector functions
Q = {q ∈ H 1 (Ω) : q(x) = 0, x ∈ ΓD },
V = {v ∈ H1 (Ω) : v(x) = 0, x ∈ ΓD }.
After multiplying (1), (2), and (3) by test functions v ∈ V and q1 , q2 ∈ Q, respectively, and integrating by parts to eliminate the second order derivatives, we come to
the following variational problem: Find u ∈ V, p1 , p2 ∈ Q such that
a(u, v) + α1 g(p1 , v) + α2 g(p2 , v) = 0, ∀v ∈ V,
(8)
!
∂u
∂p1
, q1 + α1 d
, q1 + b1 (p1 , q1 ) + γ(p1 − p2 , q1 ) = ( f1 , q1 ), ∀q1 ∈ Q, (9)
c1
∂t
∂t
!
!
∂p2
∂u
c2
, q2 + α2 d
, q2 + b2 (p2 , q2 ) + γ(p2 − p1 , q2 ) = ( f2 , q2 ), ∀q2 ∈ Q. (10)
∂t
∂t
!
Here, the bilinear forms are defined as
a(u, v) =
g(p, v) =
Z
σ(u)ε(v) dx,
Ω
Z
Ω
d(u, q) =
grad p v dx,
c1 (p, q) = β1
Z
Ω
Z
b2 (p, q) =
Z
Ω
Ω
Ω
c2 (p, q) = β2
p q dx,
b1 (p, q) =
Z
div u q dx,
Z
p q dx,
Ω
k1
grad p grad q dx,
η
k2
grad p grad q dx.
η
The initial conditions (7) are set as:
(p1 (0), q1 ) = (s1 , q1 ),
(p2 (0), q2 ) = (s2 , q2 ),
q1 , q2 ∈ Q.
(11)
The bilinear form a(·, ·) is symmetric and positive definite:
a(u, v) = a(v, u), a(u, u) ≥ δa kuk2 , δa > 0, u, v ∈ V.
Taking into account this, the form is associated with a Hilbert space Ha with the following inner product and norm:
(u, v)a = a(v, u),
kuka = (a(u, u))1/2 .
The norm of the adjoint of the space Ha is denoted by kuk∗,a . Note, that forms b1 (·, ·),
b2 (·, ·), c1 (·, ·), and c2 (·, ·) are also symmetric and positive definite.
4
According to the Green’s formula, we have
Z
Z
Z
div u p dx = −
grad p u dx + un p dx.
Ω
Ω
Γ
Then, in view of the boundary conditions (4)–(6), forms g(·, ·) and d(·, ·) are related as
d(u, p) = − g(p, u),
p ∈ Q,
u ∈ V.
Now, we derive the simplest a priori estimates for the solution of problem (8)–(11).
Setting v = ∂u/∂t in (8), q1 = p1 in (9), and q2 = p2 in (10), summing up these
equations and taking into account that, for example,
!
∂u
1d
a u,
=
a(u, u),
∂t
2 dt
we obtain
1d
a(u, u) + c1 (p1 , p1 ) + c2 (p2 , p2 ) + b1 (p1 , p1 ) + b2 (p2 , p2 )
2 dt
+ γkp1 − p2 k2 = ( f1 , p1 ) + ( f2 , q2 ).
For the right-hand side we use the following inequalities
1
( fl , pl ) ≤ kpl k2bl + k fl k2∗,bl ,
4
l = 1, 2.
Thus, we have
1
d
(kuk2a + kp1 k2c1 + kp2 k2c2 ) ≤
k f1 k2∗,b1 + k f1 k2∗,b1 .
dt
2
Using the initial condition (11), we compute the initial displacement u0 (x)
a(u0 , v) + α1 g(s1 , v) + α2 g(s2 , v) = 0,
v ∈ V.
Integration with respect to time gives the estimate:
|uk2a + kp1 k2c1 + kp2 k2c2 ≤ ku0 k2a + ks1 k2c1
Z
1 t
+ ks2 k2c2 +
k f1 (ξ)k2∗,b1 + k f2 (ξ)k2∗,b2 dξ.
2 0
(12)
using the notation f (·, t) = f (t).
A priori estimate (12) ensures the stability of the solution of problem (8)– (11) with
respect to the initial data and the right-hand side. Similar estimates can be obtained for
the solution of discrete problem.
5
4. Finite Element Approximation
Now, we approximate our problem in space using finite element methods. First, we
construct a computational mesh Ωh = {ω1 , ω2 , . . . , ωN } of domain Ω. Here, N is the
number of cells ω, h = maxω∈Ωh hω , where hω is the diameter of circle inscribed in a
cell ω.
Then, in the mesh we define spaces of conformal finite elements for scalar and
vector functions Qh ⊂ Q and Vh ⊂ V and restrict the variational problem (8)–(11) to
these spaces: find uh ∈ Vh and p1,h , p2,h ∈ Qh such that
a(uh , v) + α1 g(p1,h , v) + α2 g(p2,h , v) = 0, ∀v ∈ Vh ,
!
!
∂p1,h
∂uh
, q1 + α1 d
, q1 + b1 (p1,h , q1 )
c1
∂t
∂t
c2
+ γ(p1,h − p2,h , q1 ) = ( f1 , q1 ), q1 ∈ Qh ,
!
!
∂p2,h
∂uh
, q2 + α2 d
, q2 + b2 (p2,h , q2 )
∂t
∂t
(13)
(14)
(15)
+ γ(p2,h − p1,h , q2 ) = ( f2 , q2 ), q2 ∈ Qh ,
(p1,h (0), q1 ) = (s1 , q1 ),
(p2,h (0), q2 ) = (s2 , q2 ), q1 , q2 ∈ Qh .
(16)
For further consideration, it is convenient to use the operator formulation of problem (13)– (16). We define finite-dimensional operators A, B1 , B2 , C1 , C2 , D, G related
to the corresponding bilinear forms by setting, for example,
(Auh , v) = a(uh , v),
∀uh , v ∈ Vh .
As a result, we come from problem (13)–(16) to the following Cauchy problem for a
system of equations:
Auh + α1G p1,h + α2G p2,h = 0,
(17)
d p1,h
duh
+ α1 D
+B1 p1,h + γE(p1,h − p2,h ) = f1,h ,
dt
dt
d p2,h
duh
C2
+ α2 D
+B2 p2,h + γE(p2,h − p1,h ) = f2,h ,
dt
dt
p1,h (0) = s1,h , p2,h (0) = s2,h .
C1
(18)
(19)
(20)
where E is the identity operator and
( fl,h , ql ) = ( fl , ql ), (sl,h , q1 ) = (sl , q), ∀ql ∈ Qh , l = 1, 2.
The operator A, B1 , B2 , C1 and C2 are self-adjoint and positive definite, for example,
A = A∗ > 0,
while D and G are adjoint with to each other with an opposite sign:
D = −G∗ .
6
Note that for the operator formulation (17)–(20) we can derive a discrete analogue
of estimate (12):
kuh k2A +kp1,h kC2 1 + kp2,h kC2 2 ≤ ku0,h k2A + ks1,h kC2 1 + ks2,h kC2 2
Z
1 t
+
k f1,h (ξ)k2∗,B1 + k f2,h (ξ)k2∗,B2 dξ.
2 0
(21)
Now, introducing a vector of pressures: ph = {p1,h , p2,h }, from (17)–(20) we obtain
Auh + G ph = 0,
C
(22)
d ph
duh
+D
+ B ph = fh ,
dt
dt
ph (0) = sh .
(23)
(24)
Here
G = α1 G
C
C= 1
0
!
α1 D
D=
,
α2 D
α2 G ,
!
0
B + γE
, B= 1
C2
−γE
!
−γE
,
B2 + γE
(25)
and
fh = { f1,h , f2,h },
sh = {s1,h , s2,h }.
The operators C and B are self-adjoint and positive definite, and G∗ = −D in the
sense of the equality (G ph , uh ) = −( ph , Duh ). Then, the solution of problem (22)–(24)
satisfies the a priori estimates
Z
1 t
kuh k2A + kph k2C ≤ ku0,h k2A + ksh k2C +
k fh (ξ)k2∗,B dξ.
2 0
Note that this estimate is similar to the apriori estimates (12), (21).
5. Time Discretisation
For the discretization in time we use a uniform grid with a step τ > 0. Let un =
uh (x, tn ), pn = ph (x, tn ), where tn = nτ, n = 0, 1, . . . . To obtain an approximate solution
of problem (17)–(20) we use the standard two-level scheme with weights:
Aun+1 + G pn+1 = 0,
C
pn+1 − pn
un+1 − un
+D
+ B pn+1
= fθn+1 ,
θ
τ
τ
(26)
(27)
where
pn+1
= θ pn+1 + (1 − θ)pn ,
θ
fθn+1 = fh (tθn+1 )
and 0 ≤ θ ≤ 1. The initial condition is written as
( p0 , q) = (s, q).
7
(28)
Theorem 1. For θ ≥ 0.5 the solution of problem (26)–(28) satisfies the a priori estimate
τ
kun+1 k2A + kpn+1 k2C ≤ kun k2A +kpn k2C + k fθn+1 k2B−1 .
2
(29)
un+1 − un
and (27) by pn+1
θ , we get
τ
!
!
n+1
− un
un+1 − un
n+1 u
+
G
p
,
= 0,
Aun+1
,
θ
θ
τ
τ
!
!
pn+1 − pn n+1
un+1 − un n+1
n+1
n+1 n+1
C
, pθ + D
, pθ + (B pn+1
θ , pθ ) = ( fθ , pθ ),
τ
τ
Proof. In view if linearity, multiplying (26) by
where
un+1
= θun+1 + (1 − θ)un .
θ
Summing up these equations, we obtain
!
!
n+1
− un
pn+1 − pn n+1
n+1 u
n+1
n+1 n+1
Auθ ,
+ C
, pθ + (B pn+1
θ , pθ ) = ( fθ , pθ ).
τ
τ
Taking into account inequality
1 n+1 2
n+1 2
( fθn+1 , pn+1
θ ) ≤ kpθ k B + k fθ k B−1 ,
4
gives
Aun+1
θ ,
!
!
pn+1 − pn n+1
1
un+1 − un
+ C
, pθ
≤ || fθn+1 ||2B−1 .
τ
τ
4
Using the equality
vn+1
= θvn+1 + (1 − θ)vn =
θ
!
1 n+1
vn+1 + vn
+ θ−
(v − vn )
2
2
and taking into account that any self-adjoint operator A satisfies
(A(u + v), u − v) = (Au, u) − (Av, v),
we obtain
1
1 n+1 2
||u ||A − ||un ||2A + || pn+1 ||2C − || pn ||2C
2
2
!
τ
1 n+1
+ θ−
||u − un ||2A + ||pn+1 − pn ||2C ≤ || fθn+1 ||2B−1 .
2
4
For θ ≥ 0.5, this yields the required estimate (29).
8
Note that in the numerical implementation of weighted scheme (26)–(28) un+1 ,
and pn+1
are determined at each time level by solving the coupled system:
2
pn+1
1 ,
Aun+1 + α1G p1n+1 + α2G pn+1
= 0,
2
= ψn+1
+ α1 Dun+1 + θτ(B1 + γE)p1n+1 − θτγE pn+1
C1 pn+1
1 ,
2
1
C2 pn+1
+ α2 Dun+1 + θτ(B2 + γE)p2n+1 − θτγE pn+1
= ψn+1
2
1
2 ,
with the corresponding ψn+1
and ψn+1
1
2 . Various special numerical algorithms can be
used to solve this system [1, 19].
Another opportunity is to construct splitting schemes:
Aun+1 = ϕn+1
u ,
C1 pn+1
+ θτB1 pn+1
= ϕn+1
p1 ,
1
1
C2 pn+1
+ θτB2 pn+1
= ϕn+1
p2 ,
2
2
where the transition to a new time level involves the solution of separate equations for
displacements and pressures.
6. Incomplete Splitting Scheme
To construct splitting schemes, we express u = uh from (22) and substitute the
result into (23), which leads to a single equation for p = ph
− DA−1 G
dp
dp
+C
+ Bp = f.
dt
dt
This equation can be written as
ep = f ,
edp + A
B
dt
(30)
e = B, and the operator B
e is equal to sum of two self-adjoint, positive definite
where A
operators:
e=B
e0 + B
e1 , B
e0 = C, B
e1 = −DA−1 G.
B
(31)
eB
e are self-adjoint and positive definite.
The operators A,
Under the following constraint
e1 ≤ γB B
e0 ,
B
γB > 0,
(32)
for the numerical solution of problem (24), (30), we can use a thee-level explicitimplicit scheme with weights [18]:
n
n
n−1
n
n−1
n+1
e0 θ p − p + (1 − θ) p − p
e1 p − p + A
epn+1 = f n+1 .
+B
B
τ
τ
τ
9
(33)
To calculate the first step, we can apply the two-level scheme
p0 = p0 ,
1
0
ep − p + A
ep1 = f 1 .
B
τ
(34)
The value of θ is determined by the stability conditions for difference scheme (33),
(34).
Taking into account equalities
pn+1 − pn
pn+1 − pn−1 τ pn+1 − 2pn + pn−1
,
=
+
τ
2τ
2
τ2
pn − pn−1
pn+1 − pn−1 τ pn+1 − 2pn + pn−1
=
−
,
τ
2τ
2
τ2
pn+1 − pn
+ pn ,
τ
we can write the scheme (33) in the canonical form of three-level operator-difference
schemes
pn+1 = τ
n+1
ep
C
− pn−1 e pn+1 − 2 pn + pn−1 e n
+D
+ A p = f n.
2τ
τ2
(35)
e e
where C,
D are the self-adjoint and positive definite operators given by
e=B
e + τ A,
e
C
2
τ
e
e0 − B
e1 + τ A.
e
D=
(2θ − 1) B
2
2
(36)
Our subsequent analysis is based on the following general statement from the stability (well-posedness) theory of three-level operator-difference schemes [34, 35].
Lemma 1. For
τ2 e
e
D > A,
4
scheme (35) is unconditionally stable and its solution satisfies the estimate
En+1 ≤ En +
where
En =
pn + pn−1
2
τ n+1
f
2
2
+
e
A
2
e−1
C
pn − pn−1
τ
(37)
,
2
.
2
e
e
D− τ4 A
Then, stability condition (37) holds if
e0 − B
e1 ≥ 0.
(2θ − 1) B
In view of inequalities (32), this is achieved if we choose
2θ ≥ 1 + δ,
10
(38)
where δ is determined from (32) as the maximum eigenvalue of spectral problem δ =
νmax :
e1 p = ν B
e0 p.
B
Taking into account (31) yields
− DA−1 G p = νC p.
We set u = −A−1 G p and come to the problem
Au + G p = 0,
Du = νC p.
Excluding p, we have
GC−1 Du + νAu = 0.
Taking into account (25), we get the Cosserat spectrum problem [16, 32]
(α21 + α22 )GC−1 Du + νAu = 0.
(39)
In principle, the maximum eigenvalue in (39) depends on the physical parameters of
the differential problem (αi , βi , i = 1, 2, µ, λ) and the domain, but not on the time steps.
Combining (31) and (33), we obtain the following incomplete splitting scheme:
Aun+1 + G pn = 0,
(40)
pn+1 − pn
pn − pn−1
un+1 − un
(41)
Cθ
+ (1 − θ)
+D
+ B pn+1 = f n+1 ,
τ
τ
τ
where the solution is determined by solving separately the elasticity problem (40) and
the double porosity problem (41). The stability of this splitting scheme can be described by the following theorem:
Theorem 2. Scheme (40), (41) is unconditionally stable for 2θ ≥ 1 + δ, where δ = νmax
is the maximum eigenvalue of spectral problem (39).
7. Full Splitting Scheme
In the splitting scheme with respect to physical processes (40), (41) on each time
level we solve two separate problems: the displacement problem (40) and the coupled
problem for pressures in pores and fractures (41). It makes sense to build a splitting
scheme, where we have separate problems for pressures as well.
e as
We can split the scheme (30) further separating the diagonal part of operator A
follows
e= A
e0 + A
e1 ,
A
(42)
where
e0 = B1 + γE
A
0
!
0
,
B2 + γE
e1 =
A
0
−γE
!
−γE
.
0
Under the assumption
e0 − A
e1 ≥ 0
A
11
(43)
e we can use the following
which follows from positive-definiteness of the operator A,
explicit-implicit scheme
n
n
n−1
n+1
e0 θ p − p + (1 − θ) p − p
B
τ
τ
n
n−1
p
−
p
e1
e0 pn+1 + A
e1 pn = f n+1 .
+B
+A
τ
(44)
Similar to (33) scheme (44) can be written in the canonical form of three-level
operator difference schemes (35) with
e=B
e + τA
e0 ,
C
2
τ
e0 − B
e1 + τ A
e0 ,
e
(2θ − 1) B
D=
2
2
and the analysis of the resulting scheme can be based on Lemma 1. In this case, the
stability condition (37) takes the form
e0 − B
e1 + τ(2 A
e0 − A)
e > 0.
2 (2θ − 1) B
Taking into account (42), (43), the stability condition becomes 2θ ≥ 1 + δ, where δ is
determined from (39).
Finally, combining (31), (42), and (44), we have the full splitting scheme:
Aun+1 + α1G pn1 + α2G pn2 = 0,
α1 D
un+1 − un
+ B1 pn+1
+ γ(pn+1
− pn2 )
1
1
τ
pn+1 − pn1
pn − pn−1
1
+ C1 θ 1
+ (1 − θ) 1
= f1n+1 ,
τ
τ
(45)
(46)
un+1 − un
+ B2 pn+1
+ γ(pn+1
− pn1 )
2
2
τ
(47)
pn+1 − pn2
pn − pn−1
2
+ (1 − θ) 2
= f2n+1 .
+ C2 θ 2
τ
τ
Here, the solution is determined by solving each equation separately. We can formulate
a similar theorem, which describes the stability of this scheme:
α2 D
Theorem 3. Splitting scheme (45)–(47) is unconditionally stable for 2θ ≥ 1 + δ, where
δ = νmax is the maximum eigenvalue of spectral problem (39).
8. Numerical Experiment
We consider a two-dimensional double-porosity poroelasticity problem in a unit
square domain Ω subjected to a surface load g as in Fig. 1. At some centered upper
part of the domain Γ1 the load g = −sin(πt)n is applied in strip of length 0.2 m. The
remaining of the top boundary Γ2 is traction free. On the vertical boundaries Γ3 we set
zero horizontal displacement and vertical surface traction. The bottom of the domain
Γ4 is assumed to be fixed, i.e. the displacement vector is taken as zero. For pressures,
12
we prescribe both pore and fracture pressure to be zero at Γ2 , while the remaining
boundaries are impermeable. More precisely, the boundary conditions are given as
follows
p1 = 0, p2 = 0, σn = 0, x ∈ Γ1 ,
∂p2
∂p1
= 0,
= 0, σn = g, x ∈ Γ2 ,
∂n
∂n
∂p1
∂p2
= 0,
= 0, (σn) × n = 0, un = 0, x ∈ Γ3 ,
∂n
∂n
∂p2
∂p1
= 0,
= 0, u = 0, x ∈ Γ4 .
∂n
∂n
Similar mathematical model is commonly used to test computational algorithms for
numerical solution of poroelasticity problems [9, 10, 17].
Γ2
Γ2
Γ1
Γ3
Ω
Γ3
Γ4
Figure 1: Computational domain
To analyze the considered numerical schemes and to identify the dependence of the
stability of splitting schemes from the problem parameters we use three sets of input
parameters presented in Table 1. For convenience, we vary only values of the β1 and
β2 , and other parameters are the same for all sets.
For the numerical solution we use four computational meshes of different quality
with the local refinement in the area of application of the traction. The numbers of
vertices, cells, and degrees of freedom (dof) of u, p1 , p2 , and w = {u, p1 , p2 } for each
mesh are given in Table 2. Fig 2 shows the coarse mesh 1. Here, the quadratic vector
element and the linear scalar element are used for discretization of the displacement
and pressures, respectively. Therefore, the number of dofs for the the displacement u
is much higher than the dofs number for the pressures p1 and p2 .
The numerical implementation is performed using the DOLFIN library, which is
a part of the FEniCS project for automated solution of differential equations by finite
element methods [2, 28, 29]. The computational meshes are generated using the Gmsh
13
Table 1: Problem properties
Parameter
η
µ
λ
β1
β2
k1
k2
α1
α2
γ
Unit
Pa·s
MPa
MPa
GPa−1
GPa−1
10−15 m2
10−15 m2
−
−
10−10 kg/(m·s)
Set 1
0.001
4.2
2.4
54
14
6.18
27, 2
0.95
0.12
5
Set 2
0.001
4.2
2.4
108
24
6.18
27.2
0.95
0.12
5
Set 3
0.001
4.2
2.4
216
48
6.18
27.2
0.95
0.12
5
software [20]. To solve systems of linear equations we use LU solver provided by
PETSc [5].
Table 2: Parameters of meshes
Vertices
Cells
Number of u dof
Number of p1 dof
Number of p2 dof
Number of w dof
mesh 1
881
1654
6830
881
881
8592
mesh 2
3409
6604
26842
3409
3409
33660
mesh 3
13398
26370
106330
13398
13398
133126
mesh 4
52366
103884
417230
52366
52366
521962
Fig. 3 – 5 show the displacement and pressures in pores and fractures, respectively,
at time t = 0.5 s, which are obtained using the two-level scheme (26)–(28) with θ = 1.0
on the finest mesh 4 and time step τ = 0.0025 s. The displacement is presented on the
deformed domain (overstated for visual contrast). We see that the pressure in fractures
declines more rapidly than in pores. These results will be used as etalon solutions ue ,
p1,e , and p2,e to evaluate errors as
εu = ||ue − u||,
ε p1 = ||p1,e − p1 ||,
ε p2 = ||p2,e − p2 ||.
We analyze the dependence of the accuracy of two-level scheme with weights (26)–
(28) from computational parameters: mesh and time step sizes. On this stage, we
employ only the first set of input parameters (Table. 1), because they do not have an
impact on the computational character of two-level scheme.
Fig. 6 illustrates the dynamics of errors of the pressure in pores ε p1 for different
meshes 1, 2, and 3 with τ = 0.0025 s, while errors for different time steps τ = 0.02,
0.01, and 0.005 s on mesh 3 are shown in Fig. 7. The weight θ = 1.0 is used. We
14
Figure 2: Computational mesh 1
observe the convergence of solution when increasing quality of mesh and reducing
time step τ.
Now, we conduct experiments to demonstrate the efficiency of incomplete splitting
scheme (40)–(41). First, using the SLEPs library [21] we solve the eigenvalue problem
(39) and find the value of δ = νmax and estimate the minimum value of the weight θ
using (38), when three-level scheme is unconditionally stable. The results are presented
in Table 3. Note that the eigenvalues νmax are virtually independent from mesh size and
depend only on the problem properties.
In Figs. 8 – 10 we present the dynamics of errors ε p1 for different values of θ for
sets 1, 2, and 3, respectively. In all cases, if the condition (38) is not satisfied, then the
Table 3: Parameter δ and weight θ
δ
θ
Set 1
2.49
1.75
Set 2
1.25
1.12
15
Set 3
0.62
0.81
Figure 3: Distribution of displacement
Figure 4: Distribution of pressures in pores
16
Figure 5: Distribution of pressures in fractures
Figure 6: Coupled scheme: comparison of the errors ε p1 for different meshes
solution is unstable. When the value of the weight is taken according to Table 3, the
solution is regularized.
17
Figure 7: Coupled scheme: comparison of the errors ε p1 for different time steps
Table 4: Solve time for different meshes and numerical schemes
mesh 1
mesh 2
mesh 3
mesh 4
Full splitting
scheme
6.894
27.578
154.824
866.452
Incomplete
splitting scheme
7.045
28.452
165.958
1033.207
Coupled
scheme
12.684
55.892
305.932
2488.546
Next, we consider the full splitting scheme (45)–(47). The dynamics of errors of
the pressure in fractures ε p2 for different meshes and time steps are shown in Figs. 11
and 12. Here, we use the first set of input parameters (Table 1) and θ = 1.8. We also
see good convergence of solution.
Finally, in Table 4 we present the dependence of solve time of coupled scheme (26),
(27), incomplete splitting scheme (40), (41), and full splitting scheme (45)–(47) from
mesh size. Here, the first set of input parameters and τ = 0.005 are used. We see that
splitting schemes are significantly faster than coupled scheme. When mesh is small,
the solve times of the incomplete and full splitting schemes differ little from each other.
When mesh is big, the full splitting scheme is much faster than incomplete scheme.
9. Conclusion
In this paper, we constructed unconditionally stable three-level splitting schemes
with weights for numerical solution of double-porosity poroelasticity problems. The
18
Figure 8: Incomplete splitting scheme: comparison of the errors ε p1 for different θ and Set 1
Figure 9: Incomplete Splitting scheme. Comparison of the errors ε p1 for different θ and Set 2
19
Figure 10: Incomplete splitting scheme: comparison of the errors ε p1 for different θ and Set 3
Figure 11: Full splitting scheme: comparison of the errors ε p2 for different meshes
analysis was based on the general theory of stability and correctness of operatordifference schemes. The finite element method was used approximation in space. The
efficiency of considered schemes were verified by numerical experiments.
20
Figure 12: Full splitting scheme: comparison of the errors ε p2 for different time steps
References
[1] Alexsson, O., Blaheta, R., Byczanski, P.: Stable discretization of poroelastiicty problems and efficient preconditioners for arising saddle point type matrices.
Comput. Vissualizat. in Sci. 15(4), 191–2007 (2013)
[2] Alnæs, M.S., Blechta, J., Hake, J., Johansson, A., Kehlet, B., Logg, A., Richardson, C., Ring, J., Rognes, M.E., Wells, G.N.: The fenics project version 1.5.
Archive of Numerical Software 3(100), 9–23 (2015)
[3] Armero, F., Simo, J.C.: A new unconditionally stable fractional step method for
non-linear coupled thermomechanical problems. International Journal for Numerical Methods in Engineering 35(4), 737–766 (1992)
[4] Bai, M., Elsworth, D., Roegiers, J.C.: Multiporosity/Multipermeability Approach
to the Simulation of Naturally Fractured Reservoirs. Water Resources Research
(29), 1621–1633 (1993)
[5] Balay, S., Abhyankar, S., Adams, M.F., Brown, J., Brune, P., Buschelman, K.,
Dalcin, L., Eijkhout, V., Gropp, W.D., Kaushik, D., Knepley, M.G., McInnes,
L.C., Rupp, K., Smith, B.F., Zampini, S., Zhang, H.: PETSc Web page. http:
//www.mcs.anl.gov/petsc (2015)
[6] Barenblatt, G.I., Zheltov, I.P., Kochina, I.N.: Basic concepts in the theory of
seepage of homogeneous liquids in fissured rocks [strata]. Journal of Applied
Mathematics and Mechanics 24(5), 1286–1303 (1960)
21
[7] Beskos, D.E., Aifantis, E.C.: On the theory of consolidation with double porosityII. International Journal of Engineering Science 24(11), 1697–1716 (1986)
[8] Biot, M.A.: General theory of three dimensional consolidation. Journal of Applied Physics 12(2), 155–164 (1941)
[9] Boal, N., Gaspar, F.J., Lisbona, F.J., Vabishchevich, P.N.: Finite-difference analysis of fully dynamic problems for saturated porous media. Journal of Computational and Applied Mathematics 236(6), 1090–1102 (2011)
[10] Boal, N., Gaspar, F.J., Lisbona, F.J., Vabishchevich, P.N.: Finite-difference analysis for the linear thermoporoelasticity problem and its numerical resolution by
multigrid methods. Mathematical Modelling and Analysis 17(2), 227–244 (2012)
[11] Boal, N., Gaspar, F.J., Lisbona, F.J., Vabishchevich, P.N.: Finite difference analysis of a double-porosity consolidation model. Numerical Methods for Partial
Differential Equations 28(1), 138–154 (2012)
[12] Brenner, S.C., Scott, L.R.: The mathematical theory of finite element methods.
Springer, New York (2008)
[13] Brezzi, F., Fortin, M.: Mixed and hybrid finite element methods. Springer (1991)
[14] Bukač, M., Yotov, I., Zakerzadeh, R., Zunino, P.: Partitioning strategies for the
interaction of a fluid with a poroelastic material based on a nitsches coupling
approach. Computer Methods in Applied Mechanics and Engineering 292, 138–
170 (2015)
[15] Bukač, M., Yotov, I., Zunino, P.: An operator splitting approach for the interaction
between a fluid and a multilayered poroelastic structure. Numerical Methods for
Partial Differential Equations 31(4), 1054–1100 (2015)
[16] Bychenkov, Y.V., Chizhonkov, E.V.: Iterative methods for solving saddle problems (in Russian). Binom, Moscow (2010)
[17] Gaspar, F.J., Gracia, J.L., Lisbona, F.J., Vabishchevich, P.N.: A stabilized method
for a secondary consolidation biot’s model. Numerical Methods for Partial Differential Equations 24(1), 60–78 (2008)
[18] Gaspar, F.J., Grigoriev, A.V., Vabishchevich, P.N.: Explicit-implicit splitting
schemes for some systems of evolutionary equations. Internat. J. of Num. Analys.
Model. 11(2), 346–357 (2014)
[19] Gaspar, F.J., Lisbona, F.J.: An efficient multigrid solver for a reformulated version of the poroelasticity system. Computer Methods in Applied Mechanics and
Engineering 196(8), 1447–1457 (2007)
[20] Geuzaine, C., Remacle, J.F.: Gmsh: A 3-d finite element mesh generator with
built-in pre-and post-processing facilities. International Journal for Numerical
Methods in Engineering 79(11), 1309–1331 (2009)
22
[21] Hernandez, V., Roman, J.E., Vidal, V.: SLEPc: A scalable and flexible toolkit for
the solution of eigenvalue problems. ACM Trans. Math. Software 31(3), 351–362
(2005)
[22] Jha, B., Juanes, R.: A locally conservative finite element framework for the simulation of coupled flow and reservoir geomechanics. Acta Geotechnica 2(3), 139–
153 (2007)
[23] Khaled, M.Y., Beskos, D.E., Aifantis, E.C.: On the theory of consolidation with
double porosity–III A finite element formulation. International journal for numerical and analytical methods in geomechanics 8(2), 101–123 (1984)
[24] Kim, J., Tchelepi, H., Juanes, R.: Stability and convergence of sequential methods
for coupled flow and geomechanics: Drained and undrained splits. Computer
Methods in Applied Mechanics and Engineering 200(2324), 2094 – 2116 (2011)
[25] Kim, J., Tchelepi, H., Juanes, R.: Stability and convergence of sequential methods
for coupled flow and geomechanics: Fixed-stress and fixed-strain splits. Computer Methods in Applied Mechanics and Engineering 200(1316), 1591 – 1606
(2011)
[26] Knabner, P., Angermann, L.: Numerical methods for elliptic and parabolic partial
differential equations. Springer Verlag, New York (2003)
[27] Kolesov, A.E., Vabishchevich, P.N., Vasileva, M.V.: Splitting schemes for poroelasticity and thermoelasticity problems. Computers & Mathematics with Applications 67(12), 2185–2198 (2014)
[28] Logg, A., Mardal, K.A., Wells, G.N., et al.: Automated Solution of Differential
Equations by the Finite Element Method. Springer (2012)
[29] Logg, A., Wells, G.N.: Dolfin: Automated finite element computing. ACM Transactions on Mathematical Software 37(2) (2010)
[30] Marchuk, G.I.: Splitting and alternating direction methods. In: P. Ciarlet, J.L. Lions (eds.) Handbook of Numerical Analysis, pp. 197–462. North-Holland (1990)
[31] Mikelić, A., Wheeler, M.F.: Convergence of iterative coupling for coupled flow
and geomechanics. Computational Geosciences 17(3), 455–461 (2012)
[32] Mikhlin, S.G.: The spectrum of a family of operators in the theory of elasticity.
Russian Mathematical Surveys 28(3), 45 (1973)
[33] Phillips, P., Wheeler, M.: Overcoming the problem of locking in linear elasticity
and poroelasticity: an heuristic approach. Computational Geosciences 13(1), 5–
12 (2008)
[34] Samarskii, A.A.: The theory of difference schemes. Marcel Dekker, New York
(2001)
23
[35] Samarskii, A.A., Matus, P.P., Vabishchevich, P.N.: Difference schemes with operator factors. Kluwer (2002)
[36] Vabishchevich, P.N.: Additive operator-difference schemes. Splitting schemes. de
Gruyter (2014)
[37] Vabishchevich, P.N., Vasileva, M.V., Kolesov, A.E.: Splitting scheme for poroelasticity and thermoelasticity problems. Computational Mathematics and Mathematical Physics 54(8), 1305–1315 (2014)
[38] Wheeler, M., Xue, G., Yotov, I.: Coupling multipoint flux mixed finite element
methodswith continuous galerkin methods for poroelasticity. Computational Geosciences 18(1), 57–75 (2013)
[39] Wheeler, M.F., Gai, X.: Iteratively coupled mixed and galerkin finite element
methods for poro-elasticity. Numerical Methods for Partial Differential Equations
23(4), 785–797 (2007)
[40] Wilson, R.K., Aifantis, E.C.: On the theory of consolidation with double porosity.
International Journal of Engineering Science 20(9), 1009–1035 (1982)
24
| 5cs.CE
|
Model Selection as a Multiple Testing Procedure:
Improving Akaike’s Information Criterion
Adrien Saumard
Fabien Navarro
CREST, CNRS-ENSAI, Université Bretagne Loire
arXiv:1803.02078v2 [math.ST] 13 Apr 2018
Abstract
By interpreting the model selection problem as a multiple hypothesis testing task, we propose
a modification of Akaike’s Information Criterion that avoids overfitting, even when the sample
size is small. We call this correction an over-penalization procedure. As a proof of concept, we
show nonasymptotic optimality of our procedure for histogram selection in density estimation,
by establishing sharp oracle inequalities for the Kullback-Leibler divergence. A strong feature
of our theoretical results is that they include the estimation of unbounded log-densities. To do
so, we prove several analytical and probabilistic lemmas that are of independent interest. In an
experimental study, we also demonstrate state-of-the-art performance of our over-penalization
procedure for bin size selection.
MSC 2010 subject classifications: 62G07, 62G09, 62G10.
Key words and phrases: model selection, multiple testing, AIC, maximum likelihood, density
estimation, bin size, over-penalization, small sample size.
1
Introduction
Since its introduction by Akaike in the early seventies [1], the celebrated Akaike’s Information
Criterion (AIC) is an essential tool for the statistician and its use is nearly systematic in problems
of model choices and estimator selection for prediction. When selecting among estimators or models
constructed from finite degrees of freedom, AIC more precisely recommends maximizing the loglikelihood of the estimators penalized by their corresponding degrees of freedom. This procedure
has found pathbreaking applications in density estimation, regression, time series or neural networks
analysis, to name but a few ([26]). Due to its simplicity and negligible computational cost—whenever
the estimators are given—,it is also far from being outdated and continues to serve as one of the
most useful devices for model selection in high dimensional statistics. For instance, it can be used
to efficiently tune the Lasso ([45]).
Any substantial and principled improvement of AIC is likely to have a significant impact on the
practice of model choices and we bring in this paper an efficient and theoretically grounded solution
to the problem of overfitting that can occur when using AIC on samples of small to moderate sample
sizes.
The fact that AIC tends to be unstable and thus is perfectible in case of small sample sizes is
well-known from practitioners and has been noticed for a long time. Suguira [42] and Hurvich and
Tsai [29] have proposed the so-called AICc (for AIC corrected), which tends to penalize more than
AIC. However, the derivation of AICc comes from an asymptotic analysis where the dimension of
the models are considered to be fixed compared to the sample size. In fact, such an assumption
does not fit the usual practice of model selection, where the largest models are of dimensions close
to the sample size. Another drawback related to AICc is that it has been legitimated through
a mathematical analysis only in the linear regression model and for autoregressive models ([26]).
However, to the best of our knowledge, outside these frameworks, there is no theoretical ground for
using AICc.
Building on considerations coming from the general nonasymptotic theory of model selection
developed during the nineties (see for instance [9] and [35]) and in particular building on the analysis
1
2
Saumard and Navarro
of Castellan [24], Birgé and Rozenholc [17] have considered a modification of AIC specifically designed
for the selection of the bin size in histogram selection for density estimation. Indeed, results of [24]—
and more generally results of [9]—advocate to take into account in the design of penalty the number
of models to be selected. The importance of the cardinality of the collection of models for model
selection is in fact a very general phenomenon and one of the main outcomes of the nonasymptotic
model selection theory. In the bin size selection problem, this corresponds to adding a small amount
to AIC. Unfortunately, the theory does not specify uniquely the term to be added to AIC. On
the contrary, infinitely many corrections are accepted by the theory and in order to choose a good
one, intensive experiments were carried in [17]. The resulting correction of AIC has thus still the
drawback to be specifically designed for the task on which it has been tested.
We propose a general approach, based on a multiple testing formulation of the model selection
task, that surpasses the limits of the Unbiased Risk Estimation principle. The latter principle is
indeed at the core of Akaike’s model selection procedure and is more generally the main principle
of model selection, underlying procedures such as Stein’s Unbiased Risk Estimator (SURE, [40]) or
cross-validation ([6]). We point out that, from the multiple testing point of view that we develop,
unbiased risk estimation does not allow to control the sizes of the considered tests. In a nutshell, it is
more efficient to estimate a quantile of the risk of the estimators–the level of the quantile depending
on the size of the collection of models–than its mean. This is thus a new, very general model selection
principle that we put forward. We call it an over-penalization procedure, because it systematically
implies to add small terms to classical penalties such as AIC.
Our procedure amounts to building an estimator from multiple tests. From this perspective,
our work shares strong connections with recent breakthroughs obtained in robust estimation by
designing estimators from tests ([15], [16], [8], [31]). However, our focus and our perspectives are
noticeably different from this exiting line of works. Indeed, estimators such as T-estimators ([15])
or ρ-estimators ([8]) are quasi-universal estimator in the sense that they have very strong statistical
guarantees, but they have the drawback to be very difficult—if feasible—to compute. Here, we rather
keep the tractability of AIC procedure, but we don’t look particularly at robustness properties (for
instance against outliers). Our focus is on an improvement of AIC for small and moderate sample
sizes, that we explain in terms of the control of the sizes of our tests.
Lets us now detail our contributions.
• We propose a general formulation of the model selection task for prediction in terms of a
multiple testing procedure (understood in a non-classical way to be detailed in Section 2.3.2),
thus making a link between two major topics of contemporary research. In particular, we put
forward a generic property that the collection of tests should satisfy in order to ensure an
oracle inequality for the selected model. We call this property the “transitivity property” and
show that it generalizes penalization procedures together with T-estimators and ρ-estimators.
• Considering the problem of estimating a density by selecting an histogram, we prove a sharp
and fully nonasymptotic oracle inequality for our procedure. Indeed, we describe a control of
Kullback-Leibler (KL) divergence—also called excess risk—of the selected histogram as soon
as we have one observation. We highlight that this very strong feature may not be possible
by considering AIC. We also put on emphasis that up to our knowledge our oracle inequality
is the first nonasymptotic result comparing the KL divergence of the selected model to the
KL divergence of the oracle in an unbounded setting. Indeed, oracle inequalities in density
estimation are generally expressed in terms of Hellinger distance—which is much easier to
handle than the KL divergence, because it is bounded—for the selected model.
• In order to prove our oracle inequality, we improve upon the previously best known concentration inequality for the chi-square statistics (Castellan [24], Massart [35]) and this allows us to
gain an order of magnitude in the control of the deviations of the excess risks of the estimators.
Our result on the chi-square statistics is general and of independent interest.
• We also prove new Bernstein-type concentration inequalities for log-densities that are unbounded. Again, these probabilistic results, that are naturally linked to information theory,
are general and of independent interest.
• We generalize previous results of Barron and Sheu [10] related to the existence of margin rela-
Model Selection as a Multiple Testing Procedure
3
tions in maximum likelihood estimation (MLE). Indeed, related results of [10] where established
under boundedness of the log-densities and we extend them to unbounded log-densities with
moment conditions.
• Finally, from a practical perspective, we bring a nonasymptotic improvement of AIC that has,
in its simplest form, the same computational cost as AIC. Furthermore, we show that our
over-penalization procedure largely outperforms AIC on small and moderate sample sizes, but
also outperforms existing corrections of AIC such as AICc or Birgé-Rozenholc’s procedure.
Let us conclude this introduction by detailing the organization of the paper.
We present our over-penalization procedure in Section 2. More precisely, we detail in Sections 2.1
and 2.2 our model selection framework related to MLE via histograms. Then in Section 2.3 we define
formally over-penalization procedures and highlight their generality. We explain then from three
different angles the ideas underlying over-penalization: estimation of the ideal penalty, multiple
testing and a geometric point of view.
Section 3 is devoted to the statistical guarantees related to over-penalization. In particular, as
concentration properties of the excess risks are at the core of the design of an over-penalization, we
detail them in Section 3.1. Then we derive a general and sharp oracle inequality in Section 3.2 and
highlight the theoretical benefits compared to an analysis of AIC.
New mathematical tools that are of probabilistic and analytical nature and of independent interest are presented in Section 4. Section 5 contains the experiments, with the practical procedures
detailed. We consider two different practical variants of over-penalization and compare them to
existing penalization procedures. The superiority of our method is in particular transparent.
The proofs are gathered in Section 6, as well as in the Appendix A which provides further
theoretical developments that extend the description of our over-penalization procedure.
2
Statistical Framework and Notations
The over-penalization procedure, described in Section 2.3, is legitimated at a heuristical level in
a generic setting of selection of M-estimators. We put to emphasis in Section 2.1 a maximum
likelihood estimation (MLE) setting since, as a proof of concept for our over-penalization procedure,
our theoretical and experimental results will address the case of bin size selection for maximum
likelihood histogram selection in density estimation. In order to be able to discuss in Section 2.3
the generality of our approach in an M-estimation setting, our presentation of MLE brings notations
that directly extend to M-estimation with a general contrast.
2.1
Maximum Likelihood Density Estimation
We are given n independent observations (ξ1 , . . . , ξn ) with unknown common distribution P on a
measurable space (Z, T ). We assume that there exists a known probability measure µ on (Z, T )
such that P admits a density f∗ with respect to µ: f∗ = dP/dµ. Our goal is to estimate the density
f∗ .
R
For
an
integrable
function
f
on
Z,
we
set
P
f
=
P
(f
)
=
Z f (z) dP (z) and µf = µ (f ) =
R
Pn
f
(z)
dµ
(z).
If
P
=
1/n
δ
denotes
the
empirical
distribution
associated to the sample
n
ξ
i=1 i
Z
P
(ξ1 , . . . , ξn ), then we set Pn f = Pn (f ) = 1/n ni=1 f (ξi ). Moreover, taking the conventions ln 0 =
−∞, 0 ln 0 = 0 and defining the positive part as (x)+ = x ∨ 0, we set
Z
S = f : Z −→ R+ ;
f dµ = 1 and P (ln f )+ < ∞ .
Z
We assume that the unknown density
f∗ belongs to S.
R
Note that since P (ln f∗ )− = f∗ ln f∗ 1f∗ ≤1 dµ < ∞, the fact that f∗ belongs to S is equivalent
to ln(f∗ ) ∈ L1 (P ), the space of integrable functions on Z with respect to P .
We consider the MLE of the density f∗ . To do so, we define the maximum likelihood contrast γ
to be the following functional,
γ : f ∈ S 7−→ (z ∈ Z 7−→ − ln (f (z))) .
4
Saumard and Navarro
Then the risk P γ (f ) associated to the contrast γ on a function f ∈ S is the following,
P γ (f ) = P (ln f )− − P (ln f )+ ∈ R ∪ {+∞} .
Also, the excess risk of a function f with respect to the density f∗ is classically given in this
context by the KL divergence of f with respect to f∗ . Recall that for two probability distributions
Pf and Pt on (Z, T ) of respective densities f and t with respect to µ, the KL divergence of Pt with
respect to Pf is defined to be
(R
R
dPf
f
ln
dP
=
f
ln
g
dPg
t dµ if Pf Pg
Z
Z
K (Pf , Pg ) =
∞
otherwise.
By a slight abuse of notation we denote K (f, t) rather than K (Pf , Pt ) and by the Jensen inequality
we notice that K (f, t) is a nonnegative quantity, equal to zero if and only if f = t µ-a.s. Hence, for
any f ∈ S, the excess risk of a function f with respect to the density f∗ satisfies
Z
f∗
P (γ(f )) − P (γ (f∗ )) =
ln
f∗ dµ = K (f∗ , f ) ≥ 0
(1)
f
Z
and this nonnegative quantity is equal to zero if and only if f∗ = f µ − a.s. Consequently, the
unknown density f∗ is uniquely defined by
f∗ = arg min {P (− ln f )}
f ∈S
= arg min {P γ (f )} .
f ∈S
For a model m, that is a subset m ⊂ S, we define the maximum likelihood estimator on m, whenever
it exists, by
fˆm ∈ arg min {Pn γ (f )}
f ∈m
( n
)
1X
= arg min
− ln (f (ξi )) .
f ∈m n
(2)
i=1
2.2
Histogram Models
The models m that we consider here to define the maximum likelihood estimators as in (2) are made
of histograms defined on a fixed partition of Z. More precisely, for a finite partition Λm of Z of
cardinality |Λm | = Dm + 1, Dm ∈ N, we set
(
)
X
X
Dm +1
m= f =
βI 1I ; (βI )I∈Λm ∈ R+
, f ≥ 0 and
βI µ (I) = 1 .
I∈Λm
I∈Λm
Note that the smallest affine space contained in m is of dimension Dm . The quantity Dm can thus
be interpreted as the number of degrees of freedom in the (parametric) model m. We assume that
any element I of the partition Λm is of positive measure with respect to µ: for all I ∈ Λm , µ (I) > 0
. As the partition Λm is finite, we have P (ln f )+ < ∞ for all f ∈ m and so m ⊂ S. We state in
the next proposition some well-known properties that are satisfied by histogram models submitted
to the procedure of MLE (see for example [35, Section 7.3]).
Proposition 2.1 Let
fm =
X P (I)
1I .
µ (I)
I∈Λm
Then fm ∈ m and fm is called the KL projection of f∗ onto m. Moreover, it holds
fm = arg min P (γ(f )) .
f ∈m
Model Selection as a Multiple Testing Procedure
5
The following Pythagorean-like identity for the KL divergence holds, for every f ∈ m,
K (f∗ , f ) = K (f∗ , fm ) + K (fm , f ) .
(3)
We also have the following formula
fˆm =
X Pn (I)
1I ,
µ (I)
I∈Λm
and so the maximum likelihood estimator on m is well-defined and corresponds to the so-called
frequencies histogram associated to the partition Λm .
Remark 2.2 Histogram models are special cases of general exponential families exposed for example
in Barron and Sheu [10] (see also Castellan [24] for the case of exponential models of piecewise
polynomials). The projection property (3) can be generalized to exponential models (see [10, Lemma
3] and Csiszár [27]).
Remark 2.3 As by (1) we have
P (γ(fm ) − γ (f∗ )) = K (f∗ , fm )
and for any f ∈ m,
P (γ(f ) − γ (f∗ )) = K (f∗ , f )
we easily deduce from (3) that the excess risk on m is still a KL divergence, as we then have for any
f ∈ m,
P (γ(f ) − γ(fm )) = K (fm , f ) .
2.3
Over-Penalization
Let us now define our model selection procedure. We provide three ways to understand the benefit
of over-penalization. Of course the three point of views are related to each other but they bring
different and complementary enlightening on the behavior of over-penalization.
2.3.1
Over-Penalization as Estimation of the Ideal Penalty
We are given a collection of histogram models denoted Mn , with finite cardinality
n depending on
o the
ˆ
sample size n, and its associated collection of maximum likelihood estimators fm ; m ∈ Mn . By
taking a (nonnegative) penalty function pen on Mn ,
pen : m ∈ Mn 7−→ pen (m) ∈ R+ ,
the output of the penalization procedure (also called the selected model) is by definition any model
satisfying,
n
o
m
b ∈ arg min Pn (γ(fˆm )) + pen (m) .
(4)
m∈Mn
We aim at selecting an estimator fˆm
b with a KL divergence, pointed on the true density f∗ , as small
as possible. Hence, we want our selected model to have a performance as close as possible to the
excess risk achieved by an oracle model (possibly non-unique), defined to be,
n
o
m∗ ∈ arg min K(f∗ , fˆm )
(5)
m∈Mn
n
o
= arg min P (γ(fˆm )) .
(6)
m∈Mn
From (6), it is seen that an ideal penalty in the optimization task (4) is given by
penid (m) = P (γ(fˆm )) − Pn (γ(fˆm )) ,
6
Saumard and Navarro
since in this case, the criterion critid (m) = Pn (γ(fˆm )) + penid (m) is equal to the true risk P (γ(fˆm )).
However penid is unknown and, at some point, we need to give some estimate of it. In addition,
penid is random, but we may not be able to provide a penalty, even random, whose fluctuations at
a fixed model m would be positively correlated to the fluctuations of penid (m). This means that
we are rather searching for an estimate of a deterministic functional of penid . But which functional
would be convenient? The answer to this question is essentially contained in the solution of the
following problem.
Problem 1. For any fixed β ∈ (0, 1) find the deterministic penalty penid,β : Mn → R+ , that
minimizes the value of C, among constants C > 0 which satisfy the following oracle inequality,
n
o
≥1−β .
(7)
P K(f∗ , fˆm
K(f∗ , fˆm )
b ) ≤ C inf
m∈Mn
The solution—or even the existence of a solution—to the problem given in (7) is not easily accessible
and depends on assumptions on the law P of data and on approximation properties of the models,
among other things. In the following, we give a reasonable candidate for penid,β . Indeed, let us set
βM = β/Card(Mn ) and define
n
o
penopt,β (m) = q1−βM P (γ(fˆm ) − γ(fm )) + Pn (γ(fm ) − γ(fˆm )) ,
(8)
where qλ {Z} = inf {q ∈ R; P (Z ≤ q) ≥ λ} is the quantile of level λ for the real random variable
Z. Our claim is that penopt,β gives in (7) a constant C which is close to one, under some general
assumptions (see Section 3 for precise results). Let us explain now why penopt,β should lead to a
nearly optimal model selection.
We set
o
\ n
P (γ(fˆm ) − γ(fm )) + Pn (γ(fm ) − γ(fˆm )) ≤ penopt,β (m) .
Ω0 =
m∈Mn
We see, by definition of penopt,β and by a simple union bound over the models m ∈ Mn , that the
event Ω0 is of probability at least 1 − β. Now, by definition of m,
b we have, for any m ∈ Mn ,
Pn (γ(fˆm
b ≤ Pn (γ(fˆm )) + penopt,β (m) .
b )) + penopt,β (m)
(9)
By centering by P (γ (f∗ )) and using simple algebra, Inequality (9) can be written as,
h
i
ˆ
ˆ
P (γ(fˆm
)
−
γ
(f
))
+
pen
(
m)
b
−
(P
(γ(
f
)
−
γ(f
))
+
P
(γ(f
)
−
γ(
f
)))
∗
n
b
m
b
m
b
m
b
m
b
opt,β
i
h
≤P (γ(fˆm ) − γ (f∗ )) + penopt,β (m) − (P (γ(fˆm ) − γ(fm )) + Pn (γ(fm ) − γ(fˆm )))
+ (Pn − P ) (γ(fm ) − γ(fm
b )) .
ˆb ))) ≥ 0, so we get on Ω0 ,
Now, on Ω0 , we have penopt,β (m)
b − (P (γ(fˆm
b ) − γ(fm
b )) + Pn (γ(fm
b ) − γ(fm
P (γ(fˆm
b ) − γ (f∗ ))
h
i
≤ P (γ(fˆm ) − γ (f∗ )) + penopt,β (m) − (P (γ(fˆm ) − γ(fm )) + Pn (γ(fm ) − γ(fˆm )))
+ (Pn − P ) (γ(fm ) − γ(fm
b )) .
Specifying to the MLE context, the latter inequality writes,
K(f∗ , fˆm
b)
h
i
≤K(f∗ , fˆm ) + penopt,β (m) − (K(fm , fˆm ) + K(fˆm , fm )) + (Pn − P ) (γ(fm ) − γ(fm
b )) .
{z
}
|
{z
} |
(a)
(b)
Model Selection as a Multiple Testing Procedure
7
In order to get an oracle inequality as in (7), it remains to control (a) and (b) in terms of the
excess risks K(f∗ , fˆm ) and K(f∗ , fˆm
b ). Quantity (a) is related to deviations bounds for the true and
empirical excess risks of the M-estimators fˆm and quantity (b) is related to fluctuations of empirical
bias around the bias of the models. Suitable controls of these quantities (as achieved in our proofs)
will give sharp oracle inequalities (see Section 3 below).
Remark 2.4 Notice that our reasoning is not based on the particular value of the contrast, so that
to emphasize this point we choose to keep γ in most of our calculations rather than to specify to the
KL divergence related to the MLE case. As a matter of fact, the penalty penopt,β given in (8) is a
good candidate in the general context of M-estimation.
We define an over-penalization procedure as follows.
Definition 2.5 A penalization procedure as defined in (4) is said to be an over-penalization procedure
if the penalty pen that is used satisfies pen (m) ≥ penopt,β (m) for all m ∈ Mn and for some
β ∈ (1/2, 1).
Based on concentration inequalities for the excess risks (see Section 3.1) we propose the following
over-penalization penalty for histogram selection,
r
s
D ln(n + 1)
ln(n + 1) ln(n + 1) Dm
m
pen+ (m) = 1 + C max
;
;
,
(10)
n
Dm
Dm
n
where C is a constant that should be either fixed a priori (C = 1 or 2 are typical choices) or estimated
using data (see Section 5 for details on the choice of C). We show below nonasymptotic accuracy of
such procedure, both theoretically and practically.
2.3.2
Over-Penalization as a Multiple-Testing Procedure
Let us now introduce a multiple testing point of view on the model selection problem. The goal is
to infer the oracle model (5) in such a way that an oracle inequality of the type of (7) is ensured.
This task can be formulated through the solving of iterative statistical tests. Indeed, set the
following collection of null and alternative hypotheses indexed by pairs of models: for (m, m0 ) ∈ M2n ,
H0 (m, m0 ) : K(f∗ , fˆm ) > Cn · K(f∗ , fˆm0 )
H1 (m, m0 ) : K(f∗ , fˆm ) ≤ Cn · K(f∗ , fˆm0 )
,
where the constant Cn > 1 is uniform in m, m0 ∈ Mn . To each pair (m, m0 ) ∈ M2n , let us assume
that we are given a test T (m, m0 ) that is equal to one if H0 (m, m0 ) is rejected and zero otherwise.
It is worth noting at this point that what we just called “test” does not fall directly in the
classical theory of statistical testing, since the null and alternative hypotheses that we consider are
random events. However, as we will see, the only notion related to our tests needed for our model
selection study is the notion of the “size” of a test, that we will give in the following and that will
provide a mathematically consistent analysis of the statistical situation. Furthermore, it seems that
random hypotheses arise naturally in some statistical frameworks. For instance, in the context of
variable selection operated along the Lasso path, Lockhart et al. [33] consider sequential random
null hypotheses based on the active sets of the variable included up to a step on the Lasso path
(see especially Section 2.6 of [33] as well as Section 2 of the discussion of Bühlmann et al. [14] and
Section 2 of the rejoinder [32]).
Finally, if the testing of random hypotheses is disturbing to the reader, we suggest to take
our multiple testing interpretation of the model selection task in its minimal sense, that is as a
mathematical description aiming at investigating tight conditions on the penalty, that would allow
for near-optimal oracle guarantees for the penalization scheme (4).
8
Saumard and Navarro
mi
mj
mj+1
T (mi , mj ) = 1
mi
mj
T (mi , mj ) = 0
mj+1
mi
mj
mj+1
Figure 1: Iteration of the tests along the collection of models.
In order to ensure an oracle inequality such as in (7), we want to avoid as far as possible selecting
a model whose excess risk is far greater than the one of the oracle. In terms of the preceding tests,
we will see that this exactly corresponds to controlling the “size” of the tests T (m, m0 ).
0 ) = 1} the event where the test T (m, m0 ) rejects the null hyLet us note R (m, m0 ) = {T (m, mn
o
pothesis H0 (m, m0 ) and T (m, m0 ) = K(f∗ , fˆm ) > Cn · K(f∗ , fˆm0 ) the event where the hypothesis
H0 (m, m0 ) is true. By extension with the classical theory of statistical testing, we denote α (m, m0 )
the size of the test T (m, m0 ), given by
.
α m, m0 = P R m, m0 T m, m0
To explain how we select a model m
b that is close to the oracle model m∗ , let us enumerate the
models of the collection: Mn = {m1 , ..., mk }. Note that we do not assume that the models are
nested. Now, we test T (m1 , mj ) for j increasing from 2 to k. If there exists i1 ∈ {2, .., k} such that
T (m1 , mi1 ) = 0, then we perform the tests T (mi1 , mj ) for j increasing from i1 + 1 to k or choose mk
as our oracle candidate if i1 = k. Otherwise, we choose m1 as our oracle candidate. In general, we
can thus define a finite increasing sequence mil of models and a selected model m
b =m
b (T ) through
0
0
the use of the collection of tests {T (m, m ) ; m, m ∈ Mn }. Also, the number of tests that are needed
to define m
b is equal to Card (Mn ) − 1.
Let us denote P the set of pairs of models that have effectively been tested along the iterative
procedure. We thus have Card(P) = Card (Mn ) − 1. Now define the event ΩT under which there is
a first kind error along the collection of tests in P,
[ h
\
i
ΩT =
R m, m0
T m, m0
(m,m0 )∈P
and assume that the sizes of the tests are chosen such that
X
β=
α m, m0 < 1 .
(11)
(m,m0 )∈P
Assume also that the selected model m
b has the following Transitivity Property (TP):
(TP) If for any (m, m0 ) ∈ M2n , T (m, m0 ) = 1 then T (m,
b m0 ) = 1.
We believe that the transitivity property (TP) is intuitive and legitimate as it amounts to assume
that there is no contradiction in choosing m
b as a candidate oracle model. Indeed, if a model m is
thought to be better than a model m0 , that is T (m, m0 ) = 1, then the selected model m
b should also
be thought to be better that m0 , T (m,
b m0 ) = 1.
The power of the formalism that we just introduced lies on the fact that the combination of
Assumptions (11) and (TP) automatically implies that an oracle inequality as in (7) is satisfied.
Indeed, property (TP) ensures that T (m,
b m∗ ) ⊂ R (m,
b m∗ ) because on the event T (m,
b m∗ ), we
have m
b 6= m∗ , so there exists m such that (m, m∗ ) ∈ P and T (m, m∗ ) = 1 (otherwise m
b = m∗ )
Model Selection as a Multiple Testing Procedure
9
which in turn ensures T (m,
b m∗ ) = 1. Consequently, T (m,
b m∗ ) = T (m,
b m∗ ) ∩ R (m,
b m∗ ) ⊂ ΩT ,
which gives
ˆ
P K(f∗ , fˆm
b m∗ )) ≤ P (ΩT ) ≤ β ,
(12)
b ) > Cn · K(f∗ , fm∗ ) = P (T (m,
that is equivalent to (7).
Let us turn now to a specific choice of tests corresponding to model selection by penalization.
If we define for a penalty pen, the penalized criterion critpen (m) = Pn (γ(fˆm )) + pen (m), m ∈ Mn
and take the following tests
T m, m0 = Tpen m, m0 = 1{critpen (m)≤critpen (m0 )} ,
then it holds
m(T
b pen ) ∈ arg min
m∈Mn
n
o
Pn (γ(fˆm )) + pen (m)
and property (TP) is by consequence satisfied.
It remains to choose the penalty pen such that the sizes of the tests Tpen (m, m0 ) are controlled.
This is achieved by taking pen = penopt,β/2 as defined in (8). Indeed, in this case,
α m, m0 =P(Pn (γ(fˆm )) + pen (m) ≤ Pn (γ(fˆm0 )) + pen m0 T m, m0 )
h
i
≈P(K(f∗ , fˆm ) + penopt,β/2 (m) − (K(fm , fˆm ) + K(fˆm , fm ))
(13)
h
i
≤ K(f∗ , fˆm0 ) + penopt,β/2 m0 − (K(fm0 , fˆm0 ) + K(fˆm0 , fm0 )) T m, m0 )
β
≤
+ P K(f∗ , fˆm ) ≤ K(f∗ , fˆm0 )
2Card (Mn )
h
i
+ penopt,β/2 m0 − (K(fm0 , fˆm0 ) + K(fˆm0 , fm0 )) T m, m0
.
In line (13), the equality is only approximated since we neglected the centering of the biases of the
model by their empirical counterparts, as these centered random variables should be small compared
to the other quantities for models of interest. Now assume that
P penopt,β/2
ˆ
ˆ
ˆ
0
0
0
0
m − (K(f , fm ) + K(fm , fm )) ≥ εn K(f∗ , fm ) ≤
0
m0
β
,
2Card (Mn )
for some deterministic sequence εn not depending on m0 . Such result is obtained in Section 3.1 and
is directly related to the concentration behavior of the true and empirical excess risks. Then we get
h
i
P K(f∗ , fˆm ) ≤ K(f∗ , fˆm0 ) + penopt,β m0 − (K(fm0 , fˆm0 ) + K(fˆm0 , fm0 )) T m, m0
β
≤
+ P K(f∗ , fˆm ) ≤ (1 + εn )K(f∗ , fˆm0 ) K(f∗ , fˆm ) > Cn K(f∗ , fˆm0 )
2Card (Mn )
β
=
,
2Card (Mn )
where the last equality is valid if Cn ≥ 1 + εn . In this case,
α m, m0 ≤
β
Card (Mn )
and as we only perform Card(Mn )−1 tests, inequality (11) is satisfied and so is the oracle inequality
(12).
Remark 2.6 The Transitivity Property (TP) allows to unify all the selection rules that we are aware
of. Indeed, as soon as we want to select an estimator f˜ (or a model) that optimizes a criterion,
f˜ ∈ arg min {crit (f )} ,
f ∈F
10
Saumard and Navarro
Models that can be selected
q1−α [Pn (γ(fˆm ))] + pen(m)
E[P (γ(fˆm ))]
' E[Pn (γ(fˆm ))] + pen(m)
qα [Pn (γ(fˆm ))] + pen(m)
Pn (γ(fˆm )) + pen(m)
m∗
m̂
m
Figure 2: A schematic view of the situation cooresponding to a selection procedure based on the
unbiased risk principle. The penalized empirical risk (in red) fluctuates around the expectation of
the true risk. The size of the deviations typically increase with the model size, making the shape
of the curves possibly flat for the largest models of the collection. Consequently, the selected model
can potentially be very large and lead to overfitting.
where F is a collection of candidate functions (or models), then f˜ is also defined by the collection
of tests T (f, f 0 ) = 1{crit(f )≤crit(f 0 )} . In particular, in T-estimation ([15]) as well as in ρ-estimation
[8]) the estimator is indeed a minimizer of a criterion that is interpreted as a diameter of a subset
of functions in F. Note that in T-estimation, the criterion is itself constructed through the use of
some (robust) tests, that by consequence do not act at the same level as our tests in this case.
2.3.3
Geometric insights on over-penalization
Finally, let us bring a geometric point of view on our over-penalization procedure.
If the penalty pen is chosen accordingly to the Unbiased Risk Estimation principle, then it should
satisfy, for any model m ∈ Mn ,
h
i
h
i
E Pn (γ(fˆm )) + pen (m) ∼ E P γ(fˆm )
.
In other words, the curve Cn : m
fˆm )) + pen (m) fluctuates around its mean, which is
h 7→
Pn (γ(i
essentially the curve CP : m →
7 E P γ(fˆm ) , see Figure 2. Furthermore, the largest is the model
m, the largest are the fluctuations of Pn (γ(fˆm )) = K fˆm , fm +Pn (γ(fm )). This is seen for instance
through
the
concentration inequalities established in Section A.1 below for the empirical excess risk
ˆ
K fm , fm . Consequently, it can happen that the curve Cn is rather flat for the largest models and
that the selected model is among the largest of the collection, see Figure 2.
By using an over-penalization procedure instead of the unbiased risk estimation principle, we
will compensate the deviations for the largest models and thus obtain a thinner region of potential
selected models, see Figures 3 and 4. In other words, we will avoid overfitting and by doing so, we
will ensure a reasonable performance of our over-penalization procedure in situations where unbiased
risk estimation fails. This is particularly the case when the amount of data is small to moderate.
3
Theoretical Guarantees
We state here our theoretical results related to the behavior of our over-penalization procedure.
As explained in Section 2.3, concentration inequalities for the true and empirical excess risks are
essential tools to understand our model selection problem. As the theme of concentration inequalities
for the excess risk constitutes a very recent and exciting line of research, these inequalities also have
an interest by themselves and we state them in Section 3.1.
Model Selection as a Multiple Testing Procedure
11
q1−α [Pn (γ(fˆm ))] + pen(m)
E[P (γ(fˆm ))]
Correction (over-penalization)
qα [Pn (γ(fˆm ))] + pen(m)
Pn (γ(fˆm )) + pen(m)
m∗
m
Figure 3: The correction that should be applied to an unbiased risk estimation procedure would
ideally be of the size of the deviations of the risk for each model of the collection.
q1−α [Pn (γ(fˆm ))] + pen(m)
+corr(m)
Models that can be selected
E[P (γ(fˆm ))] + corr(m)
qα [Pn (γ(fˆm ))] + pen(m)
+corr(m)
Pn (γ(fˆm )) + pen(m) + corr(m)
m̂
m∗
m
Figure 4: After a suitable correction, the minimum of the red curve has a better shape. In addition,
the region of models that can be possibly selected is substantially smaller and in particular avoids
the largest models of the collection.
12
Saumard and Navarro
In Section 3.2, we give a sharp oracle inequality proving the optimality of our procedure. We
also compare our result to what would be achieved for AIC, thus exhibiting the superiority of overpenalization in the nonasymptotic regime—that is for a small to moderate sample sizes.
3.1
True and empirical excess risks’ concentration
In this section, we fix the linear model mmade of histograms
and we are interested by concentration
ˆ
inequalities for the true excess risk P γ(fm ) − γ(fm ) on m and for its empirical counterpart
Pn γ(fm ) − γ(fˆm ) .
Theorem 3.1 Let n ∈ N, n ≥ 1 and let α, A+ , A− and AΛ be positive constants. Take m a model
of histograms defined on a fixed partition Λm of Z. The cardinality of Λm is denoted by Dm . Assume
that 1 < Dm ≤ A+ n/(ln(n + 1)) ≤ n and
0 < AΛ ≤ Dm inf {P (I)} .
I∈Λm
(14)
√
√
If (α + 1) A+ /AΛ ≤ τ = 3 − 2 < 0.32, then a positive constant A0 exists, only depending on
α, A+ and AΛ , such that by setting
r
s
D ln(n + 1)
ln(n
+
1)
ln(n
+
1)
m
ε+
(m)
=
max
;
;
(15)
n
n
Dm
Dm
and
r
s
D ln(n + 1)
ln(n
+
1)
m
;
,
ε−
(m)
=
max
n
n
Dm
we have, on an event of probability at least 1 − 4(n + 1)−α ,
Dm
ˆm ≤ 1 + A0 ε+ (m) Dm ,
1 − A 0 ε−
(m)
≤
K
f
,
f
m
n
n
2n
2n
Dm
ˆm , fm ≤ 1 + A0 ε+ (m) Dm .
1 − A 0 ε−
(m)
≤
K
f
n
n
2n
2n
(16)
(17)
In the previous theorem we achieve sharp upper and lower bounds for the true and empirical
excess risk on m. They are optimal at the first order since the leading constants are equal in the
upper and lower bounds. They show the concentration of the true and empirical excess risks around
the value Dm /(2n). Moreover, Theorem 3.1 establishes the equivalence with high probability of the
true and empirical excess risks for models of reasonable dimension.
Concentration inequalities for the excess risks as in Theorem 3.1 is a new and exiting direction of
research related to the theory of statistical learning and to high-dimensional statistics. Boucheron
and Massart [20] obtained a pioneering result describing the concentration of the empirical excess
risk around its mean, a property that they call a high-dimensional Wilks phenomenon. Then a few
authors obtained results describing the concentration of the true excess risk around its mean [39],
[25], [38] or around its median [11], [12] for (penalized) least square regression and in an abstract Mestimation framework [44]. In particular, recent results of [44] include the case of MLE on exponential
models and as a matter of fact, on histograms. Nevertheless, Theorem 3.1 is a valuable addition to
the literature on this line of research since we obtain here nonetheless the concentration around a
fixed point, but an explicit value Dm /2n for this point. On the contrary, the concentration point is
available in [44] only through an implicit formula involving local suprema of the underlying empirical
process.
The principal assumption in Theorem 3.1 is inequality (14) of lower regularity of the partition
with respect to P . It is ensured as soon as the density f∗ is uniformly bounded from below and
Model Selection as a Multiple Testing Procedure
13
the partition is lower regular with respect to the reference measure µ (which will be the Lebesgue
measure in our experiments). No restriction on the largest values of f∗ are needed. In particular,
we do not restrict to the bounded density estimation setting.
Castellan [23] proved related, but weaker inequalities than in Theorem 3.1 above. She also asked
for a lower regularity property of the partition, as in Proposition 2.5 [23], where she derived a sharp
control of the KL divergence of the histogram estimator on a fixed model. More precisely, Castellan
assumes that there exists a positive constant B such that
inf µ (I) ≥ B
I∈Λm
(ln(n + 1))2
.
n
(18)
This latter assumption is thus weaker than (14) for the considered model as its dimension Dm is less
than the order n (ln(n + 1))−2 . We could assume (18) instead of (14) in order to derive Theorem
3.1. This would lead to less precise results for second order terms in the deviations of the excess
risks but the first order bounds would be preserved. More precisely, if we replace assumption (14) in
Theorem 3.1 by Castellan’s assumption (18), a careful look at the proofs show that the conclusions
of Theorem 3.1 are still valid for εn = A0 (ln(n + 1))−1/2 , where A0 is some positive constant. Thus
assumption (14) is not a fundamental restriction in comparison to Castellan’s work [23], but it leads
to more precise results in terms of deviations of the true and empirical excess risks of the histogram
estimator.
The proof of Theorem 3.1, that can be found in Section A.1, is based on an improvement
of independent interest of the previously best known concentration inequality for the chi-square
statistics. See Section 4.1 for the precise result.
3.2
An Oracle Inequality
Let us first state the set of five structural assumptions required to establish the nonasymptotic
optimality of the over-penalization procedure. These assumptions will be further discussed at the
end of this section, after the statement of a sharp oracle inequality.
Set of assumptions (SA)
(P1) Polynomial complexity of Mn : Card (Mn ) ≤ nαM .
(P2) Upper bound on dimensions of models in Mn : there exists a positive constant AM,+ such
that for every m ∈ Mn ,
n
≤n.
Dm ≤ AM,+
(ln(n + 1))2
(Asm) The unknown density f∗ satisfies some moment condition and is uniformly bounded from
below: there exist some constants Amin > 0 and p > 1 such that,
Z
f∗p (ln f∗ )2 ∨ 1 dµ < +∞
Z
and
inf f∗ (z) ≥ Amin > 0 .
z∈Z
(19)
(Alr) Lower regularity of the partition with respect to µ: there exists a positive finite constant AΛ
such that, for all m ∈ Mn ,
Dm inf µ (I) ≥ AΛ ≥ AM,+ (αM + 6)/τ ,
I∈Λm
√
√
where τ = 3 − 2 > 0.
(Ap) The bias decreases like a power of Dm : there exist β− ≥ β+ > 0 and C+ , C− > 0 such that
−β−
−β+
C− D m
≤ K (f∗ , fm ) ≤ C+ Dm
.
We are now ready to state our main theorem related to optimality of over-penalization.
14
Saumard and Navarro
Theorem 3.2 Take n ≥ 1 and r ∈ (0, p − 1). For some ∆ > 0, consider the following penalty,
pen (m) = 1 + ∆ε+
n (m)
Dm
,
n
for all m ∈ Mn .
(20)
Assume that the set of assumptions (SA) holds and that
β− < p (1 + β+ ) /(1 + p + r)
or
p/(1 + r) > β− + β− /β+ − 1 .
(21)
Then there exists an event Ωn of probability at least 1 − (n + 1)−2 and some positive constant A1
depending only on the constants defined in (SA) such that, if ∆ ≥ A1 > 0 then we have on Ωn ,
n
o
ˆ
K f∗ , fˆm
≤
(1
+
δ
)
inf
K
f
,
f
,
(22)
n
∗ m
b
m∈Mn
where δn = L(SA),∆,r (ln(n + 1))−1/2 is convenient.
We derive in Theorem 3.2 a pathwise oracle inequality for the KL excess risk of the selected
estimator, with constant almost one. Our result thus establishes the nonasymptotic quasi-optimality
of over-penalization with respect to the KL divergence.
It is worth noting that three very strong features related to oracle inequality (22) significantly improve upon the literature. Firstly, inequality (22) expresses the performance of the selected estimator
through its KL divergence and compare it to the KL divergence of the oracle. Nonasymptotic results
pertaining to (robust) maximum likelihood based density estimation usually control the Hellinger
risk of the estimator, [24], [35], [17], [15], [8]. The main reason is that the Hellinger risk is much easier to handle that the KL divergence from a mathematical point of view. For instance, the Hellinger
distance is bounded by one while the KL divergence can be infinite. However, from a M-estimation
perspective, the natural excess risk associated to likelihood optimization is indeed the KL divergence
and not the Hellinger distance. These two risks are provably close to each other in the bounded
setting [35], but may behave very differently in general.
Secondly, nonasymptotic results describing the performance of procedures based on penalized
likelihood, by comparing more precisely the (Hellinger) risk of the estimator to the KL divergence of
the oracle, all deal with the case where the log-density to be estimated is bounded ([24], [35]). Here
we substantially extend the setting by considering only the existence of a finite polynomial moment
for the large values of the density to be estimated.
Last but not least, the oracle inequality (22) is valid with positive probability - larger than
3/4 - as soon as one data is available. Up to our knowledge any other oracle inequality describing
the performance of penalization for maximum likelihood density estimation is valid with positive
probability only when the sample size n is greater than an integer n0 that depends on the constants
defining the problem and that is thus unknown. We highlight the fact that we control the risk of the
selected estimator for any sample size and that this property is essential in practice when dealing
with small to moderate sample sizes. Based on the arguments developed in Section 2.3, we believe
that such a feature of Theorem 3.1 is only accessible through the use of over-penalization and we
conjecture in particular that it is impossible using AIC to have a control the KL divergence of the
selected estimator for any sample size.
The oracle inequality (22) is valid under conditions (21) relating the values of the bias decaying
rates β− and β+ to the order p of finite moment of the density f∗ and the parameter r. In order
to understand these latter conditions, let us assume for simplicity that β− = β+ =: β. Then the
conditions (21) both reduce to β < p/(1 + r). As r can be taken as close to zero as we want, the
latter inequality reduces to β < p. In particular, if the density to be estimated is bounded (p = ∞),
then conditions (21) are automatically satisfied. If on the contrary the density f∗ only has finite
polynomial moment p, then the bias should not decrease too fast. In light of the following comments,
if f∗ is assumed to be α-Hölderian, α ∈ (0, 1], then β ≤ 2α ≤ 2 and the conditions (21) are satisfied,
in the case where β− = β+ , as soon as p ≥ 2.
To conclude this section, let us now comment on the set of assumptions (SA). Assumption (P1)
states that the collection of models has a polynomially increasing complexity. This is well adapted
Model Selection as a Multiple Testing Procedure
15
to bin size selection since in this case we usually select among a number of models that is strictly
bounded from above by the sample size. In the same manner, Assumption (P2) is fairly legitimate
for us and corresponds to practice, where we aim at considering bin sizes for which each element of
the partition contains a few sample points.
Assumption (Asm) imposes conditions on the density to be estimated. More precisely, Assumption (19) stating that the unknown density is uniformly bounded from below can also be found in
the work of Castellan [23]. The author assumes moreover in Theorem 3.4 where she derives an oracle
inequality for the (weighted) KL excess risk of the histogram estimator, that the target is of finite
sup-norm. Furthermore, from a statistical perspective, the lower bound (19) is legitimate since, by
Assumption (Alr), we use models of lower-regular partitions with respect to the Lebesgue measure.
In the case where Inequality (19) would not hold, one would typically have to consider exponentially
many irregular histograms to take into account the possibly vanishing mass of some elements of the
partitions (for more details on this aspect that goes beyond the scope of the present paper, see for
instance [35]).
We require in (Ap) that the quality of the approximation of the collection of models is good
enough in terms of bias. More precisely, we require a polynomially decreasing of excess risk of KL
projections of the unknown density onto the models. For a density f∗ uniformly bounded away from
zero, the upper bound on the bias is satisfied when for example, Z is the unit interval, µ = Leb is the
Lebesgue measure on the unit interval, the partitions Λm are regular and the density f∗ belongs to
the set H (H, α) of α-hölderian functions for some α ∈ (0, 1]: if f ∈ H (H, α), then for all (x, y) ∈ Z 2
|f (x) − f (y)| ≤ H |x − y|α .
In that case, β+ = 2α is convenient and AIC-type procedures are adaptive to the parameters H and
α, see Castellan [23].
In assumption (Ap) of Theorem 3.2 we also assume that the bias K (f∗ , fm ) is bounded from
below by a power of the dimension Dm of the model m. This hypothesis is in fact quite classical as
it has been used by Stone [41] and Burman [22] for the estimation of density on histograms and also
by Arlot and Massart [7] and Arlot [3], [4] in the regression framework. Combining Lemma 1 and 2
of Barron and Sheu [10] - see also inequality (35) of Proposition 4.7 below – we can show that
Z
1 −3 ln ffm∗
(fm − f∗ )2
∞
e
dµ ≤ K (f∗ , fm )
2
f∗
Z
and thus assuming for instance that the target is uniformly bounded, kf∗ k∞ ≤ A∗ , we get
Z
A3min
(fm − f∗ )2 dµ ≤ K (f∗ , fm ) .
2A4∗ Z
Now, since in the case of histograms the KL projection fm is also the L2 (µ) projection of f∗ onto
m, we can apply Lemma 8.19 in Section 8.10 of Arlot [2] to show that assumption (Ap) is indeed
satisfied for β− = 1 + α−1 , in the case where Z is the unit interval, µ = Leb is the Lebesgue measure
on the unit interval, the partitions Λm are regular and the density f∗ is a non-constant α-hölderian
function.
The proof of Theorem 3.2 and further descriptions of the behavior of the procedure can be found
in the supplementary material, Section A.2.
4
Probabilistic and Analytical Tools
We state in this section some general results that are of independent interest and that serve as tools
for the mathematical description of our statistical procedure.
The two first sections contain new or improved concentration inequalities, for the chi-square
statistics (Section 4.1) and for general log-densities (Section 4.2). We established in Section 4.3
some results that are related to the so-called margin relation in statistical learning and that are of
analytical nature.
16
Saumard and Navarro
4.1
Chi-Square Statistics’ Concentration
The chi-square statistics plays an essential role in the proofs related to Section 3.1. Let us recall its
definition.
Definition 4.1 Given some histogram model m, the statistics χ2n (m) is defined by
χ2n (m) =
Z
Z
fˆm − fm
fm
2
dµ =
X (Pn (I) − P (I))2
.
P (I)
I∈m
The following proposition provides an improvement upon the previously best known concentration inequality for the right tail of the chi-square statistics, available in [24]—see also Proposition
7.8 in [35] and Theorem 12.13 in [19].
Proposition 4.2 For any x, θ > 0, it holds
P χn (m) 1Ωm (θ) ≥
r
r !
√
Dm
2x
θ
≤ exp (−x) ,
+ 1 + 2θ +
n
6
n
(23)
T
where we set Ωm (θ) = I∈m {|Pn (I) − P (I)| ≤ θP (I)}. More precisely, for any x, θ > 0, it holds
with probability at least 1 − e−x ,
r
r
r
r r
!
Dm
2x
θ √
θ x
x
1
xDm 1/4
χn (m) 1Ωm (θ) <
+
+2
+
∧√
.
(24)
x∧
n
n
n
2
3 n
Dm
2
The proof of Theorem 4.2 can be found in Section 6.1.
Let us detail its relationship with the bound in Proposition 7.8 in [35], which is: for any x, ε > 0,
r
r !!
Dm
2x
P χn (m) 1Ωm (ε2 /(1+ε/3)) ≥ (1 + ε)
+
≤ exp (−x) .
(25)
n
n
p
√
By taking θ = ε2 / (1 + ε/3) > 0, we get ε = θ/6 + θ2 /36 + θ > θ/6 + θ > 0. Assume that
Dm ≥ 2x. We obtain by (23), with probability at least 1 − exp (−x),
χn (m) 1Ωm (θ)
r
r
√
Dm
2x
θ
<
+ 1 + 2θ +
n
6
n
r
r
√
Dm
2x
+ 1 + 2ε
≤
n
n
r
r !
Dm
2x
< (1 + ε)
+
.
n
n
So in this case, inequality (23) improves upon (25). Notice that in our statistical setting (see Theorem
3.1) the restriction on Dm is as follows Dm ≤ A+ n (ln(n + 1))−1 . Furthermore, in our proofs we
apply (23) with x proportional to ln(n + 1) (see the proof of Theorem 3.1). Hence, for a majority of
models, we have x Dm and so
r
r
2x
Dm
.
n
n
As a result, the bounds that we obtain in Theorem 3.1 by the use of Inequality (23) are substantially
better
than the bounds we would obtain by using Inequality (25) of [35]. More precisely, the term
p
Dm ln(n + 1)/n in (15) would be replaced by (Dm ln(n + 1)/n)1/4 , thus changing the order of
magnitude for deviations of the excess risks.
Model Selection as a Multiple Testing Procedure
17
Now, if Dm ≤ 2x then (23) gives that with probability at least 1 − e−x ,
r
r
r
r
Dm
2x
θ xDm 1/4 θ 2x
χn (m) 1Ωm (θ) <
+
+
+2
n
n
n
2
6 n
!r
√
√ !r
θ
Dm
θ
θ
2x
+ 1 + 3/4 +
≤ 1 + 1/4
n
6
n
2
2
r
r !
Dm
2x
< (1 + ε)
.
+
n
n
So in this case again, inequality (24) improves upon (25), with an improvement that can be substantial depending on the value of the ratio x/Dm .
The following result describes the concentration from the left of the chi-square statistics and is
proved in Section 6.1.
Proposition 4.3 Let α, AΛ > 0. Assume 0 < AΛ ≤ Dm inf I∈m {P (I)}. Then there exists a
positive constant Ag depending only on AΛ and α such that
r
s
p
ln(n
+
1)
ln(n
+
1)
D
m
P χn (m) ≤ 1 − Ag
∨
≤ (n + 1)−α .
Dm
n
n1/4
4.2
Bernstein type concentration inequalities for log-densities
The following proposition gives concentration inequalities for the empirical bias at the right of its
mean.
Proposition 4.4 Consider a density f ∈ S. We have, for all z ≥ 0,
z
≤ exp (−z) .
P Pn (ln ( f / f∗ )) ≥
n
(26)
2
R
dµ, we have
Moreover, if we can take a finite quantity v which satisfies v ≥ (f ∨ f∗ ) ln ff∗
for all z ≥ 0,
!
r
2vz 2z
P (Pn − P ) (ln ( f / f∗ )) ≥
+
≤ exp (−z) .
(27)
n
n
One can notice, with Inequality (26), that the empirical bias always satisfies some exponential
deviations at the right of zero. In the Information Theory community, this inequality is also known
as the “No Hyper-compression Inequality” ([28]).
Inequality (27) seems to be new and takes the form of a Bernstein-like inequality, even if the
usual assumptions of Bernstein’s inequality are not satisfied. In fact, we are able to recover such a
behavior by inflating the usual variance to the quantity v.
We turn now to concentration inequalities for the empirical bias at the left of its mean.
Proposition 4.5 Let r > 0. For any density f ∈ S and for all z ≥ 0, we have
P (Pn (ln ( f / f∗ )) ≤ −z/nr − (1/r) ln (P [( f∗ / f )r ])) ≤ exp (−z) .
(28)
2
R f∗r+1
Moreover, if we can set a quantity wr which satisfies wr ≥
∨
f
ln ff∗
dµ , then we
r
∗
f
get, for all z ≥ 0,
!
r
2wr z
2z
P (Pn − P ) (ln ( f / f∗ )) ≤ −
−
≤ exp (−z) .
(29)
n
nr
18
4.3
Saumard and Navarro
Margin-Like Relations
Our aim in this section is to control the variance terms v and wr , appearing respectively in Lemma
4.4 and 4.5, in terms of the KL divergence pointed on the target f∗ . This is done in Lemma 4.6
below under moment assumptions for f∗ .
Proposition 4.6 Let p > 1 and c+ , c− > 0. Assume that the density f∗ satisfies
Z
Z
(ln (f∗ ))2 ∨ 1
2
p
f∗ (ln (f∗ )) ∨ 1 dµ < +∞ and Q :=
J :=
dµ < +∞
f∗p−1
Z
Z
(30)
Take a density f such that 0 < c− ≤ inf z∈Z {f (z)} ≤ supz∈Z {f (z)} ≤ c+ < +∞. Then, for some
AM R,d > 0 only depending on J, Q, p, c+ and c− , it holds
"
2 #
f
f
1− 1
P
∨1
ln
(31)
≤ AM R,d × K (f∗ , f ) p .
f∗
f∗
1/p
p
2
2
holds. For any 0 < r ≤
More precisely, AM R,d = 4c1−p
(ln
c
)
∨
1
J
+
4c
(ln
c
)
∨
1
Q
−
+
−
+
p − 1, we have the following inequality,
"
r 2 #
f∗
f
1− r+1
p
P
∨1
ln
≤ AM R,g × K (f∗ , f )
,
(32)
f
f∗
r+1
p
2
2
available with AM R,g = 4c1−p
(ln
c
)
∨
1
J
+
2
(ln
(c
))
+
J
+
Q
.
−
+
−
Theorem 4.6 states that the variance terms, appearing in the concentration inequalities of Section 4.2, are bounded from above, under moment restrictions on the density f∗ , by a power less than
one of the KL divergence pointed on f∗ . The stronger are the moment assumptions, given in (30),
the closer is the power to one. One can notice that J is a restriction on large values of f∗ , whereas
Q is related to values of f∗ around zero.
We call these inequalities “margin-like relations” because of their similarity with the margin
relations known first in binary classification ([34], [43]) and then extended to empirical risk minimization (see [36] and [5] for instance). Indeed, from a general point of view, margin relations relate
the variance of contrasted functions (logarithm of densities here) pointed on the contrasted target
to a function (in most cases, a power) of their excess risk.
Now we strengthen the restrictions on the values of f∗ around zero. Indeed, we ask in the
following lemma that the target is uniformly bounded away from zero.
Proposition 4.7 Let p > 1 and Amin , c+ , c− > 0. Assume that the density f∗ satisfies
Z
J :=
f∗p (ln (f∗ ))2 ∨ 1 dµ < +∞ and 0 < Amin ≤ inf f∗ (z) .
Z
z∈Z
Then there exists a positive constant AM R,− only depending on Amin , J, r and p such that, for any
m ∈ Mn ,
"
2 #
fm
fm
P
∨1
ln
≤ AM R,− × K (f∗ , fm )1−1/p
(33)
f∗
f∗
and for any 0 < r ≤ p − 1,
"
r 2 #
f∗
fm
1− r+1
p
P
∨1
ln
≤ AM R,− × K (f∗ , fm )
.
fm
f∗
(34)
If moreover ln (f∗ ) ∈ L∞ (µ), i.e. 0 < Amin ≤ inf z∈Z f∗ (z) ≤ kf∗ k∞ < +∞, then there exists à > 0
only depending on r, Amin and kf∗ k∞ such that, for any m ∈ Mn ,
"
"
2 #
r 2 #
fm
fm
f∗
fm
P
∨1
ln
∨P
∨1
ln
≤ Ã × K (f∗ , fm ) .
(35)
f∗
f∗
fm
f∗
Model Selection as a Multiple Testing Procedure
∆m
0.20
â
n
−1 X
n i=1 ln
0.10
fˆm (ξi)
!
30
Dm
0.15
19
20
0.05
10
0.00
0
25
50
Dm
75
100
0
α = 25%
25
50
75
100
Number of the bigest models
Figure 5: Estimation of the over-penalization constant.
It is worth noting that Lemma 4.7 is stated only for projections fm because we actually take
advantage of their special form (as local means of the target) in the proof of the lemma. The benefit,
compared to results of Lemma 4.6, is that Inequalities (33), (34) and (35) do not involve assumptions
on the values of fm (and in particular they do not involve the sup-norm of fm ).
5
Experiments
A simulation study is conducted in order to compare the numerical performances of the model
selection procedures we have discussed. We demonstrate the usefulness of our procedure on simulated
data examples. The numerical experiments have been carried out using R.
5.1
Experimental Setup
We have compared the numerical performance of our procedure to state-of-the-art classical penalization methods of the literature on several densities. In particular we consider the estimator of [17]
and AICc ([42, 29]). We also report AIC behavior. In the sequel, we name the procedure of [17] by
BR, and our criterion AIC1 when the constant C = 1 in (10) and AICa for a fully adaptive procedure
to be detailed below. More precisely, the performance of the following four model selection methods
were compared:
1. AIC:
m
b AIC ∈ arg min {critAIC (m)} ,
m∈Mn
with
2. AICc:
with
m
b AICc
Dm
critAIC (m) = Pn (γ(fbm )) +
.
n
n
o
∈ arg min Pn (γ(fbm )) + penAICc (m) ,
m∈Mn
penAICc (m) =
3. BR:
m
b BR ∈ arg min
m∈Mn
with
Dm
.
n − Dm − 1
n
o
Pn (γ(fbm )) + penBR (m) ,
penBR (m) =
(log Dm )2.5
n
20
Saumard and Navarro
4. AIC1 :
m
b AIC1 ∈ arg min
n
o
Pn (γ(fbm )) + penAIC1 (m) ,
m
b AICa ∈ arg min
n
o
Pn (γ(fbm )) + penAICa (m) ,
m∈Mn
with
5. AICa :
r
s
D ln(n + 1)
ln(n + 1) ln(n + 1) Dm
m
penAIC1 (m) = 1 × max
;
;
,
n
Dm
Dm n
m∈Mn
with
r
s
D ln(n + 1)
ln(n + 1) ln(n + 1) Dm
m
;
penAICa (m) = Ĉ max
;
,
n
Dm
Dm n
where Ĉ = medianα∈P Ĉα , with Ĉα = medianm∈Mα |Ĉm |, where
Ĉm =
max
q
∆m
,
q
Dm
Dm
1
n ;
Dm
2
∆m is the least-squares distance between the opposite of the empirical risk −Pn (γ(fbm )) and a
fitted line of equation y = xDm /n+â (see Figure 5), P is the set of proportions α corresponding
to the longest plateau of equal selected models when using penalty (10) with constant C = Ĉα
and Mα is the set of models in the collection associated to the proportion α of the largest
dimensions.
Let us briefly explain the ideas underlying the design of the fully adaptive procedure AICa .
According to the definition of the penalty penopt,β given in (8) the constant C in our over-penalization
procedure (10) should ideally estimate some normalized deviations of the sum of the excess risk and
the empirical excess risk on the models of the collection. Based on Theorem 3.1, we may also
assume that the deviations of the excess risk and the empirical excess risk are of the same order.
Furthermore, considering the largest models of the collection allows to neglect bias issues and thus,
the median of the normalized deviations of the empirical risk around its mean for the largest models
should be a sensible estimator of the constant C. Now the problem is give a tractable definition to
the "largest models" of the collection. To do that we pick a proportion α of the largest dimensions
of the models at hand and calculate using these models an estimator Ĉα of the constant C in (10).
We proceed then for each α in a grid of values between 0 and 1 to a model selection step by overpenalization using the constant C = Ĉα . This gives us a graph of the selected dimensions with
respect to the proportions (see Figure 5) and we finally define our over-penalization constant Ĉ as
the median of the values of the constants Ĉα , α ∈ P where P is the largest plateau in the graph of
the selected dimensions with respect to proportions α.
The models that we used along the experiments are made of histograms densities defined a on a
regular partition of the interval [0, 1].
We show the performance of the proposed method for a set of four test distributions supported
on the interval [0, 1] (see Figure 6) and described in the benchden 1 R-package [37] which provides
an implementation of the distributions introduced in [13].
5.2
Results
We compared the procedures on N = 100 independent data sets of size n ranging from 50 to
1000. We estimate the quality of the model-selection strategies through the median KL divergence.
Boxplots were made of the KL risk over the N trials. The horizontal lines of the boxplot indicate the
1
Available on the CRAN http://cran.r-project.org.
Model Selection as a Multiple Testing Procedure
Infinite peak
-1.0
-0.5
0.0
0.5
1.0
(a) Isosceles triangle
1
0.0
0.2
0.4
0.6
0.8
(b) Bilogarithmic peak
1.0
0
0.0
0.0
0.0
0.2
0.5
0.5
0.4
2
1.0
0.6
3
1.0
1.5
0.8
4
1.5
Beta(2,2)
2.0
Bilogarithmic peak
1.0
Isosceles triangle
21
0.0
0.2
0.4
0.6
0.8
(c) Beta (2,2)
1.0
0.0
0.2
0.4
0.6
0.8
1.0
(d) Infinite peak
Figure 6: Test densities f .
5%, 25%, 50%, 75%, and 95% quantiles of the error distribution. The median value of AIC (horizontal
black line) is also overlayed for visualization purposes.
It can be seen from Figure 7 that, as expected, for each method and in all cases, the KL divergence
is decreasing as the sample size increases. We also clearly see that there is in general a substantial
benefit in modifying AIC for sample sizes smaller than 1000.
Furthermore, none of the methods outperforms the others in all cases. However, for virtually all
cases, AIC1 consistently showed lower median KL than the others. We thus recommend the use of
AIC1 rather than the commonly used AICc . Indeed, in comparison to AIC1 , it seems that AICc is
not penalizing enough, which is seen through a worst performance for sample sizes equal to 50 and
100. In addition, the two procedures have the exact same computational costs.
On the contrary, it seems that the criterion BR penalizes too much. Consequently, its performances are getting worst in comparison to the other methods when the sample size increases.
Finally, the behaviour of AICa is quite good in general, but it is often a little less efficient than
AIC1 . We interpret this fact by postulating that the extra computations that we engage in AICa to
remove any hyper-parameter from the procedure actually induces extra variance in the estimation.
Acknowledgements
The first author warmly thanks Matthieu Lerasle for instructive discussions on the topic of estimation
by tests and Alain Célisse for a nice discussion at a early stage of this work. He is also grateful to
Pascal Massart for having pushed him towards obtaining better oracle inequalities than in a previous
version of this work. Finally, we owe thanks to Amandine Dubois for a careful reading that helped
to correct some mistakes and improve the presentation of the paper.
6
6.1
Proofs
Proofs Related to Section 4.1
Proof of Theorem 4.2. We fix x, θ > 0 and we set z > 0 to be chosen later. Let us set for
any I ∈ m, ϕI = (P (I))−1/2 1I . The family of functions (ϕI )I∈m forms an orthonormal basis of
(m, k·k2 ). By Cauchy-Schwarz inequality, we have
!
X
(Pn − P )
aI ϕI
,
χn (m) =
sup
(aI )I∈m ∈B2
I∈m
P
2
where B2 := (aI )I∈m ;
I∈m aI ≤ 1 . Furthermore, the case of equality in Cauchy-Schwarz
inequality gives us
!
X
(Pn − P ) (ϕI )
χn (m) = (Pn − P )
a∞
with a∞
.
I ϕI
I =
χn (m)
I∈m
22
Saumard and Navarro
n = 50
n = 100
n = 500
n = 1000
0.3
0.030
0.05
0.15
0.025
0.04
43
KL
0.10
36
25 11 16 19
0.1
KL
KL
KL
0.2
0.03
23 6 16 20
22 10 0
6 10
0.020
0.015
10 5
1
1
5
BR
AIC1
AICa
0.02
0.05
0.010
AIC
AICc
BR
AIC1
AICa
AIC
AICc
methods
BR
AIC1
AICa
AIC
AICc
methods
BR
AIC1
AICa
AIC
AICc
methods
methods
(a) Isosceles triangle
n = 50
n = 100
0.150
n = 500
0.05
n = 1000
0.030
0.16
0.125
0.04
0.025
14
0.08
0.075
6
AIC
AICc
0
1
BR
AIC1
2
8
0.03
4
1
4
1
0.050
AICa
KL
KL
0.12
KL
KL
0.100
0
0
0
0
0.020
0
AIC
AICc
BR
AIC1
AICa
0
AIC
AICc
0
0
0
AIC1
AICa
0.015
0.02
methods
0
AIC
AICc
methods
BR
AIC1
AICa
methods
BR
methods
(b) Bilogarithmic peak
n = 50
n = 100
n = 500
n = 1000
0.030
0.20
0.04
0.12
0.025
0.10
0.08
6
7
8
3
0.03
7
4
4
KL
25
KL
26 18 2
KL
KL
0.15
0
3
1
0.020
4
0.015
7
6
AIC
AICc
0
2
6
AIC1
AICa
0.02
0.04
0.05
0.010
AIC
AICc
BR
AIC1
AICa
AIC
AICc
methods
BR
AIC1
AICa
AIC
AICc
methods
BR
AIC1
AICa
methods
BR
methods
(c) Beta (2,2)
n = 50
n = 100
n = 500
0.28
0.35
0.30
n = 1000
0.14
0.10
0.24
0.09
0.20
25
11 2
5
6
AIC1
AICa
0.20
27
0.16
10 1
1
3
AIC1
AICa
KL
0.25
KL
KL
KL
0.12
0.10
15 4
0
0.08
0
1
10 1
0
0
0
AIC1
AICa
0.07
0.08
0.15
AIC
AICc
BR
0.12
AIC
AICc
methods
BR
AIC
methods
AICc
BR
AIC1
AICa
AIC
AICc
methods
BR
methods
(d) Infinite peak
AIC
AICc
BR
AIC1
AICa
Figure 7: KL divergence results. Box plots of the KL divergence to the true distribution for the
estimated distribution. The solid black line corresponds to the AIC KL divergence median. The
term inside the box is the number of time the KL divergence equals ∞ out of 100.
Hence, by setting A (s) := B2
o
Tn
(aI )I∈m ; supI∈m aI (P (I))−1/2 ≤ s for any s ≥ 0, we get
χn (m) 1Ωm (θ) 1{χn (m)≥z} ≤
sup
(aI )I∈m ∈A(θ/z)
(Pn − P )
X
I∈m
aI ϕI
!
.
(36)
Model Selection as a Multiple Testing Procedure
23
By applying Bousquet’s inequality ([21]) to the supremum in the right-hand side of (36) then gives
for any δ > 0,
!
r
2 x
1 1 bm x
2σm
(37)
P χn (m) 1Ωm (θ) 1{χn (m)≥z} ≥ (1 + δ) Em +
+
+
n
δ 3
n
!
!
r
2 x
X
2σm
1 1 bm x
≤P
sup
(Pn − P )
aI ϕI
≥ (1 + δ) Em +
+
+
n
δ 3
n
(aI )I∈m ∈A(θ/z)
I∈m
≤ exp (−x) ,
with
Em ≤ E [χn (m)] ≤ E
1/2
χ2n (m)
=
r
Dm
n
;
and
bm =
2
σm
X
sup
≤
sup
(aI )I∈m ∈B2
aI ϕI
(aI )I∈m ∈A(θ/z) I∈m
∞
≤
V
X
!
aI ϕI (ξ1 )
I∈m
≤1
θ
.
z
p
p
(Dm ) /n + 2x/n, we obtain that with
Injecting the latter bounds in (37) and taking z =
probability at least 1 − exp (−x),
r
r
Dm
2x
1 1
θx
√ √ .
χn (m) 1Ωm (θ) < (1 + δ)
+
+
+
(38)
√
n
n
δ 3
Dm + 2x n
By taking δ =
√
−1/2
√
θx Dm + 2xDm
, the right-hand side of (38) becomes
r
r
r
√
θx
Dm
θx
p
√ √
+ √
√
n Dm + 2xDm 3 Dm + 2x n
r
r
r
r r
!
Dm
2x
θ √
xDm 1/4
θ x
x
1
≤
,
+
+2
x∧
+
∧√
n
n
n
2
3 n
Dm
2
Dm
+
n
2x
+2
n
which gives (24). Inequality (23) is a direct consequence of (24).
Proof of Theorem 4.3. Let us set for any I ∈ m, ϕI = (P (I))−1/2 1I . The family of functions
(ϕI )I∈m forms an orthonormal basis of (m, k·k2 ). By Cauchy-Schwarz inequality, we have
!
X
χn (m) =
sup
aI ϕI
(Pn − P )
,
(aI )I∈m ∈B2
I∈m
P
where B2 = (aI )I∈m ; I∈m a2I ≤ 1 . As
2
σm
:=
and
bm =
sup
(aI )I∈m ∈B2
sup
(aI )I∈m ∈B2
X
X
V
aI ϕI (ξ1 )
I∈m
aI ϕI
I∈m
!
∞
≤1
1
≤ sup kϕI k∞ = sup p
,
P (I)
I∈m
I∈m
we get by Klein-Rio’s inequality (see [30]), for every x, δ > 0,
P χn (m) ≤ (1 − δ) E [χn (m)] −
r
2x
−
n
1
+1
δ
bm x
n
!
≤ exp (−x) .
(39)
24
Saumard and Navarro
q
Note that we have bm ≤ Dm A−1
Λ . Now, we bound E [χn (m)] by below. By Theorem 11.10 in [19]
applied to χn (m), we get, for any ζ > 0,
bm
σ2
E χ2n (m) − E2 [χn (m)] ≤ m + 4 E [χn (m)]
n
qn
Dm A−1
1
Λ
≤ +4
E [χn (m)]
n
n
Dm A−1
1
Λ
+ ζE2 [χn (m)] .
≤ + 4ζ −1
n
n2
ζa2 +ζ −1 b2
(40)
q
applied with a = E [χn (m)] and b = 2 Dm A−1
Λ /n.
The latter inequality follows from 2ab ≤
As E χ2n (m) = Dm /n, (40) applied with ζ = n−1/2 gives
s
1
Dm
1
−1 Dm
E [χn (m)] ≥
− 4AΛ 3/2 −
n
n +
1 + n−1/2
n
r
Dm
−1/2
≥
1 − LAΛ Dm
∨ n−1/4 .
(41)
n
p
Hence, by using (41) and taking x = α ln(n + 1) and δ = n−1/4 ln(n + 1) in (39), we obtain with
probability at least 1 − (n + 1)−α ,
χn (m)
!r
p
ln(n + 1)
1
Dm
1
1 − LAΛ √
≥ 1−
∨
n
n1/4
Dm n1/4
! √
r
n1/4
2α ln(n + 1)
α Dm ln(n + 1)
− p
+1
−
n
n
ln(n + 1)
s
r
p
p
ln(n + 1)
ln(n + 1)
Dm
2α ln(n + 1)
1
≥
1 − LA Λ
− Lα
∨√
−
n
Dm
n1/4
n1/4
Dm
s
r
p
ln(n + 1)
Dm
ln(n + 1)
∨
≥
1 − LAΛ ,α
,
1/4
n
Dm
n
which concludes the proof.
6.2
Proofs related to Section 4.2
In this section, our aim is to control for any histogram model m, the deviations of the empirical bias
from its mean, which is the true bias. Concentration inequalities for the centered empirical bias are
provided in Section 4.2. In order to prove oracle inequalities for the KL divergence, we relate the
magnitude of these deviations to the true bias. This is done in Section 4.3. As we believe that the
results of this section have some interest by themselves, we have stated them, when possible, for a
general density in S rather for the projections fm .
The results in this section
based on the Cramèr-Chernoff method (see [18] for instance). Let
Pare
n
us recall that if we set S := i=1 Xi − E [Xi ], where (Xi )ni=1 are n i.i.d. real random variables, and
for any λ ≥ 0 and y ∈ R+ ,
ϕS (λ) := ln (E [exp (λS)]) = n (ln (E [exp (λX1 )]) − λE [X1 ])
and
ϕ∗S (y) := sup {λy−ϕS (λ)} ,
λ∈R+
Model Selection as a Multiple Testing Procedure
25
then
P (S ≥ y) ≤ exp (−ϕ∗S (y)) .
(42)
Proof of Proposition 4.4.
We first prove concentration inequality (26). We set Xi :=
ln ( f / f∗ ) (ξi ) and use Inequality (42). For λ ∈ [0, 1], as E [X1 ] = −K (f∗ , f ), we have
h
i
ϕS (λ)/ n = ln P ( f / f∗ )λ + λK (f∗ , f )
≤λ ln (P [ f / f∗ ]) + λK (f∗ , f ) = λK (f∗ , f ) ,
where the inequality derives from the concavity of the function x 7→ xλ . By setting K = K (f∗ , f ),
we thus get
ϕ∗S (y) ≥ sup {λy−ϕS (λ)} ≥ (y − nK)+ ,
λ∈[0,1]
which implies,
P ((Pn − P ) (ln (f /f∗ )) ≥ x) ≤ exp (−n(x − K)+ ) .
(43)
Inequality (26) is a direct consequence of (43). Moreover, we notice that for any u ∈ R, exp (u) ≤
2
1 + u + u2 exp (u+ ) and ln (1 + u) ≤ u, where u+ = u ∨ 0. By consequence, for λ ∈ [0, 1], it holds
ϕS (λ)/ n = ln (E [exp (λX1 )]) − λE [X1 ]
λ2 2
≤ ln 1 + λE [X1 ] + E X1 exp λ (X1 )+
− λE [X1 ]
2
λ2 2
λ2
λ2 2
≤
E X1 exp λ (X1 )+ ≤
E X1 exp (X1 )+ ≤
v.
2
2
2
Now, we get, for any y ≥ 0,
2
λ2 v
y
nv
1y>nv .
ϕ∗S (y) ≥ sup {λy−ϕS (λ)} ≥ sup λy−n
=
1y≤nv + y −
2
2nv
2
λ∈[0,1]
λ∈[0,1]
So, by using (42) with (44) taken with x = y/n, it holds
2
x
v
P ((Pn − P ) (ln ( f / f∗ )) ≥ x) ≤ exp −n
.
1x≤v + x − 1x>v
2v
2
(44)
(45)
To obtain (27), we notice that Inequality (45) implies by simple calculations, for any z ≥ 0,
!
r
z
2vz
v
P (Pn − P ) (ln ( f / f∗ )) ≥
1
+
+
1
≤ exp (−z) .
n z≤nv/2
n 2 z>nv/2
p
p
To conclude the proof, it suffices to remark that 2vz/n1z≤nv/2 + (z/n + v/2) 1z>nv/2 ≤ 2vz/n +
2z/n.
Proof of Proposition 4.5.
Let us first prove concentration inequality
h (28). i We set Xi :=
ln ( f∗ / f ) (ξi ) and use (42). For λ ∈ [0, r], we have by Hölder’s inequality, P ( f∗ / f )λ ≤ P [( f∗ / f )r ]λ/r .
Hence,
h
i
ϕS (λ)/ n = ln P ( f∗ / f )λ − λK (f∗ , f )
1
r
ln (P [( f∗ / f ) ]) − K (f∗ , f ) .
≤λ
r
Let us notice that by concavity of ln, we have
any y ≥ 0,
1
r
ln (P [( f∗ / f )r ]) − K (f∗ , f ) ≥ 0. Now we get, for
ϕ∗S (y) ≥ sup {λy−ϕS (λ)}
λ∈[0,r]
1
λ y−n
ln (P [( f∗ / f )r ]) − K (f∗ , f )
r
λ∈[0,r]
≥ sup
= (ry−n (ln (P [( f∗ / f )r ]) − rK (f∗ , f )))+ .
26
Saumard and Navarro
Using (42), we obtain, for any x ≥ 0,
P ((Pn − P )(ln (f /f∗ )) ≤ −x) ≤ exp (−n(rx − ln [( f∗ / f )r ] + rK (f∗ , f ))+ ) .
(46)
Inequality (28) is a straightforward consequence of (46). As in the proof of Theorem 4.4, we notice
2
that for any u ∈ R, exp (u) ≤ 1 + u + u2 exp (u+ ) and ln (1 + u) ≤ u, where u+ = u ∨ 0. By
consequence, for λ ∈ [0, r], it holds
ϕS (λ)/ n = λE [X1 ] + ln (E [exp (−λX1 )])
λ2 2
≤ ln 1 − λE [X1 ] + E X1 exp λ (−X1 )+
+ λE [X1 ]
2
λ2 2
λ2
λ2 2
≤
E X1 exp λ (−X1 )+ ≤
E X1 exp r (−X1 )+ ≤
wr .
2
2
2
Now we get, for any y ≥ 0,
≥ sup
λ∈[0,r]
ϕ∗S (y) ≥ sup {λy−ϕS (λ)}
λy−
nλ2 w
2
r
λ∈[0,r]
=
rnwr
y2
1y≤rnwr + r y −
1y>rnwr ,
2nwr
2
which gives
2
x
rnwr
P ((Pn − P )(ln (f /f∗ )) ≤ −x) ≤ exp −n
1x≤rnwr + r x −
1x>rnwr
2nwr
2
(47)
Inequality (29) is again a consequence of (47), by the same kind of arguments as those involved in
the proof of 27 in Lemma 4.4.
6.3
Proofs related to Section 4.3
Proof of Proposition 4.6. Let us take q > 1 such that 1/p + 1/q = 1. It holds
"
2 # Z
2
f
f
f
P
∨1
ln
= (f ∨ f∗ ) ln
dµ
f∗
f∗
f∗
2 ! p1 + 1q
Z
f
f∗ ∨ f
(f∗ ∧ f ) ln
dµ
=
f∗ ∧ f
f∗
2!
2 ! 1q
Z
p
f∗ ∨ f
f
f
=
(f∗ ∧ f ) ln
dµ
1 ln
f∗
f∗
(f∗ ∧ f ) q
1
p
1
!
2
Z
Z
q
2
(f∗ ∨ f )p
f
f
≤
(f∗ ∧ f ) ln
dµ
ln
dµ
p−1
f∗
f∗
(f∗ ∧ f )
|
{z
}
:=I
where in the last step we used Hölder’s inequality. Now, by Lemma 7.24 of Massart [35], it also
holds
2
Z
f
1
dµ ≤ K (f∗ , f ) .
(f∗ ∧ f ) ln
2
f∗
In order to prove (31), it thus remains to bound I in terms of p, c+ and c− only. First, we decompose
I into two parts,
2
2
2
Z
Z
Z
(f∗ ∨ f )p
f
f∗p
f∗
fp
f
dµ =
ln
1f∗ ≥f dµ +
ln
1f ≥f∗ dµ.
p−1 ln f
p−1
p−1
f
f
f∗
(f∗ ∧ f )
∗
f∗
(48)
Model Selection as a Multiple Testing Procedure
27
For the first term in the right-hand side of (48), we get
Z
2
2
Z
f∗
f∗
1−p
p
ln
1
dµ
≤
c
f
dµ
f∗ ≥f
∗ ln
−
f p−1
f
c−
Z
1−p
2
f∗p (ln f∗ )2 ∨ 1 dµ < +∞ ,
≤ 4c−
(ln c− ) ∨ 1
f∗p
(49)
where in the second inequality we used the following fact: (a + b)2 ≤ 4 a2 ∨ 1 b2 ∨ 1 , for any
real numbers a and b. The finiteness of the upper bound is guaranteed by assumption (30). For the
second term in the right-hand side of (48), it holds by same kind of arguments that lead to (49),
Z
2
c+
dµ
p−1
p−1 ln f
∗
f∗
f∗
Z
p
2
≤ 4c+ (ln c+ ) ∨ 1
f∗1−p (ln f∗ )2 ∨ 1 dµ < +∞ ,
fp
ln
f
f∗
2
1f ≥f∗ dµ ≤ cp+
Z
1
(50)
where in the last inequality we used the following fact: (a − b)2 ≤ 4 a2 ∨ 1 b2 ∨ 1 . Again, the
finiteness of the upper bound is guaranteed by (30). Inequality (31) then follows from combining
(48), (49) and (50).
Inequality (32) follows from the same kind of computations. Indeed, we have by the use of
Hölder’s inequality,
"
r 2 # Z r+1
2
f∗
f∗
f
f
P
=
∨1
ln
∨
f
ln
dµ
∗
f
f∗
fr
f∗
2 !1− r+1
2(r+1)
Z
p
r+1
r
p
f
∨
f
f
f
f
∗
∗
=
(f∗ ∧ f ) ln
dµ
ln
1− r+1
f∗
f∗
p
f r (f ∧ f )
∗
≤
r+1
p
p
rp
Z
2 !1− r+1
2
Z
p
p
r+1
f
∨
f
f r+1
f
f
∗
∗
(f∗ ∧ f ) ln
ln
dµ
dµ
p
rp
−1
f∗
f∗
f r+1 f∗r+1 ∧ f p−1
{z
}
|
.
:=Ir
In order to prove (32), it thus remains to bound Ir in terms of p, c+ and c− only. Again, we split Ir
into two parts,
Z
p
rp
f∗p ∨ f∗r+1 f r+1
rp
p
f r+1 f∗r+1
−1
∧ f p−1
2
2
2
Z
Z
f
f∗p
f∗
f
ln
dµ =
ln
1f∗ ≥f dµ+ f∗ ln
1f ≥f∗ dµ .
p−1
f∗
f
f
f∗
(51)
The first term in the right-hand side of (51) is given by (49) above. For the second term in the
right-hand side of (51), it holds
Z
2
2
Z
c
f
f∗ ln
1f ≥f∗ dµ ≤ f∗ ln +
dµ
f∗
f∗
≤2 (ln (c+ ))2 + 2P (ln f∗ )2 .
(52)
Furthermore we have f∗p−1 +f∗−p ≥ 1, so P (ln f∗ )2 ≤ P f∗p−1 (ln f∗ )2 +P f∗−p (ln f∗ )2 ≤ J +Q
and by (52),
2
Z
f
f∗ ln
1f ≥f∗ dµ ≤ 2 (ln (c+ ))2 + J + Q < +∞ ,
f∗
where the finiteness of the upper bound comes from (30). Inequality (32) then easily follows.
28
Saumard and Navarro
Proof of Proposition 4.7. Let us first prove Inequality (33). Considering the proof of Inequality
(31) of Lemma 4.6 given above, we see that it is sufficient to bound the second term in the right-hand
side of (48), applied with f = fm , in terms of Amin , J and p only. It holds
Z
fm
f∗p−1
2
2
Z
fm p
fm
fm
ln
ln
1fm ≥f∗ dµ ≤ Amin
dµ .
f∗
Amin
Amin
(53)
Now, we set h an auxilliary function, defined by h (x) = xp (ln x)2 for any x ≥ 1. It is easily seen
that h is convex. So it holds, for any I ∈ Λm ,
Z
Z
f∗
fm (I)
f∗ dµ
dµ
h
=h
≤ h
.
Amin
Amin µ (I)
I
I Amin µ (I)
From the latter inequality and from (53), we deduce
Z
fm
f∗p−1
ln
fm
f∗
2
1fm ≥f∗ dµ ≤ Amin
Z
h
fm
Amin
dµ ≤ Amin
Z
h
f∗
Amin
1−p
dµ ≤ 4Amin
(ln Amin )2 ∨ 1 J .
Inequality (33) is thus proved.
In the same manner, to establish Inequality (34) it suffices to adapt the proof of inequality (32)
given above by controlling the second term in the right-hand side of (51), applied with f = fm ,
in terms of Amin , p and P (ln f∗ )2 only. Let us notice that the function f defined on [1, +∞) by
f (x) = x (ln x)2 is convex. We have
2
2
Z
Z
fm
fm
fm
f∗ ln
1fm ≥f∗ dµ ≤ fm ln
dµ = Amin f
dµ .
f∗
Amin
Amin
R
R
(I)
dµ
f∗
dµ
∗
= f I Afmin
≤
f
Now, for any I ∈ Λm , it holds f fAmmin
Amin µ(I) . Hence,
I
µ(I)
Z
Z
2
Z
f∗
fm
f∗ ln
1fm ≥f∗ dµ ≤ Amin f
dµ ≤ 2P (ln f∗ )2 + 2 (ln Amin )2 ,
f∗
Amin
which gives the desired upper-bound and proves (34).
In the case where f∗ ∈ L∞ (µ), we have to prove (35). We have inf z∈Z f∗ (z) ≤ inf z∈Z fm (z) ≤ kfm k∞ ≤ kf∗ k∞ ,
so it holds
"
"
" #
2 #
r 2 #
2
kf∗ k∞ r∨1
fm
f∗
fm
fm
fm
P
∨1
ln
∨P
∨1
ln
≤
P
ln
.
f∗
f∗
fm
f∗
Amin
f∗
Now, Inequality (35) is a direct consequence of Lemma 1 of Barron and Sheu [10], which contains
the following inequality,
" #
2
fm
fm
P
ln
≤ 2 exp ln
K (f∗ , fm ) .
f∗
f∗
∞
This finishes the proof of Lemma 4.7.
References
[1] H. Akaike. Information theory and an extension of the maximum likelihood principle. In
Second International Symposium on Information Theory (Tsahkadsor, 1971), pages 267–281.
Akadémiai Kiadó, Budapest, 1973.
[2] S. Arlot. Resampling and Model Selection. PhD thesis, University Paris-Sud 11, Dec. 2007.
oai:tel.archives-ouvertes.fr:tel-00198803_v1.
Model Selection as a Multiple Testing Procedure
29
[3] S. Arlot. V -fold cross-validation improved: V -fold penalization, Feb. 2008. arXiv:0802.0566v2.
[4] S. Arlot. Model selection by resampling penalization. Electron. J. Stat., 3:557–624, 2009.
[5] S. Arlot and P. L. Bartlett. Margin-adaptive model selection in statistical learning. Bernoulli,
17(2):687–713, 05 2011.
[6] S. Arlot and A. Célisse. A survey of cross-validation procedures for model selection. Stat. Surv.,
4:40–79, 2010.
[7] S. Arlot and P. Massart. Data-driven calibration of penalties for least-squares regression. J.
Mach. Learn. Res., 10:245–279 (electronic), 2009.
[8] Y. Baraud, L. Birgé, and M. Sart. A new method for estimation and model selection: ρestimation. Invent. Math., 207(2):425–517, 2017.
[9] A. Barron, L. Birgé, and P. Massart. Risk bounds for model selection via penalization. Probab.
Theory Related Fields, 113(3):301–413, 1999.
[10] A. Barron and C. Sheu. Approximation of density functions by sequences of exponential families.
Ann. Statist., 19(3):1347–1369, 1991.
[11] P. Bellec and A. Tsybakov. Bounds on the prediction error of penalized least squares estimators
with convex penalty. In V. Panov, editor, Modern Problems of Stochastic Analysis and Statistics,
pages 315–333, Cham, 2017. Springer International Publishing.
[12] P. C. Bellec, G. Lecué, and A. B. Tsybakov. Towards the study of least squares estimators with
convex penalty. arXiv preprint arXiv:1701.09120, 2017.
[13] A. Berlinet and L. Devroye. A comparison of kernel density estimates. Publications de l’Institut
de Statistique de l’Université de Paris, 38(3):3–59, 1994.
[14] P. Bühlmann, L. Meier, and S. van de Geer. Discussion: “a significance test for the lasso”. Ann.
Statist., 42(2):469–477, 04 2014.
[15] L. Birgé. Model selection via testing: an alternative to (penalized) maximum likelihood estimators. Ann. Inst. H. Poincaré Probab. Statist., 42(3):273–325, 2006.
[16] L. Birgé. Robust tests for model selection. In From probability to statistics and back: highdimensional models and processes, volume 9 of Inst. Math. Stat. (IMS) Collect., pages 47–64.
Inst. Math. Statist., Beachwood, OH, 2013.
[17] L. Birgé and Y. Rozenholc. How many bins should be put in a regular histogram. ESAIM
Probab. Stat., 10:24–45 (electronic), 2006.
[18] S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities. Oxford University Press,
2013.
[19] S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A Nonasymptotic Theory
of Independence. Oxford University Press, Oxford, 2013.
[20] S. Boucheron and P. Massart. A high-dimensional Wilks phenomenon. Probab. Theory Related
Fields, 150(3-4):405–433, 2011.
[21] O. Bousquet. A Bennett concentration inequality and its application to suprema of empirical
processes. C. R. Math. Acad. Sci. Paris, 334(6):495–500, 2002.
[22] P. Burman. Estimation of equifrequency histogram. Statist. Probab. Lett., 56(3):227–238, 2002.
30
Saumard and Navarro
[23] G. Castellan. Modified Akaike’s criterion for histogram density estimation. Technical report
]99.61, Université Paris-Sud, 1999.
[24] G. Castellan. Density estimation via exponential model selection. IEEE Trans. Inform. Theory,
49(8):2052–2060, 2003.
[25] S. Chatterjee. A new perspective on least squares under convex constraint. Ann. Statist.,
42(6):2340–2381, 12 2014.
[26] G. Claeskens and N. L. Hjort. Model selection and model averaging. Cambridge Series in
Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 2008.
[27] I. Csiszár. I-divergence geometry of probability distributions and minimization problems. Ann.
Probab., 3(1):146–158, 1975.
[28] P. D. Grünwald. The minimum description length principle. MIT press, 2007.
[29] C. M. Hurvich and C.-L. Tsai. Model selection for least absolute deviations regression in small
samples. Statist. Probab. Lett., 9(3):259–265, 1990.
[30] T. Klein and E. Rio. Concentration around the mean for maxima of empirical processes. Ann.
Probab., 33(3):1060–1077, 2005.
[31] G. Lecué and M. Lerasle. Learning from mom’s principles: Le cam’s approach. 2017.
[32] R. Lockhart, J. Taylor, R. J. Tibshirani, and R. Tibshirani. Rejoinder: “a significance test for
the lasso”. Ann. Statist., 42(2):518–531, 04 2014.
[33] R. Lockhart, J. Taylor, R. J. Tibshirani, and R. Tibshirani. A significance test for the lasso.
Ann. Statist., 42(2):413–468, 04 2014.
[34] E. Mammen and A. Tsybakov. Smooth discrimination analysis. Ann.Stat., 27:1808–1829, 1999.
[35] P. Massart. Concentration inequalities and model selection, volume 1896 of Lecture Notes in
Mathematics. Springer, Berlin, 2007. Lectures from the 33rd Summer School on Probability
Theory held in Saint-Flour, July 6–23, 2003, With a foreword by Jean Picard.
[36] P. Massart and E. Nédélec. Risks bounds for statistical learning. Ann.Stat., 34(5):2326–2366,
2006.
[37] T. Mildenberger and H. Weinert. The benchden package: Benchmark densities for nonparametric density estimation. Journal of Statistical Software, Articles, 46(14):1–14, 2012.
[38] A. Muro and S. van de Geer. Concentration behavior of the penalized least squares estimator.
arXiv preprint arXiv:1511.08698, 2015.
[39] A. Saumard. Optimal upper and lower bounds for the true and empirical excess risks in heteroscedastic least-squares regression. Electron. J. Statist., 6(1-2):579–655, 2012.
[40] C. M. Stein. Estimation of the mean of a multivariate normal distribution. Ann. Statist.,
9(6):1135–1151, 1981.
[41] C. Stone. An asymptotically optimal histogram selection rule. In Proceedings of the Berkeley conference in honor of Jerzy Neyman and Jack Kiefer, Vol. II (Berkeley, Calif., 1983),
Wadsworth Statist./Probab. Ser., pages 513–520, Belmont, CA, 1985. Wadsworth.
[42] N. Sugiura. Further analysts of the data by Akaike’s information criterion and the finite corrections. Communications in Statistics - Theory and Methods, 7(1):13–26, 1978.
Model Selection as a Multiple Testing Procedure
31
[43] A. Tsybakov. Optimal aggregation of classifiers in statistical learning. Ann.Stat., 32:135–166,
2004.
[44] S. van de Geer and M. J. Wainwright. On concentration for (regularized) empirical risk minimization. Sankhya A, 79(2):159–200, Aug 2017.
[45] H. Zou, T. Hastie, and R. Tibshirani. On the “degrees of freedom” of the lasso. Ann. Statist.,
35(5):2173–2192, 2007.
32
A
Saumard and Navarro
Further Proofs and Theoretical Results
A.1
Proofs Related to Section 3.1
Most of the arguments given in the proofs of this section are borrowed from Castellan [23]. We
essentially rearrange these arguments in a more efficient way, thus obtaining better concentration
bounds than in [23] (or [35], see also [18]).
We also set, for any ε > 0, the event Ωm (ε) where some control of fˆm in sup-norm is achieved,
(
)
fˆm − fm
Ωm (ε) =
≤ε .
fm
∞
As we have the following formulas for the estimators and the projections of the target,
X Pn (I)
1I
µ (I)
fˆm =
,
fm =
I∈Λm
X P (I)
1I ,
µ (I)
I∈Λm
we deduce that,
fˆm − fm
fm
T
= sup
I∈Λm
∞
|(Pn − P ) (I)|
.
P (I)
(54)
Hence, it holds Ωm (ε) = I∈m {|Pn (I) − P (I)| ≤ εP (I)}.
Let us state the main result of this section, concerning upper and lower bounds for the true and
empirical excess risks of the histograms on each model. These bounds are optimal to the first order.
Before giving the proof of Theorem 3.1, the following lemma will be useful. It describes the
consistency in sup-norm of the histogram estimators, suitably normalized by the projections of the
target on each model.
Lemma A.1 Let α, A+ and AΛ be positive constants. Consider a finite partition m of Z, with
cardinality Dm . Assume
0 < AΛ ≤ Dm inf {P (I)}
I∈m
and
0 < Dm ≤ A+
n
≤n.
ln(n + 1)
(55)
Then by setting
Rn∞ (m) =
s
2 (α + 1) Dm ln(n + 1) (α + 1) Dm ln(n + 1)
+
,
AΛ n
AΛ n
we get
P
In other words,
fˆm − fm
fm
∞
≤
!
Rn∞ (m)
≥ 1 − 2(n + 1)−α .
(56)
(57)
P (Ωm (Rn∞ (m)))
only depending on α, A+
≥ 1 − 2(n + 1)−α . In addition,
there exists a positive constant Ac ,
q
Dm ln(n+1)
∞
and AΛ , such that Rn (m) ≤ Ac
. Furthermore, if
n
√
√
(α + 1) A+
≤ τ = 3 − 2 < 0.32 ,
AΛ
or if n ≥ n0 (α, A+ , AΛ ), then Rn∞ (m) ≤ 1/2 .
Proof of Lemma A.1. Let β > 0 to be fixed later. Recall that, by (54) we have
fˆm − fm
fm
= sup
∞
I∈Λm
|(Pn − P ) (I)|
.
P (I)
(58)
Model Selection as a Multiple Testing Procedure
By Bernstein’s inequality (see Proposition 2.9 in [35]) applied to the random variables
for all x > 0,
"
#
r
2P (I) x
x
P |(Pn − P ) (I)| ≥
≤ 2 exp (−x) .
+
n
3n
33
1ξi ∈I we get,
Taking x = β ln(n + 1) and normalizing by the quantity P (I) > 0 we get
"
|(Pn − P ) (I)|
P
≥
P (I)
s
#
2β ln(n + 1) β ln(n + 1)
≤ 2(n + 1)−β .
+
P (I) n
P (I) 3n
(59)
Now, by the first inequality in (55), we have 0 < P (I)−1 ≤ A−1
Λ Dm . Hence, using (59) we get
|(Pn − P ) (I)|
P
≥
P (I)
s
2βDm ln(n + 1) βDm ln(n + 1)
+
≤ 2(n + 1)−β ,
AΛ n
AΛ n
We then deduce from (58) and (60) that
"
fˆm − fm
P
fm
∞
≥
#
Rn∞ (m)
≤
(60)
2Dm
(n + 1)β
and, since Dm ≤ n, taking β = α + 1 yields Inequality (57). The other facts of Lemma A.1 then
follow from simple computations.
Proof of Theorem 3.1. Recall that α is fixed. By Inequality (23) in Proposition 4.2 applied with
θ = Rn∞ (m)—where Rn∞ (m) is defined in (56) with our fixed value of α—and x = α ln(n + 1), it
holds with probability at least 1 − (n + 1)−α ,
χn (m) 1Ωm (Rn∞ (m)) ≤
As Rn∞ (m) ≤ Lα,A+ ,AΛ
r
r
∞ (m)
p
Dm
R
2α ln(n + 1)
n
+ 1 + 2Rn∞ (m) +
.
n
6
n
p
Dm ln(n + 1)/n ≤ Lα,A+ ,AΛ , (61) gives
χn (m) 1Ωm (Rn∞ (m))
We set the event Ω0 on which we have
r
Dm
ln(n + 1)
≤
+ Lα,A+ ,AΛ
n
n
s
r
Dm
ln(n + 1)
=
1 + Lα,A+ ,AΛ
.
n
Dm
r
fˆm − fm
fm
χn (m) ≤
and
(61)
r
∞
≤ Rn∞ (m) ,
Dm
1 + Lα,A+ ,AΛ
n
s
χn (m) ≥ 1 − Ag
(62)
ln(n + 1)
∨
Dm
s
ln(n + 1)
,
Dm
r
p
ln(n + 1) Dm
.
n
n1/4
(63)
(64)
In particular, Ω0 ⊂ Ωm (Rn∞ (m)). By (62), Lemma A.1 and Proposition 4.3, it holds P (Ω0 ) ≥
1 − 4(n + 1)−α . It suffices to prove the inequalities of Theorem 3.1 on Ω0 . The following inequalities,
34
Saumard and Navarro
between the excess risk on m and the chi-square statistics, are shown in [23] (Inequalities (2.13)).
For any ε ∈ (0, 1), on Ωm (ε),
1−ε
2
ˆm ≤ 1 + ε χ2 (m) .
χ
(m)
≤
K
f
,
f
m
n
n
2 (1 + ε)2
2 (1 − ε)2
(65)
∞
Under the condition (α + 1) A+ A−1
Λ < τ , Proposition A.1 gives Rn (m) < 1. Hence, by applying
the right-hand side of (65) with ε = Rn∞ (m) ≤ 1/2, using (63) and the fact that (1 − )−1 ≤ 1 + 2
for ≤ 1/2, we get on Ω0 ,
K fm , fˆm ≤
1 + Rn∞ (m)
χ2n (m)
2 (1 − Rn∞ (m))2
2
s
!
r
1
Dm ln(n + 1) Dm
ln(n + 1)
≤
+ Lα,A+ ,AΛ
1 + Lα,A+ ,AΛ
.
2
n
n
Dm
Then simple computations allow to get the right-hand side inequality in (16).
By applying the left-hand side of (65) with ε = Rn∞ (m) ≤ 1/2 and using the fact that (1+)−1 ≥
1 − , we also get on Ω0 ,
K fm , fˆm ≥
1 − Rn∞ (m)
2
2 χn (m)
∞
2 (1 + Rn (m))
s
2
p
ln(n
+
1)
ln(n + 1)
Dm
≥(1 − Rn∞ (m))3 1 − Ag
∨
Dm
2n
n1/4
s
p
ln(n
+
1)
ln(n
+
1)
Dm
≥(1 − (3Rn∞ (m) ∧ 1)) 1 − 2Ag
∨
1/4
Dm
2n
n
The left-hand
side inequality
in (16) then
p
p
p follows by simple computations, noticing in particular that
−1/4
n
ln(n + 1) ≤ ln(n + 1)/Dm ∨ Dm ln(n + 1)/n.
Inequalities in (17) follow from the same kind of arguments as those involved in the proofs of
inequalities in (16). Indeed, from [35, Lemma 7.24]—or [23, Lemma 2.3] —, it holds
1
2
Z
fˆm ∧ fm
fˆm
ln
fm
!2
dµ ≤ K fˆm , fm
Z
1
≤
2
1+ε
≤
2
fˆm ∨ fm
fˆm
ln
fm
!2
dµ .
We deduce that for ε ∈ (0, 1), we have on Ωm (ε),
1−ε
2
Z
fm
fˆm
ln
fm
!2
dµ ≤ K fˆm , fm
Z
fm
fˆm
ln
fm
!2
dµ .
(66)
As for every x > 0, we have (1 ∨ x)−1 ≤ (x − 1)−1 ln x ≤ (1 ∧ x)−1 , (66) leads by simple computations to the following inequalities
1−ε
2
ˆm , fm ≤ 1 + ε χ2 (m) .
χ
(m)
≤
K
f
n
n
2 (1 + ε)2
2 (1 − ε)2
We thus have the same upper and lower bounds, in terms of the chi-square statistic χ2n (m), for the
empirical excess risk as for the true excess risk.
Model Selection as a Multiple Testing Procedure
A.2
35
Oracle inequalities and dimension guarantees
Using the notations of Section 3.2, we define the set of assumptions (SA0 ) to be the conjunction
of assumptions (P1), (P2), (Asm) and (Alr). The set of assumptions (SA) of Section 3.2 thus
consists on assuming (SA0 ) together with (Ap).
For some of the following results, we will also need the following assumptions.
√
√
(P3) Richness of Mn : there exist m0 , m1 ∈ Mn such that Dm0 ∈ [ n, crich n] and Dm1 ≥
Arich n (ln(n + 1))−2 .
(Apu ) The bias decreases as a power of Dm : there exist β+ > 0 and C+ > 0 such that
−β+
K (f∗ , fm ) ≤ C+ Dm
.
(Ap) The bias decreases like a power of Dm : there exist β− ≥ β+ > 0 and C+ , C− > 0 such that
−β−
−β+
C− D m
≤ K (f∗ , fm ) ≤ C+ Dm
.
Theorem 3.2 is a direct corollary of the following theorem.
Theorem A.2 Take n ≥ 1 and r ∈ (0, p − 1). Assume that the set of assumptions (SA0 ) holds and
that for some θ ∈ (1/2, +∞) and ∆ > 0,
Dm
pen (m) = θ + ∆ε+
,
n (m)
n
for every model m ∈ Mn . Then there exists an event Ωn of probability 1 − (n + 1)−2 and some
positive constants A1 , A2 depending only on the constants defined in (SA0 ) such that, if ∆ ≥
(θ − 1)− A1 + A2 > 0 then we have on Ωn ,
3p−1−r
n
o
(ln(n + 1)) 2(p+1+r)
ˆ
.
inf
K f∗ , fm +L(SA),θ,r
p
m∈Mn
n p+1+r
(67)
Assume furthermore that Assumption (Apu ) holds. Then it holds on Ωn ,
−1/2
1 + 2 (θ − 1) + L
(SA),θ,r (ln(n + 1))
+
K f∗ , fˆm
≤
b
1 − 2 (θ − 1)−
Dm
b ≤ L(SA),∆,θ,r n
1
2+β+ 1− r+1
p
(
)p
ln(n + 1)
1
,
Dm∗ ≤ L(SA) n 1+β+ .
In particular, if we are in the case where p < β+ then Inequality (67) reduces to
3p−1−r
(ln(n + 1)) 2(p+1+r)
ˆ
K f∗ , fm
.
p
b ≤ L(SA),θ,r
n p+1+r
(68)
Grant Assumption (Ap). Then it holds on Ωn ,
β+
(1)
L∆,θ,(SA)
1
n β− (1+β+ )
(ln(n + 1))
(1)
1
β−
≤ Dm
b ≤
2+β+ (1−
(2)
L∆,θ,(SA) n
β+
(2)
r+1
p
)p
ln(n + 1) ,
1
L(SA) n (1+β+ )β− ≤ Dm∗ ≤ L(SA) n 1+β+
and
−
β+
p
ln(n + 1)
3p−1−r
n
o
(ln(n + 1)) 2(p+1+r)
ˆ
inf
K f∗ , fm +L(SA),θ,r
p
m∈Mn
n p+1+r
(69)
Furthermore, if β− < p (1 + β+ ) /(1 + p + r) or p/(1 + r) > β− + β− /β+ − 1, then we have on Ωn ,
K f∗ , fˆm
b
1 + 2 (θ − 1)+ + L(SA),θ,r n (1+β+ )β−
≤
1 − 2 (θ − 1)−
−1/2
1 + 2 (θ − 1) + L
(SA),θ,r (ln(n + 1))
+
K f∗ , fˆm
≤
b
1 − 2 (θ − 1)−
inf
m∈Mn
n
o
K f∗ , fˆm
.
(70)
36
Saumard and Navarro
We obtain in Theorem A.2 oracle inequalities and dimension bounds for the oracle and selected
models.
In order to avoid cumbersome notations in the proofs of Theorem A.2, when generic constants L
and n0 depend on constants defined in the sets of assumptions (SA0 ) or (SA), we will note L(SA)
and n0 ((SA)). The values of the these constants may change from line to line, or even whithin one
line.
Proof of Theorem 3.2
• Proof of oracle Inequality (67):
From the definition of the selected model m
b given in (4), m
b minimizes
crit (m) := Pn γ(fˆm ) + pen (m) ,
over the models m ∈ Mn . Hence, m
b also minimizes
crit0 (m) := crit (m) − Pn (γ (f∗ ))
(71)
over the collection Mn . Let us write
K f∗ , fˆm = P γ(fˆm ) − γ (f∗ )
= Pn γ(fˆm ) + Pn γ(fm ) − γ(fˆm ) + (Pn − P ) (γ (f∗ ) − γ(fm ))
+ P γ(fˆm ) − γ(fm ) − Pn (γ (f∗ )) .
By setting
p1 (m) = P γ(fˆm ) − γ(fm ) = K fm , fˆm ,
p2 (m) = Pn γ(fm ) − γ(fˆm ) = K fˆm , fm ,
δ̄ (m) = (Pn − P ) (γ (f∗ ) − γ(fm )) = (Pn − P ) (ln ( fm / f∗ ))
and
pen0id (m) = p1 (m) + p2 (m) + δ̄ (m) ,
we have
and by (71),
K f∗ , fˆm = Pn γ(fˆm ) + p1 (m) + p2 (m) + δ̄ (m) − Pn (γ (f∗ ))
crit0 (m) = K f∗ , fˆm + pen (m) − pen0id (m) .
(72)
As m
b minimizes crit0 over Mn , itis therefore sufficient by (72) to control pen (m) − pen0id (m) in
terms of the excess risk K f∗ , fˆm , for every m ∈ Mn , in order to derive oracle inequalities. We
further set
"
"
2 #
r 2 #
fm
f∗
fm
fm
∨1
ln
and wm = P
∨1
ln
.
Km = K (f∗ , fm ) , vm = P
f∗
f∗
fm
f∗
Let Ωn be the event on which:
• For all models m ∈ Mn , we set zn = (2 + αM ) ln(n + 1) + 2 ln 2 and we have,
r
2vm zn 2zn
δ̄ (m) ≤
+
n
n
r
2wm zn 2zn
−δ̄ (m) ≤
+
n
nr
Dm
≤ p1 (m) −
2n
Dm
−L(SA),α ε−
≤ p2 (m) −
n (m)
2n
−L(SA),α ε−
n (m)
Dm
Dm
≤ L(SA),α ε+
n (m)
2n
2n
Dm
Dm
+
≤ L(SA),α εn (m)
2n
2n
(73)
(74)
(75)
(76)
Model Selection as a Multiple Testing Procedure
37
By Theorem 3.1 applied with α = 5 + αM and Propositions 4.4 and 4.5 applied with z = zn , we get
P (Ωn ) ≥ 1 −
= 1−
X
4(n + 1)
−5−αM
m∈Mn
X
m∈Mn
(n + 1)−2−αM
+
2
(n + 1)−2−αM ≥ 1 − (n + 1)−2 .
Dm
The following simple remark will be used along the proof: for any m ∈ Mn , zn /n ≤ L(SA) ε+
n (m) n .
Notice also that ε−
b ≤ ε+
b (see Theorem 3.1).
n (m)
n (m)
(1)
By using (33), (72), (73), (75) and (76), we get that on Ωn , for ∆ of the form (θ − 1)− L(SA) +
(2)
(1)
(2)
L(SA) with L(SA) and L(SA) sufficiently large,
crit0 (m)
b
r
2vm
2zn
b zn
ˆ
≥K f∗ , fm
b − p1 (m)
b − p2 (m)
b −
−
b + pen (m)
n
n
s
1−1/p
+
zn
2AM R,− Km
Dm
Dm
b −1
b
b
≥K f∗ , fˆm
+
(θ
−
1)
+
∆
−
L
ε
(
m)
b
−
b
(SA)
n
n
n
n
Dm
Dm
b −1
b
≥K f∗ , fˆm
− L(SA) ε−
b
b − 2 (θ − 1)−
n (m)
2n
n
s
1−1/p
zn
2AM R,− Km
Dm
(1)
(2)
b
b
+ ∆ − L(SA) (θ − 1)− + L(SA) ε+
(
m)
b
−
n
n
n
s
1−1/p
2AM R,− Km
zn
b
≥ 1 − 2 (θ − 1)− K f∗ , fˆm
−
.
b
n
(77)
Note that 1 − 2 (θ − 1)− > 0. Let us take η ∈ 0, 1/2 − (θ − 1)− , so that
1 − 2 (θ − 1)− − η > 1/2 − (θ − 1)− > 0 .
p−1
p
− p−1
2p
2p , b = η
2p
By Lemma A.5 applied with a = (ηKm
2AM R,− zn /n, u = p−1
and v =
b)
have
s
p p−1
1−1/p
2AM R,− Km
zn
ln(n + 1) p+1 1 p+1
b
≤ ηKm
.
b + L(SA),α
n
n
η
2p
p+1 ,
we
By using the latter inequality in (77) we obtain,
p p−1
ln(n + 1) p+1 1 p+1
ˆ
crit (m)
b ≥ 1 − 2 (θ − 1)− − η K f∗ , fm
.
b − L(SA),α
n
η
0
(78)
We compute now an upper bound on crit0 for each model m. By Lemma A.5 applied with a =
p−r−1
− p−1−r p
2p
2p
(ηKm ) 2p , b = η 2p
2AM R,− zn /n, u = p−1−r
and v = p+1+r
, we have
s
1− r+1
p
2AM R,− Km
n
zn
≤ ηKm + L(SA),r
ln(n + 1)
n
p
p+1+r
p−1−r
1 p+1+r
.
η
38
Saumard and Navarro
By (74), (72), (75), (76), (34) and by using Lemma A.5 we have on Ωn ,
0
ˆ
crit (m) = K f∗ , fm + pen (m) − p1 (m) − p2 (m) − δ̄ (m)
D
D
m
m
−
≤K f∗ , fˆm + 2 (θ − 1)+
− L(SA) εn (m)
2n
n
s
1− r+1
p
2A
K
zn 2zn
D
M R,− m
(2)
(1)
m
+ ∆ + (θ − 1)+ L(SA) + L(SA) ε+
(m)
+
+
n
n
n
nr
Dm
(1)
(2)
≤ 1 + 2 (θ − 1)+ + η K f∗ , fˆm + ∆ + (θ − 1)+ L(SA) + L(SA),r ε+
n (m)
n
p p−1−r
p+1+r
p+1+r
ln(n + 1)
1
+ L(SA),r
.
n
η
(2)
(1)
Recall that we took ∆ = L(SA) (θ − 1)− + L(SA) for some positive constants sufficiently large, so
(1)
(2)
(1)
(2)
∆ + (θ − 1)+ L(SA) + L(SA),r ≤ |θ − 1| L(SA) + L(SA),r
and we finally get,
Dm
(1)
(2)
crit0 (m) ≤ 1 + 2 (θ − 1)+ + η K f∗ , fˆm + |θ − 1| L(SA) + L(SA),r ε+
n (m)
n
p p−1−r
ln(n + 1) p+1+r 1 p+1+r
+ L(SA),r
.
n
η
Now, as m
b minimizes crit0 we get from (78) and (79), on Ωn ,
1 + 2 (θ − 1) + η
Dm ∗
+
+
ˆ
ˆ
K f∗ , fm
K f∗ , fm∗ + L(SA),θ,r εn (m∗ )
b ≤
1 − 2 (θ − 1)− − η
n
p p−1−r
ln(n + 1) p+1+r 1 p+1+r
+L(SA),θ,r
.
n
η
We distinguish two cases. If Dm∗ ≥ L(SA) (ln(n + 1))2 with a constant L(SA) chosen such that
1
1
A0 ε+
≤ √
<1,
n (m∗ ) ≤ p
2 ln 2
2 ln(n + 1)
where A0 and ε+
n (m) are defined in Theorem 3.1, then by Theorem 3.1 it holds on Ωn ,
!
1
Dm∗
K f∗ , fˆm∗ ≥ 1 − p
.
2n
2 ln(n + 1)
On the other hand, if Dm∗ ≤ L(SA) (ln(n + 1))2 then by Theorem 3.1 we have on Ωn ,
Dm ∗
(ln(n + 1))3
K f∗ , fˆm∗ + L(SA),θ,r ε+
≤
L
.
(m
)
∗
(SA),θ,r
n
n
n
Hence, in any case we always have on Ωn ,
Dm ∗
K f∗ , fˆm∗ + L(SA),θ,r ε+
n (m∗ )
n
(ln(n + 1))3
≤L(SA),θ,r
+ 1 + L(SA),θ,r (ln(n + 1))−1/2 K f∗ , fˆm∗ .
n
(79)
(80)
Model Selection as a Multiple Testing Procedure
39
By taking η = (ln(n + 1))−1/2 1/2 − (θ − 1)− and using the fact that in this case,
1 + 2 (θ − 1)+ + η
1 + 2 (θ − 1)+ + Lθ η
≤
,
1 − 2 (θ − 1)− − η
1 − 2 (θ − 1)−
(81)
we deduce that Inequality (80) gives,
−1/2
1 + 2 (θ − 1) + L
(SA),θ,r (ln(n + 1))
+
ˆ
K f∗ , fm
b ≤
1 − 2 (θ − 1)−
which is Inequality (67).
• Proof of Inequality (68):
3p−1−r
n
o
(ln(n + 1)) 2(p+1+r)
ˆ
inf
K f∗ , fm +L(SA),θ,r
p
m∈Mn
n p+1+r
1
By Lemma A.4 we know that Dm∗ ≤ L(SA) n 1+β+ on Ωn . Furthermore, we have on Ωn , by simple
computations,
K f∗ , fˆm =K (f∗ , fm ) + K fm , fˆm
Dm
−β+
+ 1 + L(SA) ε+
≤C+ Dm
n (m)
2n
D
ln(n
+
1)
m
−β+
+
≤L(SA) Dm
+
.
n
n
This yields
n
o
1
Dm ln(n + 1)
−β+
1+β+
ˆ
+
; m ∈ Mn , Dm ≤ L(SA) n
K f∗ , fm
≤L(SA) inf Dm +
inf
m∈Mn
n
n
β
≤L(SA) n
+
− 1+β
+
.
To conclude, it suffices to notice that if β+ > p then
β
n
+
− 1+β
+
≤n
p
− p+1
≤n
p
− p+1+r
,
which finally gives
1 + 2 (θ − 1)+ + L(SA),θ,r (ln(n + 1))−1/2
1 − 2 (θ − 1)−
3p−1−r
n
o
(ln(n + 1)) 2(p+1+r)
ˆ
.
inf
K f∗ , fm
≤ L(SA),θ,r
p
M ∈Mn
n p+1+r
• Proof of Inequality (69):
(1)
(2)
From (Ap), we know by Lemma A.4 that there exist L(SA) , L(SA) > 0 such that
β+
(1)
(2)
1
L(SA) n (1+β+ )β− ≤ DM∗ ≤ L(SA) n 1+β+
and so
ε+
n (m∗ )
s
)
Dm∗ ln(n + 1)
ln(n + 1) ln(n + 1)
≤ max
;
;
n
Dm∗
Dm∗
β+
β+
p
−
−
≤L(SA) max n 2(1+β+ ) ; n 2(1+β+ )β−
ln(n + 1)
(r
=L(SA) n
Assume for now that we also have
−
β+
(1+β+ )β−
p
ln(n + 1) .
−1/2
A0 ε+
,
n (m∗ ) ≤ (ln(n + 1))
(82)
(83)
40
Saumard and Navarro
where A0 and ε+
n (m) are defined in Theorem 3.1. Then by Theorem 3.1 it holds on Ωn ,
D
m∗
K f∗ , fˆm∗ ≥ 1 − (ln(n + 1))−1/2
.
2n
In this case, we deduce that
Dm ∗
K f∗ , fˆm∗ + L(SA),θ,r ε+
n (m∗ )
n
β+
p
−
1+β+ )β−
(
≤ 1 + L(SA),θ,r n
ln(n + 1) K f∗ , fˆm∗ ,
(84)
and Inequality (69) simply follows from using Inequality (80).
If Inequality (83) is not satisfied, that is
−1/2
A0 ε+
,
n (m∗ ) > (ln(n + 1))
(85)
then by (82), this means that there exists a positive constant L(SA) such that
L(SA) n
−
β+
(1+β+ )β−
p
ln(n + 1) > (ln(n + 1))−1/2 .
Consequently, this ensures that in the case where (85) is true, we also have n ≤ n0 ((SA)). Hence,
as
β+
−β−
(1+β+ )β− > 0 ,
≥
L
n
K f∗ , fˆm∗ ≥ C− Dm
(SA)
∗
this yields Inequality (84) with a positive constant L(SA),θ,r in the right-hand term sufficiently large
and then the result easily follows from using Inequality (80).
• Proof of Inequality (70):
If Dm∗ ≥ L(SA) (ln(n + 1))2 with a constant L(SA) chosen such that
A 0 ε+
n (m∗ ) ≤ 1/2 ,
where A0 and ε+
n (m) are defined in Theorem 3.1, then by Theorem 3.1 it holds on Ωn ,
Dm ∗
−β−
+
.
K f∗ , fˆm∗ ≥ C− Dm
∗
4n
β+
By Lemma A.4 we know that
(1)
L(SA) n (1+β+ )β−
(2)
1
≤ Dm∗ ≤ L(SA) n 1+β+ on Ωn . This gives
β+
β
−1+
− −
(1+β+ )β−
K f∗ , fˆm∗ ≥ L(SA) n 1+β+ + L(SA) n
and we deduce by simple algebra that if β− < p (1 + β+ ) /(1+p+r) or p/(1+r) > β+ /(β− (1 + β+ ))−
1, then
3p−1−r
(ln(n + 1)) 2(p+1+r)
−1/2
ˆ
L(SA) (ln(n + 1))
K f∗ , fm∗ ≥
.
p
n p+1+r
On the other hand, if Dm∗ ≤ L(SA) (ln(n + 1))2 , this implies in particular
β+
(1)
L(SA) n (1+β+ )β−
≤ L(SA) (ln(n + 1))2 .
Hence, there exists an integer n0 ((SA)) such that n ≤ n0 ((SA)). In this case, we can find a constant
L(SA) such that
3p−1−r
(ln(n + 1)) 2(p+1+r)
−1/2
L(SA) (ln(n + 1))
K f∗ , fˆm∗ ≥
p
n p+1+r
and through the use of inequality (69), this conclude the proof of Inequality (70).
Model Selection as a Multiple Testing Procedure
41
Lemma A.3 (Control on the dimension of the selected model) Assume that (SA0 ) holds top
gether with (Apu ). If β+ ≤ r+1
then, on the event Ωn defined in the proof of Theorem A.2, we have
Dm
b ≤ L∆,θ,r,(SA) n
1
2+β+ 1− r+1
p
(
If moreover (Ap) holds, then we get on the event Ωn ,
)p
ln(n + 1) .
β+
(1)
L∆,θ,(SA)
1
n β− (1+β+ )
(ln(n + 1))
1
β−
≤ Dm
b ≤
2+β+ (1−
(2)
L∆,θ,(SA) n
(86)
)p
ln(n + 1) .
r+1
p
(87)
Lemma A.4 (Control on the dimension of oracle models) Assume that (SA0 ) holds together
with (Apu ). We have on the event Ωn defined in the proof of Theorem A.2,
1
Dm∗ ≤ L(SA) n 1+β+ .
If moreover (Ap) holds, then we get on the event Ωn ,
(1)
β+
1
(2)
L(SA) n (1+β+ )β− ≤ Dm∗ ≤ L(SA) n 1+β+ .
(88)
Proof of Lemma A.3. Recall that m
b minimizes
crit0 (m) = crit (m) − Pn γ (f∗ ) = Km − p2 (m) − δ̄ (m) + pen (m)
over the models m ∈ Mn . Moreover, pen (m) = (θ + ∆ε+
n (m)) Dm /n. The analysis is restricted on
Ωn .
1. Upper bound on crit0 (m):
1
Dm
p2 (m) ≥
− L(SA) ε+
(m)
n
2
n
r
2wm zn 2zn
−δ̄ (m) ≤
+
.
n
nr
1− r+1
p
Moreover, by Lemma 4.6, we have wm ≤ AM R,− Km
and so,
s
1− r+1
p
2A
K
zn
1
D
M R,− m
m
crit0 (m) ≤Km + θ − + L∆,(SA) ε+
(m)
+
n
2
n
n
v
u
u −β+ 1− r+1
p
−β
Dm ln(n + 1) t Dm
ln(n + 1)
+
.
D
≤L∆,θ,(SA)
+
+
+
m
n
n
n
Now, if β+ ≤
p
r+1 ,
then for m0 such that Dm0
β
β+
Dm0
− +
− 1+β
−β+
+ ;
≤ 2n 1+β+ ; Dm
≤
n
0
n
so we get
1
= n 1+β+ we have
v
u
u −β+ 1− r+1
p
t Dm 0
n
β
crit0 (m0 ) ≤ L∆,θ,(SA) n
+
− 1+β
+
≤n
−
(
)
β+ 2− r+1
+1
p
2(1+β+ )
p
ln(n + 1) .
β
≤n
+
− 1+β
+
,
(89)
42
Saumard and Navarro
Otherwise, if β+ >
p
r+1 ,
&
then for m1 such that Dm1 = n
1
2+β+ 1− r+1
p
(
'
) , we have
v
u
1+β+ (1− r+1
u −β+ 1− r+1
(
)
p )
p
−
−
t
r+1
D
m1
2+β+ (1− p )
2+β+ (1− r+1
p ) ;
;
≤ 2n
≤n
n
1+β+ (1− r+1
β+
p )
−
−
2+β+ (1− r+1
2+β+ (1− r+1
)
p
p )
≤n
≤n
1+β+ 1− r+1
p
Dm1
n
−β+
Dm
1
and
−
0
0
crit (m1 ) ≤ L∆,θ,(SA) n
2. Lower bound on crit (m): we have
(
(
1+β+ 1− r+1
p
2+β+ 1− r+1
p
)
p
) ln(n + 1) .
(90)
1
Dm
+
p2 (m) ≤
+ L(SA) εn (m)
2
n
r
2vm zn 2zn
−δ̄ (m) ≥ −
−
.
n
n
1− 1
Moreover, by Lemma 4.6, we have for some constant AM R,− > 0, vm ≤ AM R,− Km p . For ∆
large enough we thus get,
s
1− 1
2AM R,− Km p zn
1 Dm
0
crit (m) ≥ Km + θ −
−
(91)
2
n
n
Assume that β+ ≤
p
r+1 .
p
We take Dm ≤ L (n/ ln(n + 1)) (p+1)β− for some constant L > 0. If L
is small enough, we have by (Ap) Km ≥ (8AM R,− zn /n)p/(p+1) and by (91), crit0 (m) ≥ Km /2.
Now if
β+
p
β− (1+β+ )
(1+p)β−
n
n
Dm ≤ L
∧
(92)
1
ln(n + 1)
2β−
(ln(n + 1))
with L sufficently small, only depending on ∆, θ and constants in (SA), then by (89) we
p
obtain crit0 (m) > crit0 (m0 ). As β+ ≤ r+1
< p, the upper bound in (92) reduces to Dm ≤
β+
1
Ln β− (1+β+ ) / (ln(n + 1)) β− . This proves the left-hand side of (87).
p
Assume now that β+ > r+1
. If
1+β+ (1− r+1
p )
p
r+1
(1+p)β−
n
n β− (2+β+ (1− p ))
Dm ≤ L
∧
1
ln(n + 1)
2β−
(ln(n + 1))
with L sufficiently small, only depending on ∆, θ and constants in (SA), then by (90) we
p
obtain crit0 (m) > crit0 (m1 ). As β+ ≤ r+1
< p, the upper bound in (92) reduces to Dm ≤
β+
1
Ln β− (1+β+ ) / (ln(n + 1)) β− . This proves the left-hand side of (87).
1
1
p
Assume that β+ ≤ r+1
. We take Dm ≤ L (n ln(n + 1)) 2+β+ (1− p ) for some constant L > 0. If
L is large enough, then we get by (91) and simple calculations, crit0 (m) ≥ (θ/2 − 1/4) Dm /n.
Furthermore, if
!
1
1
p
1
D ≥ L (n ln(n + 1)) 2+β+ (1− p ) ∧ n 1+β+ ln(n + 1)
(93)
m
Model Selection as a Multiple Testing Procedure
43
with L sufficiently small, only depending on ∆, θ and constants in (SA), then by (89) we
p
obtain crit0 (m) > crit0 (m0 ). As β+ ≤ r+1
< p, the lower bound in (93) reduces to Dm ≤
1
1
2+β+ 1− p
( ) . This proves the left-hand side of (87).
L (n ln(n + 1))
p
Assume now that β+ > r+1
. If
1
1
r+1 p
2+β
1−
+(
1
p )
ln(n + 1)
Dm ≥ L (n ln(n + 1)) 2+β+ (1− p ) ∨ n
(94)
with L sufficiently large, only depending on ∆, θ and constants in (SA), then by (89) we
p
obtain crit0 (m) > crit0 (m1 ). As β+ ≤ r+1
< p, the upper bound in (92) reduces to Dm ≤
β+
1
Ln β− (1+β+ ) / (ln(n + 1)) β− . This proves the left-hand side of (87). As r > 0, (94) reduces to
1
p
2+β+ (1− r+1
p )
Dm ≥ Ln
ln(n + 1), which proves (86) and the right-hand side of (87).
Proof of Lemma A.4. By definition, m∗ minimizes
K f∗ , fˆm = Km + p1 (m)
over the models m ∈ Mn
. The analysis
is restricted on Ωn .
ˆ
1. Upper bound on K f∗ , fm : we have
Dm
1
+
−β+
ˆ
+ L(SA) εn (m)
K f∗ , fm ≤C+ Dm +
2
n
Dm ln(n + 1)
−β+
+
≤L(SA) Dm
+
.
n
n
1
Hence, if m0 is such that Dm0 = n 1+β+ , then
β
− +
K f∗ , fˆm0 ≤ L(SA) n 1+β+ .
(95)
2. Lower bound on K f∗ , fˆm : there exists a constant A0 , only depending on constants in (SA),
such that
Dm
1
−
ˆ
K f∗ , fm ≥Km +
− A0 εn (m)
2
n
(
)
p
Dm ln(n + 1)
Dm
Dm 3/2 p
≥Km +
− A0 max
ln(n + 1);
.
(96)
2n
n
n
If (Ap) holds, then for
Dm ≤ L(SA) min
n
3
3+2β−
(ln(n + 1))
1
3+2β−
;
n
2
1+2β−
1
(ln(n + 1)) 1+2β−
with L(SA) sufficiently small, we have
−β−
Km /2 ≥ C− Dm
/2 ≥ A0 max
(
Dm
n
3/2
p
ln(n + 1);
)
p
Dm ln(n + 1)
.
n
−β
In this case, we have by (96), K f∗ , fˆm ≥ Km /2 + (Dm ) /2n ≥ C− Dm − /2. Moreover, if m
β+
is such that Dm ≤ L(SA)
n (1+β+ )β−
β+
Dm ≤ L(SA) n (1+β+ )β−
with L(SA) sufficiently small, we also have
3
2
β+
3+2β−
1+2β−
n
n
≤ L(SA) min n (1+β+ )β− ;
;
1
1
(ln(n + 1)) 3+2β− (ln(n + 1)) 1+2β−
44
Saumard and Navarro
and by (95) we get K f∗ , fˆm0 < K f∗ , fˆm , which gives the left-hand side of (88).
We turn now to the proof of the right-hand side of (88). Let m ∈ mn be such that Dm ≥
1
L1 n 1+β+ . By (95) we deduce that if L1 is large enough, depending only on constants in (SA),
then we have
Dm
> K f∗ , fˆm0 .
4n
In addition, if Dm ≥ L2 (ln(n + 1))2 and Dm ≤ L−1
2 n/ ln(n + 1) for some constant L2 sufficiently large, then
)
(
p
Dm ln(n + 1)
Dm
Dm 3/2 p
ln(n + 1);
≥ A0 max
4n
n
n
and by (96), we deduce that K f∗ , fˆm > K f∗ , fˆm0 . The latter inequality implies that
1
1
Dm∗ ≤ L1 n 1+β+ . Our reasoning is valid if n is such that L2 (ln(n + 1))2 ≤ L1 n 1+β+ ≤
1
2
1+β+
,
L−1
2 n/ ln(n+1). At the price of enlarging L1 , we can always achieve L2 (ln(n + 1)) ≤ L1 n
1
1+β+
with L1 not depending on n. Then if L−1
, we still have
2 n/ ln(n + 1) < L1 n
Dm∗ ≤ max Dm ≤ AM,+
m∈Mn
1
n
1+β+
≤
A
L
L
n
.
2
1
M,+
(ln(n + 1))2
In every case, there exists L > 0 only depending on constants in (SA) such that Dm∗ ≤
1
Ln 1+β+ .
Lemma A.5 Let (a, b) ∈ R2+ and (u, v) ∈ [1, ∞] such that 1/u + 1/v = 1. Then
ab ≤ max {au ; bv } ≤ au + bv .
Proof. By symmetry, we can assume au ≥ bv . Then b = (bv )1/v ≤ au/v = au−1 and so, ab ≤
aau−1 = au ≤ au + bv .
| 10math.ST
|
IEEE TRANSACTIONS ON SERVICE COMPUTING , VOL. , NO. , 2016
1
On Dynamic Job Ordering and Slot
Configurations for Minimizing the Makespan
Of Multiple MapReduce Jobs
arXiv:1604.04471v1 [cs.DS] 15 Apr 2016
Wenhong Tian, Guangchun Luo, Ling Tian, and Aiguo Chen
Abstract—MapReduce is a popular parallel computing paradigm for Big Data processing in clusters and data centers.
It is observed that different job execution orders and MapReduce slot configurations for a MapReduce workload have
significantly different performance with regarding to the makespan, total completion time, system utilization and other
performance metrics. There are quite a few algorithms on minimizing makespan of multiple MapReduce jobs. However,
these algorithms are heuristic or suboptimal. The best known algorithm for minimizing the makespan is 3-approximation by
applying Johnson rule. In this paper, we propose an approach called UAAS algorithm to meet the conditions of classical
Johnson model. Then we can still use Johnson model for an optimal solution. We explain how to adapt to Johnson model
and provide a few key features of our proposed method.
Index Terms—MapReduce; Hadoop; Batch Workload; Optimized Schedule; Minimized Makespan.
F
1
I NTRODUCTION
With the rapid increase in size and number of
jobs that are being processed in the MapReduce
framework, efficiently scheduling multiple jobs
under this framework is becoming increasingly
important. Job scheduling in MapReduce
framework brings a new challenge to Cloud
computing [1] such as minimizing the makespan,
load balancing and reduce data skew etc., it
has already received much attention. Originally,
MapReduce was designed for periodically running
large batch workloads with a FIFO (First-In-FirstOut) scheduler. As the number of users sharing
the same MapReduce cluster increased, there
are Capacity scheduler [11] and Hadoop Fair
Scheduler (HFS) [7] which intend to support more
efficient cluster sharing. There are also a few
research prototypes of Hadoop schedulers that
aim to optimize explicitly some given scheduling
metrics, e.g., FLEX [8], ARIA [4]. A MapReduce
simulator called SimMR [5] is also developed to
• Prof. Tian is in the School of Information and Software Engineering,
University of Electronic Science and Technology of China. E-mail:
tian [email protected];This research is sponsored by the National Science Foundation of China with project ID 61450110440;
• G.Luo and L. Tian and A. Chen are with the same University.
simulate different workload and performance of
MapReduce. Yao et al. [15] proposed a scheme
which uses slot ratio between Map and Reduce
tasks as a tunable knob for dynamically allocating
slots. However, as pointed out in [1], the existing
schedulers do not provide a support for minimizing
the makespan for a set of jobs.
Starfish project [9] proposes a workflow-aware
scheduler that correlate data (block) placement
with task scheduling to optimize the workflow
completion time. Zhao et al. [16] propose a
reference service framework for integrating
scientific workflow management systems into
various cloud platforms. Moseley et al. [10]
formulate MapReduce scheduling as a generalized
version of the classical two-stage flexible flow-shop
problem with identical machines; they provide a
12-approximation algorithm for the offline problem
of minimizing the total flow-time, which is the
sum of the time between the arrival and the
completion of each job. Zhu et al. [15] consider
nonpreemptive case to propose 32 -approximation
for offline scheduling regarding the makespan
where they did not considering job ordering or
applying Johnson model. In [1] and [2], the authors
propose heuristics to minimize the makespan, the
proposed algorithm called BalancedPools by
considering two pools for a Hadoop cluster. Tang
et al. [17] proposed a new algorithm called MK JR
for minimizing the makespan. The works of [1]
and [17] are closely related to our research in
minimizing the makespan. However, our present
work meets all the requirements of Johnson model
and provide optimal solution to offline scheduling
while Verma et al. [1] did not modify Johnson’s
model and provided separating pools (called
BalancedPools) for minimizing the makespan,
and BalancedPools is a heuristic approach but not
optimal in many cases. MK JR is a 3-approximation
algorithm for minimizing the makespan. There
is still room for improving the performance of
MapReduce regarding minimize the makespan.
reduce tasks all finish). The makespan for the workload {J1 , J2 , . . . , Jn } is defined as Cmax =maxi∈[n] ci .
We denote |JiM | and |JiR | as the number of tasks
in Ji ’s map stage and reduce stage, respectively.
R
Let tM
i,j and ti,j denote the execution time of Ji ’s
jth map task and jth reduce task, respectively.
Let TiM and TiR denote the execution time of Ji ’s
map and reduce stage respectively. Ji requests
SiM × SiR MapReduce slots and has Map and Reduce stage durations (TiM , TiR ) respectively. The
system scheduler can change a job’s MapReduce
slots allocation depending on available resources.
We aim to determine an order (a schedule) of
execution of jobs Ji ∈ J such that the makespan
of all jobs is minimized. Let us set the actually
R
allocated MapReduce slots for job Ji as |AM
i |×|Ai |,
the max available MapReduce slots in the Hadoop
cluster is |SiM | × |SiR |. The original Johnson Rule
[3] considers that There are n items which must
go through one production stage or machine and
then a second one. There is only one machine for
each stage. At most one item can be on a machine
at a given time. We consider MapReduce as two
non-overlapped stages, i.e., map and reduce stage
respectively, the same as in [1][17]. Also we classify
all jobs into Map type and Reduce type. For Map
type jobs, their map durations should be smaller
than reduce durations while Reduce type jobs have
longer reduce durations than map durations. Based
on these assumptions and Johnson algorithm [1],
we can obtain the optimal makespan of a set of
jobs as follows:
In summary, there is only a small number of
scheduling algorithms with regarding to minimize
the makespan of a set of MapReduce jobs in open
literature and still much room for improving the
performance of MapReduce regarding minimizing
the makespan. Therefore, we propose new modeling and scheduling approaches for offline jobs in
the following sections. The major contributions of
this paper include:
1) provided a new modeling and scheduling approach for multiple MapReduce jobs;
2) proposed an optimal algorithm for offline
scheduling considering Map and Reduce stages by
adapting to classical Johnson’s model;
3) introduced a few key features (theorems) of our
proposed algorithm (UAAS).
2
P ROBLEM F ORMULATION
Cmax =
n
X
TiR + maxnu=1 Ku
(1)
i=1
We consider the following problem as in [1] [17].
Let J= {J1 , J2 , . . . , Jn } be a set of n MapReduce jobs
with no data dependencies between them. These
jobs can be executed in any order. A MapReduce
job Ji consists of two stages, a map stage M and
reduce stage R. Each stage consists of a number of
tasks. The workload is executed on a MapReduce
cluster under FIFO scheduling by default, consisting of a set of (map and reduce) slots. Let S M
and S R denote the set of map slots and reduce
slots configured by MapReduce administrator (i.e.,
S=S M U S R ), so that the number of map slots and
reduce slots are |S M | and |S R |, correspondingly. Let
φ denote the job submission order for a MapReduce
workload. We consider the offline case in which
all the jobs are available at time 0. Let ci denote
the completion time of Ji (i.e., the time when Ji ’s
where
Ku =
u
X
i=1
TiM −
u−1
X
TiR .
(2)
i=1
Observation 1. If each job utilizes either all map
or all reduce slots during its processing, there
is a perfect match between the assumptions
of the classic Johnson algorithm for twostage production system and MapReduce job
processing, then Johnson’s algorithm can be
applied to find optimal solution for minimizing
the makespan of a set of MapReduce jobs.
Based on our observations and intensive real test
experiences, we propose a new method called
UAAS (Utilizing All Available Slots) algorithm,
with the pesudocode given in Algorithm 2.1. The
2
following theorem is the key strategy for our
results.
Theorem 1. Based on available MapReduce slots
in the system, the scheduler can increase or
decrease the number of MapReduce slots to
the job to meet the requirements of JohnSon
Rule, the result obtained by UAAS algorithm
following Johnson rule is optimal regarding to
minimize the makespan.
input : the total number of MapReduce slots
(|S M |, |S R |) for a Hadoop cluster,
estimated all Jobs’ Map and Reduce
durations (TiM , TiR ) [1] by utilizing all
available Map and Reduce slots for
each job in the cluster
output: the scheduled jobs, the makespan Cmax
1 List the Map and Reduce’s durations in two
vertical columns (implemented in a list) ;
2 for all Ji ∈ J do
3
Find the shortest one among all durations
(min (TiM , TiR ));
4
In case of ties, for the sake of simplicity,
order the item with the smallest subscript
first. In case of a tie between Map and
Reduce, order the item according to the
Map ;
5
IF it is the first job of Map type, place the
corresponding item at the first place ;
6
ELSE it is the first job of Reduce type, place
the corresponding item at the last place ;
7
IF it is Map type job (and not the first job),
place the corresponding item right next to
the previous job (i.e., in non-decreasing
order of Map durations) ;
8
ELSE it is Reduce type job (and not the
first job), place the corresponding item left
next to the previous job (i.e., in
non-increasing order of Reduce durations) ;
9
Remove both durations for that job ;
10
Repeat these steps on the remaining set of
jobs
11 end
12 Compute the makespan (Cmax )
Algorithm 2.1: Utilizing All Available Slots
(UAAS) algorithm
Proof: The original Johnson Rule [3] considers
that ”there are n items which must go through one
production stage or machine and then a second one.
There is only one machine for each stage. At most
one item can be on a machine at a given time”. To
adapt the MapReduce model, we treat the Map and
Reduce stage resources as a whole (like a single machine), i.e., to represent the resources as MapReduce
slots in the whole in our algorithm UAAS. USSA
algorithm allocates all available MapReduce slots
to each job at each stage, so that UAAS meets all
requirements of Johnson Rule. Since Johnson Rule
obtains optimal results with regarding to minimize
the makespan (the proof is provided in [3]), and our
UAAS algorithm meets all requirements of Johnson
Rule, therefore UAAS obtains the optimal result
with regard to minimizing the makespan.
map
J2
1
4
J5
2
J3
3
4
J1
J4
reduce
30X30
5
6
20X20
10X10
30
30
20X20
5
10
15
4
Time
(in
units)
20
25
30
35
40
45
BalancedPools Algorithm [1]: is way to minimize
the makespan for offline scheduling proposed
in [1], it partitions the Hadoop cluster into two
balanced pools and then allocated each job to a
suitable pool to minimize the makespan.
Observation 2. BalancedPool Algorithm does not
meet the requirement of Johnson model but just
order the job by Johnson rule and is a heuristic
algorithm with computational complexity of
O(n2 lognlogP ) where n is the number of jobs
and P is the number of MapReduce slots.
Fig. 1. Five MapReduce Jobs Execution in One
Cluster by MK JR
3
T HREE A LGORITHMS C OMPARED
In this section, we compare UAAS algorithm with
two best known algorithms (BalancedPools and
MK JR) regarding to minimize the makespan of a
set of offline MapReduce Jobs.
MK JR algorithm [17]: Divide the jobs set J into
3
5
10
15
m5
m2
m1
m4
20
25
30
MapReduce slots. There are five jobs, among them,
J1 ,J2 and J5 require 30×30 MapReduce slots
while J3 and J4 require 20×20 MapReduce slots.
The total makespan by MK JR algorithm is 47
units, visualized in Fig.1. However, if we allow
that any job can use all available MapReduce
slots in the system when execution (this can be
implemented easily in Hadoop, for example by
splitting the input files based on available number
MapReduce slots), the result is very different from
both MK JR and BalancedPools algorithms. For
the same example, in UAAS algorithm job J3 and
J4 can use all available 30 × 30 MapReduce slots,
then J3 will have Map and Reduce durations (20,
8
3 ), J4 will have Map and Reduce durations (4, 20)
respectively. Therefore the total makespan will be
35 23 as shown in Fig. 2, where X1 =1. This result is
smaller (about 31.76%) than the result (47 units)
obtained by MK JR in [1]. The makespan of Pool1
and Pool2 is 39 and 40 time units respectively by
applying BalancedPools algorithm, where Pool1
has configuration of 10×10 MapReduce slots and
Pool2 has configuration of 20×20 MapReduce slots,
and J1 ,J2 and J5 (short jobs) are with Pool1 while
J2 and J3 (longer jobs) are with Pool2. Therefore,
the UAAS result is about 12.14% smaller than the
result (40 time units) obtained by BalancedPools
algorithm.
35
map
m3
reduce
Map
1
2
X1
4
Reduce
20
4
4
3
5
20
8/3
r2
r5
r1
r4
r3
5me
(in
units)
Fig. 2. New Result of Five MapReduce Jobs Execution by UAAS
TABLE 1
The example of 5 jobs
Job ID
J1
J2
J3
J4
J5
SrM
30
30
20
20
30
SrR
30
30
20
20
30
TiM
4
1
30
6
2
TiR
5
4
4
30
3
two disjoint sub-sets JA and JB . Set JA consists
of those jobs Ji for which TiM < TiR . Set JB
contains the remaining jobs. Sequence jobs in JA
in non-decreasing order of TiM and those in JB
in non-increasing order of TiR . The job order is
obtained by appending the sorted set JB to the
end of sorted set JA .
Observation 3. MK JR algorithm does not meet
the requirement of Johnson model but just order
the job by Johnson rule after estimating the map
and reduce durations of each job.
Theorem 2. MK JR is an 3-approximation algorithm for the makespan optimization in general
case.
Proof: Applying the intermediate results from [17]
(Equ. (8) in supplementary material for proof of
THOREM 1 in [17]), we have
The reason that BalancedPools and MK JR
algorithms do not meet the requirement of
Johnson model lies that they do not utilize
all available MapReduce slots for each job
in general case, though they estimate the job
ordering by Johnson rule. Therefore, unlike UAAS
algorithm, BalancedPools and MK JR algorithms
are suboptimal.
ˆ + maxnk=1
Cmax ≤ (Cmax
k
X
n
ˆ
tM
i + maxk=1
i=1
k
X
tˆR
i )
i=1
(3)
ˆ
R
where
and ti is the estimated map and reduce duration for job Ji , respectively. Let us define
ˆ
tM
i
σ=
maxn
k=1
n
ˆ
M
i=1 ti +maxk=1
Pk
ˆ
Cmax
ˆR
i=1 ti
Pk
, the same as in [17],
ˆ is the theoretical optimal makespan
where Cmax
given by Equ. (1)-(2). Considering the worst case
that there are two jobs J1 and J2 , T1M =1, T2M =C0 ,
and T1R =C0 , T2R =1; In this case, the optimal order
Pk ˆ
Pk ˆR
n
is J1 -J2 and maxnk=1 i=1 tM
i =maxk=1
i=1 ti =C0 .
P
k
ˆ =C +2 by
M
ˆ =C opt , C opt = maxn
And Cmax
t
0
max
max
k=1
i=1 i
0 +1
UAAS algorithm, we have σ= 2C
C0 +2 ≤ 2. Therefore
Table 1 shows an example from [2], where SrM
and SrR is the requested number of slots for map
and reduce stage respectively for job Ji . Example
1. Consider a scenario shown in Table 1 from [1],
where the cluster has a configuration of 30×30
4
the approximation ratio of MK JR is
0 +1
C0 + 2 + 2C
Cmax (M K JR)
C0 + 2 + σ
C0 +2
=
=
≈3
opt
C0 + 2
C0 + 2
Cmax
(4)
It worths notice that the worst case is applied
for approximation ratio. (1+σ)-appromixation
algorithm where σ ∈ [0, 2], should be called 3approximation algorithm since σ is 3 in the worst
case.
Based on previous results, we have the following
observation.
Observation 4. BalancedPools and MK JR
algorithms
are
suboptimal
regarding to
minimizing the makespan, they may not have the
minimum makespan for a set of jobs; applying
Theorem 1 to single Hadoop cluster always has
optimal total makespan for a set of jobs.
Theorem 3. Given a homogeneous environment
where the Hadoop configurations of slave nodes
are identical, the job order φ1 produced by
UAAS for a batch of jobs are independent of
the number of slave nodes (N ) but depends on
the total number of available MapReduce slots
(|S M |, |S R |), and is stable with regarding to the
change of the total number of slave nodes.
Proof: Let us set the execution durations of map
and reduce stages for a given job Ji under a given
configuration of Hadoop cluster with |S M | × |S R |
MapReduce slots, as TiM and TiR , respectively. If the
MapReduce slots configuration of Hadoop cluster
|S M |
is changed to |SxM |×|SxR | and set ρ0 = |S
M . Applying
x |
UAAS algorithm, the execution duration of map
and reduce stage for a given job Ji will change to
0
0
TiM and TiR . And we have
0
TiM = TiM
0
TiR = TiR
|S M |
= TiM ρ0
|SxM |
(5)
|S R |
= TiR ρ0
|SxR |
(6)
This means execution duration of map and
reduce stage for each job will change proportional
to ρ0 but their relative relationship (ordering by
their durations) will not change. Therefore the job
order of UAAS is stable with regarding to the
change of the total number of slave nodes.
Observation 5. The the job ordering of MK JR
and BalancedPools is not stable when the total
number of slave nodes changes.
Let us consider the example given in [17]. There
is a Hadoop cluster with 5 nodes, each configured
with 2 map and 2 reduce slots. Let J1 be defined
as follows: Map stage duration is 9 and requires
10 map slots. Reduce stage duration is 10 and
requires 1 reduce slot. Let J2 be defined as follows:
Map stage duration is 11 and requires 8 map slots
and reduce stage duration is 15 and requires 1
reduce slot. In this case, the optimal job scheduling
order by UAAS is J2 -J1 , and their corresponding
map and reduce duration is (8.8,1.5) and (9,1)
respectively by utilizing all 10 MapReduce slots
in each stage, with the makespan of 18.8. The
job order produced by MK JR is J1 -J2 with the
makespan of 35, which is about 86.17% larger
than optimal result. Now, if one node fails, then
there are only 4 nodes left with 8 map and 8
reduce tasks available in the cluster. In this case,
the optimal job scheduling by UAAS is still J2 -J1 ,
however, their corresponding map and reduce
duration is (11.25,1.25) and (11, 1.875) respectively
by utilizing all 8 MapReduce slots in each stage,
with makespan of 23.5. The job order generated by
MK JR keeps the same, i.e., J1 -J2 , with makespan
of 43, about 82.97% larger than the optimal.
Notice that BalancedPools algorithm has following
results. When there are 5 nodes, J1 with duration
(9,10) will be put into Pool1 with 2 nodes of 4
MapReduce slots and J2 will be allocated to Pool2
with 3 nodes of 6 MapReduce slots. Then J1 will
have duration (22.5, 10) and J2 will have duration
(14.67,15). If one node fail, J1 still with Pool1
and J2 with Pool2; J1 and J2 will have duration
(32.5, 10) and (37, 10) respectively. In either case,
BalancedPools is far from optimal results.
Theorem 4. Let ρ be the ratio of map slots to
M
|
reduce slots, i.e., ρ= |S
. The optimal configu|S R |
ration of ρ for makespan Cmax depends on the
total number of slots (|S M |, |S R |), MapReduce
workload as well as its job submission order φ.
Proof:
5
[4]
Cmax
=
n
max
k=1
k
X
n
TiM + max
i=1
i=k
n
X
TiR f rom[17]
M
=
[5]
i=k
R
k |Ji |
n |Ji |
[6]
1 XX
1 XX
tM
+
tR
max( M
i,j
i,j )
R
k=1 |S |
|S |
i=1 j=1
j=1
n
i=k
=
1
1
n
max(
|S R | k=1 ρ
M
k |J
i |
X
X
i=1 j=1
tM
i,j
+
R
n |J
i |
X
X
[7]
tR
i,j )
(7)
i=k j=1
[8]
This means the optimal configuration of ρ for
makespan Cmax depends on the total number of
R
slots (|S M |,|S R |) MapReduce workload (tM
i,j , ti,j )
as well as its job submission order φ(={1,..n}).
When the workload and job order are fixed, it
is obvious that larger number of total number
of MapReduce slots will lead to smaller value of
Cmax . This is consistent with Theorem 1 and UAAS
algorithm to utilize all available MapReduce slots
(|S M |, |S R |).
4
[9]
[10]
[11]
[12]
C ONCLUSION
Observing that there are quite a few algorithms on
minimizing makespan of multiple MapReduce jobs
and these algorithms are heuristic or suboptimal.
In this paper, we proposed an optimal approach
called UAAS algorithm to minimize the makespan
of a set of MapReduce jobs. The proposed algorithm meets the requirements of classical Johnson
algorithm and therefore is optimal with regarding
to the makespan. We also conducted extensive tests
in real Hadoop environment to validate our theoretical results by benchmarks provided in [13][14].
Because this is a short paper, we do not provide the
test results yet. There are future research directions
such as considering minimizing the makespan of
online MapReduce jobs and minimizing the total
completion time and total flow time of a set of
Mapreduce jobs.
[13]
[14]
[15]
[16]
[17]
R EFERENCES
[1]
[2]
[3]
A. Verma, L. Cherkasova, R. H. Campbell, Orchestrating
an Ensemble of MapReduce Jobs for Minimizing Their
Makespan, IEEE Transactions on Dependable and Secure
Computing, April 2013 (online version).
A. Verma, L. Cherkasova, and R. H. Campbell, Two Sides
of a Coin: Optimizing the Schedule of MapReduce Jobs
to Minimize Their Makespan and Improve Cluster Performance. MASCOTS, page 11-18. IEEE Computer Society,
(2012)
S. Johnson, Optimal Two-and Three-Stage Production
Schedules with Setup Times Included,Naval Res. Log.
Quart., 1954.
6
A. Verma, L. Cherkasova, and R. H. Campbell, ARIA: Automatic Resource Inference and Allocation for MapReduce
Environments,in Proc. of ICAC,pp.235-244, 2011,Germany.
A. Verma, L. Cherkasova, and R. H. Campbell, Play It
Again, SimMR! in Proc. of Intl. IEEE Cluster’2011, pp. 253261, IEEE Computer Society Washington, DC, USA, 2011.
M. Garey and D. Johnson, Computers and Intractability: A
Guide to the Theory of NP-completeness. WH Freeman &
Co., 1979.
M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy, S.
Shenker, and I. Stoica, “Delay scheduling: A simple technique for achieving locality and fairness in cluster scheduling,in Proc. of EuroSys. ACM, 2010, pp. 265-278.
J. Wolf and et al, FLEX: A Slot Allocation Scheduling Optimizer for MapReduce Workloads, ACM/IFIP/USENIX
Intl. Middleware Conference, Lecture Notes in Computer
Science Volume 6452, pp.1-20, 2010.
H. Herodotou and S. Babu, Profiling, What-if Analysis, and
Costbased Optimization of MapReduce Programs. in Proc.
of the VLDB Endowment, pp.1111-1122, Vol. 4, No. 11,
2011.
B. Moseley, A. Dasgupta, R. Kumar, and T.Sarl’s, On
scheduling in map-reduce and flow-shops, in Proc. of
SPAA, pp.289-298, ACM New York, NY, USA, 2011.
Capacity
Scheduler
Guide.
[Online].
Available:
http://hadoop.apache.org/common/docs/r0.20.1/
capacity scheduler.html
Y. Zheng, N. B. Shroff, P. Sinha, A New Analytical Technique for Designing Provably Efficient MapReduce Schedulers, In the Proceedings of INFOCOM 2013, pp.1600-1608,
14-19 April 2013,Turin.
Wordcount, http://www.cs.cornell.edu/home/llee/data/simple/
http://sortbenchmark.org/YahooHadoop.pdf.
Y. Zhu, Y. Jiang, W. Wu, L. Ding, A. Teredesai, D. Li,
W. Lee, Minimizing makespan and total completion time
in MapReduce-like systems, In Proceedings of INFOCOM
2014, pp.2166-2174, April 27 2014-May 2 2014, Toronto, ON.
Yong Zhao ;Youfu Li ; I. Raicu ; Shiyong Lu ; Cui Lin
; Yanzhe Zhang ; Wenhong Tian ; Ruini Xue, A Service Framework for Scientific Workflow Management in
the Cloud, IEEE Trans. Services Computing 8(6): 930-944
(2015).
Shanjiang Tang, Bu-Sung Lee, Bingsheng He, Dynamic Job
Ordering and Slot Configurations for mapreduce Workloads, IEEE Trans. Services Computing 9(1): 4-17 (2016).
| 8cs.DS
|
arXiv:1711.11117v1 [cs.CV] 29 Nov 2017
Towards Alzheimer’s Disease Classification through
Transfer Learning
1st Marcia Hon
2nd Naimul Mefraz Khan
Data Science
Ryerson University
Toronto, ON
[email protected]
Department of Electrical & Computer Engineering
Ryerson University
Toronto, ON
[email protected]
Abstract—Detection of Alzheimer’s Disease (AD) from neuroimaging data such as MRI through machine learning have been
a subject of intense research in recent years. Recent success of
deep learning in computer vision have progressed such research
further. However, common limitations with such algorithms are
reliance on a large number of training images, and requirement
of careful optimization of the architecture of deep networks. In
this paper, we attempt solving these issues with transfer learning,
where state-of-the-art architectures such as VGG and Inception
are initialized with pre-trained weights from large benchmark
datasets consisting of natural images, and the fully-connected
layer is re-trained with only a small number of MRI images. We
employ image entropy to select the most informative slices for
training. Through experimentation on the OASIS MRI dataset,
we show that with training size almost 10 times smaller than the
state-of-the-art, we reach comparable or even better performance
than current deep-learning based methods.
Index Terms—Alzheimer’s, Deep Learning, Transfer Learning,
Entropy
I. I NTRODUCTION
Alzheimers Disease (AD) is a neurodegenerative disease
causing dementia in elderly population. It is predicted that
one out of every 85 people will be affected by AD by
2050 [1]. Early diagnosis of AD can be achieved through
automated analysis of MRI images with machine learning. It
has been shown recently that in some cases, machine learning
algorithms can predict AD better than clinicians [2], making
it an important field of research for computer-aided diagnosis.
While statistical machine learning methods such as Support
Vector Machine (SVM) [3] have shown early success in
automated detection of AD, recently deep learning methods
such as Convolutional Neural Networks (CNN) and sparse
autoencoders have outperformed statistical methods.
However, the existing deep learning methods train deep
architectures from scratch, which has a few limitations [4], [5]:
1) properly training a deep learning network requires a huge
amount of annotated training data, which can be a problem
especially for the medical imaging field where physicianannotated data can be expensive, and protected from crossinstitutional use due to ethical reasons; 2) training a deep
network with large number of images require huge amount of
computational resources; and 3) deep network training requires
careful and tedious tuning of many parameters, sub-optimal
tuning of which can result in overfitting/underfitting, and, in
turns, result in poor performance.
An attractive alternative to training from scratch is finetuning a deep network (especially CNN) through transfer
learning [6]. In popular computer vision domains such as
object recognition, trained CNNs are carefully built using
large-scale datasets such as ImageNet [7]. Through Transfer
Learning , these trained networks can be used with smaller
datasets with only fine-tuning the final fully-connected layers
of CNN. Transfer learning has been proven to be robust even
for cross-domain application, such as networks trained on
natural images used with medical images [8].
In this paper, we investigate the adaptation of two popular
CNN architectures into an AD diagnosis problem through
transfer learning, namely VGG16 and Inception. These two
architectures are previous winners of the ImageNet Large
Scale Visual Recognition Challenge (ILSVRC), a yearly competition that evaluates object detection in images [7]. With
the two winners being open source, the architectures as well
as the pre-trained weights are readily available. Although the
architectures are trained on a different domain (natural images
of ImageNet as opposed to MRI images in our paper), we
show that the pre-trained weights generalize very well for AD
diagnosis when the training data is picked intelligently.
Since our target was to test the robustness of transfer learning on a small training set, merely choosing training data at
random may not provide us with a dataset representing enough
structural variations in MRI. Instead, we picked the training
data that would provide the most amount of information
through image entropy. We show that through intelligent training selection and transfer learning, we can achieve comparable,
or even better performance than training a deep network from
scratch with minimal parameter optimization.
II. R ELATED W ORKS
Classical machine learning methods such as SVM and feedforward neural networks have been applied successfully to
diagnose AD from structural MRI images [3], [9]. One such
recent method is [9], where a dual-tree complex wavelet
transform is used to extract features, and a feed-forward neural
network is used to classify images. Elaborate discussion and
comparative results with other popular classical methods can
also be found in [9].
Recently, deep learning methods have outpeformed classical
methods by a large margin. A combination of patches extracted from an autoencoder followed by convolutional layers
for feature extraction were used in [10]. The method was
further improved by using 3D convolution in [11]. Stacked
autoencoders followed by a softmax layer for classification
was used in [12]. Popular CNN architectures such as LeNet
and the first Inception model were used in [13]. While these
deep learning methods provide good accuracy results, as we
show in the experiments, due to our intelligent training data
selection and use of transfer learning, we get comparable, or
even better accuracy while using only a fraction of the training
data.
Transfer learning or fine-tuning has also been explored in
medical imaging. [14] provides an in-depth discussion and
comparative results of training from scratch vs fine-tuning on
some medical applications. They show that in most cases, finetuning outperforms training from scratch. Fine-tuned CNNs
have been used to localize planes in ultrasound images [15],
classify interstitial lung diseases [8], and retrieve missing or
noisy plane views in cardiac imaging [16]. All these methods
prove that employing transfer learning in the medical imaging
domain has tremendous value, and has the potential to achieve
high accuracy in AD detection with smaller training dataset
when compared to training from scratch.
III. M ETHODOLOGY
A. Convolutional Neural Networks and Transfer Learning
The core of Convolutional Neural Networks (CNN) are
convolutional layers which can extract local features (e.g.
edges) across an input image through convolution. Each node
in a convolutional layer is connected to a small subset of
spatially connected neurons. To reduce computational complexity, a max pooling layer follows convolutional layers,
which reduces the size of feature maps by selecting the
maximum feature response in a local neighborhood. Pairs of
convolutional and pooling layers are followed by a number
of fully-connected layers, where a neuron in one layer has
connections to all activations in the previous layer. Fullyconnected layers help learning non-linear relationship among
the local features extracted by convolutional layers. Finally,
a soft-max layer follows the fully-connected layers, which
normalizes the outputs to desired levels.
CNNs are trained with the back-propagation algorithm,
where in each iteration, weights associated with the neurons in
the convolutional layers are updated in a way that minimizes
a cost function. When training from scratch, the weights
are typically initialized randomly, drawing from a normal
distribution.
However, with a small training dataset, this may result
in the cost function getting stuck in a local minima, which
may result in overfitting/underfitting. A better alternative in
such case is transfer learning, where pre-trained weights from
the same architecture, but on a different larger dataset from
the same/different domain is used to initialize the network
layers. Only the last fully-connected layer is re-trained with
the training data. This not only provides us with a robust set of
pre-trained weights to work with, it gives us the opportunity
to employ proven network architectures in our problem. The
two popular architectures that we will use in this paper are:
1) VGG16:
VGG16 is a 16-layer network built by Oxfords Visual
Geometry Group (VGG) [17]. It participated in the
ImageNet competition in ILSVRC-2014. One of the
main reasons that VGG16 won the competition, is that it
is one of the first architectures to explore network depth
by pushing to 16-19 layers and using very small (3x3)
convolution filters.
2) Inception:
The Inception architecture is a variant of deep learning
architecture built by Google. Throughout the years, there
have been different iterations. Inception V4, discussed
in this article, was introduced in [18]. The Inceptions’
breakthrough is in the realization that non-linear functions can be learned by changing how convolutional
layers are connected. Accordingly, the fully-connected
layer is discarded in preference of a global average
pooling that averages the feature maps and then connect
with a softmax layer for classification. Thus, there are
less parameters, and as a result, less overfitting.
B. Most informative training data selection
While transfer learning provides an opportunity to use
smaller set of training data, choosing the best possible data for
training is still critical to the success of the overall method.
Typically, from a 3D MRI scan, we have a large number of
images that we can choose from. In most recent methods, the
images to be used for training are extracted at random. Instead,
in our porposed method, we extract the most informative
slices to train the network. For this, we calculate the image
entropy of each slice. In general, for a set of M symbols with
probabilities p1 , p2 , . . . , pM the entropy can be calculated as
follows [19]:
H=−
M
∑
pi log pi .
(1)
i=1
For an image (a single slice), the entropy can be similarly
calculated from the histogram [19]. The entropy provides a
measure of variation in a slice. Hence, if we sort the slices
in terms of entropy in descending order, the slices with
the highest entropy values can be considered as the most
informative images, and using these images for training will
provide robustness.
IV. E XPERIMENTAL R ESULTS
In this section, we provide experimental results on the two
aforementioned models. We implemented our deep learning
methods using Keras with a TensorFlow backend, while the
training data selection method described in Section III-B was
implemented in MATLAB. Architecture models for TensorFlow and pre-trained weights were downloaded from open
source repositories of the models 1 . Our target was to differentiate AD patients from Healthy Control (HC) through
analyzing MRI scan data through the CNN architectures.
A. Dataset
(a)
(b)
Fig. 1. Images from the OASIS dataset (a) AD. (b) HC.
We used structural MRI data from the Open Access Series of Imaging Studies (OASIS), accessible at
(http://www.oasisbrains.org) [20]. Samples of MR brain images from the AD and the HC group are shown in Figure
1.
OASIS provides two types of data: cross-sectional and longitudinal. Since our target was to simply differentiate between
AD and HC patients through the images, we used crosssectional data. The dataset consists of 416 subjects whose
ages are between 18 and 96. In our experiments, we randomly
picked 200 subjects, 100 of whom were picked from the
AD group, while the other 100 from the HC group. The
grouping was determined by the Clinical Dementia Rating
(CRD) variable ranging from 0 to 2. The assumption is that a
0 implies HC and those greater than 0 are AD.
To test the generalization power of transfer learning, we
only used a limited number of images from these scans to
train our networks, especially when compared to other CNNbased Alzheimer’s detection algorithms (further comparison in
Section IV-C). We used our entropy-based sorting mechanism
to pick the most informative 32 images from the axial plane of
each 3D scan. That resulted in a total of 6400 training images,
3200 of which were AD and the other 3200 were HC.
To be compatible with the pre-trained models of VGG16
and Inception V4, the images were resized to be 150X150 for
VGG16, and 299X299 for Inception V4.
B. Accuracy results
5-fold cross-validation was used to obtain the results, with
an 80% - 20% split between training and testing. To test the
1 https://github.com/flyyufelix/cnn
finetune
power of transfer learning, we also trained a VGG16 network
from scratch for comparison. For the transfer learning models,
pre-trained network weights from ImageNet were obtained.
For VGG16, 100 epochs were used with a batch size of 40.
For Inception V4, 100 epochs with a batch size of 8 was used.
For the VGG16 model (both from scratch and with transfer
learning), the RMSProp optimization model was used, which
uses an adaptive learning rate.
For the Inception V4 model, stochastic gradient descent
optimization with a learning rate of 0.0001 was used 2 .
These hyperparameter values were the recommended settings from the respective methods. There is a possibility that
a hyperparameter search method such as grid search will result
in further improvement, but as we see from the reported results
below, the recommended settings already provide excellent
results. Since grid search is computationally very expensive,
we plan to explore this in future with cloud computing or
additional hardware resources.
TABLE I
T ESTED MODELS AND CORRESPONDING AVERAGE ACCURACY FROM
5- FOLD CROSS - VALIDATION . S TANDARD DEVIATION OVER THE 5 FOLDS
IN BRACKETS .
Model
VGG16 (from scratch)
VGG16 (transfer learning)
Inception V4 (transfer learning)
Avg. Acc. (st. dev.) (%)
74.12 (1.55)
92.3 (2.42)
96.25 (1.2)
Table I shows the accuracy results (with standard deviation) for the three models. As can be seen, VGG16, when
trained from scratch, results in poor performance. This can
be attributed towards the small training size that we used,
which can result in overfitting/underfitting. However, using
pre-trained model for transfer learning, the accuracy significantly increases. Finally, using Inception V4 with transfer
learning results in high accuracy. This signifies the value of
transfer learning, where using even a relatively smaller training
dataset can result in an excellent automated system for AD
detection.
C. Comparison with other methods
In this section, we discuss the results in relation to other
recent methods, both in terms of accuracy and training size.
It is important to note that while the other methods may have
used different datasets and/or different experimental setups, the
results are still comparable, since all the methods discussed
below report the results with structural MRI images, which
have high level of similarity across datasets, especially when
the images have been preprocessed and the brain have been
registered and segmented in the published datasets already.
We compare our results with five recent methods, four of
which are deep learning-based. For all these methods, the
training size was calculated based on the reported sample
size and the cross-validation/training-testing split (e.g. in our
2 Models, weights, dataset, and code available at https://github.com/
marciahon29/Ryerson MRP
method, there are total 6,400 images; a 5-fold cross-validation/
80% - 20% split therefore results in a training size of 5,120).
TABLE II
C OMPARISON WITH THE STATE - OF - THE - ART IN
AND TRAINING SIZE .
Model
Wavelet + NN [9]
DeepAD (Inception) [13]
3DConv [11]
Sparse autoencoder + conv [10]
Stacked autoencoders [12]
Inception V4 (transfer learning)
TERMS OF ACCURACY
Average
Acc.(%)
90.06
98.84
95.39
94.74
87.76
96.25
Training Size
(# of Images)
3,629
46,751
117,708
103,683
21,726
5,120
As can be seen from Table II, our method outperforms every
method except [13] in terms of accuracy. The most significant
effect of transfer learning can be seen when comparing the
training sizes. As we can see, only [9] used a smaller training
set than ours, which is not a deep learning method, and therefore, does not require a large training set. While comparing
with the deep learning methods, we can see that our training
dataset is considerably smaller. Especially if we compare the
training sizes between [13] and ours, both of which use the
Inception architecture, we can see that our accuracy results
are very close to theirs despite using a training size almost 10
times smaller (5,120 vs 46,751). This is due to the fact that we
not only use transfer learning, but also employ an intelligent
entropy-based method to pick the most informative training
images. The combination of these two techniques results in
very good performance with a comparatively smaller training
dataset. A smaller training dataset is significant because it frees
the method from dependence on large tediously annotated
training data, and also, improves the training computational
time significantly.
V. C ONCLUSION
In this paper, we propose a transfer learning-based method
to detect AD from structural MRI images. We test two popular
architectures, namely VGG16 and Inception V4. Through
pre-trained weights from ImageNet and fine-tuning, we can
only use a small number of training images to obtain highly
accurate results. Moreover, we employ an intelligent entropybased technique to select the training dataset in a way that
represents the most amount of information within a small set.
We test our models on images from the OASIS brain imaging
dataset, where 6,400 images extracted from MRI scans of
200 subjects are used to train the models. As we show, our
method provides performance comparable to the state-of-theart despite having a training set many times smaller.
In future, we plan to test the models with other datasets,
and employ grid search to optimize the hyperparameters of the
proposed models further to achieve better results. We also plan
to test the robustness of our models for three-way classification
with AD, HC, and Mild Cognitive Impairment (MCI).
Keeping up with the spirit of reproducible research, all
our models, dataset, and code can be accessed through the
repository at: https://github.com/marciahon29/Ryerson MRP .
R EFERENCES
[1] R. Brookmeyer, E. Johnson, K. Ziegler-Graham, and H. M. Arrighi,
“Forecasting the global burden of alzheimers disease,” Alzheimer’s &
dementia, vol. 3, no. 3, pp. 186–191, 2007.
[2] S. Klöppel, C. M. Stonnington, J. Barnes, F. Chen, C. Chu, C. D. Good,
I. Mader, L. A. Mitchell, A. C. Patel, C. C. Roberts et al., “Accuracy
of dementia diagnosisa direct comparison between radiologists and a
computerized method,” Brain, vol. 131, no. 11, pp. 2969–2974, 2008.
[3] C. Plant, S. J. Teipel, A. Oswald, C. Böhm, T. Meindl, J. MouraoMiranda, A. W. Bokde, H. Hampel, and M. Ewers, “Automated detection
of brain atrophy patterns based on mri for the prediction of alzheimer’s
disease,” Neuroimage, vol. 50, no. 1, pp. 162–174, 2010.
[4] N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B.
Gotway, and J. Liang, “Convolutional neural networks for medical image
analysis: Full training or fine tuning?” IEEE transactions on medical
imaging, vol. 35, no. 5, pp. 1299–1312, 2016.
[5] D. Erhan, P.-A. Manzagol, Y. Bengio, S. Bengio, and P. Vincent, “The
difficulty of training deep architectures and the effect of unsupervised
pre-training,” in Artificial Intelligence and Statistics, 2009, pp. 153–160.
[6] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are
features in deep neural networks?” in Advances in neural information
processing systems, 2014, pp. 3320–3328.
[7] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large
scale visual recognition challenge,” International Journal of Computer
Vision, vol. 115, no. 3, pp. 211–252, 2015.
[8] M. Gao, U. Bagci, L. Lu, A. Wu, M. Buty, H.-C. Shin, H. Roth, G. Z.
Papadakis, A. Depeursinge, R. M. Summers et al., “Holistic classification of ct attenuation patterns for interstitial lung diseases via deep
convolutional neural networks,” Computer Methods in Biomechanics and
Biomedical Engineering: Imaging & Visualization, pp. 1–6, 2016.
[9] D. Jha, J.-I. Kim, and G.-R. Kwon, “Diagnosis of alzheimers disease
using dual-tree complex wavelet transform, pca, and feed-forward neural
network,” Journal of Healthcare Engineering, vol. 2017, 2017.
[10] A. Gupta, M. Ayhan, and A. Maida, “Natural image bases to represent
neuroimaging data,” in International Conference on Machine Learning,
2013, pp. 987–994.
[11] A. Payan and G. Montana, “Predicting alzheimer’s disease: a neuroimaging study with 3d convolutional neural networks,” arXiv preprint
arXiv:1502.02506, 2015.
[12] S. Liu, S. Liu, W. Cai, S. Pujol, R. Kikinis, and D. Feng, “Early
diagnosis of alzheimer’s disease with deep learning,” in Biomedical
Imaging (ISBI), 2014 IEEE 11th International Symposium on. IEEE,
2014, pp. 1015–1018.
[13] S. Sarraf, G. Tofighi et al., “Deepad: Alzheimer s disease classification
via deep convolutional neural networks using mri and fmri,” bioRxiv, p.
070441, 2016.
[14] N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B.
Gotway, and J. Liang, “Convolutional neural networks for medical image
analysis: Full training or fine tuning?” IEEE transactions on medical
imaging, vol. 35, no. 5, pp. 1299–1312, 2016.
[15] H. Chen, D. Ni, J. Qin, S. Li, X. Yang, T. Wang, and P. A. Heng, “Standard plane localization in fetal ultrasound via domain transferred deep
neural networks,” IEEE journal of biomedical and health informatics,
vol. 19, no. 5, pp. 1627–1636, 2015.
[16] J. Margeta, A. Criminisi, R. Cabrera Lozoya, D. C. Lee, and N. Ayache, “Fine-tuned convolutional neural nets for cardiac mri acquisition
plane recognition,” Computer Methods in Biomechanics and Biomedical
Engineering: Imaging & Visualization, vol. 5, no. 5, pp. 339–349, 2017.
[17] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[18] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi, “Inception-v4,
inception-resnet and the impact of residual connections on learning.” in
AAAI, 2017, pp. 4278–4284.
[19] C. Studholme, D. L. Hill, and D. J. Hawkes, “An overlap invariant
entropy measure of 3d medical image alignment,” Pattern recognition,
vol. 32, no. 1, pp. 71–86, 1999.
[20] D. S. Marcus, A. F. Fotenos, J. G. Csernansky, J. C. Morris, and
R. L. Buckner, “Open access series of imaging studies: longitudinal mri
data in nondemented and demented older adults,” Journal of cognitive
neuroscience, vol. 22, no. 12, pp. 2677–2684, 2010.
| 1cs.CV
|
Comparison of Galled Trees
Gabriel Cardona1 , Mercè Llabrés1 , Francesc Rosselló1 , and Gabriel Valiente2
arXiv:0906.1166v1 [q-bio.PE] 5 Jun 2009
1
Department of Mathematics and Computer Science, University of the Balearic Islands,
E-07122 Palma de Mallorca, {gabriel.cardona,merce.llabres,cesc.rossello}@uib.es
2
Algorithms, Bioinformatics, Complexity and Formal Methods Research Group, Technical
University of Catalonia, E-08034 Barcelona, [email protected]
Abstract. Galled trees, directed acyclic graphs that model evolutionary histories
with isolated hybridization events, have become very popular due to both their
biological significance and the existence of polynomial time algorithms for their
reconstruction. In this paper we establish to which extent several distance measures
for the comparison of evolutionary networks are metrics for galled trees, and hence
when they can be safely used to evaluate galled tree reconstruction methods.
1
Introduction
The study of phylogenetic networks as a model of reticulate evolution began with
the representation of conflicting phylogenetic signals as an implicit splits nework [2,
22], but it was soon realized that internal nodes in a splits network did not have
any direct interpretation in evolutionary terms. Attention turned then to the study
of explicit evolutionary networks, in which the internal nodes have a direct interpretation as reticulate evolutionary events such as recombination, hybridization, or
lateral gene transfer. Unfortunately, the hardness of reconstructing an evolutionary
network with as few recombination events as possible for a set of sequences, under
the assumption of no repeated or back mutations, was soon established [6, 36, 37].
However, when the conflicting phylogenetic signals show a particular structure,
such that the conflict graph of the set of sequences is biconvex, the evolutionary
network with the smallest possible number of recombination events is unique, it
can be reconstructed in polynomial time and it is a galled tree, an evolutionary
network with hybrid nodes of in-degree 2 (because they correspond to explicit recombinations) and disjoint reticulation cycles [19]. Galled trees are also relevant
from a biological point of view because, as Gusfield et al point out in loc. cit.,
reticulation events tend to be isolated, yielding to disjoint reticulation cycles, if
the level of recombination is moderate, or if most of the observable recombinations
are recent. Actually, several slightly different notions of galled tree have been introduced so far in the literature, depending on the degree of disjointness of their
reticulation cycles. The original galled trees [19] have node-disjoint reticulation
cycles, while the nested networks with nesting level 1 [25, 27] (dubbed, for simplicity, 1-nested networks in this paper) have arc-disjoint reticulation cycles. Between
both notions lie the level-1 networks [16, 26], without biconnected components with
more than one hybrid node. We have studied the relationships among these types
of networks [35]: see Section 3 below.
Now, various algorithms are known for reconstructing galled trees from either
sequences [18–21], trees [33], distances [15], splits [23], or triplets [24], and metrics
provide a safe way to assess phylogenetic reconstruction methods [28, 32]. A few
polynomial time computable metrics, like the path multiplicity or µ-distance [14]
or Nakhleh’s metric m for reduced networks [31], are known for tree child evolutionary networks [14], which include galled trees [35]. But most distance measures
introduced so far were only known to be metrics on time consistent [4] tree-child
phylogenetic networks, including the Robinson-Foulds distance [3, 13, 10], the tripartitions distance [13, 29], the nodal and splitted nodal distances [9, 11], and the
triplets distance [11]. Since galled trees need not be time consistent, it was not
known whether these distance measures define metrics for galled trees. On the
other hand, Nakhleh gave in his PhD Thesis [30] two metrics for time consistent
galled trees (based on splits and subtrees), but they are not metrics for arbitrary
galled trees [14]. Recent simulation studies using the coalescent model with recombination show that only a small fraction of the simulated galled trees are time
consistent [1].
In this paper, we study which of the aforementioned metrics for tree-child time
consistent phylogenetic networks are also metrics for galled trees, under the various
notions of the latter. We show that the Robinson-Foulds distance is only a metric
in the binary case (in which the original galled trees, the level-1 networks and the
1-nested networks are the same objects); the tripartitions distance is a metric for
1-nested networks without any restriction on the degrees of their nodes (besides
the general restriction that hybrid nodes have in-degree 2); and the splitted nodal
distance is a metric in the semibinary (hybrid nodes of in-degree 2 and out-degree
1) case, in which the 1-nested and level-1 conditions define the same objects, but
they are strictly weaker than the node-disjoint reticulation cycles condition). On
the other hand, neither the nodal distance nor the triplets distance are metrics
even for the most restrictive case of binary galled trees.
2
Preliminaries
Given a set S, a S-rDAG is a rooted directed acyclic graph with its leaves bijectively labeled in S.
A tree node of a S-rDAG N = (V, E) is a node of in-degree at most 1, and a
hybrid node is a node of in-degree at least 2. A tree arc (respectively, a hybridization
arc) is an arc with head a tree node (respectively, a hybrid node). A node v ∈ V
is a child of u ∈ V if (u, v) ∈ E; we also say in this case that u is a parent of v.
Two nodes are sibling when they have a common parent.
We denote by u v any path in N with origin u and end v. Whenever there
exists a path u v, we shall say that v is a descendant of u and also that u is an
2
ancestor of v. The length of a path is its number of arcs, and the distance from a
node u to a descendant v of it is the length of a shortest path u v.
A node v is a strict descendant of a node u in N when every path from the root
of N to v contains the node u; thus, v is a non-strict descendant of u when it is
a descendant of u, but there exist paths from the root to v that do not contain u.
The following straightforward result, which is Lemma 1 in [10], will be used often,
usually without any further notice.
Lemma 1. Every strict ancestor of a node v is connected by a path with every
ancestor of v.
⊓
⊔
A tree path is a path consisting only of tree arcs, and a node v is a tree descendant of a node u when there is a tree path u v. The following result summarizes
Lemma 3 and Corollary 4 in [13], and it will also be used many times in this paper
without any further notice.
Lemma 2. Let u
v be a tree path in a S-rDAG.
(1) Every other path w v ending in v either is contained in u v or contains
u v. In particular, if w is a descendant of u and there exists a path w v,
then this path is contained in the tree path u v.
(2) The tree path u v is the unique path from u to v.
(3) The node v is a strict descendant of u.
⊓
⊔
Two paths in a S-rDAG are internally disjoint when they have disjoint sets of
intermediate nodes. A reticulation cycle for a hybrid node h is a pair of internally
disjoint paths ending in h and with the same origin. Each one of the paths forming
a reticulation cycle for h is called generically a merge path, their common origin is
called the split node of the reticulation cycle, and the hybrid node h, its end. The
intermediate nodes of a reticulation cycle are the intermediate nodes of the merge
paths forming it.
A subgraph of an undirected graph is biconnected when it is connected and it
remains connected if we remove any node and all edges incident to it. A subgraph
of a S-rDAG N is said to be biconnected when it is so in the undirected graph
associated to N .
3
1-nested networks
In the rest of this paper, by a hybridization network on a set S we understand
a S-rDAG without out-degree 1 tree nodes and with all its hybrid nodes of indegree 2. We shall also use the term hybridization network with n leaves to refer
to a hybridization network on a set S with n elements. A phylogenetic tree is a
hybridization network without hybrid nodes.
3
We shall say that a hybridization network is semibinary when its hybrid nodes
have out-degree 1, and that it is binary when it is semibinary and its internal tree
nodes have out-degree 2.
A hybridization network is:
– a galled tree, when every pair of reticulation cycles have disjoint sets of nodes
[19].
– 1-nested, when every pair of reticulation cycles have disjoint sets of arcs: by
[35, Prop. 12], this is equivalent to the fact that every pair of reticulation cycles
for different hybrid nodes have disjoint sets of intermediate nodes, and hence it
corresponds to the notion of nested (hybridization) network with nesting depth
1 [25, 27].
– level-1, when no biconnected component contains more than 1 hybrid node [16,
26].
To simplify the language, from now on we shall write simply 1-nested network to
mean a 1-nested hybridization network. The following two results summarize the
main results on 1-nested networks proved in [35].
Lemma 3. In a 1-nested network, every hybrid node is the end of exactly one
reticulation cycle, and all the intermediate nodes of this reticulation cycle are of
tree type.
⊓
⊔
Theorem 1. (a) Every 1-nested network is tree-child, in the sense that every
internal node has a child of tree type.
(b) For general hybridization networks,
galled tree =⇒ level-1 =⇒ 1-nested,
and these implications are strict.
(c) For semibinary hybridization networks,
galled tree =⇒ level-1 ⇐⇒ 1-nested,
and the first implication is strict.
(d) For binary hybridization networks,
galled tree ⇐⇒ level-1 ⇐⇒ 1-nested.
⊓
⊔
The fact that every 1-nested network is tree-child implies, by [13, Lem. 2] the
following result.
Corollary 1. Every node in a 1-nested network has some tree descendant leaf,
and hence some strict descendant leaf.
⊓
⊔
The following result lies at the basis of most of our proofs.
4
Proposition 1. Every 1-nested network contains some internal tree node with
all its children tree leaves, or a hybrid node with all its children tree leaves and
such that all the intermediate nodes in its reticulation cycle have all their children
outside the reticulation cycle tree leaves.
Proof. Let N be a 1-nested network. Let the galled-length of a path in N be
the number of reticulation cycles which the arcs of the path belong to, and the
galled-depth of a node in N the largest galled-length of a path from the root to
it. Notice that the galled-depth of a hybrid node is equal to the galled-depth of
the intermediate nodes of its reticulation cycle (because every arc in N belongs at
most to one reticulation cycle).
Assume that N does not contain any internal tree node with all its children tree
leaves. Let h be a hybrid node of largest galled-depth in N , and let v denote either
h or any intermediate node in the reticulation cycle K for h. It turns out that
v has no hybrid descendant other than h, because any path from v to any other
hybrid node h′ 6= h would contain arcs belonging to at least one more reticulation
cycle, making the galled-depth of h′ larger than that of v.
Let v ′ be any descendant of v not belonging to K. Then, v ′ is a tree node and
all its descendants are tree nodes, and therefore, since we assume that N does not
contain any internal tree node with all its children tree leaves, we conclude that v ′
is a tree leaf.
4
Reductions for 1-nested networks
We introduce in this section a set of reductions for 1-nested networks. Each of
these reductions, when applied to a 1-nested network with n leaves and m nodes,
produces a 1-nested network with at most n leaves and less than m nodes, and
given any 1-nested network with more than one leaf, it is always possible to apply
to it some of these reductions. We shall also show that suitable subsets of these
reductions have similar properties for binary and for semibinary 1-nested networks.
Similar sets of reductions for other types of evolutionary networks have already
been published [7, 11].
The R reductions. Let N be a 1-nested network with n leaves, and let u be an
internal node whose children are exactly the tree leaves i and j. The Ri;j reduction
of N is the network Ri;j (N ) obtained by removing the leaves i and j, together with
their incoming arcs, and labeling with i their former common parent u, which has
become now a leaf; cf. Fig. 1.3 It is clear that Ri;j (N ) is a 1-nested network on
S \ {j}, and it has 2 nodes less than N .
The T reductions. Let N be a 1-nested network with n leaves, and let u be an
internal node with two tree leaf children i, j and at least some other child. The Ti;j
3
In graphical representations of hybridization networks, we shall represent hybrid nodes by
squares, tree nodes by circles, and indeterminate (that is, that can be of tree or hybrid type)
nodes by pentagons.
5
⇒
i
i
j
Fig. 1. The Ri;j reduction.
reduction of N is the network Ti;j (N ) obtained by removing the leaf j together
with its incoming arc; cf. Fig. 2. It is clear that Ti;j (N ) is a 1-nested network on
S \ {j} with 1 node less than N .
u
j
i
⇒
...
u
i
...
Fig. 2. The Ti;j reduction.
The G reductions. Let N be a 1-nested network with n leaves. Assume that N
contains a reticulation cycle K consisting of two merge paths (u, v1 , . . . , vk , h) and
(u, v1′ , . . . , vk′ ′ , h), with k > k′ (k′ can be 0, in which case the corresponding merge
path is simply the arc (u, h), but then k > 0), such that
– the hybrid node h has only one child, and it is the tree leaf i;
– each intermediate node of K has only one child outside K, and it is a tree leaf:
the child outside K of each vj is the leaf ij and the child outside K of each vj′
is the leaf i′j .
Notice that u may have children outside K.
The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction of N is the network Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N ) obtained
k
k
by removing the nodes v1 , . . . , vk , v1′ , . . . , vk′ , h and the leaves i1 , . . . , ik , i′1 , . . . , i′k′ , i,
together with all their incoming arcs, and then adding to the node u two new
tree leaf children, labeled i and i1 ; cf. Fig. 3. Since we remove a complete reticulation cycle and all descendants of its intermediate nodes, and we replace it
by two tree leaves, it is clear that Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N ) is a 1-nested network on
k
S \ {i2 , . . . , ik , i′1 , . . . , i′k′ } (and in particular, if k = 1 and k′ = 0, it has the same
leaves as N ) with 2(k + k′ ) nodes less than N .
The G reductions. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction is the same as Gi;i1 ,...,ik ;i′1 ,...,i′ ′ ,
k
k
except for the fact that in order to apply the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction, the hybrid
k
node h must be the leaf labeled i, instead of the leaf’s parent: see Fig. 4. Then,
Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N ) is a 1-nested network on S \ {i2 , . . . , ik , i′1 , . . . , i′k′ } and it has
k
2(k + k′ ) − 1 nodes less than N .
6
=⇒
u
v1
′
v1
v2
′
v2
vk
′
vk
′
i1
i2
..
.
..
.
u
...
i1
i′1
i
...
i′2
i′k′
ik
h
i
Fig. 3. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction.
k
=⇒
u
v1
′
v1
v2
′
v2
vk
′
vk
′
i1
i2
..
.
..
.
u
...
i1
i′1
i
...
i′2
i′k′
ik
i
Fig. 4. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction.
k
Remark 1. In the G and G reductions, we leave two tree leaves attached to the
former split node of the removed reticulation cycle in order to ensure that their
application never generates an out-degree 1 tree node, while avoiding to increase
unnecessarily the number of reductions.
Now we have the following basic applicability results.
Proposition 2. Let N be a 1-nested network with more than one leaf. Then, at
least one R, T , G, or G reduction can be applied to N , and the result is a 1-nested
network.
Proof. If N contains some internal node v with at least two children that are tree
leaves, say i and j, then we can apply to N the Ri;j reduction, if the out-degree of
v is 2, or the Ti;j reduction, if its out-degree is greater than 2.
Assume now that N does not contain any internal node with more than one
tree leaf child: in particular, it does not contain any internal tree node with all its
children tree leaves. Then, by Proposition 1, it contains a hybrid node h with all
its children tree leaves (and therefore, by the current assumption on N , h either
is a leaf itself or has out-degree 1), and such that all the intermediate nodes in its
reticulation cycle K have all their children outside K tree leaves (and therefore
each one of them has exactly one child outside K, by Lemma 3 and the current
7
assumption on N ): let i1 , . . . , ik and i′1 , . . . , i′k′ , with k > k′ , be the tree leaf children of the intermediate nodes of the two respective merge paths of K, listed in
descending order of their parents along the path. Then, if h has out-degree 1 and
its child is the tree leaf i, we can apply to N the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction, while
k
if h is the leaf i, we can apply to N the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction.
k
The fact that the result of the application of a R, T , G, or G reduction to N
is again a 1-nested network has been discussed in the definition of the reductions.
Corollary 2. Let N be a semibinary 1-nested network with more than one leaf.
Then, at least one R, T , or G reduction can be applied to N , and the result is a
semibinary 1-nested network.
Proof. Since N does not contain hybrid leaves, we cannot apply to it any G reduction, and therefore, by Proposition 2, we can apply to it at least one R, T , or
G reduction.
Now, if we can apply a Ri;j or Ti;j reduction to N , the common parent of the
tree leaves i and j is a tree node, and if we can apply a Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction,
k
the split node of the reticulation cycle for the hybrid parent of i is a tree node (in
both cases because hybrid nodes in N have out-degree 1), and therefore neither
application produces a hybrid node of out-degree different from 1.
Corollary 3. Let N be a binary 1-nested network with more than one leaf. Then,
at least one R or G reduction can be applied to N , and the result is a binary
1-nested network.
Proof. Since N does not contain nodes with out-degree greater than 2, we cannot
apply to it any T reduction, and thus, by Corollary 2, we can apply to it some R
or G reduction.
Now, if we apply a R reduction to N , we replace an internal tree node with
two tree leaf children by a tree leaf, and the result is again binary. And if we apply
to N a G reduction, the split node of the reticulation cycle we remove is, as in the
semibinary case, a tree node, and in this case moreover without any child outside
the reticulation cycle (because its out-degree must be 2), and after the application
of the reduction it is still a tree node of out-degree 2.
We shall call the inverses of the R, T , G, and G reductions, respectively, the
−1
−1
−1
T −1 , G−1 , and G
expansions, and we shall denote them by Ri;j
, Ti;j
,
R−1 ,
−1
, and Gi;i1 ,...,ik ;i′1 ,...,i′ ′ . More specifically, for every 1-nested network
G−1
i;i1 ,...,ik ;i′1 ,...,i′k′
k
N:
−1
– if N contains a leaf labeled with i but no leaf labeled with j, then the Ri;j
−1
expansion can be applied to N , and Ri;j (N ) is obtained by unlabeling the leaf
i and adding to it two tree leaf children labeled i and j;
8
– if N contains a tree leaf labeled with i that has some sibling, but no leaf labeled
−1
−1
with j, then the Ti;j
expansion can be applied to N , and Ti;j
(N ) is obtained
by adding to the parent of the leaf i a new tree leaf child labeled with j;
– if N contains an internal node u with two tree leaf children i1 , i, but no leaf
expanlabeled with i2 , . . . , ik , i′1 , . . . , i′k′ (with k > k′ ), then the G−1
i;i1 ,...,i ;i′ ,...,i′
k 1
k′
(N ) is obtained by removing the
sion can be applied to N , and G−1
i;i1 ,...,ik ;i′1 ,...,i′k′
leaves i, i1 and their incoming arcs, and then starting in u two new internally
disjoint paths with k and k′ , respectively, intermediate nodes and ending in
the same hybrid node h, and then adding to each intermediate node of these
paths one new tree leaf and labeling these leaves (in descending order along
the paths) with i1 , . . . , ik and i′1 , . . . , i′k′ , respectively, and finally adding to h a
new tree leaf child labeled with i;
−1
– the application condition for the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ expansion is exactly the same
as for G−1
i;i1 ,...,i
k
−1
′
′
k ;i1 ,...,ik′
, and Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N ) is as G−1
i;i1 ,...,i
k
that the new hybrid node is itself a leaf labeled with i.
′
′
k ;i1 ,...,ik′
(N ), except
−1
¿From these descriptions we easily see that the result of a R−1 , T −1 , G−1 or G
expansion applied to a 1-nested network is always a 1-nested network.
The following result is easily deduced from the explicit descriptions of the
reductions and expansions.
Lemma 4. Let N and N ′ be two 1-nested networks. If N ∼
= N ′ , then the result of
′
−1
applying to both N and N the same R expansion (respectively, T −1 expansion,
−1
G−1 expansion or G expansion) are again two isomorphic 1-nested networks.
Moreover, if we apply a R, T , G, or G reduction to a 1-nested network N , then
we can apply to the resulting network the corresponding inverse R−1 , T −1 , G−1 ,
−1
or G expansion and the result is a 1-nested network isomorphic to N .
⊓
⊔
5
Proving metrics through reductions
Let C be throughout this section a class endowed with a notion of isomorphism ∼
=.
A metric on C is a mapping
d:C×C →R
satisfying the following axioms: for every A, B, C ∈ C,
(a)
(b)
(c)
(d)
Non-negativity: d(A, B) > 0;
Separation: d(A, B) = 0 if and only if A ∼
= B;
Symmetry: d(A, B) = d(B, A);
Triangle inequality: d(A, C) 6 d(A, B) + d(B, C).
A metric space is a pair (X, d) where X is a set and d is a metric on X, taking
as the notion of isomorphism in X the equality (that is, replacing ∼
= by = in the
separation axiom).
9
All distances for hybridization networks considered in this paper are induced
through representations, in the following sense. A representation of C in a metric
space (X, d) is a mapping
F :C→X
such that if A ∼
= B, then F (A) = F (B).
Given such a representation, the distance induced by d through F is the mapping
dF : C × C → R
defined by dF (A, B) = d(F (A), F (B)), for every A, B ∈ C.
The metric axioms for d imply that this mapping is non-negative, symmetric, it
sends pairs of isomorphic members of C to 0, and it satisfies the triangle inequality.
So, to be a metric on C, dF only needs to satisfy that dF (A, B) = 0 implies A ∼
= B.
Now, it is straightforward to prove the following result (cf. [12, Prop. 1]).
Lemma 5. The mapping dF is a metric on C if, and only if, it is injective up to
isomorphism, in the sense that, for every A, B ∈ C, if F (A) = F (B), then A ∼
= B.
⊓
⊔
Reductions as those introduced in the last section can be used to prove the
injectivity up to isomorphism of a representation F and hence, as a consequence,
that the corresponding dF is a metric; it was done for specific classes C of evolutionary networks and specific metrics in [7, 11]. Since we shall use several times
this kind of proofs in this paper, we make explicit here their general outline and
the lemma they rely on.
Let CS ′ ,m denote a class of 1-nested hybridization networks of S
some specific
type on a given set S ′ and with at most m nodes, and let CS ′ = m>|S ′ | CS ′ ,m .
Assume we have a set of reductions R1 , . . . , Rs that can be applied to members of
CS ′ , with inverse expansions R1−1 , . . . , Rs−1 . Consider the following conditions on
these reductions and expansions:
(R1) For every N ∈ CS ′ ,m with |S ′ | > 2, there exists some reduction Ri that can be
applied to N .
(R2) For every N ∈ CS ′ ,m and for every reduction Ri , Ri (N ) ∈ CSi′ ,mi for some
Si′ ⊆ S ′ and mi < m; moreover, Si′ and mi only depend on S ′ , m and Ri , not
on N .
(R3) For every N ∈ CS ′ ,m and for every reduction Ri , if Ri can be applied to N ,
then Ri−1 can be applied to Ri (N ) and Ri−1 (Ri (N )) ∼
= N.
′
′
(R4) For every reduction Ri and for every N, N ∈ CSi ,mi such that N ∼
= N ′ , if the
−1
corresponding expansion Ri can be applied to N , then it can also be applied
to N ′ and the resulting networks are isomorphic.
The definitions and results given in Section 4 imply that:
– The set of all R and G reductions satisfy conditions (R1) to (R4) for the classes
CS ′ of all binary 1-nested hybridization networks on a set S ′ .
10
– The set of all R, T , and G reductions satisfy conditions (R1) to (R4) for the
classes CS ′ of all semibinary 1-nested hybridization networks on a set S ′ .
– The set of all R, T , G, and G reductions satisfy conditions (R1) to (R4) for
the classes CS ′ of all 1-nested hybridization networks on a set S ′ .
Now, we have the following result.
Lemma 6. Let S be a given set of labels. For every S ′ ⊆ S, let CS ′ ,m and R1 , . . . , Rs
be as above, and assume that these reductions satisfy conditions (R1) to (R4). For
′
every S ′ ⊆ S, let FS ′ : CS ′ → XS ′ be a representation in a metric space (XS ′ , d(S ) ).
Then, FS is injective up to isomorphism if the following two conditions are
satisfied for every S ′ ⊆ S, for every m > |S ′ |, for every reduction Ri , and for
every N, N ′ ∈ CS ′ ,m such that FS ′ (N ) = FS ′ (N ′ ):
(A) If Ri can be applied to N , then it can also be applied to N ′ .
(R) If Ri is applied to N and N ′ , then FSi′ (Ri (N )) = FSi′ (Ri (N ′ )).
(S)
In particular, if these two conditions are satisfied, then dFS is a metric on CS .
Proof. We shall prove by induction on |S ′ | + m the following statement:
For every S ′ ⊆ S and m > |S ′ |, if N, N ′ ∈ CS ′ ,m are such that FS ′ (N ) =
FS ′ (N ′ ), then N ∼
= N ′.
The starting case, when |S ′ | + m = 2, is obvious because then S ′ must be a
singleton, and there is, up to isomorphism, only one 1-nested hybridization network
on a given singleton {i}: a single node labeled with i.
Let now N, N ′ ∈ CS ′ ,m be such that FS ′ (N ) = FS ′ (N ′ ) and |S ′ | + m > 3.
If |S ′ | = 1, we reason as in the starting case to deduce that N ∼
= N ′ , so we
′
assume that |S | > 2. By (R1), some reduction Ri can be applied to N , and
since FS ′ (N ) = FS ′ (N ′ ), by (A) it can also be applied to N ′ . Then, by (R2),
Ri (N ), Ri (N ′ ) ∈ CSi′ ,mi , with Si′ ⊆ S ′ and mi < m, and by (R) we have that
FSi′ (Ri (N )) = FSi′ (Ri (N ′ )). Therefore the induction hypothesis applies, implying
that Ri (N ) ∼
= Ri (N ′ ). But then, by (R3), Ri−1 can be applied to Ri (N ) and Ri (N ′ )
and Ri−1 (Ri (N )) ∼
=
= N ′ , while, by (R4), Ri−1 (Ri (N )) ∼
= N and Ri−1 (Ri (N ′ )) ∼
−1
′
′
∼
Ri (Ri (N )). This implies that N = N , as we wanted to prove.
Thus, in particular, we have that for every m > |S|, if N, N ′ ∈ CS,m are such
that FS (N ) = FS (N ′ ), then N ∼
= N ′ . Now notice that if N, N ′ ∈ CS , then there
′
exists some m such that N, N ∈ CS,m : take as m the largest number of nodes in
N or in N ′ . Therefore, FS is injective up to isomorphism, as we claimed.
Remark 2. If one wants to use a result like the last lemma to prove the injectivity
up to isomorphism of a certain representation of S-rDAGs more general than 1nested networks, then it may be necessary to explicitly add to (A) and (R) a third
condition that covers the starting case:
11
(S) For every i ∈ S, F{i} is injective up to isomorphism.
We shall also use a couple of times the following straightforward fact.
Lemma 7. Let F : C → X and F ′ : C → X ′ be two representations of C in metric
spaces (X, d) and (X ′ , d′ ), and assume that F (A) = F (B) implies F ′ (A) = F ′ (B).
Then, if F ′ is injective up to isomorphism, so is F .
⊓
⊔
When the hypothesis of this lemma is satisfied, we say that F refines F ′ , and
also that dF refines d′F ′ . Notice that if d′F ′ is a metric and dF refines it, then it is
also a metric.
6
Robinson-Foulds distance
Let N = (V, E) be a S-rDAG. For every node v ∈ V , the cluster of v in N is the
set C(v) ⊆ S of leaves that are descendants of v. The cluster representation of N
is the multiset
C(N ) = {CN (v) | v ∈ V },
where each member appears with multiplicity the number of nodes having it as
cluster. In particular, the cardinal of C(N ) (as a multiset, that is, every element
counted with its multiplicity) is equal to the number of nodes in N .
The Robinson-Foulds distance between a pair of S-rDAGs N, N ′ is
dRF (N, N ′ ) = |C(N ) △ C(N ′ )|,
where the symmetric difference and its cardinal refer to multisets. It is the natural generalization to S-rDAGs of the well known Robinson-Foulds distance for
phylogenetic trees [34].
Remark 3. If v is an ancestor of u in N , then C(u) ⊆ C(v), but the converse
implication is false, even in binary galled trees. See, for instance, Fig. 7 below: in
both networks, the root and its tree child have the same cluster, but the root is
not a descendant of its child.
It is known that the Robinson-Foulds distance is a metric on the class of all
regular evolutionary networks on a given set S (the networks N such that the
mapping v 7→ C(v) induces an isomorphism of directed graphs between N and
the Hasse diagram of (C(N ), ⊇)) [3] and on the class of all tree-child phylogenetic
networks on a given set S that do not contain any hybrid node with two parents
connected by a path [13]. Unfortunately, 1-nested networks, or even binary galled
trees, need not be regular (by Remark 3) and they can contain reticulation cycles
where one merge path is a single arc. So, we cannot use those results to prove that
the Robinson-Foulds distance is a metric, even on the class of all binary galled
trees.
12
As a matter of fact, the cluster representation is not injective up to isomorphism, and hence the Robinson-Foulds distance is not a metric, for 1-nested networks, or even galled trees, unless we restrict the possible in- and out-degrees of
their nodes: they cannot contain either internal tree nodes of out-degree other than
2 (see Fig. 5), or hybrid nodes of out-degree 0 (see Fig. 6) or greater than 1 (see
Fig. 7). Therefore, the Robinson-Foulds distance can only be a metric for binary
1-nested networks, that is, for binary galled trees. Now, we have the following
result.
1
2
3
4
1
2
3
4
Fig. 5. Two non-isomorphic galled trees with the same cluster representation and internal tree
nodes of out-degree 3.
1
2
1
2
Fig. 6. Two non-isomorphic galled trees with the same cluster representation and hybrid nodes
of out-degree 0.
1
2
3
4
4
1
2
3
Fig. 7. Two non-isomorphic galled trees with the same cluster representation and hybrid nodes
of out-degree 2.
13
Theorem 2. Let N, N ′ be two binary 1-nested networks on a given set S such
that C(N ) = C(N ′ ).
(A) If a specific R or G reduction can be applied to N , then it can also be applied
to N ′ .
(R) If a specific R or G reduction is applied to N and N ′ , the resulting networks
have the same cluster representations.
⊓
⊔
In order not to lose the thread of the paper, we postpone the proof of this
theorem until §A2 in the Appendix at the end of the paper. Combining Lemma 6
with this theorem, we obtain the following result.
Corollary 4. The Robinson-Foulds distance is a metric on the class of all binary
galled trees on a given set S.
⊓
⊔
7
Tripartitions distance
Let N = (V, E) be a S-rDAG. For every node v ∈ V , let A(v) ⊆ S be the set of
(labels of) strict descendant leaves of v and B(v) = C(v) \ A(v) the set of nonstrict descendant leaves of v; B(v) may be empty, but A(v) 6= ∅ by Lemma 3. The
tripartition associated to v [29] is
θ(v) = (A(v), B(v), S \ C(v)).
Notice that the tripartition associated to a node v refines its cluster C(v), by
splitting it into A(v) and B(v).
The tripartitions representation of N is the multiset
θ(N ) = {θ(v) | v ∈ V }
of tripartitions of the nodes of N . The tripartitions distance between a pair of
S-rDAGs N, N ′ is
dtri (N, N ′ ) = |θ(N ) △ θ(N ′ )|,
where the symmetric difference and its cardinal refer to multisets.
It turns out that the tripartitions distance is a metric on the class of all 1nested networks on a given set. It is a consequence of the following proposition,
whose proof we postpone until §A3 in the Appendix.
Theorem 3. Let N, N ′ be two 1-nested networks on a given set S such that
θ(N ) = θ(N ′ ).
(A) If a specific R, T , G, or G reduction can be applied to N , then it can also be
applied to N ′ .
(R) If a specific R, T , G, or G reduction is applied to N and N ′ , the resulting
networks have the same tripartitions representations.
⊓
⊔
14
So, using Lemma 6, we deduce the following result.
Corollary 5. The tripartitions distance is a metric on the class of all 1-nested
networks on a given set S.
⊓
⊔
Remark 4. Another refinement (in the sense of Lemma 7) of the Robinson-Foulds
distance, the so-called µ-distance, was introduced by Cardona et al [14] and proved
to be a metric on the class of all tree-child S-rDAGs for any given S: then, in
particular, it is a metric on the class of all 1-nested networks on a set S. Soon later,
L. Nakhleh [31] proposed a distance m that turned out to refine the µ-distance [12]
and therefore that is also a metric on the class of all 1-nested networks on a set
S. The interested reader can look up the aforementioned references for the specific
definitions of these metrics.
8
Nodal and splitted nodal distances
Let N = (V, E) be a S-rDAG; to simplify the language, throughout this section we
assume that S = {1, . . . , n} with n = |S|. Recall from [10] that the least common
semistrict ancestor, LCSA for short, of a pair of nodes u, v ∈ V is the node that is
a common ancestor of u and v and strict ancestor of at least one of them, and that
is a descendant of all other nodes in N satisfying these properties. Such a LCSA
of a pair of nodes u, v always exists and it is unique [10, §IV], and we shall denote
it by [u, v].
The LCSA of a pair of nodes in a phylogenetic tree is their lowest common
ancestor. It turns out that such a characterization extends to 1-nested networks.
Recall that a lowest common ancestor, LCA for short, of a pair of nodes u, v in
a rDAG is any common ancestor of u and v that is not a proper ancestor of any
other common ancestor of them [5].
Lemma 8. Every pair of nodes u, v in a 1-nested network has only one LCA, and
it is their LCSA.
Proof. Let x be any LCA of u and v, and let us prove that x must be a strict
ancestor of u or v. Indeed, by Lemma 9 in §A1 in the Appendix, if x is not a strict
ancestor of u, then it is intermediate in the reticulation cycle for a hybrid node
hu that is a strict ancestor of u. In a similar way, if x is not a strict ancestor of v,
then it is intermediate in the reticulation cycle for a hybrid node hv that is a strict
ancestor of v. Now, if x were not a strict ancestor either of u or of v, then it either
would happen that it is intermediate in reticulation cycles for two different hybrid
nodes, which is impossible in a 1-nested network, or that it is a proper ancestor of
a common ancestor of u and v, namely hu = hv , against the assumption that x is
a LCA of u and v.
So, x is a common ancestor of u and v and a strict ancestor of at least one of
them, and thus it is an ancestor of [u, v]. Since x cannot have proper descendants
that are common ancestors of u and v, we conclude that x = [u, v].
15
For every pair of leaves i, j ∈ S, let ℓN (i, j) and ℓN (j, i) be the distances from
[i, j] to i and to j, respectively, and let νN (i, j) = ℓN (i, j) + ℓN (j, i).
The LCSA-path lengths matrix of N is the symmetric matrix
νN (1, 1) . . . νN (1, n)
..
..
..
ν(N ) =
.
.
.
νN (n, 1) . . . νN (n, n)
and the splitted LCSA-path lengths matrix of N is the (not necessarily symmetric)
matrix
ℓN (1, 1) . . . ℓN (1, n)
..
..
..
ℓ(N ) =
.
.
.
ℓN (n, 1) . . . ℓN (n, n)
The nodal distance between a pair of S-rDAGs N, N ′ is half the Manhattan,
or L1 , distance between ν(N ) and ν(N ′ ):
dν (N, N ′ ) =
1
2
X
|νN (i, j) − νN ′ (i, j)|.
16i6=j6n
The splitted nodal distance between N and N ′ is the Manhattan distance between
ℓ(N ) and ℓ(N ′ ):
X
dℓ (N, N ′ ) =
|ℓN (i, j) − ℓN ′ (i, j)|.
16i6=j6n
Of course, instead of using the Manhattan distance on the set of n × n matrices,
one can use any other distance for real-valued matrices to compare LCSA-path
lengths, or splitted LCSA-path lengths, matrices, like for instance the euclidean
distance. The results in this section do not depend on the actual metric for realvalued matrices used.
The nodal distance dν is the natural generalization to S-rDAGs of the classical nodal metric for binary phylogenetic trees [17, 38], while the splitted nodal
distance dℓ generalizes to S-rDAGs the recently introduced homonymous metric
for arbitrary phylogenetic trees [8].
It is known [9, 11] that dν is a metric on the class of all binary tree-child time
consistent phylogenetic networks on a given set S, and that dℓ is a metric on the
class of all tree-child time consistent phylogenetic networks on a given set S, but no
binary galled tree containing a reticulation cycle with one merge path consisting of
a single arc is time consistent, and therefore we cannot use these results to prove
that dν or dℓ are metrics even for binary galled trees.
It turns out that ν is not injective up to isomorphism, and hence dν is not a
metric, even for binary galled trees, as Fig. 8 shows. As far as ℓ goes, it is not
injective up to isomorphism for 1-nested networks, or even galled trees, that are
not semibinary: if we allow hybrid nodes of out-degree 0 (see Fig. 9) or greater
16
1
2
2
1
Fig. 8. Two non-isomorphic binary galled trees on S = {1, 2} and whith the same LCSA-path
length, 3, between their only two leaves.
than 1 (see Fig. 10), there exist pairs of non-isomorphic galled trees with the same
splitted LCSA-path length matrices. Therefore, dℓ can be a metric at most on the
class of all semibinary 1-nested networks. Now, we have the following result.
1
2
1
2
Fig. 9. Two non-isomorphic galled trees with the same ℓ matrix and hybrid nodes of out-degree
0.
1
2
3
1
2
3
Fig. 10. Two non-isomorphic galled trees with the same ℓ matrix and hybrid nodes of out-degree
greater than 1.
Theorem 4. Let N, N ′ be two semibinary 1-nested networks on a given set S such
that ℓ(N ) = ℓ(N ′ ).
(A) If a specific R, T , or G reduction can be applied to N , then it can also be
applied to N ′ .
(R) If a specific R, T , G reduction is applied to N and N ′ , the resulting networks
have the same splitted LCSA-path lengths matrices.
⊓
⊔
As we did previously, we postpone the proof of this theorem until §A4 in the
Appendix at the end of the paper. Combining Lemma 6 with this theorem, we
obtain the following result.
17
Corollary 6. The splitted nodal distance is a metric on the class of all semibinary
1-nested networks on a given set S.
⊓
⊔
9
Conclusion
Several slightly different definitions of galled tree, capturing the notion of a hybridization network with isolated reticulation cycles, have been proposed so far in
the literature. The most general such definition is as a network with arc-disjoint
reticulation cycles [16, 26], called in this paper 1-nested, and the most restrictive is
Gusfield et al’s original definition of a galled tree as a network with node-disjoint
reticulation cycles [19]: in between lie the level-1 networks of Janson, Sung et al [25,
27]. In the semibinary (hybrid nodes of in-degree 2 and out-degree 1) case, level-1
and 1-nested networks are the same, and in the binary (semibinary plus tree nodes
of out-degree 2) case, galled trees, level-1 networks and 1-nested networks are the
same objects.
In this paper we have established for which classes of 1-nested networks on a
fixed set of labels, several distance measures introduced so far in the literature satisfy the axioms of metrics: actually, only the separation axiom (distance 0 means
isomorphism) is relevant here, because all other axioms of metrics are always satisfied by these distances. In summary, we have proved that:
(a) The Robinson-Foulds distance [3, 10] is a metric only for binary galled trees.
(b) The tripartitions distance [29], the µ-distance [14] and Nakhleh’s metric m for
reduced networks [31] are metrics for arbitrary 1-nested networks.
(c) The natural translation of the nodal distance for phylogenetic trees to evolutionary networks [9] is not a metric even for binary galled trees.
(d) The splitted nodal distance [9, 11] is a metric for semibinary 1-nested networks,
but not for arbitrary galled trees.
We would like to mention that the 1-nested networks turn out to form the first welldefined class of evolutionary networks where the tripartitions distance is shown to
be a metric but the Robinson-Foulds distance is not a metric.
There are other distances that have not been discussed in this paper because
they obviously fail to be metrics even for binary galled trees. This is the case of
the triplets distance [11], which cannot be a metric for binary galled trees because
there are many more binary galled trees with 3 leaves than possible triplets in the
sense defined in the aforementioned paper. And, as it was already observed in [14,
§II.D], it is also the case of any distance defined by comparing the multisets of
induced subtrees, or the multisets of splits of induced subtrees: for instance, the
pairs of galled trees depicted in Figs. 9 or 10 have the same multisets of induced
subtrees.
The splitted nodal distance and the triplets distance were introduced in [11] as
suitable generalizations of the corresponding distances for phylogenetic networks
18
with the aim of obtaining metrics on the class of tree-child time consistent phylogenetic networks, and hence they were not designed to cope with reticulation cycles
where one merge path is a single arc. This is the main reason of their failure as
metrics for arbitrary 1-nested networks. But it seems not difficult to modify them
to obtain metrics for 1-nested networks, by taking into account the restricted, and
specific, topological structure of these networks: something similar was already
done with the splitted nodal distance to make it work on tree-child time consistent
evolutionary networks with hybrid nodes of (almost) arbitrary type [9].
Galled trees, 1-nested networks, and level-1 networks are defined as having
hybrid nodes of in-degree 2, in the first case by semantical reasons and in the
other two cases for practical reasons (to guarantee that certain reconstruction
algorithms run in polynomial time), and we have kept this restriction in this paper.
But, although Gusfield et al’s node-disjoint reticulation cycles condition implies
that hybrid nodes must have in-degree 2, this restriction is not necessary in level1 and 1-nested networks, and polynomial time algorithms for the reconstruction
of level-1 or 1-nested networks with hybrid nodes of arbitrary in-degree may be
discovered in the future, in which case it would be interesting to know whether the
distance measures discussed in this paper define metrics in this more general case
and they can be used thus to assess these new algorithms.
References
1. M. Arenas, G. Valiente, D. Posada, Characterization of phylogenetic reticulate networks based
on the coalescent with recombination, Molecular Biology and Evolution 25 (12) (2008) 2517–
2520.
2. H.-J. Bandelt, A. W. Dress, Split decomposition: A new and useful approach to phylogenetic
analysis of distance data, Molecular Phylogenetics and Evolution 1 (3) (1992) 242–252.
3. M. Baroni, C. Semple, M. Steel, A framework for representing reticulate evolution, Ann.
Combin. 8 (4) (2004) 391–408.
4. M. Baroni, C. Semple, M. Steel, Hybrids in real time, Systematic Biology 55 (1) (2006) 46–56.
5. M. A. Bender, M. Farach-Colton, G. Pemmasani, S. Skiena, P. Sumazin, Lowest common
ancestors in trees and directed acyclic graphs, Journal of Algorithms 57 (2) (2005) 75–94.
6. M. Bordewich, C. Semple, Computing the minimum number of hybridization events for a
consistent evolutionary history, Discrete Applied Mathematics 155 (8) (2004) 914–928.
7. G. Cardona, M. Llabrés, F. Rosselló, G. Valiente, A distance metric for a class of tree-sibling
phylogenetic networks, Bioinformatics 24 (13) (2008) 1481–1488.
8. G. Cardona, M. Llabrés, F. Rosselló, G. Valiente, Nodal distances for rooted phylogenetic
trees, Journal of Mathematical BiologySubmitted.
9. G. Cardona, M. Llabrés, F. Rosselló, G. Valiente, Path lengths in tree-child time consistent
hybridization networks, arXiv:0807.0087v1 [q-bio.PE]; submitted (2008).
10. G. Cardona, M. Llabrés, F. Rosselló, G. Valiente, Metrics for phylogenetic networks I: Generalizations of the Robinson-Foulds metric, IEEE/ACM Transactions on Computational Biology and Bioinformatics 6 (1) (2009) 1–16.
11. G. Cardona, M. Llabrés, F. Rosselló, G. Valiente, Metrics for phylogenetic networks II: Nodal
and triplets metrics, IEEE/ACM Transactions on Computational Biology and Bioinformatics In press, DOI: 10.1109/TCBB.2008.127.
12. G. Cardona, M. Llabrés, F. Rosselló, G. Valiente, On Nakhleh’s metric for reduced phylogenetic networks, IEEE/ACM Transactions on Computational Biology and Bioinformatics In
press, DOI: 10.1109/TCBB.2009.33.
19
13. G. Cardona, F. Rosselló, G. Valiente, Tripartitions do not always discriminate phylogenetic
networks, Mathematical Biosciences 211 (2) (2008) 356–370.
14. G. Cardona, F. Rosselló, G. Valiente, Comparison of tree-child phylogenetic networks, IEEE/ACM Transactions on Computational Biology and Bioinformatics In press,
DOI: 10.1109/TCBB.2007.70270.
15. H.-L. Chan, J. Jansson, T. W. Lam, S.-M. Yiu, Reconstructing an ultrametric galled phylogenetic network from a distance matrix, Journal of Bioinformatics and Computational Biology
4 (4) (2006) 807–832.
16. C. Choy, J. Jansson, K. Sadakane, W.-K. Sung, Computing the maximum agreement of
phylogenetic networks, Theoretical Computer Science 335 (1) (2005) 93–107.
17. J. S. Farris, On comparing the shapes of taxonomic trees, Systematic Zoology 22 (1973)
50–54.
18. D. Gusfield, V. Bansal, V. Bafna, Y. S. Song, A decomposition theory for phylogenetic networks and incompatible characters, Journal of Computational Biology 14 (10) (2007) 1247–
1272.
19. D. Gusfield, S. Eddhu, C. Langley, Efficient reconstruction of phylogenetic networks with
constrained recombination, in: Proc. 2nd IEEE Computer Society Bioinformatics Conf., 2003.
20. D. Gusfield, S. Eddhu, C. H. Langley, The fine structure of galls in phylogenetic networks,
INFORMS Journal on Computing 16 (4) (2004) 459–469.
21. D. Gusfield, S. Eddhu, C. H. Langley, Optimal, efficient reconstruction of phylogenetic networks with constrained recombination, Journal of Bioinformatics and Computational Biology
2 (1) (2004) 173–213.
22. D. H. Huson, D. Bryant, Application of Phylogenetic Networks in Evolutionary Studies,
Molecular Biology and Evolution 23 (2) (2006) 254–267.
23. D. H. Huson, T. H. Klöpper, Beyond galled trees—decomposition and computation of galled
networks, in: Proc. 11th Annual Int. Conf. Research in Computational Molecular Biology,
vol. 4453 of Lecture Notes in Computer Science, Springer, 2007, pp. 211–225.
24. J. Jansson, N. B. Nguyen, W.-K. Sung, Algorithms for combining rooted triplets into a galled
phylogenetic network, SIAM Journal on Computing 35 (5) (2006) 1098–1121.
25. J. Jansson, W.-K. Sung, The maximum agreement of two nested phylogenetic networks, in:
Proc. 15th Int. Symp. Algorithms and Computation, vol. 3341 of Lecture Notes in Computer
Science, Springer, 2004, pp. 581–593.
26. J. Jansson, W.-K. Sung, Inferring a level-1 phylogenetic network from a dense set of rooted
triplets, Theoretical Computer Science 363 (1) (2006) 60–68.
27. J. Jansson, W.-K. Sung, The maximum agreement of two nested phylogenetic networks, in:
O. N. Terikhovsky, W. N. Burton (eds.), New Topics in Theoretical Computer Science, chap. 4,
Nova Publishers, 2008, pp. 119–141.
28. M. M. Miyamoto, J. Cracraft, Phylogenetic Analysis of DNA Sequences, Oxford University
Press, 1991.
29. B. M. E. Moret, L. Nakhleh, T. Warnow, C. R. Linder, A. Tholse, A. Padolina, J. Sun,
R. Timme, Phylogenetic networks: Modeling, reconstructibility, and accuracy, IEEE/ACM
Transactions on Computational Biology and Bioinformatics 1 (1) (2004) 13–23.
30. L. Nakhleh, Phylogenetic networks, Ph.D. thesis, University of Texas at Austin (2004).
31. L. Nakhleh, A metric on the space of reduced phylogenetic networks, IEEE/ACM Transactions on Computational Biology and Bioinformatics In press, DOI: 10.1109/TCBB.2009.2.
32. L. Nakhleh, J. Sun, T. Warnow, C. R. Linder, B. M. E. Moret, A. Tholse, Towards the development of computational tools for evaluating phylogenetic network reconstruction methods,
in: Proc. 8th Pacific Symp. Biocomputing, 2003.
33. L. Nakhleh, T. Warnow, C. R. Linder, K. St. John, Reconstructing reticulate evolution in
species—theory and practice, Journal of Computational Biology 12 (6) (2005) 796–811.
34. D. F. Robinson, L. R. Foulds, Comparison of phylogenetic trees, Mathematical Biosciences
53 (1/2) (1981) 131–147.
35. F. Rosselló, G. Valiente, All that glisters is not galled, arXiv:0904.2448v1 [cs.DM]; submitted
(2009).
20
36. L. Wang, B. Ma, M. Li, Fixed topology alignment with recombination, Discrete Applied
Mathematics 104 (1–3) (2000) 281–300.
37. L. Wang, K. Zhang, L. Zhang, Perfect phylogenetic networks with recombination, Journal of
Computational Biology 8 (1) (2001) 69–78.
38. W. T. Williams, H. T. Clifford, On the comparison of two classifications of the same set of
elements, Taxon 20 (4) (1971) 519–522.
Appendix: Proofs of the main theorems
A1
Some lemmas on clusters and tripartitions
We establish in this subsection some basic properties of clusters on 1-nested networks that will be used in the proofs of the next two subsections. To simplify the
notations, given a 1-nested network N on a set S, let CI (N ) denote the multiset of
clusters of its internal nodes. CI (N ) is obtained by removing from C(N ) one copy
of every singleton {i} with i ∈ S.
Lemma 9. Let N be a 1-nested network on S.
(a) For every i ∈ S and for every internal node v, C(v) = {i} if, and only if, i is
a tree leaf and v is its parent and it has out-degree 1.
(b) If two leaves i, j are such that there does not exist any member of CI (N ) containing one of them and not the other, then they are sibling.
(c) Let v be a tree node and u its only parent. If C(u) 6= C(v), then C(u) is the
only (up to multiplicities) minimal member of CI (N ) strictly containing C(v).
If C(u) = C(v) and u has out-degree greater than 1, then u is the split node of
a reticulation cycle such that one of the merge paths contains v as intermediate
node and the other merge path is a single arc.
(d) If a node v is a non-strict descendant of a node u, then u is intermediate in
the reticulation cycle for a hybrid node that is a strict ancestor of v.
Proof. (a) If v is a node with only one child and this child is the tree leaf i, then
C(v) = {i} ∈ CI (N ). Conversely, let v be an internal node such that C(v) =
{i}. Since every internal node in N has a tree descendant leaf, i must be a tree
descendant leaf of v (and in particular a tree leaf). Let w be the parent of i, and
let us prove that it has out-degree 1. Indeed, if u is a child of w other than i, it
has a tree descendant leaf j, and j 6= i, because, otherwise, the only parent w of i
would be a descendant of its child u. But then j ∈ C(v), against the assumption
that C(v) = {i}.
So, the tree path v i cannot have any intermediate node, because otherwise w
would be intermediate in this path and hence it would be a tree node, but internal
tree nodes in N have out-degree greater than 1. Therefore, v is the parent of i.
But then, as we have just seen, it must have out-degree 1.
(b) Assume that the every member of CI (N ) containing i or j contains both
of them, but that i and j are not sibling. Let v1 be a parent of i: then i ∈ C(v1 )
21
implies j ∈ C(v1 ) and, since v1 is not a parent of j, v1 is a proper ancestor of some
parent w1 of j. Then, j ∈ C(w1 ) implies i ∈ C(w1 ) and thus, since w1 is not a
parent of i, w1 is a proper ancestor of some parent v2 6= v1 (because v1 is a proper
ancestor of w1 ) of i. Iterating this process, we obtain that v2 is a proper ancestor
of another parent w2 6= w1 of j, and then that w2 is a proper ancestor of another
parent v3 6= v1 , v2 of i, which is impossible because every node in N has at most 2
parents.
(c) Let u be the parent of the tree node v. Assume that C(u) 6= C(v) and let i
be a tree descendant leaf of v, and hence also of u. For every other internal node
w, if C(v) ( C(w), then i ∈ C(w), and therefore either the path u i is contained
in the path w
i or conversely. But C(v) ( C(w) implies that w cannot be a
descendant of v, and we conclude that u i is contained in the path w i, and
hence u is a descendant of w, which implies that C(u) ⊆ C(w).
Assume now that u has out-degree greater than 1 and that C(u) = C(v). Let v ′
be another child of u and let j be a tree descendant leaf of v ′ . Then, j ∈ C(v), and
therefore either the path v j contains the path v ′ j or the path v ′ j contains
the path v j. But the last situation is impossible, because if v belongs to the path
v ′ j, its only parent u should also belong to it, and u cannot be a descendant
of its child v ′ . So, we conclude that v ′ is a descendant of v, and therefore that v ′
is a hybrid node and its reticulation cycle consists of the arc (u, v ′ ) and a path
(u, v, . . . , v ′ ).
(d) Assume that v is a non-strict descendant of u. Let u v be any path from u to
v, and r v a path from the root r of N to v not containing u. Let w be the first
node in u v contained also in r v. Since, by assumption, w 6= u and, clearly,
w 6= r, w will have different parents in both paths, which implies that it is hybrid.
Let now r u be any path from the root to u and let x be the last node in this
path belonging to the subpath r w of r v: again, u 6= x. Then, the subpath
x
w of r
v and the concatenation of the subpath x
u of r
u and the
subpath u w of u v are internally disjoint, and hence they form a reticulation
cycle for w with split node x and having u as intermediate node.
It remains to prove that w is a strict ancestor of v. But if it were not, then, as
we have just seen, w would be intermediate in a reticulation cycle for an ancestor
of v, which is impossible by Lemma 3.
Lemma 10. Let N be a 1-nested network on S, let h be a hybrid node of N with
C(h) = {i}, and let K be its reticulation cycle, with split node u.
(a) No pair of intermediate nodes of K in different merge paths are connected by
a path.
(b) Every pair of intermediate nodes in K have different clusters, and different
also from C(h).
(c) The only non-strict descendant of each intermediate node of K is i.
(d) The intersection of the clusters of any pair of intermediate nodes of different
merge paths of K is {i}.
22
(e) i is a strict descendant of u.
(f ) If v is a node outside K such that i ∈ C(v), then v is an ancestor of u and
thus C(u) ⊆ C(v).
(g) All clusters of intermediate nodes in K have multiplicity 1 in CI (N ), except
the cluster of the child other than h of u when one of the merge paths consists
of a single arc (u, h).
(h) The minimal elements of CI (N ) strictly containing C(h) are the clusters of the
parents of h that are intermediate in K.
Proof. By Lemma 9.(a), C(h) = {i} implies that either h = i or that i is a tree
child of h, and its only child.
(a) If x and y were two intermediate nodes of K belonging to different merge paths
and there existed a path x y, then the first node in this path also belonging to
the path u y would have different parents in both paths, and therefore it would
be hybrid, which is impossible by Lemma 3.
(b) Let x and y be two different intermediate nodes of K: if they belong to the
same merge path, we take them so that y is a proper descendant of x. We shall
prove that C(x) 6= C(y).
Since both nodes are of tree type, x has a child v outside K. Let l be a tree
descendant leaf of v, and assume that l ∈ C(y). Then, either the path v l contains
the path y l or vice versa. But the tree path u y contained in the merge path is
the unique path from u to y, and it does not contain v, and therefore y cannot be
a descendant of v. Thus, v is a descendant of y, and since x is not a descendant of
y by (a), we conclude that v is a hybrid node such that its parent other than x is
a descendant of y. But then, y is intermediate in the reticulation cycle of v, which
is impossible because it is already intermediate in the reticulation cycle of h. So,
we reach a contradiction that implies that l ∈
/ C(y), and hence that C(x) 6= C(y).
On the other hand, Lemma 9.(a) implies that, for every proper ancestor x of
h, C(h) = {i} ( C(x).
(c) Let x be an intermediate node of K and l a descendant leaf of x other than i. If l
were a non-strict descendant of x, then x would be intermediate in the reticulation
cycle of a hybrid ancestor of l by Lemma 9.(d), which is impossible because x is
already intermediate in K and h is not an ancestor of l. Thus, every descendant
leaf of x other than i is a strict descendant of x.
On the other hand, the fact that i is a non-strict descendant of x is obvious:
the composition of any path r u with the merge path u h not containing x,
and ending, if necessary, with the arc (h, i), yields a path r i not containing x.
(d) Let x and y be two intermediate nodes of different merge paths of K. If there
existed some leaf l 6= i in C(x) ∩ C(y), then it would be a strict descendant of both
x and y by (c), which would imply by Lemma 1 that x and y are connected by a
path, against (a).
(e) Any path r
i contains h and therefore it contains one of its parents. But
the merge path from u to any parent of h is a tree path, and hence it must be
23
contained in the subpath r
r i.
h of r
i. This implies that u belongs to the path
(f ) Let v be a node outside K such that i ∈ C(v). Then, by (e) and Lemma 1, u
and v are connected by a path. Now, since v 6= h, v will be an ancestor of one of
the parents of h, say x. But then, if v were a descendant of u, it would belong to
the only path u x, which is contained in K, against the assumption that v does
not belong to K. Thus, u is a descendant of v.
(g) Let x be an intermediate node of K and assume that there exists some w 6= x
such that C(w) = C(x). We know by (b) that w is neither h (because C(h) 6= C(x))
nor any intermediate node of K and therefore, by (f), C(u) ⊆ C(w) = C(x). Thus,
C(x) contains all clusters of nodes in K, which implies that the merge path not
containing x cannot contain any intermediate node (by (d)) and that x is the child
of u in the only merge path of K of length greater than 1 (otherwise, the cluster
of its parent in the merge path would strictly contain C(x), by (b), and would be
included in C(u)).
(h) Let v and v ′ be the parents of h. Since every proper ancestor w of h is an
ancestor of v or v ′ , and hence C(w) contains C(v) or C(v ′ ), we deduce that C(v)
and C(v ′ ) are the only possible minimal members of CI (N ) strictly containing
C(h).
Now, if C(v) and C(v ′ ) are two different such minimal members of CI (N ), then
they do not contain each other and therefore v and v ′ are not connected by a path.
This implies that neither v nor v ′ is the split node u of K, and therefore that they
are intermediate in K. Conversely, if only one of these two clusters, say C(v), is
minimal strictly containing C(h), then it is contained in the other. By (d), this
implies that v ′ cannot be intermediate in K, and therefore v ′ = u.
A2
Proof of Theorem 2
To ease the task of the reader, we split the proof of Theorem 2 into several lemmas. Throughout this subsection, N stands for a binary 1-nested network (or,
equivalently, a binary galled tree) on a fixed set S.
Lemma 11. The Ri;j reduction can be applied to N if, and only if, {i, j} ∈ CI (N )
but {i}, {j} ∈
/ CI (N ).
Proof. If N contains a node u whose children are the tree leaves i, j, then {i, j} =
C(u) ∈ CI (N ), and {i}, {j} ∈
/ CI (N ) by Lemma 9.(a).
Conversely, if {i, j} ∈ CI (N ) and {i}, {j} ∈
/ CI (N ), then i and j are tree
leaves (by the binarity of N ) and their parents have out-degree greater than 1
by Lemma 9.(a). Let now u be the parent of i. Since every internal ancestor of i
is an ancestor of its only parent u, the cluster of any internal ancestor of i must
contain the cluster of u: in particular, i ∈ C(u) ⊆ {i, j}, which implies (since
{i} ∈
/ CI (N )) that C(u) = {i, j}. But then, if i ∈ C(v) for some internal node v,
24
then {i, j} ⊆ C(v). This shows that every member of CI (N ) that contains i also
contains j. By symmetry, every member of CI (N ) that contains j also contains i.
Then, Lemma 9.(b) applies.
Lemma 12. The Gi;i1 ,...,ik ;∅ reduction can be applied to N if, and only if, the
following conditions are satisfied:
(1)
(2)
(3)
(4)
{i} ∈ CI (N ).
For every j = 2, . . . , k, {ij , . . . , ik , i} ∈ C(N ) with multiplicity 1.
{i1 , . . . , ik , i} ∈ C(N ) with multiplicity at least 2.
Any member of CI (N ) containing some label among i1 , . . . , ik and not listed in
(2)–(3), must contain {i1 , . . . , ik , i}.
Proof. If N contains a reticulation cycle K consisting of the merge paths (u, h)
and (u, v1 , . . . , vk , h) (and hence h and v1 are the only children of u), such that the
only child of the hybrid node h is the leaf i and the child outside K of each tree
node vj is the tree leaf ij , then
C(h) = {i}
C(vj ) = {ij , . . . , ik , i},
C(u) = {i1 , . . . , ik , i}
j = 1, . . . , k
and hence CI (N ) contains all clusters listed in (1)–(2), the latter with multiplicity
1 by Lemma 10.(g), as well as the cluster given in (3) with multiplicity at least
2. Now, let v be any internal node of N not belonging to K and such that C(v)
contains some label i1 , . . . , ik . If ij ∈ C(v), then ij ’s only parent vj must also be a
descendant of v. But then i ∈ C(vj ) ⊆ C(v) implies that C(u) ⊆ C(v) by Lemma
10.(f), as (4) claims.
Conversely, assume that (1)–(4) are satisfied. Then, the parent h of i has outdegree 1 and therefore it is hybrid, and, by Lemma 10.(h), its parents are connected
by a path, because there is only one minimal element of CI (N ) strictly containing
{i}, namely {ik , i}. Therefore, the reticulation cycle K for h consists of an arc
(u, h) and a tree path (u, v1 , . . . , vl , h) with l > 1. In this situation, Lemma 10
implies that:
– C(vl ) is the minimal element of CI (N ) strictly containing {i};
– C(vl ) ( C(vl−1 ) ( · · · ( C(v1 ), and then, by Lemma 9.(c), each C(vj ), j =
1, . . . , l − 1, is the minimal element of CI (N ) containing C(vj+1 );
– C(v2 ), . . . , C(vl ) appear with multiplicity 1 in CI (N );
– C(u) = C(v1 ), because the only children of u are v1 and h.
On the other hand, (1)–(4) imply that
– The minimal element of CI (N ) strictly containing {i} is {ik , i};
– {ik , i} ( {ik−1 , ik , i} ( . . . ( {i1 , . . . , ik , i}, and each {ij , . . . , ik , i}, j = 1, . . . , k−
1, is the minimal element of CI (N ) strictly containing {ij+1 , . . . , ik , i};
25
– {i2 , . . . , ik , i}, . . . , {ik , i} appear with multiplicity 1 in CI (N );
– {i1 , . . . , ik , i} appears with multiplicity at least 2 in CI (N ).
The only possibility of making these two lists of properties compatible is that k = l
and C(vj ) = {ij , . . . , ik , i} for every i = 1, . . . , k.
It remains to prove that the only child of every vj outside K is the corresponding
leaf ij . Let wj be the only parent of ij ; we want to prove that wj = vj . Since
ij ∈ C(vj ), there exists a path vj wj and hence C(wj ) ⊆ C(vj ). On the other
hand, ij ∈ C(wj ) implies, by (4), that i ∈ C(wj ) and therefore, by Lemma 10.(f),
either wj belongs to K or it is an ancestor of u. The second case cannot hold,
because wj is a proper descendant of u. Therefore, wj is a node of K that is a
descendant of vj and an ancestor of ij : it must be vj .
Lemma 13. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction, with k > k′ > 0, can be applied to N
k
if, and only if, the following conditions are satisfied:
(1)
(2)
(3)
(4)
(5)
{i} ∈ CI (N ).
For every j = 1, . . . , k, {ij , . . . , ik , i} ∈ C(N ) with multiplicity 1.
For every j = 1, . . . , k′ , {i′j , . . . , i′k′ , i} ∈ C(N ) with multiplicity 1.
{i1 , . . . , ik−1 , ik , i′1 , . . . , i′k′ −1 , i′k′ , i} ∈ C(N ).
Any member of CI (N ) containing some label among i1 , . . . , ik , i′1 , . . . , i′k′ and
not listed in (1)–(4), must contain {i1 , . . . , ik , i′1 , . . . , i′k′ , i}.
Proof. The proof that if N contains a reticulation cycle K consisting of the merge
paths (u, v1 , . . . , vk , h) and (u, v1′ , . . . , vk′ ′ , h), with k > k′ > 0, such that the only
child of the hybrid node h is the leaf i, the child outside K of each tree node vj
is the tree leaf ij , and the child of each tree node vj′ outside K is the tree leaf i′j ,
then it satisfies conditions (1) to (5), is similar to the proof of the corresponding
implication in the previous lemma, and we do not repeat it here.
As far as the converse implication goes, assume that conditions (1)–(5) in the
statement are satisfied. Then, the parent h of i has out-degree 1 and therefore it
is hybrid, and, by Lemma 10.(h), its two parents are not connected by a path,
because there are two minimal elements of CI (N ) strictly containing {i}, namely
{ik , i} and {i′k′ , i}. Therefore, the reticulation cycle K for h consists of two merge
paths (u, v1 , . . . , vl , h) and (u, v1′ , . . . , vl′′ , h) with l, l′ > 1. In this situation, Lemma
10 implies that:
– C(vl ) and C(vl′′ ) are the minimal elements of CI (N ) strictly containing {i};
– C(vl ) ( · · · ( C(v1 ), and then, by Lemma 9.(c), each C(vj ), j = 1, . . . , l − 1,
is the minimal element of CI (N ) containing C(vj+1 );
– C(vl′′ ) ( · · · ( C(v1′ ), and then, by Lemma 9.(c), each C(vj′ ), j = 1, . . . , l′ − 1,
′
is the minimal element of CI (N ) containing C(vj+1
);
′
′
– C(v1 ), . . . , C(vl ), C(v1 ), . . . , C(vl′ ) appear with multiplicity 1 in CI (N );
– the minimal element of CI (N ) strictly containing C(v1 ) is the same as the
minimal element of CI (N ) strictly containing C(v1′ ), and it is C(u).
26
On the other hand, (1)–(5) imply that:
– The minimal elements of CI (N ) strictly containing {i} are {ik , i} and {i′k′ , i};
– {ik , i} ( . . . ( {i1 , . . . , ik , i}, and each {ij , . . . , ik , i}, j = 1, . . . , k − 1, is the
minimal element of CI (N ) strictly containing {ij+1 , . . . , ik , i};
– {i′k , i} ( . . . ( {i′1 , . . . , i′k′ , i}, and each {i′j , . . . , i′k′ , i}, j = 1, . . . , k′ − 1, is the
minimal element of CI (N ) strictly containing {i′j+1 , . . . , i′k′ , i};
– {ik , i}, . . . , {i1 , . . . , ik , i}, {i′k , i}, . . . , {i′1 , . . . , i′k′ , i} appear with multiplicity 1 in
CI (N );
– the minimal element of CI (N ) strictly containing {i1 , . . . , ik , i} is the same as
the minimal element containing {i′1 , . . . , i′k′ , i}, and it is {i1 , . . . , ik , i′1 , . . . , i′k′ , i}.
The only possibility of making these two lists of properties compatible is that (up
to the interchange of k and k′ ) k = l, k′ = l′ , C(vj ) = {ij , . . . , ik , i} for every
j = 1, . . . , k, and C(vj′ ) = {i′j , . . . , i′k , i} for every j = 1, . . . , k′ .
It remains to prove that the only child of every vj (respectively vj′ ) not belonging
to the reticulation cycle for h is the corresponding leaf ij (respectively i′j ). This
fact can be proved using the same argument as in the last paragraph of the proof
of the previous lemma.
Lemmas 11 to 13 prove that the fact that a given R or G reduction can be
applied to N only depends on C(N ), from where point (A) in Theorem 2 follows.
As far as point (R) goes, it is a consequence of the following straightforward lemma
that shows that the application of a specific R or T reduction to N affects C(N )
in a way that does not depend on N itself, but only on its cluster representation;
we leave its easy proof to the reader.
Lemma 14. (a) If the Ri;j reduction can be applied to N , then C(Ri;j (N )) is
obtained by removing from C(N ) the clusters {i} and {j}, and them removing
from all remaining clusters the label j.
(b) If the Gi;i1 ,...,ik ;∅ reduction can be applied to N , then C(Gi;i1 ,...,ik ;∅ (N )) is obtained by first removing from C(N ) all clusters listed in points (1)–(2) of
Lemma 12, one copy of the cluster given in point (3) therein, and the clusters {i2 }, . . . , {ik }, and then removing the labels i2 , . . . , ik from all remaining
clusters.
(c) If the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (with k′ 6= 0) reduction can be applied to N , then
k
C(Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N )) is obtained by first removing from C(N ) all clusters
k
listed in points (1)–(3) of Lemma 13 and the clusters {i2 }, . . . , {ik }, {i′1 }, . . . , {i′k′ },
and then removing the labels i2 , . . . , ik , i′1 , . . . , i′k′ from all remaining clusters.
⊓
⊔
A3
Proof of Theorem 3
As in the previous subsection, we split the proof of Theorem 3 into several lemmas
to increase its readability. In the rest of this subsection, N stands for an arbitrary
27
1-nested network on some given set S. Since the set S is fixed, for every node v
of N , if A(v) = {i1 , . . . , ik } and B(v) = {j1 , . . . , jl }, we shall use the following
notation to denote the tripartition θ(v):
θ(v) = {i1 , . . . , ik | j1 , . . . , jl }.
To simplify the notations, we shall denote by θI (N ) the multiset of tripartitions
of its internal nodes, which is obtained by removing from θ(N ) one copy of every
tripartition {i | ∅} with i ∈ S.
Lemma 15. Two leaves i, j are tree leaves and siblings if, and only if, the following
conditions are satisfied:
(a) There exists an internal node v such that i, j ∈ A(v) and C(v) is contained in
the cluster of any internal ancestor of i or j.
(b) For every node w of N such that i, j ∈ C(w), it happens that either i, j ∈ A(w)
or i, j ∈ B(w).
Moreover, when i and j are sibling tree leaves, they are the only children of their
parent if, and only if, the node v in point (a) is such that C(v) = {i, j}.
Proof. If i, j are two sibling tree leaves and v is their common parent, then they
are strict descendants of v and C(v) is contained in the cluster of any ancestor of
i or j. Let now w be any node such that i, j ∈ C(w). Then, w is ancestor of v.
If v is a strict descendant of w, then i, j are also strict descendants of w, and if
v is a non-strict descendant of w, then i, j are also non-strict descendants of w.
Therefore, either i, j ∈ A(w) or i, j ∈ B(w). This finishes the proof of the ‘only if’
implication.
As far as the converse implication goes, the existence of the internal node v
with i, j ∈ A(v) and such that C(v) is contained in the cluster of every ancestor of i
or j implies that there does not exist any internal node whose cluster contains one
of the labels i, j but not the other, and therefore, by Lemma 9.(b), that i and j are
siblings. Let v0 be a common parent of them: then, on the one hand, i, j ∈ C(v0 )
implies that C(v) ⊆ C(v0 ), and, on the other hand, since i, j ∈ A(v), v0 must be
a descendant of v, and therefore C(v0 ) ⊆ C(v). We conclude that C(v0 ) = C(v).
Let us prove now that i, j ∈ A(v0 ). Indeed, if one of them were a non-strict
descendant of v0 , then by (b) both would be non-strict descendants of it. By Lemma
9.(d), and taking into account that v0 is a parent of i and j, this would imply that
i and j are hybrid leaves and v0 intermediate in their reticulation cycles, which
would contradict the 1-nested condition.
This implies that there would exist paths from the root of N to i and j that do
not contain v0 . This could only happen if both i and j were hybrid leaves and v0
intermediate in their reticulation cycles (if it were the split node of one of them,
the corresponding hybrid leaf would be a strict descendant of it by Lemma 10.(e)),
Let us prove now that i and j are tree leaves. Indeed, if, say, i is a hybrid leaf
28
and v0′ its other parent, then, since i ∈ A(v0 ), v0′ is a descendant of v0 and then
intermediate in the reticulation cycle for i (which would have v0 as split node).
Now, since i ∈ C(v0′ ), it must happen that j ∈ C(v0′ ) and, since v0′ cannot be an
ancestor of v0 , we conclude that j is also hybrid and that v0′ is an ancestor of its
other parent. But then, v0′ is also intermediate in the reticulation cycle for j (which
j), which is impossible.
consists of the arc (v0 , j) and the merge path v0 v0′
This shows that i and, by symmetry, j are tree leaves.
This finishes the proof that i and j are tree sibling leaves if, and only if, (a)
and (b) are satisfied; moreover, from this proof we deduce that we can take as
v in (a) the common parent of i and j. Now, as far as the last assertion in the
statement, if i and j are the only children of their common parent v, it is clear that
C(v) = {i, j}. Conversely, if v has a child u different from i and j, then u cannot
be an ancestor of i and j, and therefore any descendant leaf of it is an element of
C(v) different from i and j, which shows that {i, j} ( C(v).
As a direct consequence of this lemma we obtain the following two results.
Lemma 16. The Ri;j reduction can be applied to N if, and only if, the following
conditions are satisfied:
(1) There exists an internal node v such that θ(v) = {i, j | ∅} and C(v) is contained
in the cluster of any internal ancestor of i or j.
(2) For every node w of N such that i, j ∈ C(w), it happens either that i, j ∈ A(w)
or i, j ∈ B(w).
⊓
⊔
Lemma 17. The Ti;j reduction can be applied to N if, and only if, the following
conditions are satisfied:
(1) There exists an internal node v such that i, j ∈ A(v), {i, j} ( C(v), and C(v)
is contained in the cluster of any other internal ancestor of i or j.
(2) For every node w of N such that i, j ∈ C(w), it happens either that i, j ∈ A(w)
or i, j ∈ B(w).
⊓
⊔
Let us consider now the G and G reductions. In contrast to the corresponding
lemmas in §A2, here we do not need to distinguish between k′ = 0 and k′ > 0.
Lemma 18. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction (with k > k′ > 0) can be applied to N
k
if, and only if, the following conditions are satisfied:
(1)
(2)
(3)
(4)
{i | ∅} ∈ θI (N ).
For every j = 1, . . . , k, {ij , . . . , ik | i} ∈ θ(N ) with multiplicity 1.
For every j = 1, . . . , k′ , {i′j , . . . , i′k′ | i} ∈ θ(N ) with multiplicity 1.
For every θ(v) ∈ θI (N ), if C(v) contains some label among i1 , . . . , ik , i′1 , . . . , i′k′
and θ(v) is not listed in (2) or (3), then either i, i1 , . . . , ik , i′1 , . . . , i′k′ ∈ A(v) or
i, i1 , . . . , ik , i′1 , . . . , i′k′ ∈ B(v).
29
Proof. If N contains a reticulation cycle K consisting of the merge paths (u, v1 , . . . , vk , h)
and (u, v1′ , . . . , vk′ ′ , h) such that the only child of the hybrid node h is the tree leaf
i and each tree node vj (respectively vj′ ) has only one child outside K and it is the
tree leaf ij (respectively i′j ), then
θ(h) = {i | ∅}
θ(vj ) = {ij , . . . , ik | i}, j = 1, . . . , k
θ(vj′ ) = {i′j , . . . , i′k′ | i}, j = 1, . . . , k′
and hence θI (N ) contains all tripartitions listed in points (1)–(3). Let now v be any
internal node different from v1 , . . . , vk , v1′ , . . . , vk′ ′ that is an ancestor of some leaf
i1 , . . . , ik , i′1 , . . . , i′k′ , say that ij ∈ C(v). Then it will be an ancestor of its parent
vj and in particular i ∈ C(vj ) ⊆ C(v). Then, by Lemma 10.(f), u is a descendant
of v. Now, if u is a strict descendant of v, then the leaves i1 , . . . , ik , i′1 , . . . , i′k′ , i are
also strict descendants of v, while if u is a non-strict descendant of v, then they are
also non-strict descendants of v. This proves (4) and that all tripartitions listed in
(2) and (3) appear only once in θ(N ) (Lemma 10.(g) did not guarantee it for θ(v1 )
when k′ = 0). This finishes the proof of the ‘only if’ implication.
Conversely, assume that (1)–(4) are satisfied. Using only the information of
the clusters and arguing as in the proof of the ‘if’ implication in Lemmas 12 and
13, we already deduce that the parent h of i is hybrid, it has out-degree 1 and
i is a tree child of it (using Lemma 9.(a)), that the reticulation cycle K for h
consists of two tree merge paths (u, v1 , . . . , vl , h) and (u, v1′ , . . . , vl′′ , h), with l > k
and l′ > k′ , and that θ(vl−j ) = {ik−j , . . . , ik | i}, for every j = 0, . . . , k − 1, and
θ(vl′′ −j ) = {i′k′ −j , . . . , i′k′ | i} for every j = 0, . . . , k′ − 1. Now, if l > k, C(vl−k )
would strictly contain C(vl−(k−1) ) and i1 would be a strict descendant of vl−k
(because it is a strict descendant of its tree child vl−(k−1) ) but i would be a nonstrict descendant of it, which would contradict (4). This implies then that k = l
and C(vj ) = {ij , . . . , ik , i} for every i = 1, . . . , k, and then, by symmetry, k′ = l′
and C(vj′ ) = {i′j , . . . , i′k′ , i} for every i = 1, . . . , k′ .
It remains to prove that the only child of every vj (respectively, vj′ ) not belonging to the reticulation cycle for h is the corresponding tree leaf ij (respectively, i′j ).
Let us prove first that each ij is a tree leaf. Indeed, if ij were hybrid, then, since
ij ∈ A(vj ), vj would be an ancestor of the split node wj of the reticulation cycle
for ij . Since ij only belongs to the clusters of the nodes in K that are ancestors of
vj , we conclude that wj does not belong to K and then, since ij ∈ A(wj ), by (4)
we have that i ∈ A(wj ) and hence, by Lemma 10.(f), that wj is a strict ancestor
of u, which is impossible.
Let now uj be a child of vj different from ij and from vj ’s child in K. This node
must be internal, because the other descendant leaves ij+1 , . . . , ik , i of vj are tree
leaves and descendants of proper descendants of vj in K, and therefore vj is not
their parent. Then, since C(uj ) ⊆ C(vj ) = {ij , . . . , ik , i}, by (4) we conclude that
i ∈ C(uj ) and hence, since uj does not belong to K, by Lemma 10.(f) we conclude
that uj is an ancestor of u, which is impossible. This shows that vj does not have
30
any child outside K different from ij , and moreover, since ij is a descendant of vj ,
that it is its child.
Lemma 19. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction (with k > k′ > 0) can be applied to N
k
if, and only if, the following conditions are satisfied:
(1)
(2)
(3)
(4)
{i} ∈
/ CI (N ).
For every j = 1, . . . , k, {ij , . . . , ik | i} ∈ θ(N ) with multiplicity 1.
For every j = 1, . . . , k′ , {i′j , . . . , i′k′ | i} ∈ θ(N ) with multiplicity 1.
For every θ(v) ∈ θI (N ), if C(v) contains some label among i1 , . . . , ik , i′1 , . . . , i′k′
and θ(v) is not listed in (2) or (3), then either i, i1 , . . . , ik , i′1 , . . . , i′k′ ∈ A(v) or
i, i1 , . . . , ik , i′1 , . . . , i′k′ ∈ B(v).
Proof. If N contains a reticulation cycle K consisting of the merge paths (u, v1 , . . . , vk , h)
and (u, v1′ , . . . , vk′ ′ , h) such that the hybrid node h is the leaf i and each tree node vj
(respectively, vj′ ) has only one child outside K and it is the tree leaf ij (respectively
i′j ), then, by Lemma 9.(a), {i} ∈
/ CI (N ), and
θ(vj ) = {ij , . . . , ik | i}, j = 1, . . . , k
θ(vj′ ) = {i′j , . . . , i′k′ | i}, j = 1, . . . , k′
and hence N satisfies (1)–(3). The rest of the ‘only if’ implication can be proved
as in Lemma 18.
Conversely, assume that (1)–(4) are satisfied. To begin with, let us prove that
i is a hybrid leaf. Indeed, if it were a tree leaf, then its parent v would be a strict
ancestor of i, and therefore θ(v) would be none of the tripartitions listed in (2) or
(3). On the other hand, v would be a descendant of the node w having tripartition
{ik | i}, which would imply, since {i} =
6 C(v) by (1), that ik ∈ C(v). Then, by (4)
and since i ∈ A(v), ik would also be a strict descendant of v . This would imply
that w is a strict ancestor of v: any path r
v not containing w followed by a
path v ik (that does not contain w because w is an ancestor of v) would form
a path r ik not containing w, against the assumption that ik ∈ A(w). But then
the tree child i of v would be also a strict descendant of w, which would contradict
the assumption that i ∈ B(w).
Let us also denote by h this hybrid leaf labeled with i, so that C(h) = {i}.
Since we can still apply Lemma 10, the same argument as in the proof of the ‘if’
implication in Lemma 18 implies that the reticulation cycle K for h consists of two
merge paths (u, v1 , . . . , vk , h) and (u, v1′ , . . . , vk′ ′ , h) such that C(vj ) = {ij , . . . , ik , i},
for every j = 1, . . . , k, and C(vj′ ) = {i′j , . . . , i′k′ , i} for every j = 1, . . . , k′ .
The proof that the only child of every vj (respectively, vj′ ) not belonging to K
is the corresponding tree leaf ij (respectively, i′j ) is also similar to the one given
for the corresponding fact in Lemma 18, and we do not repeat it here.
Lemmas 16 to 19 prove condition (A) in Lemma 6 for the R, T , G, and G
reductions and the tripartitions representation. As far as point (R) goes, it is a
consequence of the following lemma.
31
Lemma 20. (a) If the Ri;j reduction can be applied to N , then θ(Ri;j (N )) is
obtained by removing the tripartitions {i | ∅} and {j | ∅} from θ(N ), and then
removing the label j from all remaining tripartitions in θ(N ).
(b) If the Ti;j reduction can be applied to N , then θ(Ti;j (N )) is obtained by removing the label j from all tripartitions in θ(N ).
(c) If the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction can be applied to N , then θ(Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N ))
k
k
is obtained by first removing from θ(N ) all tripartitions listed in points (1)–(3)
of Lemma 18 and the tripartitions {i2 | ∅}, . . . , {ik | ∅}, {i′1 | ∅}, . . . , {i′k′ | ∅},
and then removing the labels i2 , . . . , ik , i′1 , . . . , i′k′ from all remaining tripartitions.
(d) If the Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction can be applied to N , then θ(Gi;i1 ,...,ik ;i′1 ,...,i′ ′ (N )
k
k
is obtained by first removing from θ(N ) all tripartitions listed in point (2) of
Lemma 19 and the tripartitions {i2 | ∅}, . . . , {ik | ∅}, {i′1 | ∅}, . . . , {i′k′ | ∅}, and
then removing the labels i2 , . . . , ik , i′1 , . . . , i′k′ from all remaining tripartitions.
⊓
⊔
A4
Proof of Theorem 4
We also split the proof of Theorem 4 into several lemmas, in parallel to the preceding subsections. In the rest of this subsection, N stands for a semibinary 1-nested
network on S = {1, . . . , n}. Notice that all leaves in N are of tree type.
For every pair of nodes u, v in N , we shall denote by CA(u, v) the set of
common ancestors of u and v. By Lemma 8, [u, v] is the element of CA(u, v) that
is a descendant of all other nodes in this set.
The following result summarizes what Lem. 5 and Cor. 4 in [11] say about N .
Although these results were stated therein for tree-child time consistent evolutionary networks with out-degree 1 hybrid nodes, it is straightforward to check that
the time consistency is not used anywhere in their proofs, and therefore their thesis
also holds for tree-child (and, in particular, for 1-nested) semibinary hybridization
networks. In the following statement, and henceforth, by saying that a leaf j is a
quasi-sibling of a leaf i, we mean that the parent of j is a hybrid node that is a
sibling of i: cf. Fig. 11.
j
i
Fig. 11. j is a quasi-sibling of i.
Lemma 21. Let i, j be any labels in S.
(a) ℓN (i, j) = 1 if, and only if, the parent of i is an ancestor of j.
32
(b) The leaves i, j are siblings if, and only if, ℓN (i, j) = ℓN (j, i) = 1.
(c) The leaf j is a quasi-sibling of the leaf i if, and only if, ℓN (i, j) = 1, ℓN (j, i) = 2,
and ℓN (j, k) > 1 for every k ∈ S \ {i, j}.
⊓
⊔
As a consequence of this lemma, we have the following results.
Lemma 22. The Ri;j reduction can be applied to N if, and only if, ℓN (i, j) =
ℓN (j, i) = 1 and ℓN (i, k) > 1 for every k ∈ S \ {i, j}.
Proof. The Ri;j reduction can be applied to N if, and only if, i, j are sibling leaves
and their parent has out-degree 2. By the previous lemma we already know that
ℓN (i, j) = ℓN (j, i) = 1 if, and only if, i, j are sibling leaves. Thus, it only remains
to prove that the parent of i and j has out-degree 2 if, and only if, ℓN (i, k) > 1
for every k ∈ S \ {i, j}. Now, if there is a leaf k 6= j such that ℓN (i, k) = 1, then
the parent of i and j is also an ancestor of k, which means that it has out-degree
at least 3. Conversely, if the parent of i and j has out-degree at least 3 and v is a
child of it other than i, j, then ℓN (i, k) = 1 for every descendant leaf k of v.
A similar argument, using that the Ti;j reduction can be applied to N if, and
only if, i and j are tree sibling leaves and their parent has some other child, proves
the following result.
Lemma 23. The Ti;j reduction can be applied to N if, and only if, ℓN (i, j) =
ℓN (j, i) = 1 and there exists some k ∈ S \ {i, j} such that ℓN (i, k) = 1.
⊓
⊔
We have now the following lemmas for the G reductions.
Lemma 24. The Gi;i1 ,...,ik ;∅ reduction can be applied to N if, and only if, the
following conditions are satisfied:
(1)
(2)
(3)
(4)
(5)
(6)
ℓN (i, l) > 1 for every l ∈ S \ {i}.
ℓN (ik , i) = 1 and ℓN (i, ik ) = 2.
For every j = 1, . . . , k − 1, ℓN (ij , ij+1 ) = 1 and ℓN (ij+1 , ij ) = 2.
For every j = 1, . . . , k and for every l 6= ij , ..., ik , i, ℓN (ij , l) > 1.
For every j = 1, . . . , k, ℓN (i, ij ) = k − j + 2.
For every l ∈
/ {i, i1 , ..., ik }, ℓN (i, l) = ℓN (i1 , l) and ℓN (l, i) = ℓN (l, i1 ).
Proof. Assume that N contains a reticulation cycle K consisting of the merge
paths (u, v1 , . . . , vk , h) and (u, h) such that the only child of the hybrid node h is
the leaf i and each tree node vj has only one child outside K, and it is the tree leaf
ij . Then, (1) and (2) are satisfied because i is a quasi-sibling of ik , (3) is satisfied
because the parent vj+1 of each ij+1 is a sibling of ij , (4) is satisfied because the
only descendant leaves of the parent vj of ij are ij , ij+1 , . . . , ik , i, and (5) is satisfied
because [ij , i] = vj and the only path vj i has length k −j +2. As far as condition
(6) goes, let l be any label different from i, i1 , . . . , ik . Then, l is not a descendant
of v1 and therefore every common ancestor of i or i1 and l must be an ancestor of
33
u. This implies that CA(i, l) = CA(u, l) = CA(i1 , l), from where we deduce that
[i, l] = [u, l] = [i1 , l]. This clearly implies that ℓN (l, i) = ℓN (l, i1 ). On the other
hand, any shortest path [u, l] i will consist of a shortest path [u, l] u followed
by the path (u, h, i), and any shortest path [u, l] i1 will consist of a shortest path
[u, l] u followed by the path (u, v1 , i1 ), which implies that ℓN (i, l) = ℓN (i1 , l).
Conversely, assume that N satisfies conditions (1) to (6). Then, conditions (1)
and (2) imply that i is a quasi-sibling of ik : let h be the hybrid parent of i and let
vk be the parent of ik and h, which will be a tree node because it has out-degree
at least 2. Now, condition (3) implies that, for every j = 1, . . . , k − 1, the parent
of ij is also parent of the parent of ij+1 : if we let vj be the parent of ij , for every
j = 1, . . . , k − 1, we obtain a path (v1 , . . . , vk ) consisting of tree nodes (because
each node in it has out-degree at least 2) and such that each vj is the parent of
the leaf ij .
Now, vk may be either intermediate in the reticulation cycle K for h or the split
node of K (in which case one of the merge paths would be the arc (vk , h)). But, if
the latter happened, h would have another parent v and it would be a descendant
of vk , and then, any tree descendant leaf l of v would be such that ℓN (ik , l) = 1,
which would contradict (4). This implies that vk is intermediate in K.
Let now v be the other parent of h, and assume that it is intermediate in
the merge path of K not containing vk . Let l be a tree descendant leaf of v.
By Lemma 10.(d), l ∈
/ {i1 , . . . , ik }. Then, by (6), ℓN (i1 , l) = ℓN (i, l) = 2 and
ℓN (l, i) = ℓN (l, i1 ). But the latter condition implies that [l, i1 ] = [l, i] = v, and then
the former implies that v is the parent of v1 , which would imply that i1 ∈ C(v),
leading to a contradiction again by Lemma 10.(d). We conclude that the merge
path not containing vk is a single arc. In particular, this implies that no node
v1 , . . . , vk−1 is the split node of K: if vj were the split node of K, then ℓN (i, ij ) = 2,
against (5). So, the split node u of K is a proper ancestor of v1 . Let us see that
u is the parent of v1 . Indeed, if u were not the parent w of v1 , then w would be
intermediate in the merge path u v1 h: let w′ be a child of w outside K, and
let l be a tree descendant leaf of w′ . Then, since l ∈
/ {i, i1 , . . . , ik }, (6) would imply
that ℓN (i, l) = ℓN (i1 , l) = 2, while it is clear that ℓN (i, l) = k + 2 (because [l, i] = w
and the only path w i, along the merge path, has length k + 2).
In summary, we have proved so far that if N satisfies conditions (1) to (6),
then it contains a reticulation cycle for the hybrid parent h of i consisting of the
merge paths (u, v1 , . . . , vk , h) and (u, h), and that each vj is the parent of the tree
leaf ij . It remains to prove that v1 , ..., vk have out-degree 2. But, if some vj had
some child wj other than ij or its child in K, and if l were a tree descendant leaf
of wj , then l ∈
/ {i, ij , . . . , ik } but ℓN (ij , l) = 1, against (4).
Lemma 25. The Gi;i1 ,...,ik ;i′1 ,...,i′ ′ reduction (with k > k′ > 0) can be applied to N
k
if, and only if, the following conditions are satisfied:
(1) ℓN (i, l) > 1 for every l ∈ S \ {i}.
(2) ℓN (ik , i) = 1 and ℓN (i, ik ) = 2.
34
(2’)
(3)
(3’)
(4)
(4’)
(5)
ℓN (i′k′ , i) = 1 and ℓN (i, i′k′ ) = 2.
For every j = 1, . . . , k − 1, ℓN (ij , ij+1 ) = 1 and ℓN (ij+1 , ij ) = 2.
For every j = 1, . . . , k′ − 1, ℓN (i′j , i′j+1 ) = 1 and ℓN (i′j+1 , i′j ) = 2.
For every j = 1, . . . , k and for every l 6= ij , ..., ik , i, ℓN (ij , l) > 1.
For every j = 1, . . . , k′ and for every l 6= i′j , ..., i′k′ , i, ℓN (i′j , l) > 1.
ℓN (i1 , i′1 ) = ℓN (i′1 , i1 ) = 2.
Proof. Assume that N contains a reticulation cycle K consisting of the merge
paths (u, v1 , . . . , vk , h) and (u, v1′ , . . . , vk′ ′ , h), with k > k′ > 0, such that the only
child of the hybrid node h is the leaf i, each tree node vj has only one child outside
K, and it is the tree leaf ij , and each tree node vj′ has only one child outside K,
and it is the tree leaf i′j . The proof that it satisfies the conditions (1) to (4) and
(1’) to (4’) is similar to the corresponding proof in the previous lemma, and (5)
is a direct consequence of the fact that [i1 , i′1 ] = u (because v1 and v1′ are not
connected by a path by Lemma 10.(a)) .
Conversely, assume that N satisfies all conditions listed in the statement. Conditions (1), (2) and (2’) imply that i is a quasi-sibling of ik and i′k′ : let h be the
hybrid parent of i, and let vk and vk′ ′ be, respectively, the parents of ik and i′k′ .
As in the previous lemma, conditions (3) and (3’) imply the existence of paths
(v1 , . . . , vk ) and (v1′ , . . . , vk′ ′ ) consisting of tree nodes and such that each vj is the
parent of the leaf ij and each vj′ is the parent of the leaf i′j .
Now, no node v1 , . . . , vk , v1′ , . . . , vk′ is the split node of K: if, say, vj were the
split node of K, then in particular vk′ ′ , and hence i′k′ , would be a descendant of
vj , which would imply that vj = [ij , i′k′ ] and thus ℓN (ij , i′k′ ) = 1, against (4).
Therefore, the split node of K is a common ancestor of v1 and v1′ . Now, (6) implies
that [i1 , i′1 ] is simultaneously the parent of v1 and v1′ , and therefore that this parent
is the split node u of K.
Finally, the proof that the intermediate nodes of K have out-degree 2 is similar
to the proof of the corresponding fact in the previous lemma, using (4) and (4’).
Lemmas 22 to 25 imply that the possibility of applying a specific R, T or G
reduction to N depends only on ℓ(N ), from where condition (A) in Theorem 4
follows. As far as condition (R) goes, we have the following lemma.
Lemma 26. (a) If the Ri;j reduction can be applied to N , then, for every k, l ∈
S \ {j},
– ℓRi;j (N ) (i, k) = ℓN (i, k) − 1 if k 6= i.
– ℓRi;j (N ) (k, i) = ℓN (k, i) if k 6= i.
– ℓRi;j (N ) (k, l) = ℓN (k, l) if k, l 6= i.
(b) If the Ti;j reduction can be applied to N , then, for every k, l ∈ S \ {j},
ℓTi;j (N ) (k, l) = ℓN (k, l).
(c) If the Gi;i1 ,...,ik ,i′1 ,...,i′ ′ reduction (with k > k′ > 0) can be applied to N , then,
k
for every j, l ∈ S \ {i2 , . . . , ik , i′1 , . . . , i′k′ },
35
– ℓGi;i
(i, j) = ℓN (i1 , j) − 1 if j 6= i1 , i
– ℓGi;i
(j, i) = ℓN (j, i) if j 6= i1 , i
– ℓGi;i
(i1 , j) = ℓN (i1 , j) − 1 if j 6= i1 , i
– ℓGi;i
(j, i1 ) = ℓN (j, i1 ) if j 6= i1 , i
– ℓGi;i
(i, i1 ) = ℓN (i1 , i) = 1
– ℓGi;i
(j, l) = ℓN (j, l) if j, l 6= i1 , i
′ (N )
′
1 ,...,ik ,i1 ,...,ik′
′
′ (N )
1 ,...,ik ,i1 ,...,ik′
′
′ (N )
1 ,...,ik ,i1 ,...,ik′
′
′ (N )
1 ,...,ik ,i1 ,...,ik′
′
′ (N )
1 ,...,ik ,i1 ,...,ik′
′
′ (N )
1 ,...,ik ,i1 ,...,ik′
Proof. (a) Ri;j (N ) is obtained by removing the leaf j and replacing the leaf i by its
parent. This implies that, for every pair of remaining leaves, their LCA is the same
node in N and in Ri;j (N ), and that any path ending in i is shortened in one arc,
while all paths ending in any other remaining leaf are left untouched. The formulas
for ℓRi;j (N ) given in the statement follow immediately from these observations.
(b) Ti;j (N ) is obtained by removing the leaf j without modifying anything else.
This implies that, for every pair of remaining leaves, their LCA is the same node in
N and in Ti;j (N ) and no path ending in a remaining leaf is modified, and therefore
that ℓTi;j (N ) = ℓN on S \ {j}.
(c) Let us denote by N ′ the network Gi;i1 ,...,ik ,i′1 ,...,i′ ′ (N ), and let u be the split
k
node of the removed reticulation cycle. We remove all (and only) descendants of
u, and we add to u two new tree leaf children i and i1 . This implies that the LCA
in N ′ of i and i1 is u (and therefore ℓN ′ (i, i1 ) = ℓN ′ (i1 , i) = 1) and that the LCA
of any other pair of remaining leaves is the same node in N ′ as in N . On the other
hand, any path ending in i1 is shortened in one arc, the distance from any internal
node to i in N ′ is the same as its distance to i1 , and all paths ending in remaining
leaves other than i or i1 are not touched. From these observations, the formulas
for ℓN ′ given in the statement easily follow.
36
| 5cs.CE
|
Languages of Play
Towards semantic foundations for game interfaces
Chris Martens
Matthew A. Hammer
North Carolina State University
Raleigh, NC
[email protected]
arXiv:1703.05410v1 [cs.PL] 15 Mar 2017
ABSTRACT
Formal models of games help us account for and predict behavior,
leading to more robust and innovative designs. While the games
research community has proposed many formalisms for both the
“game half” (game models, game description languages) and the
“human half” (player modeling) of a game experience, little attention has been paid to the interface between the two, particularly
where it concerns the player expressing her intent toward the game.
We describe an analytical and computational toolbox based on programming language theory to examine the phenomenon sitting
between control schemes and game rules, which we identify as a
distinct player intent language for each game.
KEYWORDS
game interfaces, programming languages, formal methods
1
INTRODUCTION
To study how players interact with games, we examine both the
rules of the underlying system and the choices made by the player.
The field of player modeling has identified the value in constructing models of player cognition: while a game as a self-contained
entity can allow us to learn about its mechanics and properties
as a formal system, we cannot understand the dynamics of that
system unless we also account for the human half of the equation.
Meanwhile, Crawford [6] identifies the necessity of looking at the
complete information loop created between a player and a digital
game, defining interactivity in games as their ability to carry out
a conversation with a player, including listening, processing, and
responding, identifying the importance of all three to the overall
experience.
Given this understanding of games-as-conversation, we should
expect to discover something like a language through which games
and players converse. In Figure 1, we illustrate the game-player
loop as a process which includes an interface constituting such a language. The Game Ontology Project [22] describes game interfaces
as follows:
The interface is where the player and game meet,
the mapping between the embodied reactions of the
player and the manipulation of game entities. It
refers to both how the player interacts with the game
and how the game communicates to the player.
The first part, how the player interacts with the game, is called the input, which is further subdivided into input device and input method.
Input devices are hardware controllers (mice, keyboards, joysticks,
,
© 2017 Copyright held by the owner/author(s).
University of Colorado Boulder
Boulder, CO
[email protected]
etc.) and input methods start to brush the surface of something
more semantic: they include choices about locus of manipulation
(which game entities can the player control?) and direct versus
indirect action, such as selecting an action from a menu of options
(indirect) versus pressing an arrow key to move an avatar (direct).
However, any close look at interactive fiction, recent mobile
games, or rhythm games (just to name a few examples) will reveal
that design choices for input methods have much more variety and
possibility than these two dimensions. In this paper, we propose a
framework to support analyzing and exploring that design space.
Our first step is to refine input methods to input languages: we
are ultimately asking, how can a player communicate their intent,
and how does a digital game recognize this intent? So, in linguistic
terms, the “phonemes” of such a language are hardware controls
such as button presses and joystick movement. Then, the syntax
and semantics are defined by each game individually, depending
on what meaning they give to each control input. This language
defines the verbs of the game, which may include moving, selecting
inventory items, examining world items, applying or using items,
entering rooms, and combat actions. (Note that such a language is
also distinct from a game’s mechanics: mechanics include system
behavior which is out of the player’s control, such as falling with
gravity, non-player character actions, and other autonomous behavior.) This language is both afforded by the game designer—she must
communicate to players which verbs are available—and constrained
by her—she may declare certain expressions invalid.
Since the constraints on such a language are wholly determined
by a piece of software (the game interface), we argue that it has
more in common with a programming language1 (PL) than a natural
language. Accordingly, each game in some sense defines its own
programming language. In a slogan, we could term this project
games as programming languages. Specifically, we propose player
intent languages as a PL-inspired framework for designing playergame interfaces.
This analogy opens up a whole field of methodology to try applying to games. The PL research community has a long and deep
history of assigning mathematically formal semantics to languages
and analyzing those semantics. As games researchers become more
interested in the emergent consequences of the systems they assemble, the tools of PL theory have a lot to offer. For example, PL
theory provides an account of compositionality, i.e. how fragments
of expression fit together to form higher-level meanings. In games,
this translates into being able to understand player skills or strategies as compositions of player actions, which we demonstrate in
1 We
define programming languages broadly as formal languages whose meaning is
fully grounded in a computational system.
Chris Martens and Matthew A. Hammer
Figure 1: A process diagram of a game loop: player and game conversation as it relates to language, interface, and cognition.
this paper by using a formalized input language as a kind of “player
AI scripting language.”
Furthermore, by considering a player’s language of expression as
an object of study in its own right, we center them as a co-designer
of the experience afforded by a game. When we treat a player’s
interactions as not simply an arbitrary sequence of button presses
that advances and reveals the designer’s intent, but instead as its
own distinct voice that a game system must listen and respond to,
we enable the player to co-create with the system, potentially developing deeper systems understanding and emotional investment.
In this paper, we propose player intent languages, a programming
languages-based approach to designing player-game interfaces as
formal objects. In the remainder of the paper, we tour this approach
through concrete examples. Specifically, we consider a simple game
design space and make points in this space precise by introducing the components of a programming language: abstract syntax
(Section 4), type system (Section 6), and operational semantics (Section 5). For each, we give a corresponding concept in the game
world. By grounding these game concepts in analogous programming language concepts, we gain powerful PL reasoning tools and
design methodologies to benefit the game design process.
We demonstrate the payoff of this line of thought by extending
the metaphor with play traces as straight-line programs (Section 7),
and player skills as general programs (e.g., programs with parameters, branching and looping) (Section 8). These structures give
semantic logs and general strategies, respectively, for accomplishing a task in the game world. The framework of player intention
languages gives rise to further research directions, which we briefly
outline and discuss before concluding (Sections 9 and 10).
2
RELATED WORK
Cardona-Rivera and Young [4] detailed a conceptual framework
following the slogan games as conversation, grounding the communicative strategies of games in cognitive science for human-tohuman conversational understanding, such as Grice’s maxims [9].
They offer a linguistic and semiotic approach to understanding
how a game communicates affordances (possibilities for action) to
a player. For an account of the game’s half of the equation, which
includes the visual, textual, and audio feedback mechanisms intended to be processed by the player, this application of linguistics,
psychology, and design seems appropriate, much like the study of
cinematic language for film. On the other hand, we argue that a
PL approach better supports understanding of the player-to-game
direction, since the language the player speaks toward a digital
game is formal and unambiguous.
Researchers have previously recognized the value in formalizing
interaction vocabularies, realizing certain interaction conventions
as a single “video game description language” [7] whose implementation as VGDL [19] has been used in game AI research. We
suggest instead that the design space of player languages is as
varied as the design space of programming languages and herein
give an account of what it would mean to treat each language individually. Our project suggests that an appropriately expressive
computational framework analogous to VGDL should be one that
can accommodate the encoding of many such languages, such as
a meta-logical framework like the Twelf system for encoding and
analyzing programming language designs [18].
Any investigation into formalizing actions within an interactive
system shares ideas with “action languages” in AI extending as far
back as McCarthy’s situation calculus [13] and including planning
languages and process calculi. These systems have been studied in
the context of game design, e.g. the Ludocore system [20]; however,
AI researchers are mainly interested in these formalisms as internal
representations for intelligent systems and the extent to which
they support reasoning. Conversely, we are interested their potential to support player expression and facilitate human-computer
conversation.
Some theoretical and experimental investigations have been
carried out about differences between game interfaces along specific
axes, such as whether the interface is “integrated” (or one might say
Languages of Play
diagetic), versus extrinsic to the game world in the form of menus
and buttons [11, 12]. These investigations suggest an interest in
more detailed and formal ontologies of game interfaces, which our
work aims to provide.
From the PL research side, we note existing efforts to apply PL
methodology to user interfaces, specifically in the case of program
editors. Hazelnut is a formal model of a program editor that enforces
that every edit state is meaningful (it consists of a well-defined
syntax tree, with a well-defined type) [14]. Its type system and
editing semantics permit partial programs, which contain missing
pieces and well-marked type inconsistencies. Specifically, Hazelnut
proposes a editing language, which defines how a cursor moves and
edits the syntax tree; the planned benefits of this model range from
better editing assistance, the potential to better automate systematic
edits, and further context-aware assistance and automation based
on statistical analysis of (semantically-rich) corpora of recorded
past edits, which consist of traces from this language [15]. Likewise,
in the context of game design, we expect similar benefits from the
lens of language design.
3
A FRAMEWORK FOR
PLAYER INTENT LANGUAGES
In the formal study of a programming langugage, one may define
a language in three parts: syntax, type system, and operational
semantics.
• The syntax is written in the form of a (usually) context-free
grammar describing the allowable expressions. One sometimes distinguishes between concrete syntax, the literal program tokens that the programmer strings together in the
act of programming, and abstract syntax, the normalized
“syntax tree” structures that ultimately get interpreted.
• An operational semantics defines how runnable programs
(e.g. a function applied to an argument) reduce to values.
This part of the definition describes how actual computation takes place when programs in the language are run. It
is important to note that the operational semantics need
not reflect the actual implementation of the language, nor is
it specific to a “compiled” versus “interpreted” understanding of the language: it is simply a mathematical specification for how any compiler or interpreter for the language
should behave.
• A type system further refines the set of syntactically valid
expressions into a set of meaningful expressions, and provides a mapping between an expression and an approximation of its meaning. Type systems are usually designed
in conjunction with the operational semantics to have the
property that every expression assigned a meaning by the
type system should have a well-defined runtime behavior.
In practice, however, type systems can only approximate
this correspondence. Some err on the more permissive
side–e.g. C’s type system will permit invalid memory accesses with no language-defined behavior–and some err
on the more restrictive side, e.g. Haskell’s type system
does not permit any untracked side-effects, at the expense
of easily authoring e.g. file input/output (without first
learning the details of the type system).
Game concept
PL concept
Syntax
Recognized player intents (Section 4)
Operational semantics Game mechanics
(Section 5)
Type system
Contextual interface
(Section 6)
Straight-line programs Play traces
(Section 7)
General programs
Player skills
(Section 8)
Table 1: Player intent languages:
Formal decomposition (left) and correspondances (right).
Providing a formal language definition in programming languages research has several purposes. One is that it enables researchers to explore and prove formal properties of their language,
such as well-typed programs don’t go wrong, or in a language for
concurrency, a property like deadlock freedom. However, an even
more crucial advantage of a language specification is not mathematical rigor but human capacity to do science. A language definition
is a specification, similar to an application programmer interface
(API) or an IEEE standard: it describes an unambiguous interface
to the language along an abstraction boundary that other human
beings may access, understand, and implement, without knowing
the internals of a language implementation. It is a necessary component of reproducibility of research, and it allows researchers to
build on each other’s work. We believe that an embrace of formal specification in games research can play a similarly important
function.
Having provided loose definitions of these terms, we now wish
to draw out the analogy between a language specification and a
game specification. To treat a game in this manner, we wish to
consider player affordances and actions, as well as their behavior
(mechanics) in the context of the game’s running environment. We
summarize the components of this correspondence in Table 1.
We will use as a running example a minimal virtual environment
with two player actions: (1) movement through a discrete set of
rooms in a pre-defined map (move); (2) acquiring objects placed in
those rooms to store in a player inventory (take). We consider five
(somewhat aribitrary) possibilities in the design space of interfaces
for such a game, summarized visually in Figure 2:
• Point-and-Click: A first-person viewpoint interface where
the meaning of each click is defined based on the region the
cursor falls in. Clicking near any of the four screen edges
moves in that direction; clicking on a sprite representing
an item takes it.
• Bird’s-Eye: A top-down viewpoint interface where the
player can see multiple rooms at once, and can click on
rooms and objects that are far away, but those clicks only
do something to objects in the same room or adjacent
rooms.
• WASD+: A keyboard or controller-based interface with
directional buttons (e.g. arrow keys or WASD) move an
avatar in the correspondingi direction, and a separate key
or button expresses the take action, which takes any object
in the same room. (This interface may be used for either
of the two views described above.)
Chris Martens and Matthew A. Hammer
• Command-Line: The player interacts by typing free-form
text, which is then parsed into commands, such as take lamp
and move north.
• Hypertext: A choice-based interface where all available
options are enumerated as textual links from which the
player chooses.
In the following sections, we will consider these possibilities in
light of design choices relevant to the specified aspect of PL design.
4
PLAYER INTENT AS SYNTAX
The syntax of a game is its space of recognized player intentions.
Note that intention is different from action in the sense that we don’t
necessarily expect each well-formed intention to change anything
about the game state: a player can intend to move north, but if
there is no room to the north of the player when she expresses this
intent, no change to the game’s internal state will occur. Nonetheless, depending on the design goals of the game, we may wish to
recognize this as a valid intent so that the game may respond in
some useful way (e.g. with feedback that the player cannot move
in that direction).
In our example game, the choice of syntax answers questions
such as: can the player click anywhere, or only in regions that have
meaning? Can the player type arbitrary commands, or should we
provide a menu or auto-complete text so as to prevent the player
from entering meaningless commands? In PL, we can formalize
these decisions by describing an abstract syntax for our language,
which is typically assumed to be context-free and thus specified as
a Bachus-Naur Form (BNF) grammar. Our examples below follow
the interfaces shown visually in Figure 2.
WASD+ Interface: One way of writing the BNF for the WASD+
interface is:
syntax as follows (where direction and item are assumed to be
defined as they were in the previous examples):
intent
::=
movehdirectioni | takehitemi
Assuming the player “knows the language,” i.e. knows that move
and take are valid commands, and in fact the only valid commands,
and assuming that she knows how to map the visual affordances (e.g.
image of the flask) to the typed noun (e.g. flask), the experience
afforded by this interface is quite similar to the WASD+ interface.
The main difference is that the player must specify an argument to
the take command, asking the player to formulate a more complete
(and unambiguous) intent by actually naming the object she wishes
to take.
Hypertext interface: Finally, we consider the intent language
for the hypertext interface. This is one of the most difficult interfaces to formulate in linguistic terms, because it either requires
that we formalize link selection in an acontextual way (e.g. as a
numeric index into a list of options of unknown size) or that we formulate each link from each page as its own separate command, each
of which has meaning in only one specific game context (namely,
when the player is on the page containing that link). The former
feels like a more general formulation of hypertext that is not relevant to any particular game, and since we are aiming to provide a
correspondence between specific games and languages, we opt for
the latter:
intent
::=
selecthchoicei
choice
::=
take flask from lab
|
take book from library
|
go south from lab
|
go east from lab
direction
::=
north | south | east | west
|
go west from lab
intent
::=
movehdirectioni | collect
|
go north from library
|
...
The hardware interface maps onto this syntax quite directly: each
arrow key maps onto a move action in the corresponding direction,
and the specified other key maps onto collect.
Bird’s Eye View Mouse Interface: On the other hand, a clickingbased interface to a top-down map could enable the player to click
on any room on the map and any item within a room. This syntax
would look like:
room
::=
item
::=
flask | book
entity
::=
room | item
intent
::=
clickhentityi
courtyard | library | quarters | lab
Note that this syntax, compared to that for WASD+, describes a
larger set of possible utterances, even though it has the exact same
set of permitted game behaviors (a player may only move into
adjacent rooms and take items that share a room with them).
Command Line Interface: The command-line interface would
have an even larger space of expressible utterances if we consider
all typed strings of characters to be valid expressions, but that
syntax is too low-level for linguistic considerations. Supposing
we interpose a parsing layer between arbitrary typed strings and
syntactically-well-formed commands, we can define the abstract
Some hypertext authors put a lot of effort into scaffolding the choicebased experience with a richer language, e.g. by repeating the same
set of commands that behave in consistent ways across different
pages, or by creating menu-like interfaces where text cycles between options on an otherwise static page. In this way, hypertext
as a medium might be said as providing a platform for designers
to create their own interface conventions, rather than relying on
a set of pre-established ones; by the same token, hypertext games
created by inexperienced interface (or language) designers may feel
to players like being asked to speak a foreign language for each
new game.
Additive and subtractive properties of syntax
By now we are able to observe that, just like the rest of a game’s
rules, its syntax has both additive and subtractive properties. It
provides the menu of options for which hardware interactions are
relevant, i.e. likely to result in meaningful interaction with the game
system, but it also establishes which utterances within that set are
disallowed, or ill-formed—e.g. that it is not meaningful to say “take”
without providing an object to the command, or that “take north”
is ill-formed.
Languages of Play
Point-and-click
Bird’s eye view/WASD+
Command-line
Hypertext
Figure 2: Four different user interfaces for the move/take game.
Correspondingly, an important decision that impacts game design is (a) how discoverable the additive affordances are (e.g. can
the player determine that “examine” is a meaningful verb without
already possessing literacy in the game’s genre?) and (b) the extent to which the user interface makes meaningless expressions
impossible to form. For example, in a hypertext interface, all links
lead somewhere—so every intent the player can form, i.e. clicking a
link on the page, will get a valid response from the game, whereas
“take fnord” typed at a command-line interface may be recognized
by the parser, but meaningless to a game where “fnord” is not a
noun. Decisions about these two (related) dimensions will determine the extent to which learning the language, an exploratory but
sometimes frustrating process, is a central challenge of the game.
5
GAME MECHANICS AS
OPERATIONAL SEMANTICS
Consider the right-hand-side of Figure 1: The game parses and
interprets an unambiguous syntax of player intent, which either
advances the game state (as shown in the figure), or the game state
cannot advance as the player intended, which the game somehow
signals to the player (not shown).
We model a response to the move-take player intent with the
following BNF definition, of resp:
resp
::=
success | failure
Figure 1 shows the case where the player intent of “take flask”
(formally, the syntax “take”) leads to the game world performing
this intent as a successful action, and responding accordingly with
“Taken.” Formally, we model this resp as “success,” as defined above.
Likewise, if the flask cannot be taken (e.g., the flask is not near
the player, or is already in the player’s possession, etc.), the game
responds with failure.
As with the player’s intent, which may exist as both raw input
and as formal syntax, each formal response can be conveyed as
raw output in a variety of ways (e.g., as textual words, pictures, or
sounds). In real games there are often two two levels of game-toplayer feedback: Feedback through the game world, and feedback
outside the game world (e.g., a pop-up message with an error, guidance or advice). For simplicity, move-take gives feedback outside
of the game state, e.g., as pop-up messages.
To capture the formal relationship between player intent as
syntax, and game response as syntax, we introduce a four-place
game-step relation:
hG 1 ; intenti −→ hG 2 ; respi
This relation formalizes the dynamic behavior of the right-handside of Figure 1. It consists of four parts: An initial game state G 1 ,
a player intent intent, a resulting game state G 2 and a game response resp.
As is standard in PL formalisms, we give the rules that define
this relation as inductive inference rules, which can each be read
as a logical inference. That is, given evidence for the premises on
the top of the rule, we may conclude the bottom of the rule. For
instance, here are two example rules:
G 1 ` playerNear flask
playerTake(G 1 , flask) ≡ G 2
hG 1 ; takei −→ hG 2 ; successi
Chris Martens and Matthew A. Hammer
G ` not playerNear flask
hG; takei −→ hG; failurei
The first rule formalizes the case shown in the RHS of Figure 1.
The second rule formalizes the opposite outcome, where the flask
cannot be taken. Notably, the first rule has two premises: To be
taken by the player, it suffices to show that in the current game
state G 1 , the flask is near the player (first premise), and that there
exists an advanced game state G 2 that results from the player taking
this flask (second premise). In the second rule, there is only one
game state G, since the flask cannot be taken and consequently, the
game state does not change.
Like the syntax of player intent and game responses, these rules
are also unambiguous. Consequently, we view these rules as a
mathematical definition with an associated strategy for constructing formal (and informal) proofs about the game mechanics.
For instance, we can formally state and attempt to prove that
for all player intents intent and game worlds, G 1 , there exists a
corresponding game world G 2 and game response resp. That is, the
statement of the following conjecture:
∀G 1 , intent.
∃G 2 , resp.
hG 1 ; intenti −→ hG 2 ; respi
Using standard PL techniques, the proof of this conjecture gives rise
to an abstract algorithm that implements the game mechanics by
analyzing each possible case for the current game-state and player
intent. Indeed, this is precisely the reasoning required to show that
an implementation of the game is complete (i.e., there exists no state
and input that will lead the game into an undefined situation).
Reasoning about this completeness involves reasoning about
when each rule is applicable. For instance, the rule for a successful
player intent of take requires two premises: G 1 ` playerNear flask
and playerTake(G 1 , flask) ≡ G 2 . The first is a logical judgment about
the game world involving the proposition playerNear flask, which
may or may not be true, but which is computable. The second is a
semantic function that transforms a game state into one where the
player takes a given object; in general, this function may be undefined, e.g., if the arguments do not meet the function’s pre-conditions.
For instance, a precondition of the function playerTake(−, −) may
be that the object is not already in the possession of the player. In
this case, to show that the game mechanics do not “get stuck”, we
must show that playerNear flask implies that the player does not
already posses the flask. (Otherwise, we should add another rule to
handle the case that the player intends to take the flask but already
possesses it). The design process for programming language semantics often consists of trying to write examples and prove theorems,
and failing; these experiences inform systematic revisions to the
language definition.
In the next section, we refine intents and semantics further by
introducing the notion of a context.
6
CONTEXTUAL INTERFACES AS
TYPE SYSTEMS
Decisions about interface syntax can, to some extent, limit or expand the player’s ability to form intents that will be met with failure,
such as moving through a wall or taking an object that does not exist. But sometimes, whether an utterance is meaningful or not will
depend on the runtime game state, and can be considered a distinct
question from whether it is well-formed. For example, whether or
not we can take flask depends on whether the flask is present, but
if the flask is an object somewhere in the game, we must treat this
command as well-formed syntax and relegate its failure to integrate
with the runtime game environment to the mechanics (operational
semantics).
However, some user interfaces nonetheless restrict the set of
recognized utterances in a way that depends on current game state.
Consider a point-and-click interface that changes the shape of the
cursor to a hand whenever it hovers over an interactable object,
and only recognizes clicks when it is in this state. Alternatively,
consider the hypertext interface, which only recognizes clicks on
links made available in the current page. Providing the player with
only the option of saying those utterances that “make sense” in this
regard corresponds to a strong static type system for a programming
language.
Type systems are typically formalized be defining a relation between expressions e and contexts Γ. Contexts are sets of specific
circumstances in which an expression is valid, or well-typed. Usually, these circumstances have to do with variables in the program.
For example, the program expression x + 3 is only well-typed if x
is a number. “x is a number” is an example of a fact that would be
contained in the context. Its well-typedness could be represented
as x:num ` x + 3 ok.
In the move-take game, we can include aspects of game state in
our context, such as the location of the player and the adjacency
mapping between rooms in the world. An example of a typing rule
we might include to codify the “only present things are takeable”
rule would be:
Γ ` playerIn(R) Γ ` at(O, R)
Γ ` takehOi ok
We then need to define a relation between concrete game states
G and abstract conditions on those states, Γ. We might write this
relation G : Γ. After such rules are codified, we can refine the
“game completeness” conjecture to handle only those utterances
that are well-typed:
∀G 1 , intent.
=⇒ ∃G 2 , resp.
(G 1 : Γ) ∧ (Γ ` intent ok)
hG 1 ; intenti −→ hG 2 ; respi
This is nearly what we want to know about our game mechanics. However, we want to apply this reasoning iteratively as the
game progresses, so that we reason next about the player intention
that leads from game state G 2 to another possibly different game
state G 3 ; but what context for player intent describes state G 2 ?
For this reasoning to work, we generally need to update the original context Γ, possibly changing its assumptions, and creating Γ 0 .
We write Γ ⊆ Γ 0 to mean that Γ 0 succeeds Γ in a well-defined way.
Given that state G 1 and intent agree about the context of assumptions Γ, we wish to show that there exists a successor context Γ 0
that agrees with the new game state G 2 :
=⇒
∀G 1 , intent.
∃Γ 0, G 2 , resp.
(G 1 : Γ) ∧ (Γ ` intent ok)
hG 1 ; intenti −→ hG 2 ; respi
∧(G 2 : Γ 0 ) ∧ (Γ ⊆ Γ 0 )
This statement closely matches the usual statement of progress for
programming languages with sound type systems.
Languages of Play
7
PLAY TRACES AS
STRAIGHT-LINE PROGRAMS
If we consider the analogy of game interfaces as programming languages, the natural question arises, what is a program written in
this programming language? We want to at least consider individual, atomic player actions to be complete programs; the preceding
text provides such an account. But typical programs are more than
one line long—what does it mean to sequence multiple actions in a
game language?
In a typical account of an imperative programming language, we
introduce a sequencing operator ; where, if c 1 and c 2 are commands
in the language, then c 1 ; c 2 is also a command. The operational
semantics of such a command involves the composition of transformations on states σ :
hσ ; c 1 i −→ σ1 hσ1 ; c 2 i −→ σ2
hσ ; (c 1 ; c 2 )i −→ σ2
However, interactive software makes this account more complicated
by introducing the program response as a component. Instead of
issuing arbitrary commands in sequence, the player may wait for a
response or process responses in parallel with their decisions. In this
respect, a player’s “programming” activity more closely resembles
something like live-coding than traditional program authoring.
Execution of code happens alongside its authorship, interleaving
the two activities. If we consider a round-trip through the game
loop after each individual command issued, then what we arrive at
is a notion of program that resembles a play trace: a log of player
actions and game responses during a play session, e.g.
PLAYER: go north
GAME: failure
PLAYER: take flask
GAME: success
PLAYER: go south
GAME: success
Depending on the richness of our internal mechanics model, this
play trace may contain useful information about changes in internal
state related to the preconditions and effects of player actions. But
the main important thing to note is that, despite the informal syntax
used to present them here, these traces do not consist of strings of
text entered directly by the player or added as log information by the
game programmer—they are structured terms with abstract syntax
that may be treated to the same formal techniques of interpretation
and analysis as any program. And this syntax is at a high level of
game-relevant interactions, not at the level of hardware inputs and
engine code.
Researchers in academia and the games industry alike have recently been increasingly interested in play trace data for the sake
of analytics, such as understanding how their players are interacting with different components of the game and responding to this
information with updates that support player interest [8]. For the
most part, this trace data is collected through telemetry or other
indirect means, like game variable monitoring, after which it must
be analyzed for meaning [3]. More recently, systems of structured
trace terms that may be analyzed with logical queries have been proposed [16], identifying as a benefit an ability to support automated
Figure 3: A screenshot from BOTS, an educational game in
which players write programs to direct a player avatar.
testing at the level of design intents. Our PL analogy supports this
line of inquiry and warrants further comparison and collaboration.
8
PLAYER SKILLS AS
GENERAL PROGRAMS
While considering “straight-line” traces may have some utility in
player analytics, a more exciting prospect for formalizing game
interactions as program constructs is the possibility of encoding
parameterized sequences of actions that may carry out complex
tasks. After all, games with rich player action languages afford
modes of exploratory and creative play: consider item crafting in
Minecraft [17], puzzle solving in Zork [2], or creating sustainable
autonomous systems like a supply chain in Factorio [21], a farm in
Stardew Valley [1], or a transit system in Mini Metro [5]. Each of
these activities asks the player to understand a complex system and
construct multi-step sequences of actions to accomplish specific
tasks. From the player’s perspective, these plans are constructed
from higher-level activities, such as growing a crop or constructing
a new tool, which themselves are constructed from the lower level
game intent language.
A language, as we have formalized it, gives us the atomic pieces
from which we can construct these sequences, like Lego bricks
can be used to construct reconfigurable components of a house
or spaceship. Compositionality in language design is the principle
that we may understand the meaning and behavior of compound
structures (e.g. sequences) in terms of the meaning and behavior
of each of its pieces (e.g. actions), together with the meaning of
how they are combined (e.g. carried out one after the other, or in
parallel). In this section, we describe how we might make sense
of player skills in terms of programs written in a more complex
version of the player language.
Such programs might be integrated into a game’s mechanics
so that a player explicitly writes such programs, as in the BOTS
game, an interactive programming tutor that asks players to write
small imperative programs that direct an avatar within a virtual
Chris Martens and Matthew A. Hammer
syntax:
intent
Figure 4: A screenshot from Stardew Valley showing the
player’s farm, inventory, and avatar.
world [10] (see Figure 3), or Cube Composer2 , in which players
write functional programs to solve puzzles. However, for now, we
primarily intend this account of player skills as a conceptual tool.
8.1
Example: Stardew Valley
Our initial {move, take} example is too simple to craft really compelling examples of complex programs, so here we examine Stardew
Valley and its game language for the sake of considering player
skills. In Stardew Valley, the player has an inventory that permits
varied interactions with the world, beginning with a number of tools
for farming (axe, hoe, scythe, pickaxe) which do different things in
contact with the resources in the surrounding environment; most
include extracting some resource (wood, stone, fiber, and so on),
which themselves enter the player’s inventory and can be used
in further interaction with the game world. The player’s avatar is
shown on-screen, moved by WASD. There are also context-sensitive
interactions between the player and non-player characters (NPCs),
interfaces through which new items may be purchased (shops), and
mini-games including fishing (fish may also be sold at high value).
See Figure 4 for a typical player view of the game.
While a full account of the language that this game affords the
player is beyond the scope of this paper, we include a representative
sample of the actions and affordances found in this game that
may be used to construct player skills. These include: directional
avatar movement within and between world “rooms,” point-andclick actions for selecting items in one’s inventory, and interacting
with in-room entities.
The player avatar must be near an entity for the player to interact
with it. They can then either apply the currently-selected inventory
item to the in-world entity with a left-button click, or right-button
click, which does something based on the entity type, e.g. doors
and chests open, characters speak, and collectible items transfer to
the player’s inventory. We refer to this last action as inquire. We
also note that, for the sake of our example, movement towards an
entity and movement offscreen (towards another room) are the only
meaningful and distinct types of movement, which we refer to as
move near and move offscreen. These actions yield the following
2 http://david-peter.de/cube-composer/
::=
selecthitemi
|
applyhentityi
|
inquirehentityi
|
move nearhentityi
|
move offscreenhdirectioni
We leave the definition of items and entities abstract, but we could
imagine it to simply list all possible items and entities in the world
as terminal symbols. From these atomic inputs, we can start to construct higher-level actions performed in the game most frequently—
tilling land, planting seeds, conversing with NPCs, and so on. These
blocks of code may be assigned names like functions to be called in
many contexts:
action till
action
action
action
action
= select hoe; move_near hard_ground;
apply hard_ground
plant = select seeds; move_near tilled_ground;
apply tilled_ground
mine = select pickaxe; move_near rock; apply rock
talk = move_near npc; inquire npc
enter_shop = move_near shop; inquire door(shop)
In turn, these larger skill molecules may be combined to accomplish
specific tasks or complete quests.
8.2
Branching programs as skills and strategies
Note that we have naively sequenced actions without consideration
for the game’s response. This approach to describing player skills
does not take into account the possbility of a failed attempt, such as
attempting to mine when there are no rocks on the current screen.
One could simply assign a semantics to these sequences of action
that threads failure through the program—if we fail on any action,
the whole compound action fails.
However, we can go further with describing robust player skills
and strategies if we consider the possibility of handling failure, a
common feature of day-to-day programming and indeed of gameplay. Recall our simplified game response language consisting of
two possibilities, success and failure. We can introduce a case construct into our language to handle each of these possibilities as a
distinct branch of the program:
action mine =
response = select pickaxe;
case(response):
success => {
response' = move_near rock;
case(response'):
success => apply rock;
| failure => fail;
}
| failure => fail;
However, to avoid handling failure at every possible action, a better
approach is to explicitly indicate as parameters the world resources
that each action needs in order to complete successfully. The overall action definition for mining, or example, would require as a
precondition to the action that a pickaxe is available in the player
Languages of Play
Figure 6: Growing a crop.
Figure 5: Watering a crop until it may be harvested.
try_harvest(g)
inventory and a rock is in the same room. The actions for selecting the pickaxe and moving near the rock would depend on these
resources, and the game response language could include the resources it guarantees as outputs. Then we can write the program
using simpler notation that refers to resource dependencies of the
appropriate type (using notation resource : type):
action mine(p:pickaxe, r:rock) =
select p; move_near r; apply r
: mineral
This notation together with a branching case construct scales
to include nondeterminism in the game world, such as the fishing
minigame in Stardew Valley: the game always eventually tells the
player that something is tugging at her line, but some portion of
the time it is a fish while the rest of the time it is trash. These
constructs can also account for incomplete player mental models,
such as knowing that one must water her seeds repeatedly day after
day in order to grow a crop, but not knowing how many times.
Below, we present a notation that accounts for these aspects of
player skills: a do ... recv ... notation indicates a command and
then binds the response to a pattern, or structured set of variables,
which can then be case-analyzed. Our first example is watering a
crop until it may be harvested:
action water_until_harvestable[t](p: planted(t))
: crop(t) =
do try_harvest(p)
recv <result: crop(t) + growing(t)>.
case result of
c:crop(t) => c
| g:growing(t) =>
water(g);
wait(day);
See Figure 5 for a control flow diagram of this code.
The next example shows a parallel construct ||, which can be
used to compose actions with distinct dependencies, as well as how
an action definition may use other action definitions by threading
resource dependencies through as arguments:
fun grow_crop[t : croptype](s:soil, w:watering_can)
: crop(t) =
do
get_seeds(t) || till_soil(s)
recv <s: seeds(t), g: tilled_soil>.
do
plant(s, g)
recv <p: planted(t)>.water_until_harvestable(p)
See Figure 6 for a control flow diagram of this code.
9
DISCUSSION
Having established a vocabulary of syntax and semantics for player
languages, we can now revisit the potential benefits of this account
mentioned in the introduction and discuss them in more detail.
9.1
Composing play traces, player skills
One of the major things that a programming language account
provides is compositionality: a system for making sense of meaning
of a complex artifact in terms of the meaning of its pieces. This
comes up in two places for looking at games:
Structured play traces. With a formalized game language, the
sequence of steps along the transition system described by the
operational semantics forms a mathematical artifact that is subject
to deeper analysis than what can be gained simply from screen
or input device recordings. For example, we can carry out causal
analysis, asking “why” queries of trace data, e.g. “Why was the
Chris Martens and Matthew A. Hammer
player able to unlock the door before defeating the boss?”, as well as
filtering traces for desired properties: “Show me a play trace where
the player used something other than the torch to light the room.”
The recent PlaySpecs project [16] suggests interest in formulating
traces this way to support this kind of query.
Player skills as programs. While a play trace may be interpreted as a straight-line program, even more interesting is the idea
of latent structure in player actions, such as composing multiple
low-level game actions into a higher-level skill, following the cognitive idea of “chunking.” We map this idea onto that of functions
in the programming language that take arguments, generalizing
over state space possibilities (e.g.: the red key opened the red door,
so for all colors C, a C key will open a C door). Further reasoning
forms like case analysis to handle unpredictable game behavior
and repeating an action until a condition holds are also naturally
expressed as programming language constructs.
and concurrent systems that may benefit from the protocol-based
approach proposed here.
10
CONCLUSION
We propose player intent languages, a systematic framework for
applying programming language design principles to the design of
player-game interfaces. We define this framework through simple
examples of syntax (aka player intents), type systems (aka contextual interfaces) and operational semantics (aka game mechanics).
We show how applying this framework to a player-game interface
naturally gives rise to formal notions of play traces (as straight-line
programs) and player skills (as general programs with branching
and recursion). By defining a player intent language, game design concepts become formal objects of study, allowing existing PL
methodology to inform and guide the design process.
REFERENCES
9.2
Abstraction boundaries
between input and mechanics
Formalizing game interfaces gives us the tools to explore alternative
interfaces to the same underlying mechanics, without needing to
port game logic between different graphical interface frameworks.
For example, the interactive fiction community has been exploring
alternatives to the traditional dichotomy of “parser vs. hypertext”
for presenting text-based games and interactive story-worlds. An
abstraction boundary between the underlying mechanics, map, and
narrative of the world, and the view and input mechanisms used to
interact with it, could open the doors to research on user interfaces
that support players’ mental models of a world conveyed in text.
9.3
Enabling co-creative play
Finally, a PL formulation of player actions along with appropriate composition operators (parallel and sequential composition,
branching, and passing resource dependencies) provides a “scripting language for free” to the game environment. Such a language
can be used to test the game, provided as a game mechanic as in
BOTS, or provided as an optional augmentation to the game’s mechanics for the sake of modding or adding new content to the game
world. Especially in networked game environments, like multi-user
domains, massively multiplayer online games, and social spaces like
Second Life, the ability for the player to program not just her avatar
but parts of the game world itself introduces new opportunities
for creative and collaborative play. Our framework suggests a new
approach to designing these affordances for players in a way that is
naturally derived from the game’s existing mechanics and interface.
In the spirit of celebrating the player, this year’s conference theme,
we wish to enable the player as a co-designer of her own game
experience.
9.4
Future Work
In future work, we intend to build software for realizing game
language designs and experimenting with protocol-based game
AI developed as programs in these languages. In another direction, we aim to innovate in PL design outside of games, such as
read-eval-print loops (REPLs) for live programming that includes
rapid feedback loops motivated by gameplay, as well as distributed
[1] Eric Barone. 2016. Stardew Valley. Digital distribution. (2016).
[2] Marc Blank and David Lebling. 1980. Zork I: The Great Underground Empire.
(1980).
[3] Alessandro Canossa. 2013. Meaning in Gameplay: Filtering Variables, Defining
Metrics, Extracting Features and Creating Models for Gameplay Analysis. Springer
London, London, 255–283. DOI:http://dx.doi.org/10.1007/978-1-4471-4769-5 13
[4] Rogelio E Cardona-Rivera and R Michael Young. 2014. Games as conversation. In
Proceedings of the 3rd Workshop on Games and NLP at the 10th AAAI Conference
on Artificial Intelligence and Interactive Digital Entertainment. 2–8.
[5] Dinosaur Polo Club. 2015. Mini Metro. (2015).
[6] Chris Crawford. 2003. Chris Crawford on game design. New Riders.
[7] Marc Ebner, John Levine, Simon M Lucas, Tom Schaul, Tommy Thompson, and
Julian Togelius. 2013. Towards a video game description language. (2013).
[8] M Seif El-Nasr, Anders Drachen, and Alessandro Canossa. 2013. Game analytics.
New York, Sprint (2013).
[9] H Paul Grice. 1975. Logic and conversation. 1975 (1975), 41–58.
[10] Andrew Hicks. 2012. Creation, evaluation, and presentation of user-generated
content in community game-based tutors. In Proceedings of the International
Conference on the Foundations of Digital Games. ACM, 276–278.
[11] Kristine Jørgensen. 2013. Gameworld interfaces. MIT Press.
[12] Stein C Llanos and Kristine Jørgensen. 2011. Do players prefer integrated user
interfaces? A qualitative study of game UI design issues.
[13] John McCarthy and Patrick J Hayes. 1969. Some philosophical problems from
the standpoint of artificial intelligence. Readings in artificial intelligence (1969),
431–450.
[14] Cyrus Omar, Ian Voysey, Michael Hilton, Jonathan Aldrich, and Matthew A.
Hammer. 2017. Hazelnut: A Bidirectionally Typed Structure Editor Calculus. In
POPL.
[15] Cyrus Omar, Ian Voysey, Michael Hilton, Joshua Sunshine, Claire Le Goues,
Jonathan Aldrich, and Matthew A. Hammer. 2017. Toward Semantic Foundations
for Program Editors. In SNAPL: The 2nd Summit on Advances in Programming
Languages.
[16] J Osborn, Ben Samuel, Michael Mateas, and Noah Wardrip-Fruin. 2015. Playspecs:
Regular expressions for game play traces. Proceedings of AIIDE (2015).
[17] Markus Persson, Jens Bergensten, Michael Kirkbride, Mark Darin, and Carl
Muckenhoupt. 2011. Minecraft. (2011).
[18] Frank Pfenning and Carsten Schürmann. 1999. System description: Twelf–a
meta-logical framework for deductive systems. In International Conference on
Automated Deduction. Springer, 202–206.
[19] Tom Schaul. 2013. A video game description language for model-based or
interactive learning. In Computational Intelligence in Games (CIG), 2013 IEEE
Conference on. IEEE, 1–8.
[20] Adam M Smith, Mark J Nelson, and Michael Mateas. 2010. Ludocore: A logical
game engine for modeling videogames. In Computational Intelligence and Games
(CIG), 2010 IEEE Symposium on. IEEE, 91–98.
[21] Wube Software. 2012. Factorio. (2012).
[22] José P Zagal, Michael Mateas, Clara Fernández-Vara, Brian Hochhalter, and
Nolan Lichti. 2007. Towards an ontological language for game analysis. Worlds
in Play: International Perspectives on Digital Games Research 21 (2007), 21.
| 6cs.PL
|
Stability of Valuations and Kollár Components
Chi Li and Chenyang Xu
April 13, 2017
arXiv:1604.05398v4 [math.AG] 12 Apr 2017
Abstract
We show that among all Kollár components obtained by plt blow ups of a klt singularity
o ∈ (X, D), there is at most one that is (log-)K-semistable. We achieve this by showing that
if such a Kollár component exists, it uniquely minimizes the normalized volume function
introduced in [Li15a] among all divisorial valuations. Conversely, we show any divisorial
minimizer of the normalized volume function yields such a K-semistable Kollár component.
We also prove that for any klt singularity, the infimum of the normalized function is always
approximated by the normalized volumes of Kollár components.
Contents
1 Introduction
1.1 Kollár components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Equivariant K-semistability . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Preliminary
2.1 Normalized volume . . . . . . .
2.2 Properties of Kollár component
2.3 Deformation to cone . . . . . .
2.4 Filtrations and valuations . . .
1
3
4
5
.
.
.
.
5
5
9
10
11
3 Volume of models
3.1 Local volume of models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Approximating by Kollár components . . . . . . . . . . . . . . . . . . . . . .
12
12
14
4 K-semistability implies minimum
4.1 Degeneration to initial ideals . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Equivariant K-semistability and minimizing . . . . . . . . . . . . . . . . . . .
4.3 Proof of Theorem A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
19
22
25
5 Uniqueness
5.1 Case of cone singularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 The general case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
26
30
6 Minimizing Kollár component is K-semistable
32
7 Examples
35
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Introduction
Throughout this paper, we work over the field C of complex numbers. It has been well known
by people working in higher dimensional geometry that there is an analogue between the local
object Kawamata log terminal (klt) singularities (X, o) and log Fano varieties as the global
counterpart (cf. e.g. [Sho00, Xu14] etc.). From this comparison, since the stability theory
of Fano varieties has been a central object in people’s study in the last three decades, it is
natural to expect there is a local stability theory on singularities. The primary goal of this
1
preprint is to develope such a theory. In another word, we want to investigate singularities
using the tools from the theory of K-stability, a notion which was first defined in [Tia97] and
later algebraically formulated in [Don02]. We note that this interaction between birational
geometry and K-stability theory has been proved to significantly fertilize both subjects (cf.
[Oda12, Oda13, LX14, WX14, LWX14, Fuj15] etc.).
For the stability theory of log Fano varieties, a crucial ingredient is the CM weight.
Philosophically, the stability of log Fano varieties is equivalent to minimizing the CM weight.
In the stability theory of singularities, we fix the singularity (X, o) and look for ‘the most
stable’ valuations v. Thus the first step of establishing a local stability theory for (X, o) would
be to find the right counterpart of CM weight in the local setting. In the work [Li15a], the
c X,o on the space of valuations
first named author defined the normalized volume function vol
centered at o, which we believe should be the right invariant. In fact, its derivative at the
canonical divisorial valuation over the cone singularity along certain tangent direction gives
the the CM weight. So in the local picture, the normalized volume function indeed carries
more information than the CM weight!
By the above discussion, inspired by the global theory, we focus on studying the valuation
minimizing the normalized volume function, which is conjectured to uniquely exist and ought
to be thought as the ‘(semi-)stable’ object. This picture is understood well in the case of
Sasakian geometry where one only considers the valuations coming from the Reeb vector
fields on the torus (e.g. [MSY08, CS15]). Here we can naturally define the stability of the
singularity using the one for the base. However, this requires the extra cone structure. By
investigating the minimizer of the normalized volume function on all valuations, our plan, as
we mentioned, built on the previous work ([Li15a, Li15b, LL16]), is to establish an intrinsic
stability which only depends on the isomorphic class of the singularity. To revisit the cone
singularity case, in fact it was shown in [Li15b, LL16] that a Fano manifold X is K-semistable,
if and only if that among all valuations over the vertex o of the cone C(X) given by a multiple
of −KX , the canonical valuation defined to be the one obtained by blowing up the vertex
o ∈ C(X) minimizes the normalized volume function. This gives evidence to justify that at
least for such singularities, our study is on the right direction.
For an arbitrary klt singularity, there is no direct way to associate a global object. Nevertheless, in differential geometry, when there is a ‘canonical’ metric, the metric tangent
cone around the singularity is the stable object in the category of metric space. With a
similar philosophy, we expect the minimizer of the normalized volume function always gives
a degeneration to a K-semistable Fano cone singularity in the Sasakian setting, and such
c X,o . In the current paper, we work
degeneration should be provided by the minimizer of vol
out this picture in the case that the minimizer is divisorial, by implementing the machinery of the minimal model program (based on the foundational results in [BCHM10]). So our
treatment will be purely algebraic though it is strongly inspired by profound analytic results.
One ingredient we introduce is to define the volume associated to a birational model
and then connect it to the normalized volume of a valuation. For studying the divisorial
valuations, the class of models which play a central role here are the ones obtained by the
construction of Kollár component (cf. [Xu14]): for an arbitrary n-dimensional klt singularity
(X, o), we can use minimal model program to construct a birational model whose exceptional
locus is an (n−1)-dimensional Q-Fano variety. We will systematically develop the tools of using Kollár components to understand the normalized local volume and its minimizer. In fact,
Kollár components can be considered as the local analogue of special degenerations studied
in [LX14]. In the case of Sasakian geometry, one Reeb vector gives a Kollár component if
and only if it is rational, i.e., it is quasi-regular.
Therefore, to summarize, the aim of this paper is of twofolds: on one hand, we aim
at using the construction of Kollár components to get information of the valuations space,
especially the minimizer of the normalized volume functions; On the other hand, in the
reverse direction, we want to use the viewpoint of stability to study the birational geometry
construction of Kollár components, and search out a more canonical object under suitable
assumptions.
We also expect for any klt singularity (X, o), even when the minimizer is not necessarily
divisorial, we can still use suitable birational models to degenerate (X, o) to a K-semistable
(possibly irregular) Sasakian singularity. However, it seems to involve a significant amount
of new technical issues.
In the below, we will give more details.
2
1.1
Kollár components
Definition 1.1 (Kollár component). Let o ∈ (X, D) be a klt singularity. We call a proper
birational morphism µ : Y → X provides a Kollár component E, if µ is isomorphic over
X \ {o}, and µ−1 (o) is an irreducible divisor S, such that (Y, S + µ−1
∗ D) is pure log terminal
and −S is Q-Cartier and ample over X.
We easily see the birational model Y is uniquely determined once the divisorial valuation
S is fixed, and if we denote
(KY + S + µ−1
∗ D)|S = KS + ∆S ,
then (S, ∆S ) is a klt log Fano pair.
Given any klt singularity (X, o), after the necessary minimal model program type result
is established (see [BCHM10]), we know that there always exists a Kollár component (see
[Pro00] or [Xu14, Lemma 1]), but it is often not unique (nevertheless, see the discussion
in 7.1.4 for some known special cases for the uniqueness). From what we have discussed,
instead of an arbitrary Kollár component, we want to study those which are ‘more stable’,
and show if it exists, it provides a more canonical object. Indeed, we prove that if there is
c
a K-semistable Kollár component, then it gives the unique minimizer of vol(X,
o) among all
Kollár components.
Compared to the global theory of degeneration of Fano varieties, this fits into the philosophy that K-stability provides a more canonical degeneration(cf. [LWX14, SSY16]) and it
should minimize the CM weight among all degenerations. But one surprising thing to us is
that K-semistability is enough, instead of K-polystability which was required in the global
setting.
The following theorem is our main theorem.
Theorem 1.2. Let o ∈ (X, D) be klt singularity. A divisorial valuation ordS is a minimizer
c X,D,o if and only if the following conditions are satisfied
of vol
1. S is a Kollár component;
2. (S, ∆S ) is K-semistable.
Moreover, such a minimizing divisorial valuation is unique.
More precisely, we will prove Theorem 1.2 by proving the following four theorems. For
each of them, we need somewhat different techniques.
First we prove
Theorem A. If o ∈ (X, D) is an algebraic klt singularity. Let S be a Kollár component
c X,o is minimized at the valuation ordS .
over X. If (S, ∆S ) is (log-)K-semistable. Then vol
This extends the main theorem in [LL16] from cone singularities to the more general setting. For the proof, we indeed degenerate a general singularity to a cone singularity induced
by its Kollár components. However, instead of degenerating the valuation, we degenerate
the associated ideals and then use the result in [Liu16]. An extra subtlety is to treat the case
of the cone singularity, we can not directly use [LL16] as there they only proved the result
for the cone singularity over an analytic K-semistable Q-Fano variety, which we still do not
know to be equivalent to the algebraic definition. As a redemption, we first show that it
suffices to concentrate on the equivariant data and then use [Li15b] to finish the argument.
In Section 7, we use this criterion to find minimizers for various examples of singularities
including: quotient singularities, Ak and Ek singularities etc.
Next, we turn to the result on the uniqueness.
Theorem B. If o ∈ (X, D) is an algebraic klt singularity. Assume S is a Kollár component
over X such that (S, ∆S ) is K-semistable. Then
c X,o (ordS ) < vol
c X,o (ordT )
vol
for any other divisorial valuation T .
3
This is done by a detailed study of the geometry when the equality holds. In the cone
singularity case, we investigate the equality condition in the calculation in [Li15b]. It posts
a strong assumption which enables us to compute the corresponding invariants including nef
thresholds and pseudo-effective thresholds. The argument is partially inspired by the work
in [Liu16]. Once this is clear, the rest follows from a simple application of Kawamata’s base
point free theorem. And the general case can be again reduced to the case of cone singularity
using a degeneration process.
Now we consider the converse direction. For any klt singularity, a minimizer of the
normalized volume function always exists by [Blu16b]. The following theorem says that if a
minimizer is divisorial, it always yields a Kollár component. We indeed prove slightly more
for a general rational rank 1 minimizer.
Theorem C. Given an arbitrary algebraic klt singularity o ∈ (X, D) where X = Spec(R).
c X,o . Assume the valuation group of v is isomorphic
Let v be a valuation that minimizes vol
to Z, i.e., v has rational rank one, and one of the following two assumptions holds
1. v is a multiple of a divisorial valuation; or
2. the graded family of valuative ideals
a• = {ak } where ak = {f ∈ R | v(f ) ≥ k}
is finitely generated, i.e., there exists m ∈ N such that amk = (am )k for any k ∈ N.
Then up to a rescaling, v is given by the divisorial valuation induced by a Kollár component
S.
The fact that the assumption 1 above implies v is given by a Kollár component is also
independently proved in [Blu16b]. We note that a minimizer is conjectured to be quasimonomial ((cf. [Li15a, Conjecture 6.1.3])) and the graded family of valuative ideals for a
minimizer of the normalized volume function is conjectured to be always finitely generated
(cf. [Li15a, Conjecture 6.1.5]). So granted any one of these two conjectures, this result
should presumably characterize all the cases with minimizers of rational rank 1. The proof
uses the definition of the volumes of a model, and run the decreasing process of the volumes
given by the minimal model program as in [LX14].
Next we turn to the stability of the minimizer. By using the techniques from the toric
degeneration (see [Cal02, AB04, And13]) and the relation between CM weight and normalized
volumes, we will prove
Theorem D. We use the same notation as in Theorem C. Let µ : Y → X be the morphism
which extracts S, and write (KY + S + µ−1
∗ D)|S = KS + ∆S , then (S, ∆S ) is a K-semistable
log Fano pair.
1.2
Approximation
In a different direction, we also obtain results which describe the minimizer of the normalized volume function from the viewpoint of Kollár components. We show for a general klt
singularity, although the minimizer of its associated normalized volume function might not
be given by a Kollár component, but we can always approximate it by a sequence of Kollár
components.
Theorem 1.3. Given an arbitrary algebraic klt singularity o ∈ (X, D), and a minimizer v m
c X,o , there always exists a sequence of Kollár components {Si } and positive numbers cj
of vol
such that
lim cj · ordSj → v in ValX,o
j→∞
and
c
c m
lim vol(ord
Si ) = vol(v ).
i→∞
Here ValX,o consists of all valuations centered at o, and is endowed with the weakest
topology as in [JM12, Section 4.1].
4
1.3
Equivariant K-semistability
By relating a Fano variety and the cone over it, we can compare the calculation in [Li15b]
for a cone and [Fuj16] for its base. Then an interesting by product of our method is the
following theorem.
Theorem 1.4. Let T ∼
= (C∗ )r be a torus. Let (V, ∆) be a log Fano variety with a T -action.
Then (V, ∆) is K-semistable if and only if any T -equivariant special test configuration S → A1
of (V, ∆) has its generalized Futaki invariant Fut(V) ≥ 0.
When S is smooth and ∆ = 0, this follows from the work of [DS15] with an analytic
argument. Our proof is completely algebraic. It again uses the techniques of degenerating
any ideal to an equivariant one and showing it has a smaller invariant.
The paper is organized in the following way: In Section 2, we give some necessary backgrounds. In Section 3, we introduce one key new tool: the volume of a model. By combining
the normalized volume function on valuations with the local volume defined in [Ful13], and
applying minimal model program, we prove Theorem 1.3 and Theorem C. In Section 4, we
prove Theorem A, by connecting it to the infimum of lct(X, D; a)n · mult(a) for all m-primary
ideals a centered on o. We note that this latter invariant indeed has also been studied in
other context (cf. [dFEM04]). In Section 5, we prove Theorem B. We first prove it for
the cone singularity case, using heavily the ideas and calculations in [LL16]. Then we use
a degeneration argument to reduce the general case to the case of cone singularities. In
Section 6, we prove Theorem D, which verifies the K-semistability of a minimizing Kollár
component. In Section 7, we give some examples on how to apply our techniques to calculate
the minimizer for various classes of klt singularities.
History: Since [Li15a], the study of the minimizer of the normalized volume function moves
forward rapidly. Our preprint is inspired by the earlier work of [Li15a, Li15b] and [LL16].
It also uses ideas from [Liu16], which is written in the same period as the first version of
our preprint. After we posted our preprint, the existence of the minimizer is completely
settled in [Blu16b]. Inspired by this work, in the revision, we improve our work by showing
that for a C∗ -equivariant singularity, we only need to consider the equivariant valuations for
minimizing the normalized local volume. We also include his result on the existence of the
minimizer in the exposition. In particular, this allows us to stay with the algebraic definition
of K-semistability. Another major improvement we achieve in the revision is that we can
indeed show that any Kollár component which minimizes the normalized local volume is
always K-semistable.
Acknowledgement: We thank Yuchen Liu, Dhruv Ranganathan and Xiaowei Wang for
helpful discussions and many useful suggestions. We especially want to thank Harold Blum
and Mircea Mustaţǎ for pointing out a gap in an earlier draft. CL is partially supported by
NSF DMS-1405936. CX is partially supported by ‘The National Science Fund for Distinguished Young Scholars (11425101)’. Part of the work was done when CX visited Imperial
College London and Massachusetts Institute of Technology. He wants to thank Paolo Cascini
and Davesh Maulik for the invitation and providing a wonderful environment.
2
Preliminary
Notation and Conventiones: We follow the standard notation in [Laz04a, KM98, Kol13].
A log Fano pair (X, D) is a projective klt pair such that −KX − D is ample.
For a local ring (R, m) and a an m-primary ideal, we denote by lR (R/a) the length of
R/a.
For the K-semistability of a log Fano pair, see [Tia97, Don02] (also see [Oda13, LX14]).
2.1
Normalized volume
Let (X, o) be a normal algebraic singularity and v ∈ ValX,o , which is the space of all valuations centered on o. Let D ≥ 0 be a Q-divisor such that KX + D is Q-Cartier. We can
define the volume volX,o (v) and the log discrepancy AX,D (v) (if the context is clear, we will
5
abbreviate it as vol(v) and A(v)) following [ELS03] and [JM12] (see e.g. [Li15a, Section 1.1]).
In particular, if S is a divisor with center on X to be o, we have
AX,D (S) := AX,D (ordS ) = a(S; X, D) + 1
the same as the standard log discrepancy.
c X,D,o (v) (or
Definition 2.1. Notation as above. We denote the normalized volume by vol
c X,o (v) if D is clear or vol
c X,D (v) if o is clear or simply vol(v)
c
vol
if there is no confusion) to
be
volX,o (v) · An (v)
if An (v) < +∞; and +∞ if An (v) = +∞.
c
Recall that it was conjectured that vol(v)
achieves the minimum at a valuation v ∈ ValX,o .
In [Li15a], it is showed that the space
c
{v ∈ ValX,o | v(m) = 1, vol(v)
≤ C}
for any constant C > 0 forms a compact set. However, in general the volume function vol is
c
only upper continuous on ValX,o although we expect it is continuous at a minimizer of vol.
Proposition 2.2. If {vi } is a sequence of valuations, such that vi → v in the weakest
topology, then
vol(v) ≥ lim sup vol(vi ).
i
Proof. The valuation v determines a sequence of graded ideas
ak = ak (v) = {f ∈ R | v(f ) ≥ k}.
By [Mus02], we know that for any > 0, there exists a sufficiently large k such that
1
mult(ak ) < vol(v) + .
kn
Since R is Noetherian, we know that there exist finitely many generators fp (1 ≤ p ≤ j) of
ak = (f1 , ..., fj ). As v(fp ) ≥ k, we know that for any δ, there exists sufficiently large i0 such
that for any i ≥ i0 , vi (fp ) ≥ k − δ. Thus
(i)
ak−δ = {f ∈ R | vi (f ) ≥ k − } ⊃ ak .
Therefore,
vol(vi ) ≤
1
1
kn
(i)
mult(ak− ) ≤
mult(ak ) ≤
(vol(v) + ).
n
n
(k − δ)
(k − δ)
(k − δ)n
Proposition 2.3. Let (X, o) = (Spec(R), m) be a singularity. Let v and v 0 be two real
valuations in ValX,o . Assume
vol(v) = vol(v 0 ) > 0
v(f ) ≥ v 0 (f )
and
for any f ∈ R∗ , then v = v 0 .
Proof. We prove it by contradiction. Assume this is not true, we fix g ∈ R such that
v(g) = l > v 0 (g) = s.
Denote by r = l − s > 0. Fix k ∈ R>0 . Consider
ak := {f ∈ R| v(f ) ≥ k}
and
bk := {f ∈ R| v 0 (f ) ≥ k}.
So bk ⊂ ak , and we want to estimate the dimension of
dim(R/bk ) − dim(R/ak ) = dim(ak /bk ).
6
Fix a positive integer m <
k
l
and a set
(1)
(km )
gm
, ..., gm
∈ bk−ml
whose images in bk−ml /bk−ml+r form a C-linear basis.
We claim that
k
(j)
{f m · gm
} (1 ≤ m ≤ , 1 ≤ j ≤ km )
l
are C-linear independent in ak /bk . Granted this for now, we know since vol(v) > 0, then
1
n
lim k→∞ k
lim sup
X
X
km = lim sup
lim k→∞
1≤m≤ kl
1≤m≤ kl
1
dim(bk−ml /bk−ml+r ) > 0,
kn
which then implies vol(v) > vol(v 0 ).
Now we prove the claim.
Step 1: For any 1 ≤ m ≤ kl , 1 ≤ j ≤ km ,
(j)
v(f m · gm
)
(j)
v(f m ) + v(gm
)
=
(j)
≥ ml + v 0 (gm
)
≥ ml + k − ml
≥ k.
(j)
Thus f m · gm ∈ ak .
Step 2: If
k
, 1 ≤ j ≤ km )
l
are not C-linear independent in ak /bk , then there is an equality
X
hm = b ∈ bk ,
(j)
{f m · gm
} (1 ≤ m ≤
m
where there exists cj ∈ C, such that
X
hm = f m ·
(j)
cj gm
1≤j≤km
and some hm 6= 0. Consider the maximal m, such that hm 6= 0. Since
X
(j)
v 0 (hm ) = v 0 (f m ·
cj gm
)
1≤j≤km
0
= v (f ) + v 0 (
m
X
(j)
cj gm
)
1≤j≤km
<
ms + k − ml + r
= k − (m − 1)l + (m − 1)s,
where the third inequality follows from that
X
(j)
cj gm
∈
/ bk−ml+r .
1≤j≤km
However, we have
v 0 (hm )
= v 0 (b −
X
hj )
j<m
≥ min v 0 (b), v 0 (h1 ), ....v 0 (hm−1 )
=
min
{k, js + k − jl}
1≤j≤m−1
= k − (m − 1)l + (m − 1)s,
which is a contradiction.
7
Several results in our work depend on a relation between the normalized volume and
some normalized multiplicity. The latter quantity was first considered in the smooth case
in [dFEM04], and since then it has been studied in many other work, including positive
characteristic versions (see e.g. [TW04]). Its relevance to the normalized volume appeared
in [Li15a, Example 3.7]. We have the following more precise observation in [Liu16].
Proposition 2.4 ([Liu16, Section 4.1], cf. [Li15a, Example 3.7]). Let (X, o) = (SpecR, m)
and D ≥ 0 a Q-divisor, such that o ∈ (X, D) is a klt singularity. Then we have
c X,o (v) = inf lctn (X, D; a) · mult(a),
inf vol
v
a
(1)
where on the left hand side the infimum runs over all the valuations centered at o, and on
the right hand side it runs over all the m-primary ideals. Moreover, the left hand side can be
c X,o (v) where DivX,o denotes the space of all divisorial valuations
replaced by inf v∈DivX,o vol
with center at o.
For the reader’s convenience we provide a sketch of the proof.
Proof. We first use the same argument as in [Li15a, Example 2.7]) to prove that the left
hand side is greater or equal to the right hand side. For any real valuation v, consider the
graded family of valuative ideals
ak = ak (v) = {f ∈ R | v(f ) ≥ k}.
Then v(ak ) ≥ k and we can estimate:
n
AX,D (v)
mult(ak )
≥
· mult(ak ) ≥ lctn (X, D; ak ) · mult(ak ).
AX,D (v)n ·
kn
v(ak )
Since for a• = {ak } is a graded family of m-primary ideals on X,
mult(ak )
lR (R/ak )
= lim
.
n
k→+∞
k→+∞
k
kn
mult(a• ) = lim
c
(see e.g. [ELS03, Mus02, LM09, Cut12]), as k → +∞, the left hand side converges to vol(v)
and we get one direction.
For the other direction of the inequality, we follow the argument in [Liu16]. For any mprimary ideal a, we can choose a divisorial valuation v calculating lct(a). Then v is centered
at o. Assume v(a) = k, or equivalently a ⊆ ak (v). Then we have al ⊆ ak (v)l ⊆ akl (v) for
any l ∈ Z>0 . So we can estimate:
lctn (X, D; a) · mult(a)
=
=
n
AX,D (v)n
n mult(a)l
·
mult(a)
=
A
(v)
·
X,D
kn
(kl)n
mult(al )
mult(akl )
AX,D (v)n ·
≥ AX,D (v)n ·
.
(kl)n
(kl)n
As l → +∞, then again the right hand side converges to
c
AX,D (v)n · mult(a• (v)) = vol(v).
So we have proved the right-hand-side is bigger than the left-hand-side in (1).
The last statement follows easily from the above proof.
In a very recent preprint [Blu16b], it is proved that a minimizer always exists.
c X,D (v) always has a minimizer
Theorem 2.5 (Blum). For any klt singularity o ∈ (X, D), vol
v m in ValX,o .
8
2.2
Properties of Kollár component
The concept of Kollár component is defined in Definition 1.1. It always exists (cf. see
[Pro00] or [Xu14, Lemma 1]). In this section, we establish some of their properties using the
machinery of the minimal model program.
The following statement is the local analogue of [LX14, Theorem 1.6], which can be also
easily obtained by following the proof of the existence of Kollár component. (See e.g. the
proof of [Xu14].)
Proposition 2.6. Let o ∈ (X, D) be a klt singularlty. Let µ : Y → X be a model, such that
µ is an isomorphism over X \ {o} and (Y, E + µ−1
∗ D) is dlt where E is the divisorial part of
µ−1 (o). Then we can choose a model W → Y and run MMP to obtain W 99K Y 0 , such that
Y 0 → X gives a Kollár component S with a(Y, E; S) = −1.
We also have the following straightforward lemma.
c
Lemma 2.7. If S is a Kollár component, then vol(ordS ) = (−S|S )n−1 and vol(ord
S) =
n−1
n
(−(KY + S + µ−1
D)|
)
·
A
(S)
.
S
X,D
∗
Proof. For any k ≥ 0, we have an exact sequence,
0 → OY (−(k + 1)S) → OY (−kS) → OS (−kS) → 0.
Because −S is ample over X, we have the vanishing
R1 f∗ (OY (−(k + 1)S)) = 0,
from which we get
H 0 (S, −kS|S ) ∼
=
ak (ordS )
H 0 (Y, −kS)
=
.
+ 1)S)
ak+1 (ordS )
H 0 (Y, −(k
So we get the identity:
dimC (OX,o /am (ordS ))
=
m−1
X
dim H 0 (S, −kS|S ).
k=0
Then the result follows easily from the Hirzebruch-Riemann-Roch formula.
As KY + S + µ∗ D ∼Q,X AX,D (S) · S, the second identity is implied by the first statement.
Remark 2.8. Inspired by the above simple calculation, we indeed can extend the definition
of normalized volumes to any model f : Y → (X, o), such that f is isomorphic over X \ {o}.
See Section 3.
Lemma 2.9. Let f : (X 0 , o0 ) → (X, o) be a finite morphism, such that f ∗ (KX + D) =
KX 0 + D0 for some effective Q-divisors. We assume (X, D) and (X 0 , D0 ) are klt. If S is a
Kollár component given by Y → X over o, then Y ×X X 0 → X 0 induces a Kollár component
S 0 over o0 ∈ (X 0 , D0 ).
Conversely, if X 0 → X is Galois with Galois group G, then any G-invariant Kollár
component S 0 over o ∈ (X 0 , D0 ) is the pull back from a Kollár component over o ∈ (X, D).
0
0
Proof. The first part is standard. In fact, denote by S 0 = f −1 (S), then (Y 0 , µ0−1
∗ D + S ) is
0
log canonical, such that if we restrict to T a component of S ,
0
0
(KY 0 + µ0−1
∗ D + S )|T = KT + ∆T ,
then (T, ∆T ) is klt, which by Kollár-Shokurov connectedness theorem implies that T = S 0 .
For the converse, let
0
0
L ∼X 0 −m(KY 0 + µ0−1
∗ D +E )
1
be a divisor of general position for sufficiently P
divisible m and H := m
L, then (Y 0 , E 0 +
1
0−1 0
∗
µ∗ D + H) is plt. Replacing H by HG := |G| ( g∈G g H), we know that (X 0 , D0 + µ∗ HG )
is G-invariant, and there exists a Q-divisor HX ≥ 0, such that
f ∗ (KX + D + HX ) = KX 0 + D0 + µ∗ HG .
9
Therefore, (X, D + HX ) is plt, and its unique log canonical place is a divisor S which is
a Kollár component over o ∈ (X, D) whose pull back gives the Kollár component S 0 over
o0 ∈ (X 0 , D0 ).
We prove a change of volume formula for Kollár components under a finite map.
Lemma 2.10. With the same notation as in Lemma 2.9, then
c X,D (ordS ) = vol
c X 0 ,D0 (ordS 0 ),
d · vol
where d is the degree of X 0 → X.
Proof. Since the pull back of S is S 0 which is irreducible by Lemma 2.9, let the degree of
S 0 → S be a and the ramified degree be r, we have
ar = d
and
rAX,D (ordS ) = AX 0 ,D0 (ordS 0 )
(see [KM98, 5.20]). By Lemma 2.7, we know that
c X,D (ordS )
d · vol
=
n−1
ar · AX,D (ordS ) · ((KY + S + µ−1
∗ D)|S )
=
n−1
(rAX,D (ordS )) · (a(KY + S + µ−1
∗ D)|S )
=
0
n−1
AX 0 ,D0 (ordS 0 ) · ((KY 0 + S 0 + µ0−1
,
∗ D )|S 0 )
=
c X 0 ,D0 (ordS 0 ),
vol
where for the third equality we use the projection formula of intersection numbers.
2.3
Deformation to cone
Let (X, o) = (Spec(R), m) be an algebraic singularity such that (X, D) is klt for a Q-divisor
D ≥ 0. Let S be a Kollár component and ∆ = ∆S be the different divisor defined by
adjunction (KY + S + µ−1
∗ D)|S = KS + ∆S where Y → X is the extraction of S. Denote
v0 := ordS and
+∞
+∞
M
M
T =
ak (v0 )/ak+1 (v0 ) =
Tk
k=0
k=0
and the d-th truncation
T (d) =
+∞
M
adk (v0 )/adk+1 (v0 ) =
k=0
+∞
M
Tdk
for d ∈ N.
k=0
Now we give a more geometric description of Spec(T ) and Spec(T (d) ) using the idea
of degenerating o ∈ (X, D) to an (orbifold) cone over the Kollár component S. Assume
µ : Y → X is the extraction of the Kollár component S of (X, o). Then µA1 : Y ×A1 → X ×A1
has the exceptional divisor S × A1 . The divisor S is not necessarily Cartier, but only QCartier. Thus we can take the index 1 covering Deligne-Mumford stack π : Y → Y for S. So
π is isomorphic over Y \ S and π ∗ (S) = S is Cartier on Y.
We consider the deformation to the normal cone construction for S ⊂ Y. More precisely,
we consider the blow up π1 : Z → Y × A1 along S × {0}. Denote by E the exceptional
divisor of π1 and by T the strict transform of S × A1 . We note that E has a stacky structure
along the 0 and ∞ section, but a scheme structure at other places. Then T ⊂ Z is a
Cartier divisor which is proper over A1 and can be contracted to a normal Deligne-Mumford
stack Ψ1 : Z → W and in this way we get a flat family W → A1 such that Wt ∼
= X and
W0 ∼
= C¯ ∪ Y0 , where Y0 is the birational transform of Y × {0}. To understand C if we denote
by Z 0 := Z \ Y0 , then the fiber Z 0 over 0 is isomorphic to C which is an affine orbifold cone
over S with the polarization given by OY (−S)|S . Moreover, C¯ is the projective orbifold cone
completing C.
Let d be a positive integer such that d · S is Cartier in Y , then C (d) given by the cone
over OY (−d · S)|S is a degree d cyclic quotient of C, which is a usual (A1 -)cone over S. We
denote by C and C (d) the underlying coarse moduli space of C and C (d) . We also denote S
to be the coarse moduli space of S.
10
Applying the exact sequence,
0 → OY (−(k + 1)S) → OY (−kS) → OS (−kS) → 0,
since h1 (OY (−(k + 1)S)) = 0 by the Grauert-Riemenschneider vanishing theorem, we get:
H 0 (S, O(−kS|S )) ∼
= H 0 (OY (−kS))/H 0 (OY (−(k + 1)S)).
Notice that the right hand side is equal to:
ak (v0 )
µ∗ OY (−kS)
=
.
µ∗ OY (−(k + 1)S)
ak+1 (v0 )
In particular, C = Spec(T ). Similarly, C (d) = Spec(T (d) ).
There is also a degree d cyclic quotient morphism h : C → C (d) , and we know that
(d)
(d)
h∗ (KC̄ (d) + C∆ + CD ) = KC̄ + CD ,
(d)
where CD is the intersection of C̄ with the birational transform of D × A1 and C∆ (resp.
(d)
CD ) on C̄ (d) is the induced cone over the branched Q-divisor on S of S → S (resp. µ−1
∗ D|S ).
(d)
(d)
In particular, C∆ + CD is the cone over ∆S .
2.4
Filtrations and valuations
Assume F is a Z-graded filtration on R. We have the Rees algebra and extended Rees
algebra:
+∞
+∞
M
M
(F k R)t−k ,
(2)
(F k R)t−k , R0 := R0 (F) =
R := R(F) =
k=−∞
k=0
and the associated graded ring:
grF (R) = R0 /tR0 =
+∞
M
(F k R/F k+1 R)t−k =:
k=0
+∞
M
Ak .
(3)
k=0
Assuming R0 is finitely generated, X := SpecC[t] (R0 ) can be seen as a C∗ -equivariant flat
degeneration of X into X0 = SpecC (R0 /tR0 ) = SpecC (grF R). Denote E = Proj(grF (R)),
X̃ = ProjR R. Then the natural map X̃ → X is the filtered blow up associated with the F
such that E is the exceptional divisor. Moreover X̃ can be seen as a flat deformation of a
natural filtered blow up on X0 . Indeed following [TW89, 5.15], we have a filtration F on R0 :
( +∞
)
X
m 0
max(k,m)
−k
F R =
F
R t
.
k=−∞
The objects associated to the corresponding Rees algebra and graded algebra over R0
are:
+∞
+∞
M
M
r 0
−r
X̃ = ProjR0
(F R )T , E = ProjC
(F r R0 /F r+1 R0 )T −r .
r=0
r=0
Then we have the following commutative diagram (see [TW89, Proposition 5.17]):
C̃p ←−−−−
y
E −−−−→
y
E ←−−−−
y
E
y
X̃ −−−−→
y
X̃ ←−−−−
y
X̃0
y
Cp ←−−−− X −−−−→ X ←−−−− X0
There is a natural C∗ -action on X0Lassociated to the natural N-grading such that the
quotient is isomorphic to E. Let J = k≥0 F k+1 t−k = tR0 ∩ R so that R/J ∼
= grF (R) ∼
=
11
R0 /tR0 . Now we assume furthermore that E is a normal projective variety. This implies
both R and R0 are normal (see [TW89]). Let P be the unique minimal prime ideal of R over
J that corresponds to E, and w the valuation of K(t) attached to P. Then the restriction
of w to R is equal to b · ordE . Assume a = w(t). Thus the filtration F is equivalent to the
filtration that is given by:
(tm R0 ) ∩ R = {f ∈ R; ordE (f ) ≥ ma/b}.
Remark 2.11. There is a general Valuation Theorem about the relation between finitely
generated filtrations and valuations proved by Rees for which we refer the reader to [Ree88].
See also [BHJ15].
3
Volume of models
One very useful tool for us to study the minimzer of the normalized local volume is the
concept of a local volume of a model. It is this concept which enables us to apply the
machinery of the minimal model program to construct different models, especially those
yielding Kollár components.
3.1
Local volume of models
In this section, we extend the definition of volume to models in the ‘normalized’ sense. We
use the concept of local volumes as in [ELS03, Ful13]. Let us first recall the definition, which
is from [Ful13].
Definition 3.1 (Local volume). (cf. See [Ful13]) Let X be a normal algebraic variety of
dimension at least two and let o be a point on X. Fixing a proper birational map µ : Y → X,
for an arbitrary Cartier divisor D on Y , we define the local volume of D at x to be
volF
o (D) = lim sup
m→∞
where
h1o (mD)
:=
1
dim H{o}
(X, f∗ OY
h1o (mD)
,
mn /n!
(mD)).
We can define the volume of a Q-Cartier divisor D to be
volF
o (D) :=
volF
o (mD)
,
mn
for sufficiently divisible m.
Lemma 3.2. Let µ : Y → X be a birational morphism. If D ≥ 0 is an exceptional Q-divisor,
such that Supp(D) ⊂ µ−1 (o). Then
volF
o (−D) = lim sup
k→∞
lR (OX /ak )
,
k n /n!
where k is sufficiently divisible and ak = µ∗ (OY (−kD)).
Proof. This follows from [Ful13, Remark 1.1(ii)] (see also [Ful13, Remark 1.31 and 1.32]).
The right hand side of the above display is also the volume vol(a• ) defined in [ELS03,
Definition 3.1, Proposition 3.11]. In particular, given a prime divisor D over o with log
discrepancy a, then we see that
c
volF
o (−aD) = volX,o (ordD ).
Definition 3.3. Let o ∈ (X, D) be a klt singularity. Let µ : YP→ (X, o) be a birational
morphism, such that µ is an isomorphism over X \ {o}. Let E = i Gi be the divisorial part
of µ−1 (o). Then we define the volume vol(X,D),o (Y ) (abbreviated as volX,o (Y ) or vol(Y ) if
(X, D; o) is clear) of Y to be
X
F
−1
vol(X,D),o (Y ) := volF
−ai Gi ),
o (−KY − E − µ∗ (D)) = volo (
i
where ai = AX,D (Gi ) is the log discrepancy.
12
We mainly combine the above definition with the following construction.
Definition 3.4. For a klt pair (X, D) with an idea a, we denote by c its log canonical
threshold lct(X, D; a). We say µ : Y → X is a dlt modification of (X, D + c · a), if
P
1. denote the divisorial part of µ∗ (a) by O(− mi Gi ) and denote by µ∗ (KX + D) =
KY + DY , then
X
DY + c ·
mi Gi = µ−1
∗ (D) + E
where E is the reduced divisor on Ex(µ);
P
2. (Y, DY + c · mi Gi ) is dlt; and
3. for any divisor F ,
a(F ; Y, DY + c ·
X
mi Gi ) = −1 if and only if a(F ; X, D + c · a) = −1.
By the argument in [OX12], we know that it follows the MMP result in [BCHM10] that
the dlt modification of (X, D + c · a) always exists. More precisely, we can choose general
elements fj ∈ a (1 ≤ j ≤ l) which generate a such that l > 1c . Let Dj = (fj ), then Y is the
Pl
dlt modification of (X, D + c · 1l j=1 Dj ).
We do not need property 3 in our argument, but it seems to us it is natural to require
this by comparing with [Kol13, Theorem 1.34].
Lemma 3.5. We can indeed assume that −KY − µ−1
∗ D − E is nef over X.
Proof. Since (X, D) is klt, we know that
X
KY + µ−1
∗ D + E ∼Q,X
a i Gi
with ai = AX,D (Gi ) > 0. Running an MMP with scaling by an ample divisor, we obtain a
relative minimal model Y 99K Y 0 of
l
KY + µ−1
∗ (D + c ·
X
X
1X
Dj ) + E ∼Q,X −c ·
mi Gi +
ai Gi = 0.
l j=1
i
KY + µ−1
∗ (D + c ·
X
1X
Dj ) + E −
ai Gi = −(KY + µ−1
∗ D + E),
l j=1
So we have
l
−1
and hence −KY 0 − µ0 ∗ D − E 0 is nef over X. Furthermore, since
l
1 −1 X
KY + µ−1
Dj ,
∗ D + E ∼Q,X −c · µ∗
l
j=1
Y 0 also gives a minimal model of the dlt pair
l
Y, µ−1
∗ (D
1X
+ c(1 + ) ·
Dj ) + E ,
l j=1
−1
which implies (Y 0 , µ0 ∗ D + E 0 ) is a dlt modification of (X, D + c ·
we can replace Y by Y 0 .
1
l
Pl
j=1
Dj ). Therefore,
c X,o (ordE ). We can also generalize Lemma
When E is irreducible, then volX,o (Y ) = vol
2.7 to the dlt case.
Lemma 3.6. In the setting of Definition 3.3, if we assume that −KY − µ−1
∗ D − E is nef
over X. Then
X
n−1
volX,o (Y ) =
ai (−KY − µ−1
.
∗ D − E)|Ei
i
13
−1
Proof. Let m be sufficiently divisible such
P that L := m(KY + µ∗ D + E) is Cartier. Denote
by F the effective Cartier divisor F := i mai Gi . Then
0 → OY (L⊗−(k+1) ) → OY (L⊗−k ) → OF (L⊗−k ) → 0.
Since L−1 is nef, we know that R1 µ∗ (L⊗−(k+1) ) = 0. Thus
volF
o (L) = vol(L|F ),
then we conclude by dividing mn in both sides.
Lemma 3.7. Let a be an m-primary ideal, we define c = lct(X, D; a) and (Y, E) → X be
the dlt modification of (X, D + c · a). Then
volX,o (Y ) ≤ lctn (a) · mult(a).
P
∗
Proof. Write KY + µ−1
∗ D + E = µ (KX + D) +
i ai Gi , where E is the reduced divisor
on Ex(µ). If we denote the vanishing order of µ∗ a along Gi by mi , then since c is the log
canonical threshold and for every i, Gi computes the log canonical threshold, we know that
c · mi = ai . Thus
X
ak ⊂ µ∗ OY (−
kmi Gi ) =def bk .
i
It suffices to show that
mult(b• ) = volF
o (−
X
mi Gi ),
but this follows Lemma 3.2.
Lemma 3.8. Notations as above. Then there exists a Kollár component S, such that
c
vol(ord
S ) ≤ vol(Y ).
Proof. It follows from Proposition 2.6 that we can choose a model W → Y and run MMP
to obtain W 99K Y 0 , such that Y 0 → X gives a Kollár component S with a(Y, E; S) = −1.
If we fix a common resolution p : W 0 → Y and q : W 0 → Y 0 , then since −(KY + E) is nef
and AY,E (S) = 0, we know −p∗ (KY + E) + q ∗ (KY 0 + S) is q-nef and q-exceptional. By the
negativity lemma, we get
p∗ (KY + E) ≥ q ∗ (KY 0 + S).
Thus
c
vol(ord
S ) = vol(−KY 0 − S) ≤ vol(−KY − E) = vol(Y ).
3.2
Approximating by Kollár components
With the above discussions, we can start to prove our theorems.
Proof of Theorem 1.3. By Proposition 2.4, we know
c X,o (v) = inf lctn (a) · mult(a).
inf vol
v
a
By the above construction in Lemma 3.7 and 3.8, for any m-primary ideal a, we know that
there exists a Kollár component S, such that
n
c
vol(ord
S ) ≤ lct (a) · mult(a).
This finishes the proof of the first part of Theorem 1.3.
We continue to prove the second part of the theorem. Let {ak }k∈Φ be the associated
graded family of valuation ideals induced by v where Φ ∈ R is the value semigroup. For each
ak (k ∈ Φ), we denote by
ck := lct(X, D; ak ).
14
Let µk : Yk → X be a dlt modification of (X, D; ck · ak ) and Ek the exceptional divisor of Yk
over X. Assume the model we obtain from Lemma 3.8 is Yk0 with the Kollár component Sk .
We consider the valuation
ck · k
vk =
ordSk .
AX,D (S)
By [JM12], we know
1
AX,D (v 0 )
ck · k = lct(X, D; ak ) = inf0 1 0 k ≤ AX,D (v) < ∞.
v
k
k v (a )
Then by the Izumi type estimate in [Li15a, Proposition 1.2], we know that
vk (m)ordo ≤ vk ≤ c · ordo ,
c k)
for some positive constant c and all k. And by [Li15a, Theorem 2.3] and the fact that vol(v
is bounded from above, we know that vk (m) is bounded from below. In particular, by the
compactness result [JM12, Proposition 5.9] and Proposition 3.9, we know that there is an
infinite sequence {vkj }kj ∈Φ with kj → +∞ which has a limit in ValX,o .
Denote by
v 0 = lim vki ,
i→∞
then we know that
AX,D (v 0 ) ≤ lim inf AX,D (vki ) = lim cki · ki ≤ AX,D (v)
i→∞
i→∞
as AX,D is lower semicontinuous (see [JM12, Lemma 5.7]). We claim for any f , we have
v 0 (f ) ≥ v(f ),
c 0 ) ≤ vol(v).
c
and this clearly implies that vol(v 0 ) ≤ vol(v), which then implies vol(v
Denote by v(f ) = p. For a fixed kj , choose l such that
(l − 1)p < kj ≤ lp.
Let k = kj in the previous construction, then
v(f ) = p
=⇒ v(f l ) = pl,
=⇒ f l ∈ apl ,
=⇒ f l ∈ akj ,
=⇒ l · ordEi (f ) ≥ mkj ,i for any i,
1
,
=⇒ l · ordSkj (f ) ≥ AX,D (Skj ) ·
ckj
kj
p
=⇒ vkj (f ) ≥
>p− .
l
l
The fourth arrow is because if f l ∈ akj , then f l vanishes along mkj ,i Ei ; and the fifth arrow
is because that
X
KYkj + Ykj + µ−1
mkj ,i Ei ,
kj ∗ D + Ekj ∼Q,X ckj ·
and the pull back of KYkj + Ykj + µk−1
D + Ekj is larger than the one from
j∗
0
KYk0 + µk−1
D + Skj ∼Q,X AX,D (Skj )Skj .
j∗
j
Thus v 0 (f ) = lim vkj (f ) ≥ p = v(f ).
Proposition 3.9. Let o ∈ (X, D) be a klt singularity. Let a and b be two positive numbers.
Then the subset Ka,b of Valx,X which consists of all valuations with
a ≤ v(m)
and
is sequential compact.
15
AX,D (v) ≤ b
Proof. Let {vi } be a sequence contained in K. Let {ai,k } be its associated graded valuative
ideals for k ∈ Φi . We can find a countably generated field F ⊂ C, such that R = Spec(RF )×F
C for some finitely generated F -algebra RF and D, x are defined over F . Furthermore, we
can assume for each pair (i, k), ai,k = (ai,k )F ×F C, for some ideal (ai,k )F ⊂ RF . Denote by
XF := Spec(RF ) and DF the divisor of D descending on XF .
Now let (vi )F be the restriction of vi on RF . By our definition, we know that
ai,k = {f ∈ RF |(vi )F (f ) ≥ k},
and (vi )F ∈ (Ka,b )F where (Ka,b )F is defined for all v ∈ Valx,XF with a ≤ v(mF ) and
AXF ,DF (v) ≤ b. By [HLP12, Theorem 1.1], Valx,XF has the same topology as a set of some
Euclidean space, thus (Ka,b )F is sequential compact as it is compact by [JM12, Proposition
5.9]. Therefore after passing through a subsequence, (vi )F has a limit (v∞ )F , which can be
extended to a valuationPv∞ := (v∞ )F ⊗ C. In fact, v∞ is defined as follows: for any f ∈ R
m
it can be written f = j=1 fj ⊗F hj such that 0 6= fj ∈ R and h1 , ..., hm ∈ C are linearly
independent over F , then
m
v∞ (f ) = min(v∞ )F (fj ).
j=1
We claim vi = (vi |RF ) ⊗F C. In fact, for any f , if vi (f ) = k, then f ∈ ai,k = (ai,k )F ⊗F C,
thus (vi |RF ) ⊗F C(f ) = k.
To see that for any f , v∞ (f ) = lim vi (f ), we know for some j,
v∞ (f ) = (v∞ )F (fj ) = lim(vi |RF )(fj ) ≥ lim sup vi (f ).
i
i
For another direction, if we have a subsequence of i, such that limi vi (f ) < v∞ (f ), after
passing to a subsequence again, we can find a j, such that
lim vi (f ) = lim vi (fj ) = lim(v∞ )F (fj ) ≥ v∞ (f ),
i
i
i
a contradiction.
As Example 7.2 shows that for a general klt singularity (X, o), the minimum is not
always achieved by a Kollár component. Thus we have to take a limit process. However, if
the minimizer v has rational rank 1, then it should always yield a Kollár component. First
we have the following result which is inspired by the work in [Blu16a]. We note that it is
also independently obtained in [Blu16b].
c
Lemma 3.10. If E ∈ ValX,o such that ordE minimizes vol(X,
D). Then the Rees algebra
associated to ordE is finitely generated.
Proof. If we let {a• } be the graded valuative ideas associated to ordE , then we know that
c
vol(ord
E)
=
≥
≥
mult(ak )
kn
lim lct(X, D; ak )n · mult(ak )
lim AX,D (ordE )n ·
k→∞
k→∞
c
vol(ord
E)
c
by Proposition 2.4 and our assumption that ordE is a minimizer of vol(X,
D). So we conclude
that (see [Mus02])
lct(X, D; a• ) := lim k · lct(X, D; ak ) = AX,D (ordE ),
k→∞
which we denote by c. Therefore, we can choose sufficiently small, such that the discrepancy
a(E; X, D + (1 − )c · a• ) ∈ (−1, 0),
On the other hand, we know
lct(X, D; a• ) = lim m · lct(X, D; am ).
m→∞
So for sufficiently large m, we know that for all G, the discrepancy
a(G; X, D +
1
(1 − )c · am ) > −1.
m
16
1
(1−)c·am ) < 0. Then similar to the discussion in 3.4, we can find
We also have a(E; X, D+ m
1
1
general Q-divisor ∆, such that (X, D + m
(1−)c·∆) is klt and a(E; X, D + m
(1−)c·∆) < 0.
In particular, we can apply [BCHM10] to obtain a model µ : Y → X such that Ex(µ) = E
and −E is µ-ample, which implies the finite generation.
Proof of Theorem C. Applying Lemma 3.10, the assumption in Case 1 which says v is a
divisorial valuation implies the assumption in Case 2, thus we only need to treat the Case 2.
By the proof of Proposition 2.4, we know that if we let ak be the ideal of elements with
values at least k, then
n
AX,D (v)
n mult(ak )
≥
AX,D (v) ·
· mult(ak ) ≥ lctn (X, D; ak ) · mult(ak ).
kn
v(ak )
By the finite generation assumption, we know that akl = alk for sufficiently divisible k and
any l. So replace k by kl in the above display and let l → +∞, we know that
c X,o (v) ≥ lctn (X, D; ak ) · mult(ak ) ≥ vol
c X,o (v).
vol
Take µ : Y → X to be the dlt modification of (X, D + lct(X, D, ak ) · ak ) as given in Lemma
3.5. The above discussion then implies that
c X,o (v) = volX,o (Y ).
lctn (X, D; ak ) · mult(ak ) ≥ vol
It follows from Proposition 2.6, that we can choose a model W → Y and running MMP
to obtain W 99K Y 0 , such that µ0 : Y 0 → X gives a Kollár component S with a(S; Y, µ−1 D∗ +
E) = −1. We only need to show that if Y 0 and Y are not isomorphic in codimension 1, then
volX,o (Y 0 ) < volX,o (Y ).
This is the the local analog of the argument in [LX14, Proposition 5]. We give the details
for the reader’s convenience.
Let π : Y → Y c be the canonical model of −KY − µ−1
∗ D − E over X, which exists because
−1
−(KY + µ−1
∗ D + E) ∼Q,X KY + µ∗ (D + c ·
X
1X
Dj ) + E −
AX,D (Gi )Gi
l
i
is a klt pair for sufficiently small. The assumption that Y 0 and Y are not isomorphic in
codimension 1 implies Y c 6= Y .
Take p : Ŷ → Y and q : Ŷ → Y 0 a common log resolution, and write
−1
∗
0
p∗ (KY + µ−1
∗ D + E) = q (KY 0 + µ ∗ D + S) + G.
By negativity lemma (cf. [KM98, 3.39]), we conclude that G ≥ 0. Since
X
KY + µ−1
AX,D (Gi )Gi
∗ D + E ∼Q,X
i
and
−1
KY 0 + µ0 ∗ D + S ∼Q,X AX,D (S),
we know that
X
p∗ (
AX,D (Gi )Gi ) = q ∗ (AX,D (S)) + G.
i
For 0 ≤ λ ≤ 1, let
Lλ = q ∗ (AX,D (S)) + λG =
X
bi (λ)Fi ,
i
where Fi runs over all divisor supports on Ŷo := Ŷ ×X {o}, and −Lλ |Ŷo is nef. Define
f (λ) =
X
bi (λ)(−Lλ |Fi )n−1 ,
i
17
thus f (λ) is non-decreasing as G ≥ 0. By Lemma 2.7 and 3.6, we know that
f (1) = volX,o (Y )
f (0) = volX,o (Y 0 ).
and
Since Y 99K Y 0 are not isomorphic incodimension 1, it must contract some component
G1 of E, and the coefficient of G1 in G is
a := AY 0 ,µ0 −1
(G1 ) > 0.
∗ D+S
Then
df (λ)
|λ=1
dλ
≥
n−1
n · G · − p∗ (KY + µ−1
∗ D + E)
n−1
n · aG1 · − π∗ (KY + µ−1
∗ D + E)
>
0.
=
Thus volX,o (Y 0 ) = f (0) < f (1) = volX,o (Y ).
With all this discussion, we also obtain the following result, which characterizes the
equality condition in Proposition 2.4 and is a corresponding generalization of [dFEM04,
Theorem 1.4] (see Remark 3.12) for smooth point. See [Laz04b, 9.6] for more background.
Theorem 3.11. Let (X, o) = (Spec(R), m). Assume (X, D) is a klt singularity for a Qdivisor D ≥ 0. If there exists an m-primary ideal b, such that lctn (X, D; b) · mult(b) is equal
c
to inf v∈ValX,o vol(v),
then (X, D) is minimized at a Kollár component.
Moreover, for an ideal a,
lctn (X, D; a) · mult(a) =
inf
v∈ValX,o
c
vol(v),
if and only if there exists a Kollár component S and a positive integer k such that ordS
computes the log canonical threshold of a on o ∈ (X, D) with
c
vol(ord
S) =
inf
v∈ValX,o
c
vol(v),
and the only associated Rees valuation of ak is ordS .
Proof. By the argument in Theorem C, we see that
lctn (X, D; b) · mult(b)
c X,o if and only if the dlt modification of
reaches the minimum of vol
(X, D; c · b)
where c = lct(X, D; b)
is a model µ : Y → X which only extracts a Kollár component S of (X, D) such that ordS
c X,o .
is a minimizer of vol
Now we fix an ideal a such that
lctn (X, D; a) · mult(a) =
inf
v∈ValX,o
c
vol(v),
and S is the Kollár component we obtain as above. Let µ∗ a have the vanishing order m
along S. Since S is Q-Cartier, then let k be a positive integer such that mkS is Cartier. We
claim that
µ∗ (ak ) = OY (−mkS).
Granted this for now, then we know that Y coincides with the normalized blow up X + → X
of ak , i.e., S is the only associated Rees valuation for ak .
To see the claim, since −mkS is Cartier, we know that
µ∗ (ak ) = c · OY (−mkS)
for some ideal c ⊂ OY ,
18
and we aim to show that c is indeed trivial. If not, we take a normalized blow up φ : Y + → Y
of c, so φ∗ c = OY + (−E) for some effective Cartier divisor E. Since −S is ample over X, we
can choose l sufficiently big, such that
−D := −φ∗ (mklS) − E
on Y + is ample over X.
Since
(µ ◦ φ)∗ akl = OY + (−φ∗ (mklS) − lE) ⊂ OY + (−φ∗ (mklS) − E),
we know that
∗
mult(akl ) ≥ volF
o (−φ (mklS) − E)
=
volF
o (−D)
= mkl(−D|φ∗ S )n−1 + (−D|E )n−1
>
mkl(mkl(−S)|S )n−1
=
(mkl)n vol(ordS ).
Since lctn (X, D; a) = 1l · AX,D (ordS ), we can easily see the above inequality is contradictory
to the assumption that
c
lctn (X, D; a) · mult(a) = vol(ord
S ).
Here the inequality in the fourth row comes from a similar but easier calculation as in the
proof of of Theorem C.
For the converse direction, since a reaches the minimum if and only if ak reaches the
minimum, then we can replace a by ak and assume its only associated Rees valuation is
ordS , i.e., we know that the normalized blow up µ : X + → X, has the property that µ∗ (a) =
OX + (−mS). Then the valuative ideal
amk = ak ,
where ak means the integral closure of ak . And lct(X, D; a) =
AX,D (S)
.
m
We claim that
n! · lR (R/ak )
,
k→+∞
kn
mult(a) = lim
and this immediately implies that
n
c
vol(ord
S ) = lct(X, D; a) · mult(a).
To see the claim, if we denote by J (ak ) = J (X, D; ak ) the multiplier ideal, then we know
that
n! · lR (R/J (ak ))
mult(a) = lim
,
k→+∞
kn
by the local Skoda Theorem [Laz04b, 9.6.39]. On the other hand, since (X, D) is klt, we
have
ak ⊂ J (ak ),
thus we are done.
Remark 3.12. In the proof, we indeed show that for any k such that mkS is Cartier on Y ,
the integral closure ak coincides with the valuative ideal amk of ordS .
4
4.1
K-semistability implies minimum
Degeneration to initial ideals
Let (X, o) = (Spec(R), m) be an algebraic singularity such that (X, D) is klt for a Q-divisor
D ≥ 0. Denote by m the maximal ideal of o ∈ X. Suppose b is an m-primary ideal on X.
We consider its flat degeneration W/A1 where W is the underlying coarse moduli space of W
19
defined in Section 2.3. We keep the notation in Section 2.3 and denote by v0 the valuation
ordS .
Next we will describe explicitly a way of obtaining an ideal B on X such that B⊗OX×C∗ =
b and B ⊗ OC ∼
= in(b) by considering the closure of b × C∗ on the pull back Z.
For this we consider the extended Rees algebra (see [Eis94, 6.5]):
M
M
R0 =
R0k :=
ak t−k ⊂ R[t, t−1 ].
k∈Z
k∈Z
Notice that if k ≤ 0, then ak = R. It is well known that:
R0 ⊗C[t] C[t, t−1 ] ∼
= R[t, t−1 ],
R0 ⊗C[t] C[t]/(t) ∼
=
+∞
M
(ak /ak+1 )t−k ∼
= T.
k=0
Geometrically this exactly means W = Spec(R0 ) and
W ×A1 (A1 \ {0}) = X × (A1 \ {0}),
W ×A1 {0} = C.
Notice that there is a natural Gm -action on R0 given by the Z-grading.
For any f ∈ R, supposing v0 (f ) = k then we define
f˜ = t−k f ∈ ak t−k ⊂ R0 ,
and denote
in(f ) = [f ] = [f ]ak+1 ∈ ak /ak+1 = Tk ,
where we use [f ]a to denote the image of f in R/a. Then we define the ideal B to be the ideal
in R0 generated by {f˜; f ∈ b}, and in(b) the ideal of T generated by {in(f ); f ∈ b}. The first
two items of the following lemma is similar to (but not the same as) [Eis94, Theorem 15.17]
and should be well known to experts. Notice that here we degenerate both the ambient space
and the ideal. A version of the equality (4) was proved in [Li15b, Proposition 4.3].
Lemma 4.1.
1. There are the identities:
(R0 /B) ⊗C[t] C[t, t−1 ] ∼
= (R/b)[t, t−1 ],
(R0 /B) ⊗k[t] k[t]/(t) ∼
= T /in(b).
2. The C[t]-algebra R0 /B is free and thus flat as a C[t]-module. In particular, we have
the identity of dimensions:
dimC (R/b) = dimC (T /in(b)) .
(4)
3. If b is mR -primary, then in(b) is an mT -primary homogeneous ideal.
Proof. The statement (1) follows easily from the definition.
Next wen prove (2). Denote by
o ck = Tk ∩ in(b) the k-th homogeneous piece of in(b). We
(k)
fix a basis in(fi ); 1 ≤ i ≤ dk of Tk /ck . We want to show that
h
i
g
(k)
(k)
fi
= fi
A :=
0
B
1 ≤ i ≤ dk
⊂ R0 /B
is a C[t]-basis of R0 /B.
We first verify that A0 is a linearly independent set. To prove this, we just need to show
that A0 is a C[t, t−1 ]-linearly independent subset of (R/b)[t, t−1 ]. It is then enough to show
that
n
o
(k)
(k)
A := [fi ] = [fi ]b | 1 ≤ i ≤ dk ⊂ R/b.
(5)
is C-linearly independent, which can be verified directly as in [Li15b, Proposition 4.3]. See
also [Eis94, Proposition 15.3].
So we just need to show that A0 spans R0 /B. Equivalently, we need to show that for any
f ∈ R, [f˜] = [f˜]B ∈ R0 /B is in the C[t]-span of A0 . This can be shown again with the help
of A in (5), that is, it is enough to prove that A C-spans R/b. Indeed, assuming the latter,
20
P
(k)
for any f ∈ R, there exists a linear combination g = i,k cik fi such that f − g =: h ∈ b.
If m = v0 (f ), then
X
(k)
f˜ = t−m f =
cik t−m fi + t−m h
i,k
Because t−m h ∈ B, the above indeed implies [f˜] is in the C[t]-span of A0 .
To prove that A indeed C-spans R/b, we first claim that the following set is finite:
{v0 (g) | g ∈ R − b}.
Indeed because b is m-primary, there exists N > 0 such that mN ⊆ b ⊆ m. So R−b ⊆ R−mN .
Now the claim follows from the fact that for any element f ∈ mN ,
v0 (f ) ≤ c · A(v0 ) · N
by Izumi’s theorem, where c is a uniform constant not depending on f .
If there is [f ] 6= 0 ∈ R/b that is not in the span of A, then we can choose a maximal
k = v0 (f ) such that this happens. There are two cases:
(k)
1. If in(f ) ∈ Tk \ ck , then because in(fi ) is a basis of Tk /ck , there exists tj ∈ R such
Pdk
(k)
that in(f ) − j=1
tj in(fj ) = in(g) ∈ ck for some g ∈ b. So we get:
v0 f −
dk
X
(k)
tj fj
− g > k.
j=1
Pdk
Pdk
(k)
(k)
By maximality of k, [f − j=1
tj fj − g] = [f ] − j=1
tj [fj ] and hence [f ] is in the
span of A. Contradiction.
2. If in(f ) ∈ ck = in(b) ∩ Tk . Then in(f ) = in(g) for some g ∈ b. So v0 (f − g) > k
and hence [f − g] is in the span of A by the maximal property of k. But then [f ] =
[f − g] + [g] = [f − g] is in the span of A. Contradiction.
To prove part 3 of the Lemma, we need to show that there exists N ∈ Z>0 such that
N1
mN
T ⊆ in(b) ⊆ mT . Because b is mR primary, there exists N1 ∈ Z>0 such that mR ⊆ b ⊆ mR .
m
By Izumi’s theorem, there exists l ∈ Z>0 such that alm ⊆ mR for any m ∈ Z>0 . By letting
N = lN1 , it’s easy to see that mN
T ⊆ in(b) ⊆ mT .
Lemma 4.2. If b• = {bk } is a graded family of ideals of R, then in(b• ) := {in(bk )} is also
a graded family of ideals of T .
Proof. We need to show that:
in(bk ) · in(bl ) ⊆ in(bk+l ).
If v0 (f ) = k and v0 (g) = l, then v0 (f g) = k + l.
in(f ) · in(g) = [f ]ak+1 · [g]al+1 = [f g]ak+l+1 = in(f · g).
Lemma 4.3. If b• is a graded family of ideals, then
lctn (b• ) · mult(b• ) ≥ lctn (in(b• )) · mult(in(b• )).
(6)
Proof. By the flatness of B and the lower semicontinuity of log canonical thresholds, we
have lct(bk ) ≥ lct(in(bk )). Therefore, by (4)
lctn (bk ) · lR (R/bk ) ≥
lctn (in(bk )) · lT (T /in(bk )).
Taking limits as k → +∞, we then get the inequality (6).
21
4.2
Equivariant K-semistability and minimizing
In this section, we will take a detour to show the discussion in Section 4.1 can be used to
study the equivariant K-semistability. Here for a Q-Fano variety (X, D) with an action by
an algebraic group G, we call it G-equivariantly K-semistable (resp. Ding semistable) if for
any G-equivariant test configuration, its generalized Futaki (resp. Ding) invariant is nonnegative. Let T = (C∗ )k be a torus. First we improve the two approximating results to the
equivariant case.
Lemma 4.4. Let (X, o) = (SpecR, m) and D ≥ 0 a Q-divisor, such that o ∈ (X, D) is a klt
singularity. Assume x ∈ (X, D) admits a T -action. Then we have
c X,o (v) = inf lctn (X, D; a) · mult(a) = inf vol
c X,o (ordS ),
min vol
v
a
S
(7)
where on the left hand side the infimum runs over all the valuations centered at o, and on
the middle it runs over all the T -equivariant m-primary ideals; and at the end, it runs over
all T -equivariant Kollár components.
Proof. Let {a• } = {ak } be a graded sequence for an m-primary ideal a. The discussion in
Section 4.1 implies that for {b• } =defn in(a• )
lctn (X, D; b• ) · mult(b• ) ≤ lctn (X, D; a• ) · mult(a• ).
Since
lctn (X, D; b• ) · mult(b• ) = lim lctn (X, D; bm ) · mult(bm ),
m
we conclude the first inequality as a corollary of Proposition 2.4.
For the second equality, we just need to show that the construction in Section 3.1 can be
established T -equivariantly. This is standard, which relies on two facts: first, we can always
take an equivariant log resoltuion of (X, D, a) (see [Kol07]); second, as T is a connected group,
for any curve C in a T -variety and any t ∈ T , t·C will always be numerically equivalent to C;
as the minimal model program only depends on the numerical class [C], we know that any
MMP sequence is automatically T -equivariant. Therefore, for any T -equivariant m-primary
ideal a, we can find a T -equivariant dlt modication Y → X and then a T -equivariant Kollár
component S, such that
c
lctn (X, D; a) · mult(a) ≥ vol(Y ) ≥ vol(ord
S ).
In the below, we also need use the main idea from Fujita’s work (see [Fuj15, Fuj16]).
Proposition 4.5 ([Fuj16]). Let (V, ∆) be an (n − 1)-dimensional log-Fano pair which is T equivariantly log-Ding-semistable. Let δ be a positive rational number such that −δ −1 (KV +
∆) is Cartier. Let IM ⊂ · · · ⊂ I1 ⊂ OV be a sequence of T -invariant coherent ideal sheaves
and assume
I := IM + IM −1 t1 + · · · + I1 tM −1 + (tM ) ⊂ OV ×C
form a flag ideal. Let Π : V → V × C be the blowup along I. Denote D = Π∗ (∆ × C). Let
E ⊂ X be the Cartier divisor defined by OV (−E) = I · OV , and let
L := Π∗ OV ×C −δ −1 (KV ×C/C + ∆ × C) ⊗ OX (−E).
Assume that L is semi ample over C. Then (V, D; L) is naturally seen as a (possibly nonnormal) semi test configuration of (S, ∆; −δ −1 (KV + D)). Under these conditions, ((V ×
C1 , ∆ × C); I δ · (t)d ) must be sub log canonical, where
d := 1 +
δ n L̄n
.
n(−KV − ∆)n−1
The following argument is essentially in [Li15b].
22
Proposition 4.6. If (V, ∆) is a projective log Fano variety and o ∈ (X, D) is the cone
of (V, ∆) induced by some ample Cartier divisor L = −r−1 (KV + ∆). Then the canonical
c X,D if and only if (V, ∆) is Kvaluation obtained by blowing up the vertex minimizes vol
semistable.
Proof. First we assume (S, ∆) is (n − 1)-dimensional log-K-semistable and prove the volume
minimizing. By Lemma 4.4, we only need to prove that for any C∗ -invariant divisorial
valuation v over (X, o),
c
c
vol(ord
V ) ≤ vol(v).
L+∞
L+∞ 0
Let c1 = v(V ) > 0 and R = k=0 Rk = k=0 H (V, kL) thus X = Spec(R).
On R, we define a filtration
M
FR(t) =
F kt Rk ,
k
where
F x Rk := H 0 (V, L⊗k ⊗ ax ),
and ax = {f ∈ OV | v(f ) ≥ x}.
Then the volume is defined to be
vol(FR(t) ) := lim sup
m→∞
dimC F mt Rm
.
mn /n!
By [Li15b, (18)], we get a formula for vol(v):
vol(v)
=
=
=
n!
dimC R/am (v)
mn
Z
+∞
dt
Ln−1
−
vol FR(t) n+1
n
c1
t
c1
Z +∞
(t)
dvol FR
.
−
tn
c1
lim
m→+∞
Then we consider the following function
Z +∞
Ln−1
λsdt
(t)
Φ(λ, s) =
−
n
vol
FR
n
(λc1 s + (1 − s))
(1 − s + λst)n+1
c1
Z +∞
−d vol(FR(t) )
=
.
((1 − s) + λst)n
c1
Φ(λ, s) satisfies the following properties:
1. For any λ ∈ (0, +∞), we have:
Φ(λ, 1) = vol(λv) = λ−n vol(v),
Φ(λ, 0) = vol(v0 ) = Ln−1 .
2. For any λ ∈ (0, +∞), Φ(λ, s) is continuous and convex with respect to s ∈ [0, 1].
3. The directional derivative of Φ(λ, s) at s = 0 is equal to:
Z +∞
1
vol FR(t) dt .
Φs (λ, 0) = nλLn−1 λ−1 − c1 − n−1
L
c1
r
Let λ∗ = A(X,C(D))
(v) . By Item 1 and 2, we just need to prove Φs (λ∗ , 0) ≥ 0. Let v̄ = v|C(S)
be the restriction of v under the inclusion C(S) ,→ C(X). It is known that v̄ = b · ordE where
b > 0 and ordE is a divisorial valuation on C(S).
Using Adjunction formula, it is easy to show that:
λ−1 − c1 =
A(X,D) (v)
− c1 = AV,∆ (v̄) = b · AV,∆ (E).
r
By change of variables we get:
Z +∞
Z
(t)
vol FR
dt =
c1
0
23
+∞
vol Fv̄ R(t) dt.
where
Fv̄ R(t) =
M
H 0 (V, L⊗k ⊗ akt ),
and akt = {f ∈ OV | v̄(f ) ≥ kt}.
k
So we get:
Φs (λ∗ , 0)
Z +∞
1
vol Fv̄ R(t) dt
= nλ∗ Ln−1 A(V,∆) (v̄) − n−1
L
0
Z +∞
1
n−1
(t)
= nλ∗ L
b A(V,∆) (E) − n−1
dt .
vol FordE R
L
0
By the valuative criterion of (log-)K-semistability derived in [Li15b, Fuj16, LL16], we get
Φs (λ∗ , 0) ≥ 0 (see e.g. Proposition 4.5).
Conversely the statement follows immediately by the valuative criterion for log-K-semistability
in [Li15b, Fuj16, LL16] by choosing a family of valuation vs emanating from v0 = ordS in
the direction of ordE as in [Li15b, Section 7].
Remark 4.7. The fact that K-semistability implies the canonical valuation is a minimizer
was proved in [LL16], where (V, ∆) is assumed to be analytic K-semistable, i.e., if there
exists a special test configuration (X , D) that degenerates (X, D) to a log Fano (X0 , D0 ) with
a conic Kähler-Einstein metric. If (V, ∆) is analytic K-semistable, then we know (V, ∆) is
K-semistable. The converse implication conjecturally true, and known for smooth X when
D = 0 (see [CDS15, Tia15]).
With all the techniques we have, we can prove Theorem 1.4.
Proof of Theorem 1.4. Let (X, D) be a cone of −r(KV + ∆) over (V, ∆) for some sufficiently
divisible positive integer r. And we consider the minimizing problem of the normalized
local volume at the T -singularity o which is the vertex. We aim to show that if (V, ∆) is
c X,D .
T -equivariantly K-semistable then ordV minimizes vol
Following the proof of Lemma 4.4, by degenerating the ideal r times, we can find a
sequence of T -equivariant ideals {a}i∈I such that
inf lctn (X, D; a) · mult(a) =
i∈I
c
min vol(v).
v∈ValX,o
Then using the equivariant resolution and running an MMP process, we can find a sequence
of T -equivariant Kollár components Si such that
c
inf vol(ord
Si ) =
i
c
min vol(v).
v∈ValX,o
Since we assume that for any T -equivariant special degeneration of (S, ∆), the generalized
Futaki invariant is non-negative, and for special test configuration, the generalized Futaki
invariant is the same as Ding invariant, we know that the Ding invariant for any T -equivariant
special test configuration is non-negative. Using the fact that MMP decreases Ding invariant
(see [Fuj16, Corollary 3.4] and its proof), we know this implies that the Ding invariant for
any T -equvariant test configuration is nonnegative.
Then for any T -equivariant Kollár component Si , we consider v = ordSi ∈ ValX,o . Denote
its induced divisorial valuation b · ordE on S. Restricting the calculation in [Fuj16] to the
equivariant setting, we conclude that
Z +∞
1
vol FordE R(t) dt ≥ 0.
A(V,∆) (E) − n−1
L
0
Running the construction as in the proof of Proposition 4.6, then we know that Φs (λ∗ , 0) ≥ 0
which gives that
c ∗ ) ≤ vol(ord
c
vol(v
Si ).
This implies that the canonical component is a minimizer which then implies that (V, ∆) is
K-semistable by Proposition 4.6 .
24
4.3
Proof of Theorem A
Let (X, o) = (Spec(R), m) be an algebraic singularity such that (X, D) is klt for a Q-divisor
D ≥ 0. Let S be a Kollár component and ∆ = ∆S be the different divisor defined by
adjunction (KY + S + µ−1
∗ D)|S = KS + ∆S where Y → X is the extraction of S. We follow
the notation in Section 2.3.
Lemma 4.8. Let b• be a graded sequence of m0 -primary ideal. If (S, ∆S ) is K-semistable,
then we have
c C,C ,o (ordS ).
lctn (b• ) · mult(b• ) ≥ vol
∆
Proof. Using the result in [JM12], we have
lctn (b• ) · mult(b• )
=
=
n mult(bk )
k · lct(bk ) ·
kn
n
lim lct (bk ) · mult(bk ).
lim
k→+∞
k→+∞
c C,C ,o (ordS ) is equal to
By Proposition 2.4, it suffices to show that vol
∆
c C,C ,o (v)
min vol
∆
v
for v runs over valuations centered on o.
It follows from Proposition 4.6 that if we choose d sufficiently divisible, such that C (d) =
C(S, H) is constructed as the cone over S with some ample Cartier divisor H proportional
c (d) (d) (d) . By
to −(KS + ∆S ), then the canonical valuation ordS is a minimizer of vol
C ,C∆ +CD
Proposition 4.9, this implies the same for C.
c (C,C ) minimizes at ordS if and only if
Proposition 4.9. Under the above notation, vol
D
c (d) (d) (d) minimizes at ordS (d) .
vol
(C ,C +C )
∆
D
Proof. The degree d cover h : C → C (d) is a fiberwise map and the Galois group G =defn Z/d
is a natural subgroup C∗ . Let E be a Kollár component over C (d) , by Lemma 2.9 we know
h∗ (E) is a Kollár component over C, and it follows from Lemma 2.10 (or [Li15b, Lemma
6.9]) that
c
c ∗
d · vol(ord
E ) = vol(h E).
c C,C , then the corresponding canonical valuation also minimizes
So if ordS minimizes vol
∆
c (d) (d) (d) .
vol
C ,C∆ +CD
For the converse, let E be a T -invariant Kollár component over C, since it is G-invariant,
by Lemma 2.9 we know it is a pull back of a Kollár component F over C (d) . If the canonc (d) (d) (d) , then we see over C, vol(ord
c
ical valuation minimizes vol
S ) is less or equal to
C ,C∆ +CD
c
c
vol(ord
E ) for any T -equivariant Kollár component E, therefore ordS is a minimizer of volC,C∆
by Lemma 4.4.
Proof of Theorem A. With the notation as in Section 2.3, we denote by D the birational
transform of D × A1 on Z 0 and write aT ∼Q,X KZ 0 + T . Restricting over a general fiber
and taking the coarse moduli spaces, we obtain
aS ∼X,Q KY + S + µ−1
∗ (D),
then a = AX,D (S). Similarly, over the central fiber over 0, we get
aS0 ∼X,Q KY0 + S0 + µ−1
0,∗ CD ,
thus a = AC,CD (S0 ), where µ0 : Y0 → C is the blow up of the vertex with the exceptional
divisor S0 .
We also know that
volX,o (ordS ) = (−S|S )n−1 = (−S0 |S0 )n−1 = volC,p (ordS0 ).
25
Combining all the above, we know that for any ideal b on X, if we let b• = {bk }, then
c X,o (ordS )
vol
=
volX,o (ordS ) · AnX,D (S)
=
volC,p (ordS0 ) · AnC,CD (S0 )
=
inf volC,p (ordv ) · AnC,CD (v)
v
≤ lctn (in(b• )) · mult(in(b• ))
≤ lctn (b) · mult(b),
where the last two inequalities follow from Lemma 4.3 and 4.8. Thus we conclude that
c X,o (ordS ) ≤ inf lctn (b) · mult(b) = inf vol
c X,o (v),
vol
v
b
where the second equality follows from Proposition 2.4.
5
Uniqueness
In this section, we will prove Theorem B about the uniqueness of the minimizers among all
Kollár components. There are two steps: first we prove this for cone singularities; then for
a general singularity, we combine the deformation construction with some results from the
minimal model program to essentially reduce it to the case of cone singularities.
5.1
Case of cone singularity
We first settle the case of cone singularities. It can be proved using Proposition 5.9 and
[Li15b, Theorem 3.4]. Here we give a different proof, which analyzes the geometry in more
details. A similar argument in the global case appears in the proof of [Liu16, Theorem
3], where a characterization of quotients of Pn is given by achieving the maximal possible
volumes among all K-semistable Q-Fano varieties with only quotient singularities.
Let (V, ∆) be an (n − 1)-dimensional log Fano variety and −(KV + ∆) = rH for an ample
Cartier divisor H and r ≤ n. Assume X 0 := C(V, H) is the affine cone over the base V with
the vertex o and let X be the projective cone and D be the cone divisor over ∆ on X. In
the below, for a variety •, we denote the product • × A1 by •A1 .
Consider a Kollár component S over o ∈ (X, D) with the extraction morphism µ : Y → X.
Let
µA1 (= µ × id) : YA1 → XA1
be the extraction of SA1 . We carry out the deformation process in Section 2.3 with respect to
S. Here X is a projective variety instead of a local singularity, but the construction is exactly
the same. We denote by Z (resp. W ) the coarse moduli space of Z (resp. W), so there are
morphisms, ψ1 : Z → W , φ1 : Z → YA1 and π : W → XA1 . We denote by φ = µA1 ◦ φ1 .
26
Then we have the following equalities:
∗
1. KYA1 + (µA1 )−1
∗ (DA1 ) = µA1 (KXA1 + DA1 ) + aSA1 with a = AX,D (S) − 1;
−1
∗
2. KZ + φ−1
∗ (DA1 ) = φ1 KYA1 + (µA1 )∗ DA1 + F;
−1
∗
3. KZ + φ−1
∗ DA1 = ψ1 KW + (DA1 )W + aSA1 , where (DA1 )W := ψ1∗ φ∗ (DA1 ).
The first two equalities imply:
KZ + φ−1
∗ (DA1 )
= φ∗1 KYA1 + (µA1 )−1
∗ DA1 + F
= φ∗1 µ∗A1 (KXA1 + DA1 ) + aφ∗1 SA1 + F
= φ∗ (KXA1 + DA1 ) + aSA1 + (a + 1)F.
So AXA1 ,DA1 (F) = a + 2 = AX,D (S) + 1. This implies:
KW + (DA1 )W = π ∗ KXA1 ,DA1 + AX,D (S)C̄.
Recall that −KX − D = (1 + r)L, where L = O(1) for the cone construction, so we get:
KW + (DA1 )W = −(1 + r)ρ∗ L + AX,D (S)C̄,
where ρ : W → X the composite of π : W → XA1 with the first projection XA1 → X.
Remark 5.1.
1. As in [LL16], we define the cone angle parameter β = nr . Denote by V∞
the section at infinite place of X and V∞ the birational transform of (V∞ )A1 . Then
−(KX + D + (1 − β)V∞ ) = δL
with δ = r n+1
n , we get:
KW + (DA1 )W + (1 − β)V∞
=
π ∗ KXA1 + DA1 + (1 − β)V∞ + AX,D (S)C̄
=
−δρ∗ L + AX,D (S)C̄.
(8)
(9)
∗
2. Because KY + µ−1
∗ D = µ (KX + D) + (AX,D (S) − 1)S, we have:
∗
KY + µ−1
∗ D + (1 − β)V∞ = µ KX + D + (1 − β)V∞ + (AX,D (S) − 1)S
= −δµ∗ L + (AX,D (S) − 1)S.
The above construction works for any Kollár component. From now on we assume that
(V, ∆) is K-semistable and S minimizes the normalized volume, i.e. it satisfies
n
n−1
c
c
vol(ord
),
S ) = vol(ordV0 ) = r (H
(10)
where V0 denotes the canonical divisor obtained by blowing up the vertex of the cone and
we aim to show S = V is the canonical component. We note that by Proposition 4.6 , we
c
know that vol(ord
V0 ) is the minimal normalized volume.
Then we have:
c
rn (H n−1 )
vol(ord
S)
=
.
vol(ordS ) =
AX,D (S)n
AX,D (S)n
In Section 4.2, we have used the filtration induced by a valuation (see also [BHJ15, Fuj15]).
Here we use the same construction but for sections on the projective cone instead of the base.
Definition 5.2 (Filtration by valuation). Fixed a valuation v. Let Sm = H 0 (X, L⊗m ).
Define F x Sm ⊂ Sm to be a decreasing filtration (with respect to x) as follows:
F x Sm := H 0 (X, L⊗m ⊗ ax ),
where ax = {f ∈ OX | v(f ) ≥ x}.
L
L
On m=0 Sm , we define FS (t) =
F kt Sk . Then the volume is defined to be
vol(FS (t) ) := lim sup
m→∞
27
dimC (F mt Sm )
.
mn /n!
Proposition 5.3. With the above notation, if the case (V, ∆) is log K-semistable, then
(X, D + (1 − β)V∞ ) is log K-semistable. As a consequence,
Z +∞
δ
vol(FS (x) )dx ≥ 0.
AX,D (S) − n
(L ) 0
Proof. By Theorem 1.4, we only need to check this is true for C∗ -equvariant special test
configuration (X , D + (1 − β)V)/P1 of (X, D + (1 − β)V∞ ), where we consider the test
configuration over P1 by adding a trivial fiber over {∞}. Consider the closure of V ⊃
V∞ ×(P1 \{0}). As (1+r)V ∼Q −(KX +D), we know that if we denote by L the polarization
on the test configuration extending L, then
(1 + r)V ∼Q,P1 −KX − D = (1 + r)L
as the fiber over 0 is irreducible. Let ∆∞ (= ∆) = V∞ ∩ D and E be the closure of ∆∞ ×
(P1 \ {0}), then
(KX + D + V)|V = KV + E
as X is smooth along the codimension 2 points over 0 and so there is no different divisor.
Then the generalized Futaki invariant of (X , D + (1 − β)V)/P1 is
Fut(X ) = −
1
(−KX /P1 − D − (1 − β)V)n+1 .
(n + 1)(δL)n
Since V∞ ∼ L and δ = r n+1
n , it’s easy to verify that the generalized Futaki invariant of the
induced test configuration of (V, ∆) is
Fut(V) = −
1
nrn−1 H n−1
((−KV/P1 − E)|V )n = Fut(X ).
Recall the log-K-semistability is equivalent to the log-Ding-semistablity (see e.g. [Fuj16]).
Then the second part is a standard generalization of [LL16, Proposition 4.5] to the log setting,
and we skip it.
The key calculations are given by the following results proved in in [LL16].
Proposition 5.4. Suppose (V, ∆) is log-K-semistable. If S is a Kollár component obtaining
c over (X, o), then the graded filtration induced by S satisfies the following
the minimum of vol
two conditions:
1.
AX,D (S) −
2. Denote τ :=
q
n
(Ln )
vol(ordS ) .
δ
(Ln )
Z
+∞
vol(FS (x) )dx = 0.
0
We have:
vol FS (x) = volY (µ∗ L − xS) = (Ln ) − vol(ordS )xn for any x ∈ [0, τ ].
Proof. By Proposition 5.3, we know (X, D + (1 − β)V∞ ) is log K-semistable. Then it follows
from precisely the equality case of the Formula (25) and (29) in [LL16].
Lemma 5.5. We have τ =
AX,D (S)
.
r
Proof. Combining 1 and 2 in Proposition 5.4, we know that
Z
r(1 + n) τ n
AX,D (S) −
L − vol(ordS )xn dx = AX,D (S) − r · τ = 0.
n · Ln 0
By arguing as in [Fuj15] (see also [Liu16]), we know that:
Lemma 5.6. We know τ is the nef threshold, i.e.
τ = sup {x| µ∗A1 L − xS is ample } .
28
Proof. When the point is smooth, this follows from [Fuj15, Theorem 2.3(2)]. Exactly the
same argument can be used to treat the current case.
Theorem 5.7. If S is a Kollár component obtaining the minimum of the normalized volume,
then S is the canonical component V0 .
We first show the following statements.
Lemma 5.8.
1. ρ∗ L − τ C̄ is base-point-free, and contracts Y to S∞ (∼
= S) ⊂ C̄ as the
section at the infinite place.
2. AX,D (S) = r and there is a special test configuration degenerating (V, ∆) into (E, ∆E ).
Moreover, there is a special test configuration of (X, D + (1 − β)V∞ ; L) to (C̄, DC̄ +
(1 − β)S∞ ; L0 ) which is indeed an isomorphism, where DC̄ is the intersection of C̄ with
(D × A1 )W .
Proof. The proof of this part is along the similar line in [Liu16, Proof of Lemma 33]. First
we observe the following restrictions of ρ∗ L − xC̄:
1. ρ∗ L − xC̄
Xt
∗
2. ρ L − xC̄
3. ρ∗ L − xC̄
Y0
C̄
= L, t 6= 0;
= µ∗ L − xS;
= −xC̄|C̄ = xY0 |C̄ = xS∞ = xOC̄ (1).
So by Lemma 5.6, it is easy to see that ρ∗ L − xC̄ is ample when x ∈ (0, τ ). To show that
ρ∗ L − τ C̄ is base-point-free, we calculate by using (8):
m(ρ∗ L − xC̄) − KW − (DA1 )W
Notice that:
= m(ρ∗ L − xC̄) + (1 + r)ρ∗ L − AX,D (S)C̄
mx + AX,D (S)
∗
= (m + 1 + r) ρ L −
C̄ .
m+1+r
mx + AX,D (S)
AX,D (S)
<τ =
m+1+r
r
if and only if
x<
1 AX,D (S)
1+
.
m
r
Because this is satisfied for
x=τ =
AX,D (S)
r
for any m > 0,
the first statement holds by base-point-free theorem [KM98, Theorem 3.13]. Next we claim
that
H 0 (Y, m(µ∗ L − τ S)) ∼
(11)
= H 0 (S, −mτ S)
for any m sufficiently divisible. To see this, we consider the exact sequence:
0 → OY (m(µ∗ L − τ S) − S) → OY (m(µ∗ L − τ S)) → OY (m(µ∗ L − τ S)) ⊗ OS → 0, (12)
and its associated long exact sequence of cohomology groups. By the above discussion, and
m(µ∗ L − τ S) − S − KY = m(µ∗ L −
AX,D (S)
S) + (1 + r)µ∗ L − AX,D (S)S
r
is ample, it follows from the Kawamata-Viehweg vanishing theorem that
H 1 Y, m(µ∗ L − τ S) ⊗ O(−S) = 0 for any m ≥ 0.
We also have
H 0 Y, m(µ∗ L − τ S) ⊗ O(−S) = 0 for any m ≥ 0.
as τ is also the pseudo-effective threshold. Thus we know |m(ρ∗ L − τ C̄)| contracts the fiber
W ×A1 {0} to C̄ for sufficiently divisible m. This finishes the proof of (1). We denote by
θ : W → X the induced morphism and there is an ample line bundle L on X such that
θ∗ L = ρ∗ L − τ C̄.
29
Next we prove (2). Let (DA1 )X be the push forward of (DA1 )W on X . Then −KX −
(DA1 )X and (1 + r)L coincide outside X0 , they must be linearly equivalent on the whole X
because X0 is irreducible. In particular, they are linearly equivalent when restricted to X0 .
Since
(KY + µ−1
∗ D + S)|S = KS + ∆S ∼Q AX,D (S) · S|S ,
we know that
−KX − (DA1 )X |X0 = −KC̄ − DC̄ ∼Q (1 + AX,D (S))S∞ .
Similarly, we have L|X0 ∼Q τ S with τ =
AX,D (S)
.
r
Therefore,
1 + AX,D (S) = (1 + r)
AX,D (S)
,
r
which implies AX,D (S) = r and τ = 1.
The degree of V∞ under L is
L|n−1
X0 · V ∞
= Ln−1 · V∞
= Ln ,
while the degree of S is
L|n−1
X0 · S
= τ −1 Ln0 = Lnt = Ln0 .
The restriction θ|V∞ : V∞ → S is finite since
(ρ∗ L − τ C̄)|V∞ = L|V∞
is ample. And the degree is 1 by the above calculation on degrees, which implies this is
an isomorphism. We claim Y is indeed the P1 -bundle over V∞ induced by blowing up
the vertex of X, S is a section and the morphism θ is just contracting the P1 -bundle.
Granted this for now, we then indeed have an isomorphism from (X, D + (1 − β)V∞ ; L)
to (C̄, DC̄ + (1 − β)S∞ ; L0 ).
To see the claim, let l be a curve contracted by θ, we want to show that it is the birational
transform of a ruling line of X. To see this, since (ρ∗ (L)− C̄)·l = 0, we know that ρ∗ (L)·l = 1.
So the image ρ∗ l of l in X is a line, and it passes through the vertex. Therefore, it is a ruling
of the cone.
By the above proof, let V be the birational transform of (V∞ )A1 on X , and H the extension
of HA1 \ {0} on X , we know that:
!
+∞
M
X = ProjV
Sk ,
k=0
where
Sk =
k
M
(H 0 (V, iH) · uk−i ).
i=0
From this we easily see that S and V give the same component over the vertex.
5.2
The general case
In this section, we prove Theorem B in the general case. We first show that the cone case
we prove in Section 5.1 can be generalized to orbifold cone. Let T = C∗ .
c X,x is
Proposition 5.9. Let x ∈ (X, D) be a T -singularity. Assume a minimizer v of vol
given by a rescaling of ordF for a Kollár component F , then v is T -equivariant.
30
Proof. Let a be an ideal whose normalized blow up gives the model of extracting the Kollár
component F (see the proof of Proposition 3.11). Denote the degeneration of {a• } := {ap }
by b := {in(ap )} (which in general is not necessarily equal to but only contains (in(a))p ).
Since
n
c
mult(a) · lctn (X, D; a) = vol(ord
F ) ≥ mult(b• ) · lct (X, D; b• ),
But mult(a) = mult(b• ) and lct(X, D, a) ≥ lct(X, D, b• ), we know that
1
lct(X, D, a) = lct(X, D, b• ) = lim lct(X, D, bk ),
k→∞
k
which we denote by c. In particular, for we can choose sufficiently small, and k sufficiently
large, such that the log discrepancy
al (F, X, D + (c − )a) < 1
and (X, D + (c − ) k1 bk ) is klt. Thus by [BCHM10] we can construct a model Z → W =defn
X × A1 extracting only the irreducible divisor FZ which gives F over the generic fiber.
Furthermore, we can assume −FW is ample over W and we denote by W0 = X × {0}.
Therefore, W and Y ×A1 where Y = W ×A1 {t} are isomorphic incodimension 1, with the
exceptional divisors antiample over X × A1 . Thus we conclude that they are isomorphic.
Proposition 5.10. Under the notation in Section 2.3, S is the unique minimizer among all
c (C,C ) if and only if the same holds for C (d) on vol
c (d) (d) (d) .
Kollár components for vol
∆
(C ,C +C )
∆
D
c X,D is T -equivalent,
Proof. By Proposition 5.9, any minimizing Kollár component E of vol
therefore it is G = Z/d equivalent. So E is the pull back of a Kollár component on C (d) by
Lemma 2.9, which can be only the canonical component obtained by blowing up the vertex
by our assumption and Lemma 2.10.
Proof of Theorem B. We first notice that by Theorem 5.7 and Proposition 5.10, we know
that for the coarse moduli space of an orbifold cone over a K-semistable log Fano pair, the
only Kollár component which minimizes the normalized volume function is given by the
canonical component.
Now we consider the case of a general klt singularity o ∈ (X, D). Let us assume that
there is another component F such that
c X,o (ordS ) = vol
c X,o (ordF ),
vol
c X,o by Theorem A.
in particular, they take the minimal value of vol
Let W → B be the family which degenerates X to X0 = Y0 ∪ C̄, where Y0 ∼
= Y extracting
S over X and C̄ is the the coarse moduli space of the orbifold cone over S = C̄ ∩ Y0 . Then
we argue as in the proof Proposition 5.9: Let a be an ideal whose normalized blow up gives
the model of extracting the Kollár component F (see the proof of Proposition 3.11). Denote
the degeneration of {a• } := {ap } by b• := {in(ap )}.
Since
c X (ordF ) = vol
c Y (ordS ) ≤ mult(b• ) · lctn (C̄, b• ),
mult(a) · lctn (X, a) = vol
0
0
0
where the last inequality is from the assumption that S0 is K-semistable and Theorem A.
But mult(a) = mult(b• ) and lct(X, a) ≥ lct(C̄, b• ), we know that
1
lct(X, a) = lct(C̄, b• ) = lim lct(C̄, bk ),
k→∞
k
which we denote by c. In particular, for we can choose sufficiently small, and k sufficiently
large, such that the log discrepancy
al (F, X, D + (c − )a) < 1
and (C, CD + (c − ) k1 bk ) is klt. Thus by [BCHM10] we can construct a model Z → W
extracting only the irreducible divisor FZ which gives F over the generic fiber. Furthermore,
31
we can assume −FW is ample over W and after a base change, we can assume FW |X0 is
reduced.
We claim that Z0 → W0 also only extracts a Kollár component. In fact, locally over the
vertex v of C̄, since
(−FW |X )n = (−FW |X0 )n ,
then by Lemma 3.6,
c X (ordF ) = vol(Z
c 0 /W0 ) = vol
c Y (ordS ).
vol
0
0
c X,o (F ), which is equal to
Thus we conclude that the volume of the model Z0 is equal to vol
c C̄,v . It follows from the argument in the proof of
the minimum of the normalized volume vol
Theorem C that over v, Z0 → W0 yields a Kollár component F0 . By the proof in the cone
case Theorem 5.7, F0 has to be the same as the canonical component S0 . In particular, this
implies that the birational transform P of C̄ in Z is the extract of the canonical component.
Thus there is a morphism P → S. Let l be the fiber class of P → S. Consider KZ +
µ−1
∗ (DA1 ) + FZ , which satisfies that
KZ + µ−1
∗ (DA1 ) + FZ |F0 = KF0 + ∆F0
is anti-ample,
KZ + µ∗−1 (DA1 ) + FZ |S = KS + ∆S
is anti-ample, and
KZ + µ−1
∗ (DA1 ) + FZ · l = 0.
Thus l is an extremal ray in N1 (Z/XA1 ).
Hence we know that there is a morphism Z → W 0 which contracts P and W 0 admits
a morphism χ : W 0 → XA1 . Restricting over 0, the central fiber is the birational model
µ : Y → X which extracts S. On the other hand, let µF : YF → X be the birational model
which extracts the Kollár component F .
As YF × A1 and W 0 is isomorphic in codimension 1, if we denote by FW 0 the push forward
of FZ on W 0 , we have
M
W 0 = Proj
µ1 ∗ OW 0 − mFW 0
m
=
Proj
M
µ2 ∗ OYF ×A1 − m(F × A1 )
m
= YF × A1 .
Consider the central fiber over 0, this implies that YF = Y .
6
Minimizing Kollár component is K-semistable
In this section, we aim to prove the a Kollár component is minimizing only if it is Ksemistable. The method used in the proof of this result is motivated by a method used in
the study of toric degenerations (see [Cal02, Section 3.2], [AB04, Proposition 2.2] and [And13,
Proposition 3]). In particular this method allows us to reduce two-step degenerations to a
one-step degeneration.
(d)
Proof of Theorem D. By Proposition 4.6 we know that the canonical valuation of (C (d) , C∆ +
(d)
CD ) minimizes the normalized local volume if and only if (S, ∆S ) is K-semistable. Thus
by Proposition 4.9, to show that (S, ∆S ) is K-semistable, it suffices to show that the canonc X ,D for (X0 , D0 ) =defn (C, CD ), which is the special
ical component is a minimizer of vol
0
0
degeneration associated to S.
Let (X, D) be a Q-Gorenstein klt singularity with X = SpecC (R). Assume that S is a
c (X,D;o) and appears as the exceptional divisor in a plt
Kollár component that minimizes vol
blow-up X 0 → X. Let ∆S be the divisor on S satisfying KX 0 + S|S = KS + ∆S . By Theorem
1.3 (and Lemma 4.4), to show that (S, ∆S ) is K-semistable, it suffices to show that
c X ,D (ordS ) ≤ vol
c X ,D (ordF )
vol
0
0
0
0
32
for any C∗ -invariant Kollár component F over o0 ∈ (X0 , D0 ). Let (Y, E) be the associated
special degeneration which degenerates (X0 , D0 ) to a pair (Y0 , E0 ) where Y0 is an orbifold
cone over (F, ∆F ). Then we have a Z≥0 × Z≥0 -valued function on R:
w:R
f
−→
7→
Z≥0 × Z≥0
(13)
(ordS (f ), ordF (in(f ))) .
We give Z≥0 × Z≥0 the following lexicographic order: (m1 , u1 ) < (m2 , u2 ) if and only if
m1 < m2 , or m1 = m2 and u1 < u2 . If we denote by Γ the valuative monoid and denote the
associated graded ring by
M
grw R =
R≥(m,u) /R>(m,u) ,
(m,u)∈Γ
then it’s easy to see that Y0 = SpecC (grw R). We will also denote:
M
M
A=
R≥m /R>m =
Am .
m∈N
m∈N
Then SpecC (A) = X0 = Y . Moreover if we define the extended Rees ring of A with respect
to the filtration associated to ordF :
M
M
A0 :=
Ak :=
bk t−k ⊂ A[t, t−1 ],
k∈Z
k∈Z
where bk = {f ∈ A; ordF (f ) ≥ k}. Then the flat family Y → C is given by the SpecC[t] (A0 ).
In particular, we have
A0 ⊗C[t] C[t, t−1 ] ∼
= A[t, t−1 ],
A0 ⊗C[t] C[t]/(t) ∼
= grw R.
Pick up a set of homogeneous generators f¯1 , . . . , f¯p for grw R with deg(f¯i ) = (mi , ui ). Lift
them to generators f1 , . . . , fp for A such that fi ∈ Ami . Set P = C[x1 , . . . , xp ] and give P
the grading by deg(xi ) = (mi , ui ) so that the surjective map
P → grw R
given by
xi 7→ fi
is a map of graded rings. Let ḡ1 , . . . , ḡq ∈ P be a set of homogeneous generators of the kernel
and assume deg(ḡj ) = (nj , vj ).
Since ḡj (f¯1 , . . . , f¯p ) = 0 ∈ grw R, it follows
ḡj (f1 , . . . , fp ) ∈ (Anj )>vj
for each j.
By the flatness of A over C[t], there exist liftings gj ∈ ḡj + (Pnj )>vj of the relation ḡj such
that
gj (f1 , . . . , fp ) = 0 for 1 ≤ j ≤ q.
So gj 0 s form a Gröbner basis of J with respect to the order function ordV , where J is the
kernel of the surjection P → A. In other words, if we let K = (ḡ1 , . . . , ḡq ) denote the kernel
P → A0 , then K is the initial ideal of J with respect to the order determined by ordF . As
a consequence, we have:
A = P [τ ]/(g̃1 , . . . , g̃q ),
where g̃j = τ vj gj (τ −u1 x1 , . . . , τ −up xp ).
Now we lift f1 , . . . , fp further to generators F1 , . . . , Fp of R. Then we have:
gj (F1 , . . . , Fp ) ∈ R>mj .
By the flatness of R over C[t], there exist Gj ∈ gj + P>nj such that
Gj (F1 , . . . , Fp ) = 0.
33
Let I be the kernel of P → R. Then Gj 0 s form a Gröbner basis with respect to the order
function w in (13) and the associated initial ideal is I. As a consequence, we have:
R = P [ζ]/(G̃1 , . . . , G̃q )
where G̃j = ζ nj Gj (ζ −m1 x1 , . . . , ζ −mp xp ).
In summary we have (C∗ )2 action on Cp generated by two 1-parameter subgroups λ0 (t) =
m
t and λ0 (t) = tu . λ0 degenerates (X, D) to (X0 , D0 ) and λ0 degenerates (X0 , D0 ) further
to (Y0 , E0 ).
We now claim that, for 0 < 1, there is a family of one parameter subgroups
λ (t) : C∗ → (C∗ )2 such that the following conditions are satisfied:
1. λ (t) degenerates X to Y0 . For this to happen, we need to make sure that the initial
term of Gj with respect to the weight function π defined by λ (t) is exactly ḡj . For
the latter condition to hold it suffices:
π (n0j , vj0 ) > π (nj , vj ) where (n0j , vj0 ) = deg(Gj − ḡj ).
2. As → 0, λ → λ0 in the sense that λ (t) → λ0 (t) for any t ∈ C∗ .
3. For 0 < 1 ∈ Q, S = Y0 /λ (t) is a Kollár component over (X, o). Moreover as
→ 0, ordT → ordS .
Denote by B ⊂ Z × Z be the finite set consisting of the differences (n0j , vj0 ) − (nj , vj ),
together with 0 and the two generators of N × N. Let M be a positive integer that is larger
than all coordinates of (m, u) − (n, v) for all pairs of elements (m, u), (n, v) ∈ B and let be
sufficiently small such that 1 > M . Define
π = e∗0 − e∗1 .
Then for sufficiently small, π is a linear projection π : Z × Z → Z such that 0 < π(nj , vj ) <
π(n0j , vj0 ).
So by choosing M 1 so that 1 we see that indeed there is a family of linear
projections π approaches π0 which corresponds to ordS . We can define λ (t) to be one
parameter subgroup corresponding to π .
Let λ = λ (t) be the C∗ -action induced by a choice of rational . Now we claim that
Y0 /(λ (t)) yields a Kollár component S over both (X, o) and (X0 , D0 ). Moreover the associated special degeneration has the central fibre equal to (Z, λ ). Because ordS is a minimizer
c X,D , we have:
on vol
c X,D (ordS ) ≥ vol
c X,D (ordS ).
vol
On the other hand, we have
c X,D (ordS ) = vol
c X ,D (ordS ) = vol
c Y ,E (ordS ).
vol
0
0
0
0
c Y ,E is a convex function with respect to . So we conclude vol
c Y ,E (ordF ) ≥
It’s known that vol
0
0
0
0
c Y ,E (ordT ). As a consequence we have:
vol
0
0
0
c X ,D (ordF ) = vol
c Y ,E (ordF ) ≥ vol
c Y ,E (ordT ) = vol
c X ,D (ordS ).
vol
0
0
0
0
0
0
0
0
0
To see the construction of S , we define a filtration:
(
!
)
p
X
a1
N
ap
F R = SpanC f1 . . . fp ; π
ai (mi , ui ) ≥ N
i=1
= {f ; there exists F ∈ P such that F |X = f and λ (F ) ≥ N }.
b p → Cp and the
Then {F N R} is the weighted filtration induced by the weighted blow up C
associated graded ring of the above filtration is isomorphic to grw R with the grading given by
the weight function λ . Denote the strict transform of X by X̂. Then the exceptional divisor
X̂ → X is isomorphic to S = Y0 /λ by the discussion in Section 2.4. By Proposition 6.1,
(S , ∆ ) = (Y0 , E0 )/λ is a klt log-Fano-variety and a Kollár component over (X, D; o).
34
Proposition 6.1. For any 0 < 1 with ∈ Q+ , let (S , ∆ ) = Y0 /λ . Then S is a
Kollár component over o ∈ (X, D).
Proof. For 0 < 1 with ∈ Q+ , ξ generates a C∗ -action. We have a log orbifold C∗ bundle π : (Y0◦ , E0◦ ) := (Y0 \ {v}, E0 \ {v}) → (S , ∆ ) where v is the vertex of Y0 . The Chern
class of this orbifold C∗ -bundle, denoted by c1 (Y0◦ /S ), is contained in Pic(S ) and is ample.
Y0◦ can be compactified by adding the zero section S so that we get a birational morphism
µ : Y → Y0 with the exceptional divisor isomorphic to be S .
Because Y0 has klt singularity, by [Kol04, 40-42] (see also [BG08]) we know that c1 (Y0◦ /S ) =
−1
−r (KS + ∆ ) and (S , ∆ ) has klt singularities. So (S , ∆ ) is a Kollár component over
v ∈ (Y0 , E0 ).
To transfer this to (X, o), we notice that the graded ring of w is isomorphic to grwt C[Y0 ].
The exceptional divisor of the filtered blow-up associated to w is isomorphic to Proj(grwt C[Y0 ])
which is isomorphic to (S , ∆ ). Since (S , ∆ ) is klt, by inversion of adjunction we know
that the filtered blow up is indeed a plt blow up and hence S is a Kollár component over
(X, o).
7
Examples
In this section, we find out the minimizer for some examples of klt singularities (X, o) =
(SpecR, m). We note that by Proposition 2.4 and 3.11, this also explicit calculates
inf lct(X, a)n · mult(a)
a
for all m-primary ideals a and gives the equality condition, which generalizes the results in
[dFEM04] on a smooth point.
Example 7.1. We explicitly compute the minimizer for quotient, Ak , Ek and weakly exceptional singularities in the below.
1. (cf. [LL16, Example 4.9]) Let (X, o) = (Cn , 0)/G be an n-dimensional quotient singularity. Let E ∼
= Pn−1 be the exceptional divisor over Cn obtained by blowing up 0.
Then denote by S the valuation over (X, o) which is the quotient of E by G. Applying
Lemma 2.9 to the pull back of Kollár components on X, we know that
c X,o (ordS ) ≤ vol
c ord
vol
E
c X,o minimizes at ordS with
for any Kollár component E over (X, o). So vol
n
c X,o (ordS ) = n .
vol
|G|
2. Consider the n-dimensional Ak−1 singularity:
k
X = Ank−1 := {z12 + · · · + zn2 + zn+1
= 0}.
We consider cases when k > 2(n−1)
n−2 (for other cases, see [LL16, Example 4.7]). We want
to show that the valuation corresponding to the weight w∗ = (n − 1, · · · , n − 1, n − 2)
c X,o . In [Li15a, Example 2.8], these are
is a minimizer among all valuations in vol
computed out as the minimizer among all valuations obtained by weighted blow ups
on the ambient space Cn+1 .
We notice that under the weighted blow up corresponding to w∗ , we have a birational
morphism Y → X with exceptional divisor S isomorphic to the weighted hypersurface
S := {Z12 + · · · + Zn2 = 0} ⊂ P(n − 1, · · · , n − 1, n − 2) =: Pw∗ .
Because Pw∗ ∼
= P(1, · · · , 1, n − 2), it is easy to see that S is isomorphic to C̄(Q, −KQ )
−1
where Q = Qn−2 = {Z12 + · · · + Zn2 = 0} ⊂ Pn−1 (notice that KQ
= (n − 2)H).
On the other hand, because Pw∗ is not well-formed, we have codimensional 1 orbifold
locus along the infinity divisor Q∞ ⊂ S with the isotropy
group Z/(n− 1)Z. So the
corresponding Kollár component is the log Fano pair S, (1 −
35
1
n−1 )Q∞
. Because Q∞
has KE, by [LL16] there is a conical KE on the pair S, (1 −
1
n−1 )Q∞
. So by Theorem
c that is the unique minimizer
A and Theorem B, ordS is indeed a global minimizer of vol
among all Kollár components. Notice that for any higher order klt perturbation of these
singularities, w∗ is also a minimizer.
3. We can also use Theorem A to verify that the valuations in [Li15a, Example 2.8] for Ek
c X,o , which are unique among Kollár components.
(k=6,7,8) are indeed minimizers in vol
To avoid repetition, we will only do this for E7 singularities. The argument for other
two cases are similar. So consider the (n + 1)-dimensional E7 singularity:
3
3
X n+1 = {z12 + z22 + · · · + zn2 + zn+1
zn+2 + zn+2
= 0} ⊂ Cn+2 .
(a) If n + 1 = 2, then X 2 is a quotient singularity C2 /E7 and so we get the unique
polystable component by [LL16, Example 4.9] and example 1 above.
(b) If n + 1 = 3, then X 3 = {z12 + z22 + z33 z4 + z43 = 0} ⊂ C4 ∼
= {w1 w2 + w33 w4 + w43 =
4
0} ⊂ C by the change of variables. This singularity has a (C∗ )2 -action and is an
example of T -variety of complexity one. By the recent work in [CS15, Theorem
7.1 (II)], X 3 indeed has a Ricci flat cone Kähler metric associated to the canonical
C∗ -action associated to w∗ . So by [LL16, Theorem 1.7], the unique K-polystable
Kollár component is given by the orbifold X 3 /hw∗ i.
(c) n + 1 = 4, then under the weighted blow up corresponding to w∗ = (9, 9, 9, 5, 6),
we have a birational morphism X̂ → X with exceptional divisor E isomorphic to
the weighted hypersurface
E = {z12 + z22 + z32 + z53 = 0} ⊂ P(9, 9, 9, 5, 6) = P(w∗ ).
Since P(w∗ ) is not well-formed, we have:
E∼
= {z12 + z22 + z32 + z53 = 0} ⊂ P(3, 3, 3, 5, 2) = P0 .
with orbifold locus of isotropy group Z/3Z along
V = {z12 + z22 + z32 + z53 = 0} ⊂ P(3, 3, 3, 2).
Alternatively, E is a weighted projective cone over theweighted hypersurface. It is
easy to see that as an orbifold (V, ∆) ∼
= P2 , (1 − 31 )Q where Q = {z12 + z22 + z32 =
2
0} ⊂ P . By [LS14], there exists an orbifold Kähler-Einstein metric on (V, ∆).
Notice that −(KV + ∆) = 3L − 43 L = 53 L where L is the hyperplane bundle of P2 .
Denoting by H the hyperplane bundle of P0 , then H|V = L/3. If V is considered
as a divisor of E, then
5
V |V = {z4 = 0} ∩ E = 5H|V = L.
3
So −(KV + ∆) = V |V .
Then by [LL16, Theorem 1.7], there exists an orbifold Kähler-Einstein metric on
E because the cone angle at infinity is β = 1/3. Thus the unique log-K-semistable
(actually log-K-polystable) Kollár component is given by the pair E, 1 − 13 V .
(d) n + 1 = 5, under the weighted blow up corresponding to w∗ = (3, 3, 3, 3, 2, 2), we
have a birational morphism X̂ → X with exceptional divisor E isomorphic to the
weighted hypersurface:
E = {z12 + z22 + z32 + z42 + z63 = 0} ⊂ P(3, 3, 3, 3, 2, 2) =: P(w∗ ).
This is a weighted projective cone over the weighted hypersurface:
V = {z12 + z22 + z32 + z42 + z63 = 0} ⊂ P(3, 3, 3, 3, 2).
As orbifold, we have (V, ∆) = P3 , (1 − 31 )Q . By [LS14, Li13], (V, ∆) is log-Ksemistable and degenerates to a conical Kähler-Einstein pair. So by [LL16], we
know that (E, (1 − β)V∞ ) is log-K-semistable. To determine β, we notice that
4
8
2
−(KV + ∆) = 4L − L = L = 4 · L = 4 · V∞ |V .
3
3
3
So β = 1 and we conclude that the unique (strictly) K-semistable Kollár component is indeed the Q-Fano variety E.
36
(e) n + 1 ≥ 6. Under the weighted blow up corresponding to w∗ = (n − 1, . . . , n −
1, n − 2, n − 2), we have a birational morphism X̂ → X with exceptional divisor
E isomorphic to the weighted hypersurface:
E = {z12 + · · · + zn2 = 0} ⊂ P(n − 1, · · · , n − 1, n − 2, n − 2) =: P(w∗ ).
This is the weighted projective cone over
V = {z12 + · · · + zn2 = 0} ⊂ P(n − 1, · · · , n − 1, n − 2).
By the discussion in the above Ank−1 singularity case, we know that as an orbifold,
1
)Q∞ , which has an orbifold Kähler-Einstein met(V, ∆) = C̄(Q, −KQ ), (1 − n−1
ric. Notice that
−(KV + ∆) = (n(n − 1) + n − 2)H|V − 2(n − 1)H|V = n(n − 2)H|V .
By [LL16, Theorem 1.7], the Q-Fano variety E indeed has an orbifold KählerEinstein metric (β = n/n = 1 at infinity) and hence by Theorem A is the unique
K-semistable (actually K-polystable) Kollár component.
We remark that, however, in the case of Dk+1 singularities, since the valuations computed out in [Li15a, Example 2.8] could be irrational, our method can not directly tell
whether it is a minimizer in ValX,o .
4. A notion called weakly-exceptional singularity is introduced in [Pro00]. As the name
suggested, this is a weaker notion than the exceptional singularity introduced in [Sho00],
which forms a special class of singularities in the theory of local complements. In our
language, a singularity (X, o) is weakly-exceptional if and only if it has a unique Kollár
component S. We know that if a singularity is weakly-exceptional, then the log αinvariant for the log Fano (S, ∆S ) is at least 1 (see [Pro00, Theorem 4.3] and [CS14]).
In particular, we know that (S, ∆S ) is K-semistable (see [OS14, Theorem 1.4] or [Ber13,
Theorem 3.12]). And by Theorem A and B, we know ordS is the unique minimizer of
c
vol(S)
among all Kollár components. See [CS14] for examples of weakly exceptional
singularities.
The following example is indeed a prototype of our study.
Example 7.2. If a log terminal singularity o ∈ X has a quasi-regular Sasakian-Einstein
c x,X .
metric, then the C∗ -quotient provides a Kollár component which minimizes vol
Assume X = SpecR is an affine variety with effective torus action and Q-Gorenstein klt
isolated singularity at o ∈ X. Denote the torus by T := (Gm )d . Then we have a weight
decomposition:
M
R=
Rα ,
α∈Γ
where
• For any α ∈ Zd , Rα = {f ∈ R ; t · f = tα f for any t ∈ (Gm )d }.
• Γ = {α ∈ Zd ; Rα 6= 0}.
The Reeb cone of X is the following conic subset of the Lie algebra of T = (Gm )d :
RC := {ξ ∈ Rd ; hα, ξi > 0 for any α ∈ Γ}.
Notice that the elements in thePReeb cone RC can be considered as holomorphic vector fields
d
on X via the map ξ 7→ σξ := i=1 ξi σi where σi is the infinitesimal generator of the action
of the i-th factor of (Gm )d . If we denote by {ei } the standard basis of Rd , then
σi (p) =
d
ds
exp(sei ) ◦ p −
√
s=0
−1
d
ds
√
exp( −1sei ) ◦ p.
s=0
Suppose that there is a Kähler cone metric gξ = dr2 + r2 gL on X where
37
1. r : X − {o} → R>0 is a smooth radius function, L = {r = 1} ⊂ X is the link of the
isolated singularity (X, o).
√
∂
∂
2. r ∂r
− −1J(r ∂r
) = σξ .
Now assume that gξ is a Ricci-flat Kähler metric, equivalently the link L has a Sasaki-Einstein
metric (see [BG08]). There are two different cases:
∂
) is closed, gL is called a quasi1. (quasi-regular case) If the generic orbit of ∂θ := J(r ∂r
regular Sasaki-Einstein metric. σξ generates a Gm -subgroup of T. the quotient X/Gm
is an Fano orbifold (V, ∆) with orbifold Kähler-Einstein metric. By [LL16] ordV is a
c X,o . By Theorem B, this is the unique minimizer among all Kollár
minimizer of vol
components.
2. (irregular case) If the generic orbit of ∂θ is not closed, gL is called an irregular SasakiEinstein metric. Explicit examples can be found in [GMSW04a, GMSW04b, FOW09].
Since ξ ∈ Rd is an irrational vector, we can approximate it by a sequence of rational
vectors {ξk } ⊂ RC ∩ Qd . Each ξk generates a Gm -subgroup Gk of T, and we get
a sequence of quotients X/Gk = (Vk , ∆k ). It is easy to see that Vk are all Kollár
components. It is possible to show that a suitable rescaling ck · ordVk converges to
c (X,o) , which corresponds to the vector field σξ . The details will be
a minimizer of vol
discussed in a future work.
As we mentioned, there is a related differential geometry study on the metric tangent
cone.
Example 7.3. In the notation of [DS15], let p ∈ Z be a singularity appearing on the GromovHausdorff limit of a sequence of KE Fano manifolds. Assume the Reeb vector of the weight
tangent cone W is quasi-regular, then the C∗ -quotient of W indeed gives a minimizer of
p ∈ Z.
Assume (X, o) is an algebraic germ on a normal affine variety that is embedded into CN .
Assume that we have a weight vector w = (a1 , . . . , aN ) ∈ ZN
>0 satisfying that there is no
common factor of ai . Then it defines a filtration on R = OX,o :
F k = {f ∈ R; there exists g ∈ OCN ,o with g|X = f and w(g) ≥ k}.
(14)
From this we can form the Rees algebra and extended Rees algebra:
R=
+∞
M
F k t−k ,
R0 =
k=0
M
F k t−k ,
k∈Z
where F k = R if k ≤ 0. We also have the associated graded ring of (R, F):
G=
+∞
M
F k /F k+1 · t−k ∼
= R0 /tR0 .
k=0
In our discussion, we always assume that G is a finitely generated C-algebra that is also a
normal domain. Then we can get a normal affine variety W = Spec(G). As in [DS15], we
call W to be a weighted tangent cone of (X, o). Notice that there is a natural C∗ -action
on G and hence it is an orbifold cone over the base E = Proj(G) (by Pinkham-Demazure’s
construction).
d
N → CN given by:
On the other hand, the weight w defines a weighted blow up πw : C
!
+∞
M
−k
d
N
C = Proj
Ak · t
,
k=0
where
Ak = {g ∈ C[z1 , · · · , zn ]; w(g) ≥ k}.
Notice that the exceptional divisor of πw is the weighted projective space with weight w. We
have the following observation, which implies that the discussion can be put in the setting
of filtered blowing-ups as studied in [TW89].
38
Lemma 7.4.
1. If Y denotes the strict transform of X under πw , then Y = Proj(R). In
particular the Rees algebra R is finitely generated. Moreover, the exceptional divisor of
ψ
the natural birational morphism Y −→ X is isomorphic to E.
2. The filtration F = {F k }k≥0 is the same as the filtration induced by the divisorial
valuation ordE .
Proof. Denoting by ιX : X → CN the embedding morphism, then F k in (14) coincides with
the inverse image ideal ι−1
X Ak · OX,o (cf. [Har77, 7.15]). So Y is indeed the strict transform
of X under the weighted blow up πw . The second statement is standard (see [TW89, 1.2.1]).
Item 2 follows from the discussion in Section 2.4 (see [TW89, (2.2)]).
If (X, o) is a klt singularity appearing in a Gromov-Hausdorff limit of a sequence of KählerEinstein Fano manifolds, Donaldson-Sun ([DS15]) constructed an affine variety W using the
metric structure on the Gromov-Hausdorff limit and showed that it specially degenerates to
the metric tangent cone C(Y ). The degeneration can be realized under a common embedding
of W and C(Y ) into some ambient CN . More precisely W is a weighted tangent cone
associated to some weight under the common embedding. The weight is determined by the
Reeb vector field of the singular Ricci flat metric on C(Y ). If the Reeb vector field is quasiregular, that is, if it generates a C∗ -action, then the weight can be normalized to have integer
components so that we are in the situation discussed above. Moreover it is shown in this case
E = W \ {0}/C∗ specially degenerates to C(Y ) \ {0}/C∗ and C(Y ) \ {0}/C∗ admits a weak
Kähler-Einstein metric (see [DS15]). This implies E is analytically K-semistable, which then
implies it is (algebraically) K-semistable. So we can apply our result Theorem A to see that
c over (X, o), which is the unique minimizer among
ordE is indeed a global minimizer of vol
all Kollár components.
References
[AB04]
V. Alexeev, M. Brion, Toric degenerations of spherical varieties. Selecta Math.
(N.S.) 10 (2004), no. 4, 453–478.
[And13] D. Anderson, Okounkov bodies and toric degenerations. Math. Ann. 356 (2013),
no. 3, 1183–1202.
[BCHM10] C. Birkar, P. Cascini, C. D. Hacon, and J. Mc Kernan; Existence of minimal
models for varieties of log general type, J. Amer. Math. Soc. 23 (2010), 405-468.
[Ber13]
R. Berman; A thermodynamic formalism for Monge-Ampère equations, MoserTrudinger inequalities and Kähler-Einstein metrics, Adv. Math. 248 (2013), 12541297.
[Ber15]
R. Berman; K-polystability of Q-Fano varieties admitting Kähler-Einstein metrics,
Invent. Math. 203 (2015), no. 3, 973-1025.
[BG08]
C. Boyer and K. Galicki; Sasakian Geometry, Oxford Mathematical Monographs,
Oxford University Press, Oxford, 2008.
[BHJ15] S. Boucksom, T. Hisamoto and M. Jonsson. Uniform K-stability, DuistermaatHeckman measures and singularities of pairs, arXiv:1504.06568.
[Blu16a] H. Blum; On Divisors Computing MLD’s and LCT’s. arXiv:1605.09662.
[Blu16b] H. Blum; Existence
arXiv:1606.08894.
[Cal02]
of
valuations
with
smallest
normalized
volume,
P. Caldero. Toric degenerations of Schubert varieties. Transform. Groups 7 (2002),
51-60.
[CDS15] X. Chen, S. Donaldson and S. Sun; Kähler-Einstein metrics on Fano manifols, I-III,
J. Amer. Math. Soc. 28 (2015), 183-197, 199-234, 235-278.
[CS14]
I. Chelsov and C. Shramov; Weakly-exceptional singularities in higher dimensions,
J. Reine Angew. Math., 689 (2014), 201-241.
[CS15]
T. Collins and G. Székelyhidi;
arXiv:1512.07213
39
Sasaki-Einstein metrics and K-stability,
[Cut12] S.D. Cutkosky; Multiplicities associated to graded families of ideals, Algebra Number Theory 7 (2013), no. 9, 2059-2083.
[DS15]
V. Datar and G. Székelyhidi; Kähler-Einstein metrics along the smooth continuity
method, arXiv:1506.07495.
[Don02] S. Donaldson; Scalar curvature and stability of toric varieties, J. Diff. Geom. 62
(2002), no. 2, 289-349.
[dFEM04] T. de Fernex, L. Ein and M. Mustaţă; multiplicities and log canonical threshold,
J. Alg. Geom. 13 (2004), 603-615.
[DS15]
S. Donaldson and S. Sun; Gromov-Hausdorff limits of Kähler manifolds and algebraic geometry, II, to appear in J. Diff. Geom., arXiv:1507.05082.
[Eis94]
D. Eisenbud; Commutative algebra with a view toward algebraic geometry, Grad.
Texts in Math. 150, Springer-Verlag, 1994.
[ELS03] L. Ein, R. Lazarsfeld and K. E. Smith; Uniform approximation of Abhyankar valuation ideals in smooth function fields, Amer. J. Math. 125 (2003), 409-440.
[FOW09] A. Futaki, H. Ono and G. Wang; Transverse Kähler geometry of Sasaki manifolds
and toric Sasakian-Einstein manifolds, J. Diff. Geom., 83 (2009) 585-635.
[Fuj15]
K. Fujita; Optimal bounds for the volumes of Kähler-Einstein Fano manifolds, to
appear in Amer. J. Math., arXiv:1508.04578.
[Fuj16]
K. Fujita; A valuative criterion for uniform K-stability of Q-Fano varieties, to
appear in J. Reine Angew. Math., arXiv:1602.00901.
[Ful13]
M. Fulger; Local volumes of Cartier divisors over normal algebraic varieties, Ann.
Inst. Fourier (Grenoble) 63 (2013), no. 5, 1793-1847.
[GMSW04a] J.P. Gauntlett, D. Martelli, J. Sparks and W. Waldram; Sasaki- Einstein metrics on S2 × S3 , Adv. Theor. Math. Phys. 8 (2004), 711- 734.
[GMSW04b] J.P. Gauntlett, D. Martelli, J. Sparks and W. Waldram; A new infinite class
of Sasaki-Einstein manifolds, Adv. Theor. Math. Phys. 8 (2004), 987-1000.
[Har77] R. Hartshorne; Algebraic Geometry. Springer-Verlag, New York, 1977, Graduate
Texts in Mathematics, No. 52.
[HLP12] E. Hrushovski, F. Loeser, B. Poonen. Berkovich spaces embed in Euclidean spaces.
arXiv:1210.6485.
[HX13]
C. Hacon and C. Xu; Existence of log canonical closures, Invent. Math. 192 (2013),
no. 1, 161-195.
[JM12]
M. Jonsson and M. Mustaţă; Valuations and asymptotic invariants for sequences
of ideals, Ann. Inst. Fourier (Grenoble) 62 (2012), no. 6, 2145-2209.
[KM98] J. Kollár and S. Mori; Birational geometry of algebraic varieties, Cambridge Tracts
in Math. 134, Cambridge Univ. Press, Cambridge, 1998.
[Kol04] J. Kollár, Seifert Gm -bundles, arXiv:0404386.
[Kol07] J. Kollár; Lectures on resolution of singularities. Annals of Mathematics Studies,
166. Princeton University Press, Princeton, NJ, 2007
[Kol13] J. Kollár; Singularities of the Minimal Model Program, Cambridge Tracts in Math.,
200, Cambridge Univ. Press, Cambridge, 2013.
[Laz04a] R. Lazarsfeld; Positivity in algebraic geometry, I: Classical setting: line bundles and
linear series, Erger Mathematik und ihrer Grenzgebiete. (3) 48, Springer-Verlag,
Berlin, 2004.
[Laz04b] R. Lazarsfeld; Positivity in algebraic geometry, II. Positivity for vector bundles,
and multiplier ideals. Erger Mathematik und ihrer Grenzgebiete. (3) 49, SpringerVerlag, Berlin, 2004.
[Li11]
C. Li; Remarks on logarithmic K-stability, Commun. Contemp. Math. 17 (2015),
no. 2, 1450020, 17 pp.
[Li13]
C. Li; Yau-Tian-Donaldson correspondence for K-semistable Fano manifolds, to
appear in J. Reine Angew. Math., arXiv:1302.6681.
40
[Li15a]
[Li15b]
[Liu16]
[LL16]
C. Li; Minimizing normalized volumes of valuations, arXiv:1511.08164.
C. Li; K-semistability is equivariant volume minimization, arXiv:1512.07205v3.
Y. Liu; The volume of singular Kähler-Einstein Fano varieties, arXiv:1605.01034.
C. Li and Y. Liu; Kähler-Einstein metrics and volume minimization,
arXiv:1602.05094.
[LM09] R. Lazarsfeld and M. Mustaţă; Convex bodies associated to linear series, Ann. Sci.
Éc. Norm. Supér. 42 (2009), no. 5, 783-835.
[LS14] C. Li, S. Sun; Conical Kähler-Einstein metrics revisited, Comm. Math. Phys. 331
(2014), no. 3, 927-973.
[LWX14] C. Li, X. Wang and C. Xu; Degeneration of Fano Kähler-Einstein manifolds,
arXiv:1411.0761.
[LX14] C. Li, and C. Xu; Special test configurations and K-stability of Fano varieties, Ann.
of Math. (2) 180 (2014), no.1, 197-232.
[MSY08] D. Martelli, J. Sparks and S.-T. Yau; Sasaki-Einstein manifolds and volume minimisation, Comm. Math. Phys. 280 (2008), 611-673.
[Mus02] M. Mustaţă; On multiplicities of graded sequences of ideals, Journal of Algebra 256
(2002), 229-249.
[Oda12] Y. Odaka; The Calabi conjecture and K-stability. Int. Math. Res. Not. (2012), no.
10, 2272-2288.
[Oda13] Y. Odaka; The GIT stability of polarized varieties via discrepancy, Ann. of Math.
(2) 177 (2013), no. 2, 645-661.
[OS14] Y. Odaka and Y. Sano; Alpha invariant and K-stability of Q-Fano varieties. Advances in Mathematics 229, Issue 5, 2818-2834.
[OX12] Y. Odaka and C. Xu; Log-canonical models of singular pairs and its applications,
Math. Res. Lett. 19 (2012), no.2, 325-334.
[Pro00] Y. Prokhorov; Blow-ups of canonical singularities, Algebra (Moscow, 1998), 301317, de Gruyter, Berlin, 2000.
[Ree88] D. Rees, Lectures on the asymptotic theory of ideals, London Mathematical Society
Lecture Notes Series 113, Cambridge University Press, 1988.
[Sho00] V. Shokurov; Complements on surfaces, Algebraic geometry, 10. J. Math. Sci. (New
York) 102 (2000), no. 2, 3876-3932.
[SSY16] C. Spotti, S. Sun and S. Yao; Existence and deformations of Kähler-Einstein metrics
on smoothable Q-Fano varieties, Duke Math. J. 165 (2016), no. 16, 3043–3083.
[Sze16] G. Székelyhidi, The partial C 0 -estimate along the continuity method. J. Amer.
Math. Soc. 29 (2016), no. 2, 537–560.
[Tia97] G. Tian; Kähler-Einstein metrics with positive scalar curvature, Invent. Math. 130
(1997), no. 1, 1-37.
[Tia15] G. Tian; K-stability and Kähler-Einstein metrics, Comm. Pure and Applied Math.,
68 (2015), no. 7, 1085-1156.
[TW89] M. Tomari and K. Watanabe; Filtered rings, filtered blowing-ups and normal twodimensional singularities with “star-shaped” resolution, Publ. RIMS. Kyoto Univ.
25 (1989), 681-740.
[TW04] S. Takagi and K. Watanabe; On F -pure thresholds, J. Algebra 282 (2004), 278-297.
[WX14] X. Wang and C. Xu; Nonexistence of asymptotic GIT compactification, Duke Math.
J. 163 (2014), no. 12, 2217-2241.
[Xu14] C. Xu; Finiteness of algebraic fundamental groups, Compos. Math. 150 (2014), no.
3, 409-414.
Chi Li
Department of Mathematics, Purdue University, West Lafayette, IN 47907-2067
[email protected]
Chenyang Xu
Beijing International Center for Mathematical Research, Beijing 100871, China
[email protected]
41
| 0math.AC
|
PRODUCTS OF IDEALS MAY NOT BE GOLOD
arXiv:1506.09129v2 [math.AC] 12 Sep 2016
ALESSANDRO DE STEFANI
Abstract. We exhibit an example of a product of two proper monomial ideals such that the
residue class ring is not Golod. We also discuss the strongly Golod property for rational powers of
monomial ideals, and introduce some sufficient conditions for weak Golodness of monomial ideals.
Along the way, we ask some related questions.
1. Introduction
Let k be a field, and let (R, m, k)
Ldenote a Noetherian positively graded k-algebra, with R0 = k
and irrelevant maximal ideal m = i>1 Ri . Consider the Poincaré series of R
X
i
PR (t) =
dimk TorR
i (k, k)t
i>0
which is, in general, not rational [Ani82]. If n = dimk m/m2 is the embedding dimension of R, Serre
showed that PR (t) is bounded above term by term by the following rational series
(1 + t)n
X
.
1−t
dimk (Hi (R)) ti
i>1
Here Hi (R) is the i-th homology of the Koszul complex on a minimal homogeneous generating
set of m, over the ring R. The ring R is called Golod if equality holds. As a consequence, Golod
rings have rational Poincaré series. The main purpose of this article is to answer, in negative, the
following question
Question 1.1. [MP13, Problem 6.18] Let k be a field, and let (R, m, k) be positively graded
k-algebra. Let I, J be two proper homogeneous ideals in R. Is the ring R/IJ always Golod?
As reported in [MP13], Question 1.1 was first asked by Volkmar Welker. The general belief,
supported by strong computational evidence, was that this question had positive answer. The first
result in this direction is a theorem of Herzog and Steurich [HS79]: let S be a polynomial ring over
a field, and let I,J be two proper homogeneous ideals of S. If I ∩ J = IJ, then S/IJ is Golod.
Another reason to believe that Question 1.1 had positive answer comes from a result of Avramov
and Golod [AG71], which says that Golod rings are never Gorenstein, unless they are hypersurfaces.
This is consistent with a result of Huneke [Hun07], according to which S/IJ is never Gorenstein,
unless I and J are principal. More recently, Herzog and Huneke show that, if I is a homogeneous
ideal in a polynomial ring S over a field of characteristic zero, then, for all d > 2, the ring S/I d is
Golod [HH13, Theorem 2.3 (d)]. In [SFW14, Theorem 1.1] Seyed Fakhari and Welker write that
any product of proper monomial ideals in a polynomial ring over a field is Golod. The key step in
their proof is to show that products of monomial ideals always satisfy the strong-GCD condition.
This condition is the existence of a linear order on a minimal monomial generating set of the ideal,
satisfing certaintain properties [Jöl06, Definition 3.8]. The fact that monomial ideals that satisfy
the strong-GCD condition are Golod is first stated by Jöllenbeck in [Jöl06, Theorem 7.5], provided
2010 Mathematics Subject Classification. Primary 13A02; Secondary 13D40.
Key words and phrases. Golod rings; product of ideals; Koszul homology; Koszul cycles; strongly Golod.
1
2
ALESSANDRO DE STEFANI
an extra assumption, called Property (P), is satisfied, and then by Berglund and Jöllenbeck in
[BJ07, Theorem 5.5], where the extra assumption is removed.
In Section 2, we provide examples of products of proper monomial ideals in a polynomial ring S
over a field, such that the residue class ring is not Golod. For instance, Example 2.1:
Example. Let k be a field, and let S = k[x, y, z, w] be a polynomial ring, with standard grading.
Consider the monomial ideals m = (x, y, z, w) and J = (x2 , y 2 , z 2 , w2 ) inside S. Let
I := mJ = (x3 , x2 y, x2 z, x2 w, xy 2 , y 3 , y 2 z, y 2 w, xz 2 , yz 2 , z 3 , z 2 w, xw2 , yw2 , zw2 , w3 )
be their product, and set R = S/I. Then, the ring R is not Golod.
Our example satisfies the strong-GCD condition. Indeed, the argument of [SFW14, Theorem 1.1]
is correct, but it only shows that products of monomial ideals satisfy the strong-GCD condition.
We have not been able to locate specifically where the mistake in [Jöl06] or [BJ07] may be.
In Section 3 we study the strongly Golod property for rational powers of monomial ideals. Let
S = k[x1 , . . . , xn ] be a polynomial ring over a field k of characteristic zero, and let I ⊆ S. In [HH13]
Herzog and Huneke introduce the following notion: I is called strongly Golod if ∂(I)2 ⊆ I, where
∂(I) is the ideal of S generated by the partial derivatives of elements in I. The main point of this
definition is that, if an ideal I is strongly Golod, then the ring S/I is Golod [HH13, Theorem 1.1].
Among other things, in Section 3 we show that if I is a strongly Golod monomial ideal, then so is
I p/q , for any p > q. This generalizes [HH13, Proposition 3.1].
It is easy to find examples of ideals that are Golod, but not strongly Golod. In [HH13], Herzog
and Huneke introduce the notion of squarefree strongly Golod ideal, that applies to squarefree
monomial ideals. This is a weakening of the strongly Golod definition, but it still implies that
the multiplication on the Koszul homology is identically zero. We will say that a ring is weakly
Golod if the multiplication on Koszul homology is trivial. In [BJ07, Theorem 5.1], Berglund and
Jöllenbeck show that, in case the ideal in question is monomial, weak Golodness and Golodness are
equivalent notions. Herzog and Huneke use this result in [HH13, Theorem 3.5] to conclude that
squarefree strongly Golod ideals are Golod. See Section 5 for more discussions and questions about
this topic. In Section 4 we introduce lcm-strongly Golod monomial ideals, which are a more general
version of squarefree strongly Golod ideals. We show that lcm-strongly Golod ideals are weakly
Golod. In Section 5, we give some sufficient conditions for an ideal to be strongly Golod, and we
ask several related questions. In Appendix A we record a minimal free resolution for Example
2.5, for convenience of the reader. All computations are made using the computer software system
Macaulay2 [GS].
2. Examples of products that are not Golod
Golod rings were named after Evgenii S. Golod, who proved that the upper bound in Serre’s
inequality is achieved if and only if the Eagon resolution is minimial [Gol62]. This happens if and
only if all the Massey operations of the ring vanish. Since the vanishing of the second Massey
operation means that every product of Koszul cycles of positive homological degree is a boundary,
Golod rings have, in particular, trivial multiplication on the positive degree Koszul homology. We
will use this fact in the proofs of our examples in this section. See [GL69, Chapter 4] or [Avr98,
Section 5.2] for details and more general statements.
If (R, m, k) is a Noetherian positively graded algebra over a field k, we can write R ∼
= S/I, where
S = k[x1 , . . . , xn ] is a polynomial ring, and I ⊆ S is a homogeneous ideal. If m = (x1 , . . . , xn )
denotes the irrelevant maximal ideal of S, we can always assume that I ⊆ m2 . Let K• be the
Koszul complex on the elements x1 , . . . , xn of S, which is a minimal free resolution of k over S. We
have that K1 is a free
V S-module of rank n, and we denote by {ex1 , . . . , exn } a basis. In addition,
∼
we have that Ki = i K1 for all i = 1, . . . , n, and the differential δi : Ki → Ki−1 on a basis element
PRODUCTS OF IDEALS MAY NOT BE GOLOD
3
is given by
δi (et1 ∧ . . . ∧ eti ) =
i
X
(−1)j−1 tj et1 ∧ . . . ∧ etj−1 ∧ etj+1 ∧ . . . ∧ eti ,
j=1
and extended by linearity to Ki . Let K• (R) = K• ⊗S R be the Koszul complex on R. We denote
by Z• (R) the Koszul cycles, and by H• (R) the Koszul homology on R.
We are now ready for the first example.
Example 2.1. Let k be a field, and let S = k[x, y, z, w], with the standard grading. Let m =
(x, y, z, w) be the irrelevant maximal ideal, consider the monomial ideal J = (x2 , y 2 , z 2 , w2 ) and let
I := mJ = (x3 , x2 y, x2 z, x2 w, xy 2 , y 3 , y 2 z, y 2 w, xz 2 , yz 2 , z 3 , z 2 w, xw2 , yw2 , zw2 , w3 ).
Then, the ring R = S/I is not Golod.
Proof. Golod rings have trivial multiplication on H• (R)>1 . Therefore, to show that R is not Golod,
it is enough to show that there exist two elements α, β ∈ H• (R)>1 such that αβ 6= 0. Consider the
element u = (ex ∧ ey ) ⊗ xy ∈ K2 (R). It is a Koszul cycle:
δ2 (u) = ey ⊗ x2 y − ex ⊗ xy 2 = 0 in K1 (R),
because x2 y ∈ I and xy 2 ∈ I. Then, let α := [u] ∈ H2 (R) be its residue class in homology.
Similarly, let v = (ez ∧ ew ) ⊗ zw ∈ Z2 (R), and let β := [v] ∈ H2 (R). We want to show that
uv = (ex ∧ ey ∧ ez ∧ ew ) ⊗ xyzw ∈ Z4 (R) is not a boundary, so that αβ = [uv] 6= 0 in H4 (R). Note
that K5 (R) = 0, hence such a product is zero in homology if and only if xyzw ∈ I. But xyzw ∈
/ I,
as every monomial generator of I contains the square of a variable.
Remark 2.2. We keep the same notation as in Example 2.1. Using Macaulay2 [GS], one can
compute the first Betti numbers of k over R:
...
/ R11283
/ R2312
/ R493
/ R98
/ R22
/ R4
/R
/k
/ 0.
Therefore the Poincaré series of R is
PR (t) = 1 + 4t + 422t2 + 98t3 + 493t4 + 2312t5 + 11283t6 + . . .
On the other hand, the upper bound given by Serre’s inequality is
(1 + t)4
= 1 + 4t + 22t2 + 98t3 + 493t4 + 2313t5 + 11288t6 + . . .
1 − 16t2 − 30t3 − 20t4 − 5t5
Since the two series are not coefficientwise equal, R is not Golod. We also checked that R is not
Golod using the Macaulay2 command isGolod(S/I) which computes the generators of all the
Koszul homology modules, and determines whether their products are zero.
Example 2.3. If one is looking for an example where the ideals are generated in higher degrees, for
j > 1 one can consider, along the lines of Example 2.1, the following family of products, suggested
to us by Aldo Conca:
(xj+1 , y j+1 , z j+1 , wj+1 )(xj , y j , z j , wj ) ⊆ k[x, y, z, w].
As in Example 2.1, one can show that the product of cycles
((ex ∧ ey ) ⊗ xj y j ) · ((ez ∧ ew ) ⊗ z j wj )
is not zero as a cycle and, hence, in homology.
Remark 2.4. We want to point out that Example 2.1 is not the first example of a non-Golod
product of ideals that we discovered. In fact, Example 2.1 was suggested to the author by Srikanth
Iyengar, after some discussions about Example 2.5. Given the proof of Example 2.1, it becomes
easy to show that the ring of Example 2.5 is not Golod. In fact, going modulo a regular sequence
of linear forms in the ring of Example 2.5, one obtains a ring isomorphic to the one of Example 2.1.
4
ALESSANDRO DE STEFANI
Then, one can use [Avr98, Proposition 5.2.4 (2)], adapted to the graded case. The original proof
that Example 2.5 is not Golod is much more involved. Nonetheless, since this was the first example
discovered by this author, we want to briefly describe the argument in the rest of this section.
The original proof relies on lifting Koszul cycles. More specifically, we use the double-complex
proof of the fact that TorS• (k, S/I) can be computed in two ways, to lift a Koszul cycle to a specific
element of a finitely generated k-vector space. The results that we use are very well known, so we
will not explain all the steps. We refer the reader to [Wei94] or [Rot09] for more details.
Let S = k[x1 , . . . , xn ] be a polynomial ring over a field k, not necessarily standard graded, and
let m be the irrelevant maximal ideal. Let I ⊆ m2 be a homogeneous ideal in S, and consider
the residue class ring R = S/I. Since K• is a free resolution of k over S, we have that Hi (R) :=
Hi (K• ⊗S R) ∼
= TorSi (k, R), and its dimension as a k-vector space is the i-th Betti number, βi , of
R as an S-module. On the other hand, if F• → R → 0 is a minimal free resolution of R over S,
then Hi (k ⊗S F• ) ∼
= k ⊗ Fi is also isomorphic to TorSi (k, R). There is map ψ : Zi (R) → k ⊗ Fi ,
which is constructed by ”lifting cycles”. Since the boundaries map to zero via ψ, this induces a
map ψ : Hi (R) → k ⊗S Fi , which is an isomorphism. See [Her92] for a canonical way to construct
Koszul cycles from elements in k ⊗ Fi (that is, a canonical choice of an inverse for ψ).
We are now ready to illustrate the example. We refer the reader to Appendix A for an explicit
expression of the differentials in a resolution of R = S/I as a module over S.
Example 2.5. Let k be a field, and let S = k[a, b, c, d, x, y, z, w]. Consider the monomial ideals
I1 = (ax, by, cz, dw) and I2 = (a, b, c, d) inside S. Let
I := I1 I2 = (a2 x, abx, acx, adx, aby, b2 y, bcy, bdy, acz, bcz, c2 z, cdz, adw, bdw, cdw, d2 w)
be their product, and set R = S/I. Then, the ring R is not Golod.
Proof. Let 0 → F4 → F3 → F2 → F1 → F0 → R → 0 be a minimal free resolution of R over S,
with maps ϕj : Fj → Fj−1 , j = 1, . . . , 4, and ϕ0 : F0 = S → R being the natural projection. For
(i)
each i = 0, . . . , 4 and each free module Fi = S βi fix standard bases Ej , j = 1, . . . , βi . In this way,
the differentials can be represented by matrices (see Appendix A for an explicit description). We
have the following staircase:
S ⊗S S 5
δ0 ⊗1S 5
1S ⊗ϕ4
K1 ⊗S S 20
δ1 ⊗1S 20
1K1 ⊗ϕ3
K2 ⊗S S 30
δ2 ⊗1S 30
1K2 ⊗ϕ2
K3 ⊗S S 16
δ3 ⊗1S 16
1K3 ⊗ϕ1
K4 ⊗S S
δ4 ⊗1S
1K4 ⊗ϕ0
K4 ⊗S R
/ K3 ⊗S S
/ K2 ⊗S S 16
/ K1 ⊗S S 30
/ S ⊗S S 20
/ k ⊗S S 5
PRODUCTS OF IDEALS MAY NOT BE GOLOD
5
Let u = (ex ∧ey )⊗(ab), and v = (ez ∧ew )⊗(cd), inside K2 (R) = K2 ⊗S R. As they are cycles, we can
consider their classes α = [u] and β = [v] in homology. We want to construct a lifting ψ(uv) of the
Koszul cycle uv = (ex ∧ ey ∧ ez ∧ ew ) ⊗ (abcd) ∈ K4 (R). Given (ex ∧ ey ∧ ez ∧ ew ) ⊗ (abcd) ∈ K4 ⊗S R
(0)
we consider the lift (ex ∧ ey ∧ ez ∧ ew ) ⊗ (abcd E1 ) ∈ K4 ⊗S S, and then apply the differential
δ4 ⊗ 1S :
(0)
+(ey ∧ ez ∧ ew ) ⊗ (abcdxE1 )
(0)
−(e
(0)
x ∧ ez ∧ ew ) ⊗ (abcdy E1 )
(δ4 ⊗ 1S )((ex ∧ ey ∧ ez ∧ ew ) ⊗ (abcd E1 )) =
(0)
+(ex ∧ ey ∧ ew ) ⊗ (abcdz E1 )
(0)
−(ex ∧ ey ∧ ez ) ⊗ (abcdw E1 )
This is now a boundary, and, in fact, it is equal to
(1)
+(ey ∧ ez ∧ ew ) ⊗ (cd E2 )
(1)
−(ex ∧ ez ∧ ew ) ⊗ (cd E5 )
(1K3 ⊗ ϕ1 )
(1)
+(ex ∧ ey ∧ ew ) ⊗ (ab E12
)
(1)
−(ex ∧ ey ∧ ez ) ⊗ (ab E15 )
Now we apply δ3 ⊗ 1S 16 to this element:
(1)
(1)
+(ez ∧ ew ) ⊗ (cdy E2 − cdx E5 )
(1)
(1)
(1)
+(ey ∧ ez ∧ ew ) ⊗ (cd E2 )
−(ey ∧ ew ) ⊗ (cdz E2 − abx E12 )
(1)
(1)
(1)
−(ex ∧ ez ∧ ew ) ⊗ (cd E5 ) +(ey ∧ ez ) ⊗ (cdw E2 − abx E15 )
=
(δ3 ⊗ 1S 16 )
(1)
(1)
(1)
+(ex ∧ ey ∧ ew ) ⊗ (ab E12
) +(ex ∧ ew ) ⊗ (cdz E5 − aby E12 )
(1)
(1)
(1)
−(ex ∧ ey ∧ ez ) ⊗ (ab E15 )
−(ex ∧ ez ) ⊗ (cdw E5 − aby E15 )
(1)
(1)
+(ex ∧ ey ) ⊗ (abw E12 − abz E15 )
This is a boundary. Namely, it is equal to
(2)
−(ez ∧ ew ) ⊗ (cd E13 )
(2)
(2)
(2)
+(ey ∧ ew ) ⊗ (dz E3 + bx E17 + bd E20 )
−(ey ∧ ez ) ⊗ (cw E (2) + bx E (2) + bc E (2) )
5
23
28
(1K2 ⊗ ϕ2 )
−(e ∧ e ) ⊗ (dz E (2) + ay E (2) + ad E (2) )
x
w
8
18
21
(2)
(2)
(2)
+(ex ∧ ez ) ⊗ (cw E10
+ ay E24 + ac E29 )
(2)
−(ex ∧ ey ) ⊗ (ab E30 )
We now apply the map δ2 ⊗ 1S 30 to such a lift:
(2)
−(ez ∧ ew ) ⊗ (cd E13 )
(2)
(2)
(2)
+(ey ∧ ew ) ⊗ (dz E3 + bx E17 + bd E20 )
−(ey ∧ ez ) ⊗ (cw E (2) + bx E (2) + bc E (2) )
5
23
28
(δ2 ⊗ 1S 30 )
−(e ∧ e ) ⊗ (dz E (2) + ay E (2) + ad E (2) ) =
x
w
8
18
21
(2)
(2)
(2)
+(ex ∧ ez ) ⊗ (cw E10
+ ay E24 + ac E29 )
(2)
−(ex ∧ ey ) ⊗ (ab E30 )
(2)
(2)
(2)
(2)
(2)
(2)
+ex ⊗ (dzw E82 − czw E10 + ayw E18 + adw E21 − ayz E24 − acz E29 + aby E30 )
(2)
(2)
(2)
(2)
(2)
(2)
(2)
−ey ⊗ (dzw E3 − czw E5 + bxw E17 + bdw E20 − bxz E23 − bcz E28 + abx E30 )
=
(2)
(2)
(2)
(2)
(2)
(2)
(2)
+ez ⊗ (−cyw E5 + cxw E10 + cdw E13 − bxy E23 + axy E24 − bcy E28 + acx E29 )
(2)
(2)
(2)
(2)
(2)
(2)
(2)
−ew ∧ (−dyz E3 + dxz E8 + cdz E13 − bxy E17 + axy E18 − bdy E20 + adx E21 )
6
ALESSANDRO DE STEFANI
Again, this element is a boundary. In fact, it is equal to
(3)
(3)
+ex ⊗ (zw E7 + a E20 )
(3)
(3)
−e ⊗ (zw E4 + b E19 )
(1K1 ⊗ ϕ3 ) y
(3)
(3)
+ez ⊗ (xy E13 + c E18 )
(3)
(3)
−ew ⊗ (xy E10 + d E17 )
One more time, we apply δ1 ⊗ 1S 20 , to get
(3)
(3)
+ex ⊗ (zw E7 + a E20 )
(3)
(3)
−e ⊗ (zw E4 + b E19 )
(δ1 ⊗ 1S 20 ) y
(3)
(3) =
+ez ⊗ (xy E13
+ c E18 )
(3)
(3)
−ew ⊗ (xy E10 + d E17 )
(3)
(3)
(3)
(3)
(3)
(3)
= 1 ⊗ (−yzw E4(3) + xzw E7(3) − xyw E10
+ xyz E13 − dw E17 + cz E18 − by E19 + ax E20 ).
(4)
This is a boundary: it is equal to (1S ⊗ ϕ4 )(1 ⊗ E5 ). When applying δ0 ⊗ 1S 5 to the lift, we finally
get the image of uv under the map ψ : Z4 (R) → k ⊗S S 5 . Namely:
(4)
(4)
ψ(uv) = (δ0 ⊗ 1S 5 )(1 ⊗ E5 ) = 1 ⊗ E5 ∈ k ⊗S S 5
and since the latter is non-zero, because it is part of a k-basis of k ⊗S S 5 , we obtain that uv is not
a boundary of the Koszul complex. Thus, αβ is non-zero in H4 (R), and R is not Golod.
Remark 2.6. As pointed out in Remark 2.4, with the same notation as in Example 2.5, we have
that x − a, y − b, z − c, w − d is a regular sequence modulo I. Going modulo such a regular sequence
of linear forms, one recovers the ring of Example 2.1.
Recall that a monomial ideal I satisfies the strong-GCD condition (see [Jöl06, Definition 3.8])
if there exists a linear order ≺ on the set MinGen(I) of minimal monomial generators of I such
that, for any two monomials u ≺ v in MinGen(I), with gcd(u, v) = 1, there exists a monomial
w ∈ MinGen(I), v 6= w, with u ≺ w and such that w divides uv. The ring in Example 2.5 satisfies
the strong-GCD condition, being a product (see [SFW14, Theorem 1.1] and the discussion in the
Introduction). We present here another ideal that satisfies the strong-GCD condition, that is not
Golod. Although it is not a product, it has the advantage of having fewer generators than our
previous ideals. Another example has been discovered by Lukas Katthän [Kat], who considers the
ideal I = (x1 x2 y, x2 x3 y, x3 x4 y, x4 x5 , x5 x1 ) in the polynomial ring k[x1 , x2 , x3 , x4 , x5 , y].
Example 2.7. Let S = k[x, y, z], and let I = (x2 y, xy 2 , x2 z, y 2 z, z 2 ). Set R = S/I. The ideal
I satisfies the strong-GCD condition, for example choosing x2 y ≺ xy 2 ≺ x2 z ≺ y 2 z ≺ z 2 . Using
Macaulay2 [GS], we checked that the Poincaré series of R starts as
PR (t) = 1 + 3t + 8t2 + 21t3 + 55t4 + 144t5 + 377t6 + . . .
and that the right-hand side of Serre’s inequality is
(1 + t)3
= 1 + 3t + 8t2 + 21t3 + 56t4 + 148t5 + 393t6 + . . .
1 − 5t2 − 5t3 − t4
Therefore, R is not Golod. Alternatively, one can use the Macaulay2 command isGolod(S/I), or
one can show, with arguments similar to the ones used above, that the product of Koszuyl cycles
((ex ∧ ey ) ⊗ xy) · (ez ⊗ z) ∈ K3 (R)
is not zero in homology. Looking for a squarefree example, using polarization, one obtains that
I ′ = (axy, bxy, axz, byz, cz) ⊆ k[a, b, c, x, y, z] satisfies the strong GCD condition, and is not Golod.
PRODUCTS OF IDEALS MAY NOT BE GOLOD
7
3. Strongly Golod property for rational powers of monomial ideals
Let k be a field, and let S = k[x1 , . . . , xn ], with deg(xi ) = di > 0. We recall the definition of
rational powers of an ideal.
Definition 3.1. For an ideal I ⊆ S and positive integers p, q define the ideal
I p/q := {f ∈ R | f q ∈ I p }.
The integral closure of I p inside the definition is needed in order to make the set into an ideal,
and to make it independent of the choice of the representation of p/q as a rational number.
Remark 3.2. We would like to warn the reader about a potential source of confusion. When
p = q, the ideal I p/q = I 1/1 is the integral closure I of I, and should not be regarded as the ideal
I 1 = I, even though the exponents 1/1 and 1 are equal.
Remark 3.3. If I ⊆ S is a monomial ideal, then so is I p/q .
P
Proof. Let f = di=1 λi ui ∈ I p/q , where 0 6= λi ∈ k and ui are monomials. Since I p is monomial,
and we have that f qr ∈ I pr for all integers r ≫ 0 [HH11, Theorem 1.4.2]. Also, I pr is monomial,
therefore every monomial appearing in f qr belongs to I pr , and in particular for any i = 1, . . . , d we
pr for all r ≫ 0. This shows that uq ∈ I p , that is u ∈ I p/q for all i = 1, . . . , d, and
have that uqr
i
i ∈I
i
hence I p/q is monomial.
In the rest of the section, we assume that the characteristic of k is zero.
Definition 3.4 ([HH13]). A proper homogeneous ideal I ⊆ S is called strongly Golod if ∂(I)2 ⊆ I.
Here, ∂(I) denotes the ideal of S generated by the partial derivatives of elements in I. By [HH13,
Theorem 1.1], if I is strongly Golod, then S/I is Golod. This condition, however, is only sufficient.
For example, the ideal I = (xy, xz) ⊆ k[x, y, z] is Golod [Sha69], or [Avr98, Proposition 5.2.5].
However, it is not strongly Golod. This example is not even squarefree strongly Golod (see Section
4 for the definition). In case I is monomial, being strongly Golod is equivalent to the requirement
that, for all minimal monomial generators u, v ∈ I, and all integers i, j such that xi divides u and
xj divides v, one has uv/xi xj ∈ I.
The following argument is a modification of [HH13, Proposition 3.1].
Theorem 3.5. Let I ⊆ S be a strongly Golod monomial ideal. If p > q, then I p/q is strongly
Golod.
Proof. Let u ∈ I p/q be a monomial generator, then uqr ∈ I pr for all r ≫ 0. Let j be an index such
that xj | u, we claim that (u/xj )qr ∈ I pr/2 for all even r ≫ 0. Notice that if x2j | u then, for any
even r, r ≫ 0, we have
!qr/2
qr
u
u
= uq(r/2)
∈ I p(r/2) ,
xj
x2j
as desired. Now suppose that xj divides u, but x2j does not. Since for any r ≫ 0 we have that
uqr ∈ I pr , we can write
uqr = m1 m2 · · · mpr ,
where mi ∈ I for all i. Again, we can assume that r is even. For i = 1, . . . , pr let di be the
maximum non-negative integer such that xdj i divides mi . Then we can rewrite
uqr = m1 · · · ma ma+1 · · · ma+b ma+b+1 · · · mpr ,
8
ALESSANDRO DE STEFANI
where di = 0 for 1 6 i 6 a, di = 1 for a + 1 6 i 6 a + b and di > 2 for a + b + 1 6 i 6 pr. Because
of the assumption x2j 6 | u we have that
qr =
pr
X
di = b +
i=1
pr
X
di > b + 2(pr − b − a).
i=a+b+1
But we assumed that p > q, therefore pr > b + 2(pr − b − a), which gives a + b/2 > pr/2 and also
a + ⌊b/2⌋ > pr/2 because pr/2 is an integer. Write
qr
u
uqr
ma+b ma+b+1 · · · mpr
ma+1
,
= qr = m1 · · · ma
···
xj
xj
xj
xj
xjqr−b
b
then ma+1 . . . ma+b /xbj ∈ I ⌊b/2⌋ because I is strongly Golod, so that ∂(I)b ⊆ I ⌊ 2 ⌋ . Furthermore,
m1 · · · ma ∈ I a . Therefore
qr
u
∈ I a+⌊b/2⌋ ⊆ I pr/2 .
xj
Now let v ∈ I p/q be another monomial generator, and assume that xi |v. Then, for all even r ≫ 0
we have
uv qr
∈ I pr
xj xi
which implies that uv/xj xi ∈ I p/q . Since u and v were arbitrary monomial generators, I p/q is
strongly Golod.
Corollary 3.6. [HH13, Proposition 3.1] Let I ⊆ S be a monomial strongly Golod ideal, then I is
strongly Golod.
Proof. Choose p = q in Theorem 3.5.
Proposition 3.7. Let I ⊆ S be a monomial ideal. If p > 2q, then I p/q is strongly Golod.
Proof. Since the ideal I p/q does not depend on the representation of p/q as a rational number,
without loss of generality we can assume that p is even. Let u ∈ I p/q be a monomial generator, so
that uqr ∈ I pr for all r ≫ 0. Let j be such that xj | u, then we can write
qr
m1 . . . mpr
mpr−qr+1 . . . mpr
u
uqr
= m1 . . . mpr−qr ·
,
= qr =
qr
xj
xj
xj
xqr
j
for some mi ∈ I. But then (u/xj )qr ∈ I pr−qr ⊆ I pr/2 because q 6 p/2. Let v ∈ I p/q be another
monomial generator and assume that xi | v. For r ≫ 0 we have again that (v/xi )qr ∈ I pr/2 , so that
uv qr
∈ I pr
xj xi
for all r ≫ 0, and thus uv/xj xi ∈ I p/q . Since u and v were arbitrary, we have that I p/q is strongly
Golod.
If I is not strongly Golod and 2q > p > q, it is not true in general that I p/q is strongly Golod,
as the following family of examples shows.
Example 3.8. Let 2q > p > q be two positive integers and consider the ideal I = (xy, z q ) inside
the polynomial ring S = k[x, y, z], where k is a field of characteristic zero. Then
(xy)q (z q )p−q = (xyz p−q )q ∈ I p ,
2
/ Ip
that is xyz p−q ∈ I p/q . Thus, u := y 2 z 2p−2q ∈ ∂(I p/q )2 . On the other hand, uq = y 2q z 2pq−2q ∈
p
p
because the only monomial generator of I that can appear in an integral relation for u is z q. But
y 2qn z 2pqn−2q
2n
∈
/ (z pqn )
PRODUCTS OF IDEALS MAY NOT BE GOLOD
9
for any n because pqn > 2pqn − 2q 2 n ⇐⇒ p < 2q, and we have the latter by assumption. As a
consequence, u ∈
/ I p/q , and thus I p/q is not strongly Golod.
Remark 3.9. If we choose p = q = 2 in Example 3.8, we have in addition that I p/q = I = I =
(xy, z 2 ) is not even Golod, because it is a complete intersection of height two.
As a consequence, not all integrally closed ideals, even if assumed monomial, are Golod. A more
trivial example is the irrelevant maximal ideal m. However, as noted above in Corollary 3.6, if I is
a strongly Golod monomial ideal, then I is strongly Golod. More generally, if I ⊆ S = k[x1 , . . . , xn ]
is homogeneous, then I j is strongly Golod for all j > n + 1 [HH13, Theorem 2.11]. It is still an
open question whether I j is strongly Golod, or, at least, Golod, for any ideal I and j > 2. Since
for j > 2, the ideal I j is strongly Golod, one can ask the following more general question, which
has already been raised by Craig Huneke:
Question 3.10. [MP13, Problem 6.19] Let I ⊆ S be a strongly Golod ideal. Is I [strongly] Golod?
Remark 3.11. We checked with Macaulay2 [GS] that the ideal I of Example 2.5 is integrally
closed. Therefore, the integral closure of a product of ideals, even monomial ideals in a polynomial
ring, may not be Golod.
We end the section with a more generic question about Golodness of the ideal I 3/2 . Note
that for each ideal I = (xy, z q ) of the family considered in Example 3.8, the rational power I 3/2
is not strongly Golod. However, it is Golod. In fact, it is not hard to see that I 3 = I 3 =
q
3q
(x3 y 3 , x2 y 2 z q , xyz 2q , z 3q ). As a consequence, we have I 3/2 = (x2 y 2 , xyz ⌈ 2 ⌉ , z ⌈ 2 ⌉ ). Consider the
linear form x − y ∈ m r m2 , which is a non-zero divisor modulo I 3/2 . The image of I 3/2 in the
q
3q
polynomial ring S ′ = S/(x − y) ∼
= k[x, z] is (x4 , x2 z ⌈ 2 ⌉ , z ⌈ 2 ⌉ ). Such an ideal is easily seen to be
strongly Golod, hence Golod. By [Avr98, Proposition 5.2.4 (2)], the ideal I 3/2 is then Golod.
Question 3.12. Let I ⊆ S be a proper homogeneous ideal. Is I 3/2 always Golod? Is it true if I is
monomial?
4. lcm-strongly Golod monomial ideals
Let k be a field, and let S = k[x1 , . . . , xn ], with deg(xi ) = di > 0.
Definition 4.1. Let m ∈ S be a monomial, and let I ⊆ S be a monomial ideal. Define Im ⊆ I to
be the ideal of S generated by the monomials of I which divide m. We say that I is m-divisible if
I = Im .
Remark 4.2. Note that, choosing m = x1 · · · xn , then m-divisible simply means squarefree.
We now recall the Taylor resolution of a monomial ideal. Let I ⊆ S be a monomial ideal,
with minimal monomial generating set {m1 , . . . , mt }. For each subset Λ ⊆ [t] := {1, . . . , t} let
LΛ := lcm(mi | i ∈ Λ). Let aΛ ∈ Nn be the exponent vector of the monomial LΛ , andLlet S(−aΛ ) be
the free module, with generator in multi-degree aΛ . Consider the free modules Ti := |Λ|=i S(−aΛ ),
with basis {eΛ }|Λ|=i . Also, set F0 := S. The differential τi : Ti → Ti−1 acts on an element of the
basis eΛ , for Λ ⊆ [t], |Λ| = i, as follows:
X
LΛ
·e
τi (eΛ ) =
sign(j, Λ) ·
LΛr{j} Λr{j}
j∈Λ
Here sign(j, Λ) is (−1)s+1 if j is the s-th element in the ordering of Λ ⊆ [t]. The resulting complex
is a free resolution of S/I over S, called the Taylor resolution. The following was already noted in
[BH95, Corollary 3.2], and [Iye97, Corollary to Theorem 1].
10
ALESSANDRO DE STEFANI
Remark 4.3. [BH95, Corollary 3.2] Let I be an m-divisible monomial ideal. Then, the Koszul
homology H• (S/I) is Zn -multigraded, and it is concentrated in multidegrees aΛ′ ∈ Nn such that
(a′ )n
(a )
the monomial x1 Λ′ 1 · · · xn Λ = LΛ′ divides m.
In [HH13], given a squarefree monomial ideal, Herzog and Huneke introduce the notion of squarefree strongly Golod monomial ideal. Given Remark 4.2, we generalize it to the notion of lcm-strongly
Golod. Let I be a monomial ideal, and let m := lcm(I) be the least common multiple of the monomials appearing in the minimal monomial generating set of I. By definition, I is always m-divisible.
Also, if I is m′ -divisible for some other monomial m′ , then m divides m′ .
In what follows, we assume that the characteristic of k is zero.
Definition 4.4. Let I ⊆ S be a monomial ideal, and let m := lcm(I) be as defined above. Let
∂(I)[2] denote the ideal (∂(I)2 )m . We say that I is lcm-strongly Golod if ∂(I)[2] ⊆ I.
We also make the following definition
Definition 4.5. Let I ⊆ S be a proper homogeneous ideal. We say that R = S/I is weakly Golod
if the multiplication on Koszul homology is identically zero.
The following is the main result of the section, and justifies the previous definition. It is a
generalization of [HH13, Theorem 3.5].
Theorem 4.6. Let I ⊆ S be an lcm-strongly Golod monomial ideal. Then, S/I is weakly Golod.
Proof. Let m := lcm(I), so that I is m-divisible and ∂(I)[2] ⊆ I. By Remark 4.3, we can choose
a k-basis of H• (S/I) consisting of elements of multidegrees αΛ , where xαΛ divides m. Let a, b be
two such elements. If ab has multidegree α ∈ Nn , such that xα does not divide m, then necessarily
ab = 0 because of the multigrading on H• (S/I). So assume that the multidegree α of ab is such
that xα divides m. By [Her92], a and b can be represented by cycles whose coefficients are k-linear
combinations of elements in ∂(I). Since I is monomial, so is ∂(I). Because of the multidegree of ab,
we then have that a and b can be represented by cycles whose coefficients are k-linear combinations
of monomials u, v ∈ ∂(I), such that the products uv divide m. Then uv ∈ ∂(I)[2] ⊆ I for each
product uv appearing in these sums, and, as a consequence, ab = 0 in H• (S/I).
Discussion. It is easy to see that being lcm-strongly Golod is only sufficient to be weakly Golod.
For example, the ideal (xy, xz) ⊆ k[x, y, z] is even Golod [Sha69], but not lcm-strongly Golod. The
proof of Theorem 4.6, as well as the proofs of [HH13, Theorem 1.1] and [HH13, Theorem 3.5], are
based on a canonical description of Koszul cycles whose residue classes form a k-basis for the Koszul
homology H• (S/I) [Her92]. We want to suggest a slightly different definition of strong Golodness:
Potentially, one has to check that ∂i (f )∂j (g) ∈ I for any f, g ∈ I, and any i, j = 1, . . . , n, where
∂i = ∂/∂xi and ∂j = ∂/∂xj . However, by [Her92], each ∂i (f ) appears as a factor in some coefficient
of a Koszul cycle, which has the form (ei ∧ . . .) ⊗ ∂i (f ) ∈ K• ⊗ S/I = K• (S/I). Therefore, the
corresponding product ∂i (f )∂j (g) will appear inside some coefficient of the form
(ei ∧ ej ∧ . . .) ⊗ ∂i (f )∂j (g).
For i = j, we have that ei ∧ ej = 0. Hence we may consider only products ∂i (f )∂j (g), for i 6= j, in
the definition of strongly Golod and lcm-strongly Golod. With this modification, the ideal (xy, xz)
becomes lcm-strongly Golod. The ideal (x2 , xy) in the polynomial ring k[x, y], which is lcm-strongly
xy
2 / (x2 , xy) is the product
Golod, with this modification becomes strongly Golod. In fact, xy
x · x =y ∈
that is preventing it from being strongly Golod. However, the partial derivatives, in this case, are
both with respect to x, so we can disregard such a product.
Here follows an example of a non-squarefree ideal which is lcm-strongly Golod, but not strongly
Golod, even with the modified definition.
PRODUCTS OF IDEALS MAY NOT BE GOLOD
11
Example 4.7. Let k be a field of characteristic zero, and let I = (x2 y 2 , x2 z, y 2 z) ⊆ k[x, y, z]. Then
I is not strongly Golod, even in the definition suggested above. In fact, xz, yz ∈ ∂(I) come from
taking derivative with respect to x and y, respectively, but their product is xz · yz ∈
/ I. However,
2
2
such an element does not divide lcm(I) = x y z, therefore it can be disregarded when looking at
the lcm-strongly Golod condition. In fact, one can check that ∂(I)[2] ⊆ I, that is, I is lcm-strongly
Golod in this case.
As shown in [HH13, Proposition 3.7] for the squarefree part, if m is a monomial in S, then the
m-divisible part of a strongly Golod monomial ideal is lcm-strongly Golod. We record it in the
next proposition.
Proposition 4.8. Let I ⊆ S be a strongly Golod monomial ideal, and let m be a monomial. Then
Im is lcm-strongly Golod. In particular, I is lcm-strongly Golod.
Proof. We have that
∂(Im )[2] = (∂(Im )2 )m ⊆ (∂(I)2 )m ⊆ Im .
As mentioned in Section 3, if I is a strongly Golod monomial ideal, then I is strongly Golod. It
is natural to ask the following question:
Question 4.9. If I ⊆ S is an lcm-strongly Golod monomial ideal, is I (lcm-strongly) Golod? For
integers p > q, is the ideal I p/q (lcm-strongly) Golod?
The inequality p > q seems reasonable to require, given previous results.
5. Golodness of products and further questions
Throughout this section, unless otherwise specified, k is a field of characteristic zero, and S =
k[x1 , . . . , xn ] is a polynomial ring over k, with deg(xi ) = di > 0. It is easy to see that arbitrary
intersections of strongly Golod ideals are strongly Golod [HH13, Theorem 2.3 (a)]. Given a proper
homogeneous ideal I ⊆ S, one may ask what is the intersection of all the strongly Golod ideals
containing I. In other words, what is the smallest ideal that contains I and that is strongly Golod.
Clearly, such an ideal must contain I + ∂(I)2 . On the other hand, note that ∂(∂(I)2 )) ⊆ ∂(I),
therefore
∂(I + ∂(I)2 )2 ⊆ I + ∂(I)2 .
Thus, I + ∂(I)2 is strongly Golod, and it is indeed the smallest strongly Golod ideal containing I.
We now introduce a sufficient condition, which is far from being necessary, for the product of
two ideals to be strongly Golod.
Definition 5.1. Let S = k[x1 , . . . , xn ] and let I, J ⊆ S be two ideals. (I, J) is called a strongly
Golod pair if ∂(I)2 ⊆ I : J and ∂(J)2 ⊆ J : I.
Note that, for examples of small size, the conditions from Definition 5.1 can easily be checked
with the aid of a computer. The following proposition is the main motivation behind the definition.
Proposition 5.2. If (I, J) is a strongly Golod pair, then IJ is strongly Golod.
Proof. We noted above that the smallest strongly Golod ideal containing IJ is IJ + ∂(IJ)2 . In our
assumptions, we have
∂(IJ)2 ⊆ (∂(I)J + I∂(J))2 ⊆ ∂(I)2 J 2 + I 2 ∂(J)2 + IJ ⊆ IJ.
Therefore, IJ + ∂(IJ)2 = IJ, which is then strongly Golod.
12
ALESSANDRO DE STEFANI
Note that, looking at the proof of Proposition 5.2, one may notice that the conditions ∂(I)2 ⊆
IJ : J 2 and ∂(J)2 ⊆ IJ : I 2 are sufficient in order for the product IJ to be strongly Golod.
However, when studying properties of the product IJ, one can replace the ideal I with IJ : J
without affecting the product. In fact:
IJ ⊆ (IJ : J)J ⊆ IJ,
forcing equality. Repeating the process, one gets an ascending chain of ideals containing I, that
eventually stabilizes. Therefore one can assume that IJ : J = I. Similarly, one can assume that
IJ : I = J. Therefore the conditions above become
∂(I)2 ⊆ IJ : J 2 = (IJ : J) : J = I : J,
which is precisely the requirement in the definition of strongly Golod pair. Similarly for the other
colon ideal. Of course, as long as one can write an ideal in terms of a Golod pair, one gets that the
ideal is strongly Golod. Therefore, one may keep in mind the weaker colon conditions that come
from the proof of Proposition 5.2. Examples of strongly Golod pairs include:
(1) (I r , I s ), for any proper ideal I ⊆ S and any integers r, s > 1.
(2) If I and J are strongly Golod, then (I, J) is a strongly Golod pair.
(3) If I ⊆ J and I is strongly Golod, then (I, J) is a strongly Golod pair.
(4) (I, I : ∂(I)2 ) is a strongly Golod pair for any proper ideal I ⊆ S.
Remark 5.3. Let I1 , . . . , In be proper ideals in S. Assume that, for all i = 1, . . . , n there exists
j 6= i such that (Ii , Ij ) is a strongly Golod pair, then the product I := I1 · · · In is strongly Golod.
In fact
∂(I1 I2 . . . In ) ⊆ ∂(I1 )I2 . . . In + I1 ∂(I2 ) . . . In + . . . + I1 I2 . . . ∂(In ).
Thus
∂(I)2 ⊆ ∂(I1 )2 I22 · · · In2 + I12 ∂(I2 )2 · · · In2 + . . . + I12 I22 · · · ∂(In )2 + I.
By assumption, for each i there exists j 6= i such that ∂(Ii )2 Ij ⊆ Ii , and the claim follows. More
generally, one could define (I1 , . . . , In ) to be a strongly Golod n-uple provided
∂(Ii )2 ⊆ I : (I1 · · · Ii−1 · Ii+1 · · · In )
for all i = 1, . . . , n. Then, the above argument shows that if (I1 , . . . , In ) is a strongly Golod n-uple,
the product I1 · · · In is strongly Golod.
All the conditions discussed above are sufficient, but evidently not necessary, for a product of
two ideals to be Golod. We raise the following general question:
Question 5.4. Is there some relevant class of [pairs of] ideals for which products are [strongly]
Golod?
In particular, note that in all the examples of Section 2, the ideals appearing in the product are
not Golod. It is then natural to ask:
Question 5.5. If one of the two ideals I1 , I2 [or both] is Golod, is then S/I1 I2 Golod?
Another problem relating Golod rings to products is the following. Let I, J be two proper
homogeneous ideals in a polynomial rings S = k[x1 , . . . , xn ], with m = (x1 , . . . , xn ). Suppose that
S/IJ is Cohen-Macaulay. In [Hun07], Huneke asks whether the Cohen-Macaulay type, that is,
t(S/IJ) = dimk Extdepth(S/IJ) (k, S/IJ), is always at least the height of IJ. This was motivated by
the fact that Gorenstein rings are never products, unless they are hypersurfaces. Thus, when S/IJ
is Cohen-Macaulay and not a hypersurface, the type is always at least two. As noted in [Hun07],
the case when I = m and J is m-primary, follows by Krull’s height theorem. In our context, it
seems natural to ask the following question:
PRODUCTS OF IDEALS MAY NOT BE GOLOD
13
Question 5.6. Let I ⊆ S be a homogeneous ideal such that S/I is Cohen-Macaulay and Golod.
Is it true that the Cohen-Macaulay type t(S/I) is always at least ht(I)? Is the Cohen-Macaulay
assumption needed?
As pointed out by Frank Moore, there is still no example of a ring which is not Golod, and for
which the second Massey operation is trivial. With our notation, this means that it is not known
whether weakly Golod rings are always Golod. As explained earlier in this article, it has been
claimed in [BJ07, Theorem 5.1] that this is true in case the ideal is monomial. Given the examples
in Section 2, and our previous discussions, we do not know whether such an argument is still valid.
We decided to raise the general question here again:
Question 5.7. Are weakly Golod rings always Golod? Is it true for monomial ideals? Is there any
relevant class of ideals for which this holds true?
In [HH13, Proposition 2.12], Herzog and Huneke show that the Ratliff Rush filtration of a strongly
Golod ideal is strongly Golod. We obtain a similar statement, using Golod pairs.
Proposition 5.8. If (I, J) is a strongly Golod pair, then the ideal
[
I n+1 J : I n
n>0
is strongly Golod.
Proof. Let f ∈ S be such that f I n−1 ⊆ I n J for some n. Then f I n ⊆ I n+1 J. Let ∂ denote a partial
derivative with respect to any variable. Taking partial derivatives, from the containment above we
obtain that
∂(f )I n ⊆ f I n−1 ∂(I) + I n ∂(I)J + I n+1 ∂(J) ⊆ I n ∂(I)J + I n+1 ∂(J).
Let f, g ∈
S
n>0 I
n+1 J
: I n and choose n ≫ 0 such that f I n−1 ⊆ I n J and gI n−1 ⊆ I n J. Then
∂(f )∂(g)I 2n ⊆ I 2n ∂(I)2 J 2 + I 2n+2 ∂(J)2 + I 2n+1 J ⊆ I 2n+1 J
because ∂(I)2 J ⊆ I and I∂(J)2 ⊆ J.
In particular, Proposition 5.8 shows that the Ratliff-Rush closure of any power I d , d > 2, is
strongly Golod. In fact, it is enough to apply Proposition 5.8 to the strongly Golod pair (I d−1 , I).
This already follows from [HH13, Proposition 2.12], since I d is strongly Golod for any d > 2.
Given that the Ratliff-Rush closure of a strongly Golod ideal is strongly Golod we ask:
Question 5.9. Given a strongly Golod ideal I ⊆ S, is every coefficient ideal of I [strongly] Golod?
Question 5.9 is a more general version of Question 3.10. In fact, both the integral closure and
the Ratliff-Rush closure are coefficient ideals. See [Sha91] for details about coefficient ideals.
We conclude the section with two questions regarding the notion of strongly Golod ideal. The
definition of strongly Golod ideals is restricted to homogeneous ideals in a polynomial ring S =
k[x1 , . . . , xn ], with k a field of characteristic zero. This is because Herzog’s canonical lift of Koszul
cycles [Her92] can be applied only under these assumptions.
Question 5.10. Is there a suitable definition of strongly Golod for local rings, at least when the
ring contains a field?
Question 5.11. Is there a notion of strongly Golod that does not require the characteristic of k
to be zero?
14
ALESSANDRO DE STEFANI
Acknowledgements
We would like to thank Craig Huneke for stimulating discussions about this project, and for constant encouragement and support. We also thank David Eisenbud for significant help with some
Macaulay2 computations and for valuable suggestions. We are indebted to Srikanth Iyeangar, for
fruitful conversations that led to Example 2.1. These took place at the AMS-MRC workshop in
Commutative Algebra, Snowbird 2015. Therefore, we would like to thank the American Mathematical Society for providing such a great opportunity to this author. We thank Adam Boocher, Aldo
Conca, Lukas Katthän and Volkmar Welker for very helpful comments and suggestions. We also
thank the anonymous referee, for helpful suggestions that improved the exposition of this article.
A. The minimal free resolution of Example 2.5
Let k be a field, and let S = k[a, b, c, d, x, y, z, w]. Consider the monomial ideals I1 = (ax, by, cz, dw)
and I2 = (a, b, c, d) inside S. Let I := I1 I2 be their product, and set R = S/I. Let T =
Z[a, b, c, d, x, y, z, w], and let J be the ideal I inside T . Then, using the Macaulay2 command
res J, we get a resolution of J over T
ϕ4
/ T5
F• : 0
ϕ3
/ T 20
ϕ2
/ T 30
ϕ1
/ T 16
ϕ0
/T
/ T /I
/ 0.
Assume that char(k) = p > 0. We checked with Macaulay2 [GS] that (a2 x) ⊆ I(ϕ1 ), where I(ϕ1 )
is the Fitting ideal of the map ϕ1 . This is still a regular element after tensoring with − ⊗Z Z/(p),
so that grade(I(ϕ1 ⊗ 1Z/(p) )) > 1. Similarly, one can see that
(a12 x3 , b12 y 3 ) ⊆ I(ϕ2 )
(a15 x3 , b15 y 3 , c15 z 3 ) ⊆ I(ϕ3 )
(a5 x, b5 y, c5 z, d5 w) ⊆ I(ϕ4 )
Since the former stay regular sequences after tensoring with − ⊗Z Z/(p) we obtain that
grade(I(ϕi ⊗ 1Z/(p) )) > i
for all i = 1, . . . , 4. In addition, the ranks of the maps add up to the correct numbers after tensoring.
By Buchsbaum-Eisenbud’s criterion for exactness of complexes [BE73], F• ⊗Z Z/(p) is a minimal
free resolution of J ⊗Z Z/(p) as an ideal of T ⊗Z Z/(p) ∼
= Z/(p)[a, b, c, d, x, y, z, w]. Finally, since the
map T ⊗Z Z/(p) → S is faithfully flat, tensoring with (F• ⊗Z Z/(p)) ⊗Z/(p) S gives a minimal free
resolution of I over S, using Buchsbaum-Eisenbud’s criterion for exactness of complexes [BE73]
once again. When char(k) = 0, one can use Q instead of Z/(p) and the same arguments can be
applied.
Therefore we get a resolution
ϕ4
/ S5
F• ⊗Z S : 0
ϕ3
/ S 20
/ S 30
ϕ2
/ S 16
ϕ1
/S
ϕ0
/R
/ 0.
Lβi
(i)
Letting Ej be the canonical bases of the modules Fi ∼
=
j=1 T , for i = 0, . . . , 4, the matrices
representing the differentials of the minimal free resolution of J over T are the same as the ones of
a minimal free resolution of I over S. Here follows a description of such matrices. All the missing
entries should be regarded as zeros:
ϕ1
(0)
E1
(1)
E1
(1)
E2
(1)
E3
(1)
E4
(1)
E5
(1)
E6
(1)
E7
(1)
E8
(1)
E9
(1)
(1)
(1)
(1)
(1)
(1)
(1)
E10 E11 E12 E13 E14 E15 E16
a2 x abx acx adx aby b2 y bcy bdy acz bcz c2 z cdz adw bdw cdw d2 w
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
(2)
ϕ2 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 E11 E12 E13 E14 E15 E16 E17 E18 E19 E20 E21 E22 E23 E24 E25 E26 E27 E28 E29 E30
(1)
E1
(1)
E2
(1)
E3
(1)
E4
(1)
E6
(1)
E7
(1)
E8
(1)
E9
(1)
E10
(1)
E11
(1)
E12
(1)
E13
(1)
E14
(1)
E15
(1)
E16
a
−d
−c
a
−d
b
−y
−d
a
b
−z
c
−w
−b −c
a
−d
−c
a
x
−d
b
−d
a
b
−z
c
−w
−b −c
a
−d
−c
a
x
−d
b
y
−d
a
b
c
−w
−b −c
a
−d
−c
a
x
−d
b
y
−d
a
PRODUCTS OF IDEALS MAY NOT BE GOLOD
(1)
E5
−b −c
b
z
c
15
16
ALESSANDRO DE STEFANI
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
ϕ3 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 E11 E12 E13 E14 E15 E16
(2)
E1
c
d
(2)
E2 −b
d
(2)
E3
a
d
(2)
E4
−b −c
(2)
E5
a
−c
(2)
E6
a
b
(2)
E7
c
d
(2)
E8
−b
d
(2)
E9
a
d
(2)
E10
−b −c
(2)
E11
a
−c
(2)
E12
a
b
(2)
E13
(2)
E14
c
d
(2)
E15
−b
d
(2)
E16
a
d
(2)
E17
−b −c
(2)
E18
a
−c
(2)
E19
a
b
(2)
E20
(2)
E21
(2)
E22
c
d
(2)
E23
−b
d
(2)
E24
a
d
(2)
E25
−b −c
(2)
E26
a
−c
(2)
E27
a
b
(2)
E28
(2)
E29
(2)
E30
(3)
(3)
(3)
(3)
E17 E18 E19 E20
−yz
−yw
−zw
xz
xw
−zw
cz dw
−xy
xw
yw
−by
ax
dw
dw
−xy
−xz
−yz
−by −cz
ax
−cz
ax by
PRODUCTS OF IDEALS MAY NOT BE GOLOD
ϕ4
(4)
E1
(3)
−d
(3)
c
(3)
−b
(3)
a
E1
E2
E3
E4
(4)
E2
−d
(3)
c
(3)
−b
(3)
a
E6
E7
E8
−d
(3)
c
(3)
−b
(3)
a
E10
E11
E12
−xyw
(3)
−d
(3)
c
(3)
−b
(3)
a
E13
E14
E15
E16
(4)
E5
xzw
(3)
E9
(4)
E4
−yzw
(3)
E5
(4)
E3
xyz
(3)
−dw
(3)
cz
(3)
−by
(3)
ax
E17
E18
E19
E20
17
18
ALESSANDRO DE STEFANI
References
[AG71]
L. L. Avramov and E. S. Golod. The homology of algebra of the Koszul complex of a local Gorenstein ring.
Mat. Zametki, 9:53–58, 1971. 1
[Ani82] David J. Anick. A counterexample to a conjecture of Serre. Ann. of Math. (2), 115(1):1–33, 1982. 1
[Avr98] Luchezar L. Avramov. Infinite free resolutions. In Six lectures on commutative algebra (Bellaterra, 1996),
volume 166 of Progr. Math., pages 1–118. Birkhäuser, Basel, 1998. 2, 4, 7, 9
[BE73] David A. Buchsbaum and David Eisenbud. What makes a complex exact? J. Algebra, 25:259–268, 1973. 14
[BH95] Winfried Bruns and Jürgen Herzog. On multigraded resolutions. Math. Proc. Cambridge Philos. Soc.,
118(2):245–257, 1995. 9, 10
[BJ07]
Alexander Berglund and Michael Jöllenbeck. On the Golod property of Stanley-Reisner rings. J. Algebra,
315(1):249–273, 2007. 2, 13
[GL69] Tor H. Gulliksen and Gerson Levin. Homology of local rings. Queen’s Paper in Pure and Applied Mathematics, No. 20. Queen’s University, Kingston, Ont., 1969. 2
[Gol62] E. S. Golod. Homologies of some local rings. Dokl. Akad. Nauk SSSR, 144:479–482, 1962. 2
[GS]
Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry.
Available at http://www.math.uiuc.edu/Macaulay2/. 2, 3, 6, 9, 14
[Her92] Jürgen Herzog. Canonical Koszul cycles. In International Seminar on Algebra and its Applications (Spanish)
(México City, 1991), volume 6 of Aportaciones Mat. Notas Investigación, pages 33–41. Soc. Mat. Mexicana,
México, 1992. 4, 10, 13
[HH11] Jürgen Herzog and Takayuki Hibi. Monomial ideals, volume 260 of Graduate Texts in Mathematics. SpringerVerlag London, Ltd., London, 2011. 7
[HH13] Jürgen Herzog and Craig Huneke. Ordinary and symbolic powers are Golod. Adv. Math., 246:89–99, 2013.
1, 2, 7, 8, 9, 10, 11, 13
[HS79] Jürgen Herzog and Manfred Steurich. Golodideale der Gestalt a ∩ b. J. Algebra, 58(1):31–36, 1979. 1
[Hun07] Craig Huneke. Ideals defining Gorenstein rings are (almost) never products. Proc. Amer. Math. Soc.,
135(7):2003–2005 (electronic), 2007. 1, 12
[Iye97] Srikanth Iyengar. Shifts in resolutions of multigraded modules. Math. Proc. Cambridge Philos. Soc.,
121(3):437–441, 1997. 9
[Jöl06]
Michael Jöllenbeck. On the multigraded Hilbert and Poincaré-Betti series and the Golod property of monomial rings. J. Pure Appl. Algebra, 207(2):261–298, 2006. 1, 2, 6
[Kat]
Lukas Katthän. Personal communication. 6
[MP13] Jason McCullough and Irena Peeva. Infinite graded free resolutions. Available online at
http://www.math.cornell.edu/ irena/papers/infiniteRes.pdf, 2013. 1, 9
[Rot09] Joseph J. Rotman. An introduction to homological algebra. Universitext. Springer, New York, second edition,
2009. 4
[SFW14] S. A. Seyed Fakhari and V. Welker. The Golod property for products and high symbolic powers of monomial
ideals. J. Algebra, 400:290–298, 2014. 1, 2, 6
[Sha69] Jack Shamash. The Poincaré series of a local ring. J. Algebra, 12:453–470, 1969. 7, 10
[Sha91] Kishor Shah. Coefficient ideals. Trans. Amer. Math. Soc., 327(1):373–384, 1991. 13
[Wei94] Charles A. Weibel. An introduction to homological algebra, volume 38 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 1994. 4
Department of Mathematics, University of Virginia, Charlottesville, VA 22903
Email address: [email protected]
| 0math.AC
|
arXiv:1404.4885v1 [math.AC] 18 Apr 2014
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR
VALUATIONS
BY G. ROND AND M. SPIVAKOVSKY
Dedicated to the memory of Shreeram Abhyankar and David Rees.
Abstract. A well known theorem of Shuzo Izumi, strengthened by David Rees, asserts
that all the divisorial valuations centered in an analytically irreducible local noetherian
ring (R, m) are linearly comparable to each other. This is equivalent to saying that any
divisorial valuation ν centered in R is linearly comparable to the m-adic order. In the
present paper we generalize this theorem to the case of Abhyankar valuations ν with
archimedian value semigroup Φ. Indeed, we prove that in a certain sense linear equivalence of topologies characterizes Abhyankar valuations with archimedian semigroups,
centered in analytically irreducible local noetherian rings. In other words, saying that R
is analytically irreducible, ν is Abhyankar and Φ is archimedian is equivalent to linear
equivalence of topologies plus another condition called weak noetherianity of the graded
algebra grν R.
We give some applications of Izumi’s theorem and of Lemma 2.7, which is a crucial step
in our proof of the main theorem. We show that some of the classical results on equivalence
of topologies in noetherian rings can be strengthened to include linear equivalence of
topologies. We also prove a new comparison result between the m-adic topology and the
topology defined by the symbolic powers of an arbitrary ideal.
1. Introduction
Let (R, m, k) be a local noetherian domain with the maximal ideal m and residue field k.
Let K denote the field of fractions of R. Consider a valuation ν : K∗ ։ Γ of K with value
group Γ. We denote by Rν its valuation ring and by mν its maximal ideal.
Definition 1.1. We say that ν is centered in R if ν is non-negative on R and strictly
positive on m.
Consider a valuation ν : R −→ Γ, centered in R. Then mν ∩ R = m; thus k is a subfield of
Rν
.
kν := m
ν
Definition 1.2. We say that ν is a divisorial valuation if its value group Γ = Z and
tr.degk kν = dim R − 1.
The purpose of this paper is to generalize the following theorem of Shuzo Izumi and David
Rees (often called Izumi’s Theorem for short) to a larger class of valuations than just the
divisorial ones.
Theorem 1.3. [Iz85], [Re89] Let R be an analytically irreducible local domain. Then for
any two divisorial valuations ν and ν ′ , centered in R, there exists a constant k > 0 such that
ν(f ) ≤ k ν ′ (f )
∀f ∈ R.
2000 Mathematics Subject Classification. Primary: 13A18. Secondary: 13A15, 16W60.
1
2
BY G. ROND AND M. SPIVAKOVSKY
This result has played a central role in the study of ideal-adic topologies and other questions
about commutative rings during the last decades.
To highlight its applications, we start with some basic definitions and then recall two related
theorems due to David Rees.
Let R be a commutative noetherian ring and I an ideal of R. For an element f ∈ R the
I-order of f is defined to be
I(f ) := max{n ∈ N | f ∈ I n }.
This function takes its values in N ∪ {∞} and I(f ) = ∞ if and only if f ∈ I n for all n ∈ N.
Moreover it is easy to see that I(f g) ≥ I(f ) + I(g) for all f , g ∈ R since I is an ideal.
Next we introduce a more invariant notion, defined by David Rees and Pierre Samuel, namely
the reduced order:
n
¯ ) := lim I(f ) .
I(f
n→∞
n
¯ ) is a rational number, or even finite. The fact that I(f
¯ )
A priori, it is not obvious that I(f
is always rational in a noetherian domain R is a consequence of the following theorem of D.
Rees:
Theorem 1.4. [Re55][Re56a] For any ideal I in a noetherian domain R there exists a unique
finite set of valuations {νi }1≤i≤r of R (with values in Z) such that
¯ ) = min νi (f )
I(f
1≤i≤r νi (I)
and this representation is irredundant. These valuations νi are called the Rees valuations of
the ideal I.
Remark 1.5. This Theorem has been stated and proved by Rees assuming only that R is
noetherian (not necessarily a domain); here we restrict ourselves to the case of domains in
order to simplify the exposition.
Remark 1.6. In view of Theorem 1.4, Theorem 1.3 can be reformulated as follows: every
divisorial valuation ν is linearly equivalent to the m-adic order. This means that there exists
a positive integer k such that ν(f ) < km(f ) for all f ∈ R.
In the case R is a local domain and I = m, the valuations νi appearing in the statement of
Theorem 1.4 are divisorial valuations centered in R.
¯ ) ≥ I(f ) for all f . We will need the following result
From the definitions we see easily that I(f
of D. Rees in order to derive some corollaries of Theorem 1.3 about ideal-adic topologies in
noetherian rings:
Theorem 1.7. [Re56b] Let R be an analytically unramified local ring. Then there exists a
constant C > 0 such that
¯ ) ≤ I(f ) + C
I(f
∀f ∈ R.
The goal of this paper is to generalize Theorem 1.3 to a larger class of valuations: the Abhyankar valuations whose semigroup is archimedean.
Let us begin with some definitions. Let K denote the field of fractions of a local domain R.
Consider a valuation ν : K∗ ։ Γ of K with value group Γ, centered in R. Let Rν denote
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
3
the valuation ring of ν and mν the maximal ideal of Rν . Since ν is centered in R, we have a
ν
natural injection k ⊂ R
mν . The three basic invariants associated with ν are
Rν
tr.degk ν := tr.deg
(1)
/k
mν
(2)
rat.rk ν := dimQ (Γ ⊗Z Q)
(3)
rk ν := dimRν ,
where “dimension” means the Krull dimension. In 1956 S. Abhyankar proved that
(4)
rat.rk ν + tr.degk ν ≤ dimR
(cf. [Ab56], Theorem 1, p. 330). This inequality is called the Abhyankar Inequality. An
Abhyankar valuation is a valuation centered in R such that the above inequality is an equality.
Any divisorial valuation centered in R satisfies rat.rk ν = 1 and tr.degk ν = dimR − 1; hence
it is an Abhyankar valuation.
Let
Φ := ν(R \ {0}) ⊂ Γ.
Then Φ is an ordered semigroup contained in Γ. Since R is noetherian, Φ is a well-ordered
set. For α ∈ Φ, let
pα := {x ∈ R | ν(x) ≥ α}
and
pα+ := {x ∈ R | ν(x) > α}
(here we adopt the convention that ν(0) > α for any α ∈ Γ). Ideals of R which are
contractions to R of ideals in Rν are called ν-ideals. All of the pα and pα+ are ν-ideals and
{pα }α∈Φ is the complete list of ν-ideals in R. Specifying all the ν-ideals in R is equivalent
to specifying ν (see [ZS60], Appendix 3). The following is a characterization of ν-ideals: an
ideal I ⊂ R is a ν-ideal if and only if, for any elements a ∈ R, b ∈ I such that ν(a) ≥ ν(b)
we have a ∈ I.
We associate to ν the following graded algebra:
M p
α
.
grν R :=
pα+
α∈Φ
Definition 1.8. We say that Φ is archimedian if for any α, β ∈ Φ, α 6= 0, there exists r ∈ N
such that rα > β.
This is equivalent to saying that every ν-ideal in R is m-primary and weaker than saying
that rk ν = 1.
For l ∈ N, let Ql denote the ν-ideal
(5)
Ql := {x ∈ R | ν(x) ≥ lν(m)}.
Of course, ml ⊂ Ql for all l ∈ N.
Definition 1.9. We say that the ν-adic and the m-adic topologies are linearly equivalent
if there exists r ∈ N such that
Qrl ⊂ ml
for all l ∈ N.
4
BY G. ROND AND M. SPIVAKOVSKY
Thus Theorem 1.3 is equivalent to saying that for any divisorial valuation ν of R the ν-adic
topology M
is linearly equivalent to the m-adic topology.
Let A =
Aα be a Φ-graded k-algebra. Assume that A0 = k. By abuse of notation, let
α∈Φ
us denote
1 := ν(m)
and for l ∈ N
l := l · 1.
Definition 1.10. We say that A is weakly noetherian of dimension d if A contains d
algebraically independent elements and the function
X
F (l) =
dimk Aα
0≤α≤l
is bounded above by a polynomial in l of degree d.
If A is weakly noetherian then dimk Aα < ∞ for all α ∈ Φ. We now state the main theorem
of this paper.
Theorem 1.11. Let (R, m, k) be a local noetherian domain with field of fractions K. Let ν
be a valuation of K centered in R with value semigroup Φ. Then the following two conditions
are equivalent:
(1) R is analytically irreducible, Φ is archimedian and
(6)
rat.rk ν + tr.degk ν = dim R
(2) the ν-adic and the m-adic topologies are linearly equivalent and grν R is weakly noetherian.
This theorem is proved in Part 2. Part 3 is devoted to applications of Izumi’s Theorem and
of Lemma 2.7.
The most difficult part in the proof of Theorem 1.11 is the implication (1)=⇒ "linear equivalence between the ν-adic and the m-adic topologies". Let us mention that this implication
has been proved by D. Cutkosky [ELS03] in the case R contains a characteristic zero field.
More precisely he proved in this case that an Abhyankar valuation is quasi-monomial after a sequence of blow-ups. Thus, by Theorem 1.3, the ν-adic and the m-adic topologies
are linearly equivalent. This implication has also been proved in the case when (R, m) is a
complete local ring containing a field of positive characteristic and ν is the composition of a
R
, and of the
morphism R −→ S, where (S, m0 ) is a regular complete local ring with mS0 ≃ m
m0 -adic valuation of S (see [Ro09]).
Theorem 1.11 has been announced by the second author and a proof was sketched in [Sp90]
without details but the entire proof was never published. The proof presented here follows
the sketched proof announced in [Sp90].
1.1. Conventions. Let R be an integral domain with field of fractions K. Let ν be a
valuation of K with value group Γ.
The notation Rν will stand for the valuation ring of ν (the ring of all the elements of K
whose values are non-negative) and mν the maximal ideal of Rν (the elements with strictly
positive values). If p ⊂ R is an ideal, we put
ν(p) := min {ν(x) | x ∈ p}.
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
5
If R is noetherian, this minimum is always achieved.
If x is an element of R, x̄ will always mean the natural image of x in grν R.
We freely use the multi-index notation: if x = (x1 , · · · , xn ), α = (α1 , · · · , αn ) ∈ Nn0 , then xα
n
n
P
Q
i
αi .
xα
stands for
i . The syumbol |α| will stand for
i=1
i=1
Let R be a regular local ring with regular system of parameters x = (x1 , · · · , xn ). A valuation
ν, centered at R, is said to be monomial with respect to x if all the ν-ideals of R are generated
by monomials in x.
2. Proof of Theorem 1.11
The purpose of this section is to prove Theorem 1.11. We start with a few remarks.
Remark 2.1. Suppose the m-adic and the ν-adic topologies in R are equivalent. Then Φ
is archimedian. Indeed, to say that Φ is not archimedian is equivalent to saying that there
exists a ν-ideal p in R which is not m-primary. But then p is open in the ν-adic topology,
but not in the m-adic topology, which is a contradiction.
Moreover, the equivalence of topologies implies that R is analytically irreducible. Indeed,
let {an }, {bn } be two Cauchy sequences for the m-adic topology in R such that
lim an 6= 0
(7)
n→∞
lim bn 6= 0,
(8)
n→∞
but
lim an bn = 0.
n→∞
By the equivalence of topologies, (7) and (8), ν(an ) and ν(bn ) are independent of n for
n ≫ 0, hence so is ν(an bn ), which contradicts the fact that
lim an bn = 0
n→∞
in the ν-adic topology.
Remark 2.2. The fact that Φ is archimedian together with the equality (6) implies that
the m-adic and the ν-adic topologies are equivalent (see [Z49], pp. 63–64).
Lemma 2.3. Let R be any local domain whatsoever and ν any valuation of the field of
fractions, centered in R. Let x1 , · · · , xr be elements of R such that {ν(xi )}1≤i≤r are linearly
independent over Z. Let y1 , · · · , yt be elements of Rν such that the natural images ȳi of the
Rν
yi in m
are algebraically independent over k. Assume that there exists a monomial b = xω
ν
such that byi ∈ R for all i, 1 ≤ i ≤ t. Then the natural images of x1 , · · · , xr , by1 , · · · , byt in
grν R are algebraically independent over k.
Proof. The algebra grν R is an integral domain, on which ν induces a natural valuation,
which we shall also denote by ν. Let
zi := byi ,
Consider an algebraic relation
X
(9)
c̄α,β x̄α z̄ β = 0,
α,β
1 ≤ i ≤ s.
α ∈ Nr0 , β ∈ Ns0 , c̄α,β ∈ k,
6
BY G. ROND AND M. SPIVAKOVSKY
where the c̄α,β are not all zero. Here x̄i , z̄i denote the natural images in grν R of xi and zi ,
respectively. We may assume that (9) is homogeneous with respect to ν. At least two of the
c̄α,β must be non-zero. Take a pair (α, β), (γ, δ) ∈ Ns+r
such that
0
c̄α,β 6= 0
(10)
c̄γ,δ 6= 0.
(11)
Then
x̄α b̄|β| = x̄γ b̄|δ|
otherwise the equality
ν(c̄α,β x̄α z̄ β ) = ν(c̄γ,δ x̄γ z̄ δ )
would give a rational dependence between the ν(xi ). This implies that (9) can be rewritten
in the form
X
x̄λ
c̄α,β ȳ β = 0
α,β
and hence
X
(12)
c̄α,β ȳ β = 0,
α,β
since grν R is an integral domain. But this contradicts the choice of the yi and the Lemma
is proved.
Corollary 2.4. Under the assumptions of Lemma 2.3, let
(13)
r : = rat.rk ν
(14)
t : = tr.degk ν.
Then tr.degk grν R = r + t. Here we allow the possibility for both sides to be infinite. In
particular, if grν R is weakly noetherian, its dimension must be r + t.
Proof. We work under the assumption that r and t are finite and leave the general case as
an easy exercise. Let y1 , · · · , yt be a maximal set of elements of Rν such that the ȳi are
algebraically independent over k. Let x1 ∈ R be any element of strictly positive value such
that x1 yi ∈ R for all i, 1 ≤ i ≤ t. Choose x2 , · · · , xr in such a way that {ν(xi )}1≤i≤r
form a basis for Γ ⊗Z Q over Q. By Lemma 2.3, x̄1 , · · · , x̄r , x̄1 ȳ1 , · · · , x̄1 ȳt are algebraically
independent over k in grν R. Hence,
tr.degk grν R ≥ r + t.
On the other hand, take any z ∈ R. By the choice of the xi , there exist l ∈ N, ci ∈ Z such
that
!
r
Y
ci
l
xi .
ν z =ν
i=1
l
Qr z̄
Writing down the alBy the choice of the yi ,
ci is algebraic over k(ȳ1 , · · · , ȳt ).
i=1 x̄i
gebraic dependence relation and clearing denominators, we get that z̄ is algebraic over
k[x̄1 , · · · , x̄r , x̄1 ȳ1 , · · · , x̄1 ȳt ] and the proof is complete.
Proof of Theorem 1.11 (2) =⇒ (1). By Remark 2.1 we only have to prove that
rat.rk ν + tr.degk ν = dim R.
Let
d := rat.rk ν + tr.degk ν
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
and for l ∈ N
7
R
Ql
where Ql is as in (5). By Corollary 2.4, grν R is weakly noetherian of dimension d. Hence N (l)
is bounded above by a polynomial in l of degree d. By the linear equivalence of topologies,
there exists r ∈ N such that
Qrl ⊂ ml
for all l ∈ N.
N (l) := length
Hence
length
R
R
≤ length
,
l
m
Qrl
and so length mRl is bounded above by a polynomial of degree d in l. Therefore dim R ≤ d,
hence dim R = d by Abhyankar’s Inequality (4).
(1) =⇒ (2). Let d be as above. By Corollary 2.4, grν R contains d algebraically independent
elements over k. For l ∈ N, ml ⊂ Ql , so that
length
R
R
≥ length .
ml
Ql
Therefore length QRl is bounded above by a polynomial in l of degree d and grν R is weakly
noetherian.
We have now come to the hard part of the Theorem: proving the linear equivalence of
topologies.
b be the m-adic completion of R. Since R is analytically irreducible and Φ archimedian,
Let R
b (see [Z49], pp. 63–64). Moreover, νb has the same
there exists a unique extension νb of ν to R
Rνb
Rν
value group as ν and mνb = mν . Since
b = (mR)
b n,
mn R
b
it is sufficient to prove that the νb-adic and the mR-adic
topologies are linearly equivalent in
b Thus we may assume that R is complete.
R.
Claim. There exists a system of parameters (x1 , · · · , xd ) of R such that x̄1 , · · · , x̄d are
algebraically independent in grν R over k, and if char(R) = 0 and char(k) = p > 0 we have
x1 = p.
Proof of Claim. We construct the xi recursively. First of all we choose any non-zero element
x1 in m except in the case char(R) = 0 and char(k) = p > 0 where we choose x1 := p.
Assume that we already constructed elements
x1 , · · · , xi ∈ m
such that x̄1 , · · · , x̄i are algebraically independent over k and ht(x1 , · · · , xi ) = i < d. By
Corollary 2.4 there exists y ∈ R such that ȳ is transcendental over k[x̄1 , · · · , x̄i ] in grν R.
Let P1 , · · · , Ps denote the minimal prime ideals of (x1 , · · · , xi )R. Renumbering the Pj , if
necessary, we may assume that there exists j ∈ {0, · · · , s} such that y ∈ Pl , 1 ≤ l ≤ j and
y∈
/ Pl , j < l ≤ s.
j
j
s
s
T
S
S
T
Pl = R if j = s),
Pl = ∅ if j = 0 and
Pl (where we take
Pl \
Take any z ∈
l=j+1
l=1
l=1
and let
xi+1 := y + z N ,
l=j+1
8
BY G. ROND AND M. SPIVAKOVSKY
where N is an integer such that N ν(z) > ν(y). We have constructed elements
x1 , · · · , xi , xi+1 ∈ m
such that x̄1 , · · · , x̄i+1 are algebraically independent over K and
ht(x1 , · · · , xi+1 ) = i + 1.
For i = d we obtain the desired system of parameters (x1 , · · · , xd ). The Claim is proved.
If R is not equicharacteristic, it contains a complete non-equicharacteristic Dedekind domain
W (cf. [Ma80], Theorem 84) whose maximal ideal is generated by p = x1 . Since R is an
integral domain, R is finite over k[[x1 , · · · , xd ]] or W [[x2 , · · · , xd ]], depending on whether R
is equicharacteristic or not (cf. [Ma80], Theorem 84). Let
S := k[[x1 , · · · , xd ]] or W [[x2 , · · · , xd ]],
depending on which of the two cases we are in. Let L denote the field of fractions of S and
m0 := m ∩ S. Let t1 , · · · , tn be a system of generators of R as an S-algebra. Let T1 , · · · , Tn
be independent variables and write
S[T1 , · · · , Tn ]
,
p
R=
where p is the kernel of the natural map S[T ] → R, given by Ti → ti , 1 ≤ i ≤ n. Let
xd
x2
,··· ,
S̃ = S
x1
x1
and let S̄ be the localization of S̃ at the prime ideal (x1 )S̃:
S̄ = S̃(x1 ) .
The ring S̄ is the local ringh of the generic
i point of the exceptional divisor of the blowing-up
xd
x2
of σ S at m0 . Let R̄ := R x1 , · · · , x1 . Then R̄ is a semi-local integral domain, finite over
S̄ and birational to R. Let q0 := m0 S̄ = (x1 )S̄; q := mR̄.
Let ν0 denote the restriction of ν to L.
Lemma 2.5. Let (S, m0 , k) be a regular local ring with regular system of parameters
x = (x1 , · · · , xd )
and field of fractions L. Let ν0 be a valuation of L, centered at S, such that x̄1 , · · · , x̄d are
algebraically independent in grν0 S over k. Then the valuation ν0 is monomial with respect
to x.
Proof. In what follows, "monomial” will mean a monomial in x and "monomial ideal” - an
ideal, generated by monomials in x. For an element f ∈ S, let M (f ) denote the smallest (in
the sense of inclusion) monomial ideal of S, containing f . The ideal M (f ) is well defined: it
is nothing but the intersection of all the monomial ideals containing f . Let M on(f ) denote
the minimal set of monomials generating I(f ). In other words, M on(f ) is the smallest set of
monomials in x such that f belongs to the ideal of S generated by M on(f ). The set M on(f )
can also be characterized as follows. It is the unique set {ω1 , · · · , ωs } of monomials, none of
which divide each other and such that f can be written as
(15)
f=
s
X
i=1
ci ω i
with ci units of S.
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
9
A key point is the following: since x̄1 , · · · , x̄n are algebraically independent in grν0 S by
assumption, (15) implies that
(16)
ν0 (f ) = min {ν(ωi )}.
1≤i≤s
If I is an ideal of S, let M (I) denote the smallest monomial ideal, containing I. The ideal
M (I) is generated by the set {M on(f ) | f ∈ I}.
We want to prove that all the ν0 -ideals of S are monomial. Let I be a ν0 -ideal of S. It is
sufficient to prove that I = M (I). Obviously, I ⊂ M (I). To prove the opposite inclusion,
take a monomial ω ∈ M (I). By the above, there exists f ∈ I and ω ′ ∈ M on(f ) such that
(17)
ω ′ | ω.
By (16), we have ν0 (ω ′ ) ≥ ν0 (f ). Since I is a ν0 -ideal, this implies that ω ′ ∈ I. Hence ω ∈ I
by (17). This completes the proof.
Lemma 2.6. Let (S, m0 , k) be a regular local ring with regular system of parameters
x = (x1 , · · · , xd )
and field of fractions L. Let ν0 be a monomial valuation of L, centered at S, such that the
semigroup ν0 (S \ {0}) is archimedian. Then the ν0 -adic topology on S is linearly equivalent
to the m0 -adic topology.
Proof. Renumbering the xi , if necessary, we may assume that ν0 (x1 ) ≤ ... ≤ ν0 (xd ). Since
ν0 (S \ {0}) is archimedian, there exists a natural number N such that ν0 (xd ) ≤ N ν0 (x1 ).
For l ∈ N, let Ql denote the ν0 -ideal
(18)
Ql := {x ∈ S | ν0 (x) ≥ lν0 (m0 )}.
Then for all l ∈ N we have QN l ⊂ ml0 ⊂ Ql . This completes the proof of the Lemma.
Next, R′ be the normalization of R. Since R is complete, R′ is a local ring [Na62], (37.9).
Let m′ be the maximal ideal of R′ . If the m′ -adic topology in R′ is linearly equivalent to the
ν-adic one then the same is true for the m-adic topology in R. Hence we may assume that
R is normal.
Let K1 be a finite extension of K which is normal over L (in the sense of field theory). Then
there exists a valuation ν1 of K1 whose restriction to K is ν [ZS60], Chapter VI, 4,6,11. Let
R1 be the integral closure of R in K1 . Then R1 is a product of complete local rings and
since it is an integral domain (it is a subring of K1 ) it is a complete local ring. Then ν1 is
centered
in thei maximal ideal m1 of R1 . We have dim(R1 ) = dimR, rat.rk ν1 = rat.rk ν
h
R
Rν
and mνν1 : m
< ∞. The ring R1 is analytically irreducible by [Na62], (37.8). Finally,
ν
1
since R1 is algebraic over R, for any x ∈ R1 there exist n ∈ N and y ∈ R such that
nν1 (x) ≥ ν1 (y) = ν(y). Hence ν1 is archimedian on R1 . Therefore ν1 satisfies (1) of
Theorem 1.11. To prove that the m-adic topology on R is linearly equivalent to the ν-adic
one, it is sufficient to prove that the same is true of the m1 -adic topology on R1 . Thus we
may assume that the field extension L ֒→ K is normal. Let p = char K if char K > 0 and
p = 1 otherwise. Let pn denote the inseparability degree of K over L.
By Lemmas 2.5–2.6 the ν0 -adic topology on S is linearly equivalent to the m0 -adic topology.
Now, let f ∈ m. Since R is assumed to be integrally closed in K, it equals the integral
closure of S in K. Therefore R is mapped to itself by all the automorphisms of K over L. In
10
BY G. ROND AND M. SPIVAKOVSKY
particular, for every σ ∈ Aut(K/L) we have ν(σf ) > 0. Then
Y
n
σf p = ν(NK/L (f )) = ν0 (NK/L (f )).
ν(f ) < ν
σ∈Aut(K/L)
By the linear equivalence of topologies on S, there exists r ∈ N, such that
ν0 (NK/L (f )) ≤ rm0 (NK/L (f )).
Now Theorem 1.11 follows from the next Lemma (we state Lemma 2.7 in greater generality
than is necessary for Theorem 1.11 for future reference).
Lemma 2.7. Let S ⊂ R be two noetherian domains with fields of fractions L and K,
respectively. Let m be a maximal ideal of R and m0 := m ∩ S. Assume that Sm0 and Rm
are analytically irreducible and that Rm is finite over Sm0 . Assume that at least one of the
following conditions holds:
(1) for any f ∈ R, NK/L (f ) ∈ S.
(2) The m0 -adic order on S is a valuation
(so that the expression m0 (NK/L (f )) makes sense). Then there exists r ∈ N such that for
any f ∈ R
m0 (NK/L (f )) ≤ rm(f ).
Proof. Arguing as above, we reduce the problem to the case when R and S are complete
local rings, R is integrally closed in K and the field extension L ֒→ K is normal. We will
work under all these assumptions from now on.
First, we prove the equivalence of topologies under the additional assumptions that K is
separable (hence Galois) over L and S is regular. We fix a regular system of parameters
x = (x1 , · · · , xd ) of S. Let G := Gal(K/L). Then G acts on R.
Let R̄ be as above and let R̄′ denote the integral closure of R̄ in K. Then R̄′ is a 1-dimensional
semi-local ring, finite over S̄. Let m1 ,..., ms denote the maximal ideals of R̄′ . Write
qR̄′ =
s
\
mki i =
s
Y
mki i .
i=1
i=1
Then
(19)
qn R̄′ =
s
\
i
=
mnk
i
s
Y
i
mnk
i .
i=1
i=1
Each mi defines a divisorial valuation of K, centered in R. We denote it by νi . The group
G acts on R̄′ and permutes the mi . By Theorem 1.3, all the νi are linearly comparable in R
to the m-adic pseudo-valuation of R. Hence there exists r ∈ N such that for any i, 1 ≤ i ≤ s
and any f ∈ R,
νi (f ) ≤ rm(f ).
(20)
Since S is a UFD, q0 (f ) = m0 (f ) for all f ∈ S. Now, take any f ∈ R. Without loss of
generality, assume that
ν1 (f ) = max νi (f ).
1≤i≤s
l
Finally, let l ∈ N be such that q ∩ S ⊂ q0 . Then for any g ∈ S, q0 (g) ≤ lq(g). We have
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
(21)
(22)
11
m0 (NK/L (f )) = q0 (NK/L (f )) ≤ lq(NK/L (f )) ≤
(23)
(24)
≤ l max ki max νi (NK/L (f ))
1≤i≤s
1≤i≤s
Y
≤ l max ki max
νi (σf )
1≤i≤s
1≤i≤s
σ∈G
≤ l max ki [K : L]ν1 (f ) ≤ rl max ki [K : L]m(f ).
1≤i≤s
1≤i≤s
This proves Lemma 2.7 in the case when K is Galois over L and S is regular.
Continue to assume that S is regular, but drop the assumption of separability of K over
L. Let Ks denote the maximal separable extension of L in K. Let Rs := R ∩ Ks . Suppose
n
char K = p > 0. Then there exists n ∈ N such that Rp ⊂ Rs . Let ms denote the maximal
ideal of Rs . Since ms ⊂ m, for any g ∈ Rs we have ms (g) ≤ m(g). By the separable case
there exists r ∈ N such that for any f ∈ Rs
ν(f ) ≤ rms (f ).
Then for any f ∈ R,
(25)
(26)
n
n
r
1
ν(f p ) ≤ n ms (f p )
n
p
p
r
pn
≤ n m(f ) ≤ rm(f ).
p
ν(f ) =
Hence ν(f ) and m(f ) are linearly comparable, as desired. This proves Lemma 2.7 assuming
S is regular.
Finally, drop the assumption that S is regular. There exists a complete regular local ring
T ⊂ S such that S is finite over T . Since Lemma 2.7 is already known for T , it is also
true for S by the multiplicativity of the norm. This proves Lemma 2.7 and with it Theorem
1.11.
3. Applications
The rest of the paper is devoted to the applications of Izumi’s Theorem and of Lemma 2.7.
The first application is to rewrite some of the classical theorems on comparison of topologies
in noetherian rings (which were traditionally proved by Chevalley lemma) to include linear
equivalence of topologies.
The following observation will be useful in the sequel.
Lemma 3.1. Let R be a noetherian ring, m a maximal ideal of R. Let ϕ : R → Rm denote
the localization homomorphism. Then
ϕ−1 (mn Rm ) = mn + Ker ϕ = mn .
In other words, the symbolic powers of m coincide with the usual powers. In particular, the
m-adic topology on R coincides with the restriction of the mn Rm -adic topology to R.
Proof. Consider an element
x ∈ ϕ−1 (mn Rm ).
Then there exists u ∈ R \ m such that ux ∈ mn . For every natural number n, we have
(u) + mn = R. Then there exist vn ∈ R, mn ∈ mn such that uvn + mn = 1. We have
12
BY G. ROND AND M. SPIVAKOVSKY
x = x · 1 = x(uvn + mn ) = xuvn + xmn ∈ mn . This proves that ϕ−1 (mn Rm ) ⊂ mn for all
natural numbers n, and the Lemma follows.
The following Corollary is a partial generalization of Corollary 2, [ZS60] p. 273:
Corollary 3.2. Let R be a noetherian ring and m a maximal ideal of R, such that Rm is
an analytically irreducible local domain. Let R̄ be a finitely generated R-algebra, containing
R. Let p be a prime ideal of R̄, lying over m. Then the p-adic topology on R is linearly
equivalent to the m-adic topology. In other words, there exists r ∈ N such that for any n ∈ N
prn ∩ R ⊂ mn .
Proof. By Lemma 3.1, we may assume that R is an analytically irreducible local noetherian
b denote
domain with maximal ideal m. Let K be the field of fractions of R. Let ψ : R → R
the m-adic completion of R. The homomorphism ψ is faithfully flat, hence so is the induced
b which lies over p. The ring
b Then there exists a prime ideal b
p in R̄ ⊗R R
map R̄ → R̄ ⊗R R.
b is finitely generated over R,
b so we may assume that R is m-adically complete.
R̄ ⊗R R
If R̄ is a purely transcendental extension of R, then pn ∩ R = mn and the Corollary is trivial.
The normalization R′ of R is a complete local domain, finite over R ([Na62], Corollary 37.9).
Hence the Corollary is true when R̄ = R′ . Replacing (R, R̄) with (R′ , R′ ⊗R R̄), we may
assume that R is normal, hence analytically normal.
Since we know the Corollary for the case of purely transcendental extensions, we may replace
R by the normalization of a maximal purely transcendental extension of R, contained in R̄.
Thus we may assume that the total ring of fractions of R̄ is finite over K.
Let q be a minimal prime of R̄ such that q ⊂ p. Since R is a domain, p ∩ R = (0). Replacing
R̄ with R̄
q , we may assume that R̄ is a domain. Let K̄ denote its field of fractions. Let
π : X → Spec R̄ be the normalized blowing-up along p. Since R is Nagata, so is R̄ and π
is of finite type. Let E be any irreducible component of π −1 (p). Let ν denote the divisorial
valuation of K̄ associated to E, and let
pl := {x ∈ R̄ | ν(x) ≥ l},
l ∈ N.
Then pl ⊂ pl for all l ∈ N. Hence it is sufficient to show that the ν-adic and the m-adic
topologies in R are linearly equivalent. Since [K̄ : K] < ∞, ν induces a divisorial valuation
of K, centered in R. Now the Corollary follows from Theorem 1.11.
The above Corollary can be strengthened as follows.
Corollary 3.3. Let R be a noetherian ring and m a maximal ideal of R such that Rm is an
analytically irreducible local domain. Let R̄ be a finitely generated R-algebra, containing R.
Let p be a prime ideal of R̄ lying over m and I an ideal of R̄ such that I ⊂ p and I ∩ R = (0).
Then there exists r ∈ N such that for any n ∈ N,
(I + prn ) ∩ R ⊂ mn .
Proof. Replacing R̄ by
Corollary 3.2.
R̄
I
does not change the problem. Now Corollary 3.3 follows from
Remark 3.4. In terms of the analogy with functional analysis, this Corollary says that if
I ∩ R = (0), then I is “transversal” to R.
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
13
Corollary 3.5. Let R be a noetherian ring and T = (T1 , · · · , Tn ) independent variables. Let
m be a maximal ideal of R. Let R̄ be a finitely generated extension of R and p̄ ⊂ m̄ ⊂ R̄[[T ]]
a pair of prime ideals of R̄[[T ]] such that m̄ ∩ R[[T ]] = (m, T ). Let
(27)
Assume that
n∈N
R[[T ]](m,T )
p
p := p̄ ∩ R[[T ]] ⊂ (m, T ).
is analytically irreducible. Then there exists r ∈ N such that for all
(m̄rn + p̄) ∩ R[[T ]] ⊂ (m, T )n + p.
Remark 3.6. In particular, we can apply this Corollary to the following situation. Let R,
T , m, R̄, m̄ be as in Corollary 3.5. Assume, in addition, that R is a UFD. Let p = (F ) be a
principal prime ideal generated by a single irreducible power series F ∈ R[[T ]]. Let
(28)
pR̄[[T ]] = q1 ∩ · · · ∩ qs
√
be a primary decomposition of p in R̄[[T ]]. We must have qi ⊂ m̄ for some i. Let
√
p̄ := qi . Finally, assume that the equality (27) is satisfied. The equality (28) corresponds
to a factorization F in R̄[[T ]]. Say, F = F1 F2 in R̄[[T ]]. Then Corollary 3.5 says that there
exists r ∈ N such that if
F̃ ≡ F1 F̃2 mod m̄rn ,
where F̃2 ∈ R̄[[T ]], F̃ ∈ R[[T ]], then
F̃ ∼
= F g mod (m, T )n
for some g ∈ R[[T ]]. Thus F1 (F2 g − F̃2 ) ∈ (m, T )n ⊂ m̃n and by Artin-Rees Lemma there
exists a constant c depending only on F1 such that F̃2 ∼
= F2 g mod m̄n−c . In other words,
approximate factorization of an element of R[[T ]] in R̄[[T ]] is close to an actual factorization,
and the estimate is linear in n.
Proof of Corollary 3.5. Let R̃ := R̄[[T ]]m̄ and let
S = R[[T ]](m,T ) ⊗R R̄m̄∩R̄ ⊂ R̃.
Let m0 := m̄R̃ ∩ S. Then R̃ is the (T )-adic completion of S, followed by localization at m̄.
We can decompose the injective homomorphism S → R̃ in two steps: S → Sm0 → R̃, where
the second arrow is a faithfully flat homomorphism and the first a localization with respect
to a maximal ideal. By faithful flatness, we have m̄n ∩ Sm0 = mn0 R̃ ∩ Sm0 = mn0 Sm0 , so that
the m̄-adic topology on Sm0 is linearly equivalent to the m0 -adic topology. Combining this
with Lemma 3.1, we see that the m̄-adic topology on S is linearly equivalent to the m0 -adic
topology. The ring S is a localization of a finitely generated R[[T ]]-algebra, and we apply
]]
S
֒→ p̄∩S
. This completes the proof.
Corollary 3.2 to the ring extension R[[T
p
Remark 3.7. Corollary 3.5 can be strengthened as follows. Let R be a noetherian ring,
T1 , · · · , Tn independent variables and A a noetherian ring such that
R[T ] ⊂ A ⊂ R[[T ]].
Let m be a maximal ideal of R. Let R̄ be a finitely generated R-algebra and let B be a
noetherian R̄-algebra such that
A ⊗R R̄ ⊂ B ⊂ R̄[[T ]].
Assume that (m, T ) is a maximal ideal of A and that both A ⊗R R̄ and B have R̄[[T ]] as
their (T )-adic completion. Let m̄ be any prime ideal of R̄ lying over (m, T ) and p̄ a prime
ideal of B such that p̄ ⊂ m̄.
14
BY G. ROND AND M. SPIVAKOVSKY
Let p := p̄ ∩ A. Assume that
that for any n ∈ N
Am̄∩A
pAm̄∩A
is analytically irreducible. Then there exists r ∈ N such
(m̄rn + p̄) ∩ A ⊂ (m, T )n + p.
The proof is exactly the same as for Corollary 3.5.
The following result is not a corollary of Theorem 1.11 but uses Theorem 1.4:
Proposition 3.8. (cf. [Na62] Theorem 3.12, p. 11). Let R be a noetherian domain and I
an ideal of R. Let x be a non-zero element of R. Assume that R is analytically unramified.
Then there exists r ∈ N such that for any k, n ∈ N
I k : xn ⊂ I k−rn .
Here we adopt the convention that I n = R if n ≤ 0.
Proof. By Theorem 1.4, there exist valuations ν1 , · · · , νs such that for any f ∈ R
¯ ) = min νi (f ) .
I(f
1≤i≤s νi (I)
By Theorem 1.7 there exists r1 ∈ N such that for any n ∈ N and any f ∈ R
(29)
¯ ) ≤ r1 + I(f ).
I(f
¯ ) ≥ r1 + l then f ∈ I l . For any y ∈ I k : xn we must have
Hence, for any l ∈ N, if I(f
νi (y) + nνi (x) ≥ kνi (I)
for all i, 1 ≤ i ≤ s.
Now take a positive integer R such that
r ≥ max
1≤i≤s
νi (x)
+ r1 .
νi (I)
Then for any y ∈ I k : xn and any i ∈ {1, · · · , s} we have
νi (y) ≥ kνi (I) − nνi (x) ≥ kνi (I) − n(r − r1 )νi (I) ≥ (k − nr + r1 )νi (I).
By (29) this implies that y ∈ I k−nr , as desired.
The following corollary is a generalization of the main result of [Mo13]:
Corollary 3.9. Let R be an analytically irreducible noetherian local ring. Then there exists
a ∈ N such that for any proper ideal I of R we have:
I (ac) ⊂ mc ∀c ∈ N.
Here, if W denotes the complement of the union of the associated primes of I, I (n) is the
contraction of I n RW to R where RW denotes the localization of R with respect to the multiplicative system W . The ideal I (n) is called the n-th symbolic power of I.
Let us mention that it is known that if (R, m) is a regular local ring of dimension d then
I (dc) ⊂ I c for any ideal I of R and any integer c [HH02]. If (R, m) has is an isolated singularity
ring then I (kc) ⊂ I c for any ideal I of R and any integer c [HKV09] for some constant k
independent on p. For a general local ring R and for any ideal I, there exists a constant k
depending on I such that I (kc) ⊂ I c for any c [Sw00] but it is still an open question to know
if such a k may be chosen independently of I in general.
THE ANALOGUE OF IZUMI’S THEOREM FOR ABHYANKAR VALUATIONS
15
Proof. First let us prove the result when R is a complete local domain and I = p is a prime
ideal. By Cohen’s structure theorem R is finite over a ring of power series over a field or
over a compete Dedekind domain S. We denote by m0 the maximal ideal of S. Let K (resp.
L ) denote the field of fractions of R (resp. S).
First let us assume that K/L is Galois and R is normal. Let q := p ∩ S. Since K/L is
Galois and R is the integral closure of S in K, there exists an integer l which is independent
of p such that for any integer N , x ∈ p(N ) implies NK/L (x)l ∈ q(N ) (see Proposition 3.10 of
[Ho71]). Since S is a regular local ring we have q(k) ⊂ mk0 for any k (see [Ho71] p. 9). By
Lemma 2.7 there exists an integer r ∈ N such that for any f ∈ R, m0 (NK/L (f )) ≤ rm(f ).
rk
k
Thus if x ∈ p(rkl) then NK/L (x)l ∈ mrkl
0 , hence NK/L (x) ∈ m0 and we have x ∈ m . Finally
we obtain
p(rlc) ⊂ mc ∀p ⊂ R prime and c ∈ N.
Next, keep the assumptions that R is complete and p is prime, but drop the assumptions
that R is normal and that the extension L → K is Galois. Let p = char K if char K > 0 and
p = 1 otherwise. Let pn be the inseparability degree of K over L. Let Ks be the maximal
separable extension of L in K and set Rs := R ∩ Ks . Then Rs is a complete local domain
n
whose maximal ideal ms equals m ∩ Rs and Rp ⊂ Rs . The ideal ps := p ∩ Rs is a prime
n
ideal of Rs and pp ⊂ ps . For any element y ∈ p(c) there exists a ∈ R\p such that ay ∈ pc .
n
n
n
(c)
Thus ap y p ∈ pcs and y p ∈ ps .
n
n
n
(αpn c)
(αc)
⊂ mps c . Thus
If ps ⊂ mcs for any integer c, then for any x ∈ p(αp c) we have xp ∈ ps
n
n
xp ∈ mp c and by Rees theorem there exists a constant c0 depending only on R such that
x ∈ mc−c0 . Thus we may assume that K/L is separable.
In this case let us denote by K1 a finite separable field extension of K which is normal over
L and let R1 be the integral closure of R in K1 . Then R1 is a direct sum of complete local
rings and since R1 is a domain (it is a subring of a field) it is a complete local domain.
Let m1 be the maximal ideal of R1 . By Lemma 2.4 [Ro09] there exists α ∈ N such that
c
mαc
1 ∩ R ⊂ m for any integer c. Since R → R1 is finite there exists a prime ideal p1 of R1
lying over p. Thus by replacing R and p by R1 and p1 , we may assume that K/L is Galois
and R is normal and this case has been proved above.
Now let us assume that R is an analytically irreducible local ring and I = p is a prime
b
ideal of R. Let us consider an irredundant primary decomposition of pR:
b = q1 ∩ · · · ∩ qs
pR
b denotes the completion of R and the qi are primary ideals of R.
b Let pi be the
where R
b
radical of qi for all i and set W := R\ ∪i pi . Since pi ∩ R = p for any i, we have an inclusion
of multiplicative systems: R\p ⊂ W . Thus we have for any integer n:
(n)
b
bb
bW ∩ R
b ⊂ p(n) R
bR\p ∩ R ⊂ pn R
p(n) = pn RR\p ∩ R ⊂ pn R
1
R\p ∩ R = p1 .
1
By the previous case there exist a and b such that
b c ∩ R = mc the theorem is proved in this case.
m
(ac+b)
p1
bc
⊂ m for all integers c. Since
Finally let us assume that R is an analytically irreducible local ring and I is any ideal of R,
not necessarily prime. Let p1 ,..., ps be the associated primes of I and set W = R\ ∪i pi .
16
BY G. ROND AND M. SPIVAKOVSKY
Then we have for any integer n:
(n)
I (n) = I n RW ∩ R ⊂ pn1 RW ∩ R ⊂ pn1 RR\p1 ∩ R = p1 .
Since the theorem is proved for the symbolic powers of p1 , this proves the theorem for any
ideal I.
References
[Ab56] S. S. Abhyankar, On the valuations centered in a local domain, Amer. J. Math, 78, 1956, 321-348.
[ELS03] L. Ein, R. Lazarsfeld, K. Smith, Uniform approximation of Abhyankar valuation ideals in smooth
function fields, Amer. J. Math, 125, 2003, 409-400.
[Ho71] M. Hochster, Symbolic powers in Noetherian domains, Illinois J. Math., 15, 1971.
[HH02] M. Hochster, C. Huneke, Comparison of symbolic and ordinary powers of ideals, Invent. Math., 147,
2002, 349-369.
[HKV09] C. Huneke, D. Katz, J. Validashti, Uniform equivalence of symbolic and adic topologies, Illinois J.
Math, 53, 2009, 325-338.
[HS0] C. Huneke, I. Swanson, Integral closure of ideals, rings and modules, London Mathematical Society
Lecture Note Series, 336, Cambridge University Press, Cambridge, 2006.
[Iz85] S. Izumi, A mesure of integrity for local analytic algebras, Publ. RIMS, Kyoto Univ., 21, (1985),
719-736.
[Ma80] H. Matsumura, Commutative Algebra, Mathematics Lecture Note Series, 56, Benjamin/Cummings
Publishing Co., Inc., Reading, Mass., 1980.
[Mo13] A. More, Uniform bounds on symbolic powers, J. of Algebra, 383, 2013, 29-41.
[Na62] M. Nagata, Local rings, Interscience Tracts in Pure and Applied Mathematics, 13, New York, (1962).
[Re55] D. Rees, Valuations associated with a local ring (I), Proc. London Math. Soc., 5, 1955, 107-128.
[Re56a] D. Rees, Valuations associated with ideals (II), J. London Math. Soc., 31, 1956, 221-228.
[Re56b] D. Rees, Valuations associated with a local ring (II), J. London Math. Soc., 31, 1956, 228–235.
[Re89] D. Rees, Izumi’s theorem, Commutative algebra (Berkeley, CA, 1987), 407–416, Math. Sci. Res. Inst.
Publ., 15, (1989).
[Ro09] G. Rond, Homomorphisms of local algebras in positive characteristic, J. of Algebra, 322, no. 12,
(2009), 4382-4407.
[Sp90] M. Spivakovsky, Valuations, the linear Artin approximation theorem and convergence of formal functions, Proceedings of the II SBWAG, Santiago de Compostela, Felipe Gago and Emilio Villanueva,
editors, ALXEBRA, 54, (1990), 237-254.
[Sw00] I Swanson, Linear equivalence of topologies, Math. Z., 234, 2000, 755-775.
[Z49] O. Zariski, A simple analytical proof of a fundamental property of birational transformations, Proc.
Nat. Acad. Sci., 35, 1949, 62-66.
[ZS60] O. Zariski, P. Samuel, Commutative Algebra II, D. Van Nostrand, Springer, Princeton, New Jersey,
1960.
Institut de Mathématiques de Marseille, Université d’Aix-Marseille, Campus de Luminy, Case
907, 13288 Marseille Cedex 9
E-mail address: [email protected]
Institut de Mathématiques de Toulouse Université Paul Sabatier, 118 route de Narbonne,
31062 Toulouse Cedex 9 , France
E-mail address: [email protected]
| 0math.AC
|
A validation methodology aid for improving a thermal
building model : case of diffuse radiation accounting in a
tropical climate
A.J.P. LAURET, T. A. MARA, H. BOYER, L. ADELARD, F. GARDE
Université de La Réunion, Laboratoire de Génie Industriel, Equipe Génie Civil, BP 7151
15 avenue René Cassin, 97705 Saint-Denis Cedex, Ile de la Réunion, FRANCE
__________________________________________________________________________________________
Abstract
As part of our efforts to complete the software CODYRUN validation, we chose as test building a block of
flats constructed in Reunion Island, which has a humid tropical climate. The sensitivity analysis allowed us to
study the effects of both diffuse and direct solar radiation on our model of this building.
With regard to the choice and location of sensors, this stage of the study also led us to measure the solar
radiation falling on the windows. The comparison of measured and predicted radiation clearly showed that our
predictions over-estimated the incoming solar radiation, and we were able to trace the problem to the algorithm
which calculates diffuse solar radiation. By calculating view factors between the windows and the associated
shading devices, changes to the original program allowed us to improve the predictions, and so this article shows
the importance of sensitivity analysis in this area of research.
keywords: model validation; shadow mask; sensitivity analysis; diffuse solar radiation; time-frequency analysis
__________________________________________________________________________________________
obtained for direct radiation is applied to the total
incident radiation.
1. Introduction
This article treats the solar radiation which enter
In recent years, major efforts have been made to
a building through windows (or other glass
develop methods for estimating the radiation flux
frontages), and which are protected by a veranda,
falling on various surfaces (walls of buildings, solar
which is a very common set-up in tropical
and photo-voltaic panels, etc.). Many theoretical
architecture. In this case, the simplifications
models are therefore available, covering both short
mentioned earlier would lead us to overestimate the
and long wave radiation and also radiation sources
indoor solar radiation, and we will see that certain
(such as direct or diffuse radiation from the sky [1]
improvements should be made to the theoretical
or ground reflected diffuse radiation[2-3])
model.
With regard to the interaction of these fluxes
with the envelope of a building, their coupling with
2. The context of the study
other transfer phenomena has also been studied
intensively [4-5]. These aspects can be successfully
2-1 The Climate
incorporated into program codes, provided we take
into account accepted (and justified) assumptions
concerning the solar radiation (such as the partition
Reunion Island is situated at a latitude of 21°
between short and long wavelengths, the
South and a longitude of 55° East. The climate is
‘transparency’ of the air as far as radiation are
humid tropical. There is a dry season (May to
concerned, etc.) and also concerning the walls and
October)- mainly cool and dry, predominated by the
windows
(unidirectional
conduction
heat
trade winds and a wet season (November to April)exchanges, constant radiation properties, etc.).
hotter and wetter with light winds. In a humid
The calculation of the impact of near or further
tropical climate, the source of discomfort arises
shading devices is also well-known, carried out
from a temperature increase due to bad architectural
using either trigonometry [6-7] or other methods
design, as far as insulation is concerned. Solar
[8]. Design of shading devices is easily
radiation accounts for 80% with the rest due to
accomplished by classical methods [9-10], but only
conduction exchanges. Under the humid tropical
in the case of direct beam radiation. However, most
climate, the readings made on the short wave
of the program codes do not calculate in detail the
diffuse radiation show the importance of this kind
influence of near shading devices on diffuse
of input. In particular, a quick analysis of the daily
radiation, apart from certain reference programs
graphs on Reunion Island shows often significant
such as DOE-2 [11] and S3PAS [12]. In
diffuse radiation when the direct beam is low. The
TRNSYS[13] for example, the shading reduction
diffuse radiation therefore represents a significant
Buildenv12
fraction of the total (40% on the coast and about
50% for the interior of the island).
As a result, the solar protection of windows is
fundamental, not only because this represents 1520% of the thermal gains, but also because the
windows contribute to the increase in the
discomfort experienced by the occupant, due to the
instant heating of the ambient air temperature and
exposure to direct or reflected sunlight. All the
windows must therefore be protected by some sort
of shading device (overhanging horizontal
canopies, screens, shadow masks etc.).
The measurement programme lasted throughout
January 1998. For the meteorological data,
equipment was set-up on the roof of the dwelling,
giving half-hourly data for outdoor dry temperature,
sky temperature, relative humidity, global and
diffuse horizontal radiation and also wind speed
and direction.
Ambient temperatures (dry and resultant) and
relative humidity were measured in two places and
at different heights in each room of the dwelling.
Furthermore, some interior and exterior wall and
roof surface temperatures were included in the
experimental results.
2-1 The chosen building
2-2 The software CODYRUN
We have experimented with a typical dwelling
of collective housing in Reunion island. This
assessment is part of a technical evaluation of the
building standards [14] for French overseas
territories. The dwelling, represented in Fig. 1,
includes three bedrooms and a living room. A
veranda in effect enlarges the living room through
the use of large bay windows.
Our analysis is based on the temperatures measured
in the living-room, as the standard ECODOM
concentrates exclusively on comfort improvement
of living areas.
Fig.1. Typical dwelling
CODYRUN is a detailed building thermal
simulation software regrouping design and research
aspects. This software has already been covered in
various publications [15-18]. Essentially, it
involves multiple-zone software including natural
ventilation and humidity transfer. Another key
feature of CODYRUN is its multiple-model aspect.
Indeed, before one simulation, the expert user has
the possibility to choose between different models.
The counterpart of this flexibility is the necessity
for the building physics researcher to set up various
models of different accuracy for each physical
phenomenon.
The multiple-model aspect is detailed in [16],
the thermal model in [17], the pressure airflow in
[18], the data-structuration and the description of
the front-end in [15]. Paper [14] focuses on both
guidelines for tropical climate based on simulations
with CODYRUN and split system modelling
embedded in the software.
With regard to the validation of the program,
CODYRUN was successfully ‘BESTESTed’ [21]
and compared with other in situ measurements.
For the building chosen for this study, because
of the large surface area of the bay windows
(particularly on the West side), the largest energy
source is from solar radiation. The performance of
the building is therefore heavily influenced by the
incoming radiation. It is recognised that for homes
in cold climates with moderately sized windows
and high levels of foundation insulation, energy
simulation software should be fairly accurate [22].
The afore-mentioned article emphasises that for
homes in hot climates, with large solar apertures
and thermal mass, the situation is less clear.
Simulation results for these situations should be
used with caution.
Bearing in mind these elements, a recognised stepby-step validation procedure was used, including in
particular a sensitivity analysis which is discussed
in the following section.
Buildenv12
Sensitivity analysis (SA) of the model output is
a very important stage in model building and
analysis. In the building thermal simulation field,
SA is being applied more and more [23-26]. Indeed,
SA can help increase the reliability of the
predictions made by building thermal simulation
software.
This method identifies the most influential
factors and inputs and evaluates their effect. In the
validation methodology, the results are helpful as
they give precise information on how to plan
experiments. In particular, SA pinpoints the
parameters or inputs that must be known accurately.
In this way, we know which special measurements
should be performed to ensure reliable predicted
results.
capacity. These latter three parameters represent, in
effect, the thermal inertia of the building. As far as
the heat sources are concerned, this study clearly
shows the importance of properly taking into
account solar heat gains (both direct and diffuse).
1
0.9
0.8
0.7
%of total variance
3 Use of sensitivity analysis to plan the
experiment
0.6
Thickness
Diffuse solar rad.
Direct solar rad.
outdoor air Temp.
Spec. Heat cap..
density
0.5
0.4
0.3
0.2
0.1
0
5
10
15
20
Time (h)
3.1. The methodology
Fig.2. Assessment of the most important parameters
Using a MATLAB program which was
specially developed for the chosen building, we
conducted a sensitivity study, with the aim of
determining the most significant parameters and
input data for the prediction of the indoor dry air
temperature. The analysis was complete, in the
sense that it treated both parametric variables
(material properties associated with the building
envelope) and non-parametric (meteorological
data). We will not give details of the method, as
these can be found in another article [27].
The proposed method involves performing
several simulation runs by oscillating each input
sinusoidally over its range of interest. Analysing the
spectrum (Fourier transform or power spectral
density) of the output reveals the most influential
factors.
1024 simulations were performed by making
each factor (input or parameter) vary as a sinusoid
ranging over 10% with respect to its base-case
value. In the following study, we are looking for the
most important parameters for the predicted indoor
air temperature ; once the simulations are achieved
we can calculate the power spectrum density (PSD)
of the residual.
As PSD is defined as the spectral decomposition
of a signal’s variance, we evaluate the influence of
each parameter upon the output calculating the
variance of the temperature differences ( var( Ti ) )
due to each parameter.
We note in particular that the contribution due to
diffuse solar radiation increases at 3p.m.
As the absorptance of the walls for short
wavelengths do not appear in the preceding
analysis, the sensitivity of the apartment to shortwave radiation depends on solar gains through the
windows.
Therefore, in order to determine the interaction
between the building envelope and the outside solar
radiation, in addition to the measurements
mentioned earlier, we introduced two pyranometers
on each side of the largest bay window. This was in
the living room, facing West and so heavily
exposed in the afternoon. In the photograph below,
one can see the bay window as well as the shading
screen. In this way, we obtained a measurement of
the global incident solar radiation (without the
shading effect, because this was placed at the edge
of the plant tub) and of that transmitted by the bay
(after the effect of the shading device and the
glass).
3.2. Results
The figure below shows that the input data and
parameters which account for 90% of the residual
variance are : direct and diffuse solar radiation,
concrete walls thickness, density and specific heat
Fig.3. The dwelling (West side)
Buildenv12
The
initial
detailed
simulation
software
(CODYRUN) was modified to include the exterior
global and interior radiation calculations (for each
window) in the output file.
This study may guide the building physics
researcher in the right direction for improving the
model.
4.1. Use of time-frequency analysis
4. Initial comparison of model and
measurements
For this ‘initial’ comparison, the diffuse exterior
are considered to be unaffected by the presence of
near-field shading. For the weather conditions
experienced at the time of the measurements, all the
exterior and interior openings were closed and
cautiously sealed. The building was therefore
behaving in a thermally pure mode, with no
incoming airflow. For modelling purposes, five
zones were considered, including the zone ‘living
room’. All our analysis is centred on comparisons
of measurements with the model for dry air
temperatures in the living room.
MEASUREMENT
Temp (°C)
CODYRUN
35
34.5
34
33.5
33
32.5
32
31.5
31
30.5
30
1
3
2
4
Time ( Day)
Fig.4. Dry air temperature of the living room
Figures 4 and 5 show the comparison between the
predicted indoor dry air temperature in the living
room and its measurement. The maximum
acceptable measurement error is 0.5C . As the
residual does not fall within these limits, the model
is rejected. The mean and standard deviation are the
following :
Mean
-0.39
Std
0.369
Residual ( °C)
1
0.5
0
-0.5
-1
-1.5
1
2
Time (Day)
3
4
Fig.5. Residual
To find the source of the error, one has to search for
correlation between the inputs and the residual.
Instead of classical spectral analysis [28-29] that
shows some limitations when the inputs are
correlated (as in the case of meteorological data),
time-frequency analysis is a powerful tool capable
of de-segregating the actual effects of several inputs
at different time-scales.
Time-frequency analysis is a modern technique
to extract information from signals. The first and
natural time-frequency tool is the short-scale
Fourier Transform (SFTT). The key feature of these
tools is that they simultaneously decompose a
signal in time and frequency. One can imagine that
the information obtained is richer than that given in
only spectral or in time domain analysis. The
representation of an SFTT is called the
spectrogram. The goal of our study is to check
whether the spectrogram of the residual is similar to
the one of the inputs.
In order to obtain useful spectograms, low
frequency data must be filtered out. For this
purpose, a fourth-order Butterworth is used to
remove frequencies lower than 0.02h -1. Filtering is
justified in our case as the range of interest is
around 1 day-1 harmonic (~0.041h -1).
Unlike classical spectral analysis, spectograms
of climatic data show that the time-frequency
signature of the outdoor air-temperature is very
different from that of the solar radiation. It is
possible here to distinguish outdoor air temperature
from the other inputs. The main drawback of
classical frequency analysis is avoided.
4.2. Application to our case
Spectograms of diffuse and direct solar
radiation illustrate the complementarity of these
two inputs. Consequently, it will be difficult to
identify which of these 2 inputs is really correlated
with the residual. Nevertheless, physically, the
influence of the diffuse component seems more
relevant. Comparison of the 2 spectograms clearly
shows a correlation between diffuse solar radiation
and the residual at times 90 and 130. The source of
the error seems to be in the way the model account
for diffuse solar radiation
Buildenv12
Signal in time
-1
Wind speed (m.s )
Residual (°C)
Signal in time
0.5
0
-0.5
2
1
0
-1
Spectrogram of wind speed
0.9
0.9
0.8
0.8
Energy spectral density
0.6
0.5
0.4
0.3
0.7
-1
-1
Frequency [h ]
0.7
Frequency [h ]
Energy spectral density
Spectrogram of the residual
0.2
0.1
0.1
0
80
100
120
Time [h]
140
1.8
160
0.9
0
60
x 10 4
Dif. solar rad. (W\.m-²)
60
Out. air temp. (°C)
0
Signal in time
4
2
0
-2
0
0.8
0.6
0.5
0.4
0.7
-1
Frequency [h ]
-1
Frequency [h ]
160
200
0.8
0.3
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0.1
0
0
60
80
x 104
100
120
Time [h]
140
160
3
2
1
0
60
x 10 8
80
100
120
Time [h]
140
160
Signal in time
-
Dir. solar rad. (W\.m ²)
140
Spectrogram of diffuse solar radiation
Energy spectral density
Energy spectral density
100
120
Time [h]
Signal in time
0.9
0
80
400
0.9
0.7
4
0.4
0.3
Spectrogram of outdoor air temperature
8
0.5
0.2
0
1114 557
0.6
800
600
400
200
0
-200
Spectrogram of direct solar radiation
0.9
Energy spectral density
0.8
-1
Frequency [h ]
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
8
4
0
x 108
60
80
100
120
Time [h]
140
160
Fig.6. Spectograms of the residual and the inputs
5. Improvement of the model
5.1. Comparison of incoming radiation
Because of the two pyranometers installed after
the sensitivity analysis, we are able to compare the
global exterior solar radiation with the interior one,
both for the actual measurements and for the model
predictions. Given that the agreement for the
exterior values is excellent, we clearly see that our
initial model overestimates by 30 to 40% the
interior radiation contributions in the living room
(see fig.7).
Considering that the measurement error is
approximately 6%, and that the error due to
imprecise positioning of the pyranometers is
neglectable, an error of 30 to 40% is evidently very
significant. Further, as the sensitivity analysis had
shown that the transmittance of the glass was not an
Buildenv12
important factor, the near shading device formed by
the protection around the west bay window is
clearly responsible for the overestimation of the
solar gains. As CODYRUN has already been tested
with respect to the effect of shading devices on the
direct radiation [21], we decided to improve the
model by introducing the possibility of reducing the
diffuse solar radiation with near shading devices.
Vert. Glob. Solar
-2
Radiation (W.m )
CODYRUN
MEASUREMENT
Fig.9. The bay window in its egg-crate
250
200
150
100
50
0
1
3
2
4
Time (Day)
Fig.7. Comparison of the global solar radiation
5.2 The chosen method
As a near shading device is generally composed
of several elements (top-flap, vertical fins, low
walls,…), the calculation of the analytic shape
factor involving the window and the shading device
proved complex. We therefore preferred a
simplified approach which uses properties of the
view factors. By using projections perpendicular to
the planes containing the windows and each
element of the shade, the calculation of the global
shape factor can be achieved by combining several
view factors of the same type ; these view factors
are those obtained for orthogonal parallelepipeds
with a common edge [29]
.
l
A
h
w
90°
As a first step, we limited ourselves to one
dimension per element of the shade. In this way the
left side was represented by a parallelepiped of
width equal to the top side. An alternative, more
comprehensive approach has been tried, but was not
integrated into the program, because of the small
(approximately 1%) reduction obtained.
One should note that our method does not take
into account the solar radiation reflected off the
shade. This has been studied by Barozzi [6] for
example, and could prove useful for a more detailed
calculation of the radiation diffused around the
window. Finally, the calculation for a near shading
device uses a constant percentage of reduction for
the diffuse solar radiation.
5.3. Integration into CODYRUN
Figure 10 describes the data inputs associated to
the West bay window. The dimensions of the
elements composing the shade are given in a
standard way. Some parts of the data input window
are disabled because some of these dimensions are
considered infinite (the width of the top flap, for
example).
Given the importance of the effect of the near
shading devices on the incident diffuse around the
window (see the preceding method), the percentage
of reduction of diffuse is now indicated and taken
into account at each step of simulation.
B
Fig.8. Calculation of the shape factor
For the West bay window which concerns us
here (see Fig. 3), this approach requires us to find,
in the following figure 9, the view factor between
the window (coloured grey) and the elements
composing the shade (hatched).
Fig. 10. Shadow mask data capture window
Buildenv12
In the case which interests us, 60% of the incident
diffuse radiation on the West wall are blocked by
the near shading device before reaching the glass.
.
MEASUREMENT
CODYRUN
Temp (°C)
35
34.5
34
33.5
6. New comparison
33
32.5
32
6.1. Comparison of fluxes
31.5
31
30.5
Vert. Glob. Solar
-2
Radiation (W.m )
CODYRUN
MEASUREMENT
30
1
2
3
4
Time (Day)
250
200
Fig.12. New comparison of the dry air temperature
150
100
50
0
1
2
3
4
Regarding the validation, the new model has
allowed us to increase the number of points for
which the residual is included within the maximum
acceptable error limits, that is -0.5 to +0.5 °C.
Time (Day)
Fig.11. New comparison of the global solar radiation
In the interior, the new model is manifestly
better than the old one at calculating the radiation
transmitted into the zone. The relative average daily
error in the density of transmitted flux has changed
from 22% to 4% over the whole period. In
particular, the prediction due to the model is much
improved during days 4 and 5, which correspond to
the periods when the diffuse radiation was
dominant (the relative error is reduced from 32% to
1% in days 4 and 5).
It should be noted that the model used here to
represent the diffuse radiation is known as
‘isotropic’. Other simulations have shown us that,
as could be expected, the choice of a non-isotropic
model (in which some of the diffuse radiation is
directional) modifies slightly the incoming energy
contributions. In fact, the direct and diffuse parts of
the flux falling on the bay window are respectively
reduced by the direct solar factor associated with
the shading device and the reduction factor for the
diffuse part. Our building model is therefore
sensitive to the diffuse radiation reconstitution.
6.2 Comparison of temperatures
The variable which interested us most was the
dry air temperature in the living room, and the
modification of the model led to a better prediction
of this temperature (fig. 12). Thus, over the 5 day
sequence considered, the average residual changed
from about 0.4°C to 0°C, with a standard deviation
of about 0.3°C slightly improved (from 0.37 to
0.34). Further, by concentrating the study on one
day when the diffuse solar radiation was more
important, an even better improvement was
observed.
Fig.13. New and old residual
As the residual curve obtained was nearly
always within these limits , the new building model
is therefore considered as improved.
.
Mean
-0.04
Std
0.336
We should note however that the method only
leads to an improvement in the prediction of the
steady state (correction of the mean). It is
nevertheless an important step, which has to be
completed before moving on to improve the
prediction of the dynamic response.
7. Conclusion
We have used a step-by-step approach to
validate the thermal behaviour of a simulation
program. We have seen the advantages of a
systematic application of the validation procedure.
The sensitivity analysis combined with appropriate
sensors has allowed us to highlight the weaknesses
of the model for calculating the energy
contributions due to diffuse solar radiation from the
windows. In the case of experimental validation,
sensitivity analysis has once again proved its ability
to be a powerful tool for improving simulation
models.
Buildenv12
Regarding the aspects as basic as the calculation
of the solar gains due to the windows, it is
noteworthy that the use of computer codes written
for an architecturally and climatically different
context should be done with care. In conclusion,
this paper illustrates that under our tropical climate
and for the architectural rules currently applied, an
adapted model must be used.
[12] S3PAS 1.0, Simulacion de Sistemas Solares
Pasivos. Manual de Usuario. Catédra de
Termotecnia, ESII, Sevilla, 1988
[13] TRNSYS, A transient system simulation
program. Solar Energy Laboratory, University of
Wisconsin, Madison, WI (1990)
[14] F. Garde, H. Boyer, J.C. Gatina, Demand side
REFERENCES
[1] A. Lebru, Estimation des irradiations solaires
horaires dans un plan quelconque à partir de la
donnée de l'irradiation horaire globale (et
éventuellement diffuse) horizontale. Cahiers du
Centre Scientifique et Technique du Bâtiment,
1983, n°239, Cahier 1847. 39 p.
[2] P. Ineichen, R. Perez, R. Seals, The importance
of correct albedo determination for adequately
modeling energy received by tilted surfaces, Solar
Energy, 1987, Vol. 39, n°4. pp. 301-305.
[3] P. Ineichen, A. Zelenka, O. Guisan, A.
Razafindraibe, Solar radiation transposition models
applied to a plane tracking the sun. Solar Energy,
1988, Vol. 41, n°4. pp. 371-377
[4] A.J. Ness., Solution of a thermal network on a
digital computer, Solar Energy, 3, 1, 37-37 (1959)
[5] B.A. Price, T.F. Smith, Thermal response of
composite building envelopes accounting for
thermal radiation, Energy Conversion and
Management ,36, 1, 23-33 (1995)
[6] G. S. Barozzi, R. Grossa., Shading effect of
eggcrate devices on vertical windows of arbitrary
orientation, Solar Energy, 39, 4, 329-342 (1987)
[7] R.E. Jones Jr, Effects of overhang shading of
windows having arbitrary azimuth, Solar Energy,
24, 4, 305-312 (1980)
[8] D. Caccavelli, Modélisation du comportement
thermique des bâtiments multizones : adaptation à
un processus de conception. Thèse : Sci. : Institut
National des Sciences Appliquées de Lyon, 1988.
477 p.
[9] A. Olgyay, V Olgyay, Solar control and shading
Design, Wiley, New York (1977)
[10] E. Mazria, The passive Solar Energy book.
RodalePress, Emmaus, PA (1979)
[11] D. York, C. Cappiello, DOE-2 Engineer’s
Manual, eds Berkeley, CA : Laurence Berkeley
Laboratory, Nov. 1981.
management in tropical island buildings.
Elaboration of global quality standards for natural
and low energy cooling in buildings, Building and
Environment, 34 (1998), 1, pp. 71-84
[15] H. Boyer, Conception thermo-aeraulique de
batiments multizonzes.proposition d’un outil à
choix multiples de modèles. PhD thesis. Institut
National des Sciences Appliquées de Lyon, France
1993.
[16] H. Boyer, F. Garde, J.C. Gatina J. Brau, A
multi model approach of thermal building
simulation for design and research approach.
Energy and Building, 28, pp. 71-78 (1998)
[17] H. Boyer, J.P. Chabriat , B. Grondin Perez, C.
Tourrand, J. Brau, Thermal building simulation and
computer generation of nodal models, Building and
Environment, Vol. 31, n° 3, 1996, pp. 207-214
[18] H. Boyer, A.P. Lauret, L. Adelard, T.A. Mara,
Building ventilation : a pressure airflow model
computer generation and elements of validation,
Energy and Buildings, 29 (1999),pp. 283-292
[20] F. Garde, F. Lucas, H. Boyer, J. Brau.
Multiple model approach of a residential heat pump
for integration in a building thermal simulation
code and comparison with experimental results,
Accepted at ASHRAE Transactions, Janv. 2000.
[21] T. Soubdhan, T.A. Mara, H. Boyer, A.
Younès, use of BESTEST procedure to improve a
building thermal simulation program, World
Renewable Energy Congress, Brighton, U.H., pp.
1800-1803 Part III, 1-7, July 2000
[22] Energy Design update, April 1996,pp 5-15
(Source: AIVC recent additions to airbase, N° 73,
Sept. 96, p.3
[23] K. Lomas, H. Eppel, 1992, Sensitivity analysis
techniques for building thermal simulation
programs, Energy and Buildings ,Vol. 19, pp. 21-44
[24] J.M. Furbringer., Sensitivity Analysis for
modellers, Air Infiltration Review, Vol. 17, n° 4,
Sept. 1996.
[25] N. Rahni, N. Ramdani, Y. Candau, G. Guyon,
Applicatio of group screening to dynamic building
Buildenv12
energy simulation models, J. Statist.Comput.Simul,
Vol.57, pp.285-304,1997.
[26] P. Aude, C. Beghein, P. Depecker, C. Inard,
perturbation of the input data of models used for the
prediction of turbulent air flow in an enclosure,
Numerical Heat Transfer part B, Vol. 34, Issue 2.
pp. 139-164.
[27] T.A . Mara, F. Garde, H. Boyer, M. Mamode,
Empirical Validation of the thermal model of a
passive Solar test-cell, submitted at Energy and
Buildings, Aug. 2000
[28] N. Ramdani, Y. Candau, S. Dautin, S. Delille,
N. Rahni, P. Dalicieux, How to improve building
thermal simulation programs by use of spectral
analysis, Energy and buildings 25 (1997) 223-242
[29] F.P. Incropera, D.P. DeWitt, Introduction to
Heat transfer, 3rd Edition, J. Wiley and Sons, 1996.
| 5cs.CE
|
arXiv:1606.01888v2 [math.GR] 6 Feb 2017
IRREDUCIBLE p-CONSTANT CHARACTERS OF
FINITE REFLECTION GROUPS
MARCO ANTONIO PELLEGRINI
Abstract. A complex irreducible character of a finite group G is said to be
p-constant, for some prime p dividing the order of G, if it takes constant value
at the set of p-singular elements of G. In this paper we classify irreducible pconstant characters for finite reflection groups, nilpotent groups and complete
monomial groups. We also propose some conjectures about the structure of
the groups admitting such characters.
1. Introduction
Let G be a finite group. Fixing a prime p that divides the order of G, we denote
by Σp (G) the set of the p-singular elements of G, i.e. the elements whose orders
are divisible by p. We say that a (complex) irreducible character χ of G is pconstant if it takes constant value at Σp (G), that is, if there exists a constant cχ
such that χ(g) = cχ for all g ∈ Σp (G). By [9, Lemma 2.1], the constant cχ must
be a rational integer. It is easy to see that the irreducible characters χ of p-defect
zero (i.e. whose degree equals the order of a Sylow p-subgroup of G) are precisely
the irreducible p-constant characters with cχ = 0. In view of this, we can look
at p-constant characters as a generalization of p-defect zero characters (a different
generalization to the reducible case was considered in [8]). In this paper we will
always consider the case cχ 6= 0: for future convenience, we denote by Υp (G) the
set of the irreducible p-constant characters of G of non-zero p-defect.
The problem of classifying irreducible p-constant characters was firstly introduced in [9], where finite quasi-simple groups have been studied. In particular, the
following result was obtained.
Theorem 1.1. [9, Theorem 1.5] Let G be a finite simple group admitting an irreducible p-constant character χ for some prime p dividing the order of G. Then,
one of the following holds:
(i) cχ ∈ {0, ±1};
(ii) G = M22 , p = 3, χ(1) = 385 and cχ = −2;
(iii) G is a group of Lie type of characteristic r 6= p with a non-cyclic Sylow
p-subgroup.
In the same paper, also the p-constant characters of the symmetric groups were
classified. Starting from that result, in the present paper we study the other finite
2010 Mathematics Subject Classification. 20C15, 20F55, 20D15.
Key words and phrases. Coxeter groups; Nilpotent groups; Irreducible characters; p-singular
elements; wreath products.
This work was supported by the “National Group for Algebraic and Geometric Structures, and
their Applications” (GNSAGA-INDAM).
1
2
MARCO ANTONIO PELLEGRINI
reflection groups as well as other notable classes of groups as nilpotent groups,
direct and wreath products of groups.
In Section 2 we prove in particular that any non-linear irreducible p-constant
character of a finite nilpotent group must be of p-defect zero. In Section 3 we
classify the finite reflection groups admitting p-constant characters proving the
following results.
Theorem 1.2. Let G be a finite reflection group. Then, for any prime p dividing
the order of G and for any χ ∈ Υp (G) we have cχ = ±1.
Theorem 1.3. Let G be a finite reflection group. Then, Υ2 (G) 6= {1G} if and only
if G is a dihedral group of order 2n with n odd. In this case, Υ2 (G) consists of the
two linear characters of G.
The classification of p-constant characters for groups of type Bn relies on a combinatorial version of Murnaghan-Nakayama rule given in [7]. The same techniques
used here can be also applied for studying endotrivial modules of such groups (as
done for symmetric groups in [5]). In Section 4 we present some open problems
and conjectures that we hope can attract the interest on this subject. For any
odd prime p, we construct a Frobenius group admitting an irreducible p-constant
character χ with cχ = 2. We also classify the p-constant characters of PSL3 (q)
and PSU3 (q), proving that case (iii) of Theorem 1.1 contains an infinite number
of groups. It is worth of observe that by [6, Theorem 6.1] case (iii) may contain
only groups whose Sylow p-subgroups have rank ≤ 2, with the possible exceptions
of PSL5 (q), PSU5 (q) and E8 (q).
Notations. To avoid misunderstanding with the reflection groups, we denote the
cyclic group of order n by Cyc(n). By Alt(n) and Sym(n) we mean, respectively,
the alternating and the symmetric group of degree n. Given a finite group G and a
positive integer n, Gn denotes the direct product of n copies of G. If p is a prime,
|G|p denotes the order of a Sylow p-subgroup of G.
2. Linear characters and nilpotent groups
Let G be a finite group. Clearly, the principal character 1G of G is p-constant
for any prime p dividing the order of G (i.e., 1G ∈ Υp (G) whenever |G|p > 1). More
in general, for linear characters we prove the following.
Proposition 2.1. Let χ be a linear character of a finite group G. Then, χ is
p-constant for some prime p if, and only if, one of the following cases occurs:
(i) cχ = +1 and Σp (G) ⊆ ker(χ);
(ii) cχ = −1, p = 2 and G = ker(χ) ⋊ hxi, where ker(χ) has odd order and x is
an involution.
Proof. Since χ is linear, if χ is p-constant then cχ ∈ Z and so cχ = ±1. Clearly,
if cχ = 1 then Σp (G) ⊆ ker(χ). If cχ = −1 then p = 2 and ker(χ) has odd order.
Working in H = G/ ker(χ), consider the linear character χ̄ induced by χ on H.
Since χ̄ is a faithful linear character of H, then H is cyclic, say H = hx̄i. From
χ̄(x̄) = −1 we get x̄2 ∈ ker(χ̄) and so |H| = 2. Hence, G = ker(χ) ⋊ hxi for some
involution x.
We start our analysis considering finite p-groups and nilpotent groups. We first
recall the following result.
IRREDUCIBLE p-CONSTANT CHARACTERS OF FINITE REFLECTION GROUPS
3
Lemma 2.2. [9, Lemma 2.2] Let G be a finite group and let 1G 6= χ ∈ Υp (G) for
some prime p dividing the order of G. Then, one of the following holds:
(i) p does not divide |Z(G)|, Z(G) ≤ ker(χ) and the corresponding character
χ̄ of G/Z(G) is an irreducible p-constant character;
(ii) p = 2, |G|2 = 2 and G = ker(χ) × O2 (G).
As consequence of the previous lemma we get the following.
Corollary 2.3. Let p be a prime and let G be a finite p-group. Then Υp (G) 6= {1G }
if and only if |G| = p = 2.
Proposition 2.4. Let H, K be two finite groups and let G = H ×K. Let χ ∈ Υp (G)
for some prime p dividing the order of G. If p divides both |H| and |K|, then χ = 1G .
If p divides |H| but does not divide |K|, then χ = ψ ⊗ 1K for some ψ ∈ Υp (H).
Proof. Suppose that p divides the order of H. Since G is a direct product of groups,
χ = ψ ⊗ ϕ for some ψ ∈ Irr(H) and ϕ ∈ Irr(K). Fix an element h ∈ Σp (H): for
any k ∈ K, we have hk ∈ Σp (G) and χ(hk) = ψ(h)ϕ(k) = cχ 6= 0. Hence, ϕ(k)
has the same value for all k ∈ K and so ϕ = 1K , i.e. χ = ψ ⊗ 1K . It follows that
for any h ∈ Σp (H) we have χ(h) = ψ(h) = cχ , proving that ψ is p-constant with
cψ = cχ 6= 0.
Now, assume that p also divides the order of K. Proceeding as before, by
symmetry we obtain that χ = 1H ⊗ 1K = 1G .
Corollary 2.5. Let G = H × K, where H is a finite p-group and K is a finite
r-group for some primes p 6= r. If there exists 1G 6= χ ∈ Υp (G), then p = |H| = 2
and χ = ǫ ⊗ 1K , where ǫ is the non-principal character of H.
Proof. By Proposition 2.4, χ = ψ ⊗ 1K for some ψ ∈ Υp (H). Now, since H is a
p-group and χ 6= 1G , by Corollary 2.3, we have p = |H| = 2 and ψ = ǫ.
Corollary 2.6. Let G be a finite nilpotent group and let χ ∈ Υp (G). Then χ is
linear and either χ = 1G or p = |G|2 = 2.
Proof. The result follows from Corollary 2.5, writing G as direct product of its
Sylow pi -subgroups.
3. Finite reflection groups
In this section we consider finite groups generated by reflections (in real euclidean
spaces). We will refer to [2] for construction, properties and notation concerning
this class of groups. We recall in particular that the finite reflection groups are
precisely the finite Coxeter groups. Completely classified by Coxeter himself, they
are direct products of members of four infinite families (An , Bn , Dn , I2 (n)) and six
‘exceptional’ groups (E6 , E7 , E8 , F4 , H3 , H4 ). Therefore, by Proposition 2.4 we
may restrict our attention to the ‘irreducible’ finite reflection groups.
The p-constant characters of exceptional reflection groups can be classified using
the character tables available, for instance, in [1]. In the following proposition, χm
denotes the m-th irreducible character, according to the ordering of [1, CTblLib
1.2.2].
Proposition 3.1. Let G be a reflection group of type E6 , E7 , E8 , F4 , H3 or H4 and
let 1G 6= χm ∈ Irr(G). Then χm ∈ Υp (G) if and only if one of the following cases
occurs:
4
MARCO ANTONIO PELLEGRINI
(i) G = E6 , p = 5 and (m, χm (1)) ∈ {(5, 6), (13, 24), (23, 64), (25, 81)};
(ii) G = E7 and
p = 5: (m, χm (1)) ∈ {(9, 56), (11, 84), (18, 189), (22, 216)} or
p = 7: (m, χm (1)) ∈ {(3, 15), (6, 27), (15, 120), (22, 216), (28, 405), (30, 512)};
(iii) G = E8 , p = 7 and (m, χm (1)) ∈ {(9, 50), (18, 300), (36, 972), (61, 3200),
(62, 4096), (66, 6075)};
(iv) G = H3 ∼
= Cyc(2) × Alt(5) and
p = 3: (m, χm (1)) ∈ {(4, 4), (5, 5)} or
p = 5: (m, χm (1)) = (4, 4).
In all these cases we have cχm = ±1.
We now consider the four infinite families An , Bn , Dn and I2 (n). Groups of type
An are isomorphic to the symmetric group Sym(n + 1). The irreducible characters
of Sym(n + 1) are parametrized by the partitions λ of n + 1, see [3]. From [9,
Theorem 1.3, Lemma 3.1, Lemma 3.3 and Proposition 3.4] we get the following
result.
Proposition 3.2. Let G be a group of type An with n ≥ 2. Let χλ be the irreducible
character of G associated to the partition λ of (n+1). Suppose that χλ is non-linear
of non-zero p-defect. Then, χ is p-constant for some prime p ≤ n + 1 if, and only
if, one of the following conditions holds:
(i) n = p − 1 ≥ 2 and λ = (b, 1p−b ), where 2 ≤ b ≤ p − 1;
(ii) n = p ≥ 3 and λ = (b + 1, 2, 1p−b−2 ), where 1 ≤ b ≤ p − 2;
(iii) n + 1 = p + r ≥ 5, 2 ≤ r ≤ p and λ is one of the following partitions
(p − a, r + 1, 1a−1 ), 1 ≤ a ≤ p − r − 1;
(r, b, 1p−b ), 1 ≤ b ≤ r.
In all these cases cχλ = ±1.
Applying Lemma 2.1, it is easy to see that the linear character χ(1n ) of Sym(n)
is p-constant if and only if n = p, p + 1.
Finite reflection groups of type Bn , n ≥ 2, are isomorphic to the wreath products
Cyc(2) ≀ Sym(n). We then consider the more general case of wreath products G =
H ≀ Sym(n) = H n ⋊ Sym(n), where H is any finite group. The classification of
the irreducible p-constant characters of H ≀ Sym(n) (which is sometimes called the
complete monomial group) shall be obtained applying the combinatorial version
of the Murnaghan-Nakayama rule given in [7]. We recall here some of the basic
properties of G with respect to conjugacy classes and irreducible characters. For
further details, see [7].
We can think G as the group of the n × n monomial matrices whose entries are
elements of H. In particular, if π is the k-cycle (i1 , i2 , . . . , ik ) of Sym(n), we denote
by (h1 i1 , h2 i2 , . . . , hk ik ) the permutation matrix corresponding to the cycle π whose
entries 1 are replaced by the elements h1 , . . . , hk ∈ H. If C1 = {1}, C2, . . . , Cr are the
r conjugacy classes of H, an element (h1 i1 , h2 i2 , . . . , hk ik ) is said to be a Cj -cycle
if hk · · · h2 h1 ∈ Cj .
If γ = (γ1 , . . . , γt ) is a partition of n, we write γ ⊢ n (or |γ| = n) and ℓ(γ) = t
→
−
for the number
γ we denote a r-tuple (γ 1 , . . . , γ r ) of partitions
P of iparts of γ. By−
→
i
γ such that i |γ | = n (we use γ ⊢ n to emphasize the integer n). We also allow
→
that some of the components of −
γ can be the empty partition ∅. Then, the set of
→
−
−
−
conjugacy classes of G is given by {C→
γ is the set of elements of
γ : γ ⊢ n}, where C→
G whose Cj -cycles have a cyclic decomposition γ j , for all j = 1, . . . , r. Similarly, we
IRREDUCIBLE p-CONSTANT CHARACTERS OF FINITE REFLECTION GROUPS
5
→
− →
have Irr(G) = {χ γ : −
γ ⊢ n}. In particular, the irreducible characters of G obtained
→
−
→
by inflation from Irr(Sym(n)) are the characters χ γ , where −
γ = (γ, ∅, . . . , ∅) with
((n),∅,...,∅)
γ ⊢ n (so, 1G = χ
).
→
−
χ
The Murnaghan-Nakayama rule gives the value χ→
− of the irreducible character
δ
→
−
→
−
→
−
− . Following [7], given γ ⊢ n, we denote by ⋆ γ =
χ γ on the conjugacy class C→
δ
γ 1 ⋆ . . . ⋆ γ r the skew shape formed by placing the Ferrers diagrams of γ 1 , . . . , γ r
corner to corner (the Ferrers diagram of a partition γ is constructed taking γ1 boxes
in the first row from the top, γ2 boxes in the second row and so on). Let T be a
→
−
→
→
⋆-rim hook tableau of shape −
γ and type δ , i.e. a rim hook tableaux of shape ⋆−
γ,
→
−
where the lengths of the rim hooks are found in the parts of the components of δ .
Let η j,i (T ) be the partition formed by the lengths of the rim hooks of δ j placed in
the Ferrers diagram of γ i in T . Set
⋆ω(T ) = sgn(T )
r
Y
j,i
(ψji )ℓ(η (T )) ,
i,j
where ψji is the value of the i-th irreducible character of H on the class Cj (taking
ψ 1 = 1H ) and sgn(T ) is the product of the signs of the rim hooks in T (i.e., (−1)t−1
where t is the number of rows occupied by the rim hook). Then
X
→
−
γ
χ→
⋆ω(T ),
− =
δ
−
→
−
where the sum runs over all ⋆-rim hook tableaux of shape →
γ and type δ .
Theorem 3.3. Let H be a finite group. An irreducible character χ of H ≀ Sym(n)
of non-zero p-defect is p-constant if and only if p does not divide |H|, H n ≤ ker(χ)
and the corresponding character χ̄ of Sym(n) is p-constant.
→
−
Proof. Let χ = χ γ ∈ Υp (G) for some prime p dividing the order of G = H ≀Sym(n).
→
We want to prove that p ≤ n and −
γ = (γ, ∅, . . . , ∅) for some γ ⊢ n.
Assume first that p divides the order of H. For any conjugacy class Cj of H
→
−
− , C−
→ of G, where δj is the
contained in Σp (H), consider the conjugacy classes C→
η−
j,k
δj
r-tuple whose unique non-empty component is the partition (n) at j-th position and
−
→ is the r-tuple having only two non-empty components: the partition (n − 1) at
ηj,k
→, ⊆
− , C−
j-th position and the partition (1) at k-th position (j 6= k). Clearly, C→
η−
j,k
δj
→
−
→
−
→
−
Σp (G). Since χ γ is non-zero on δj , then γ has a unique non-empty component,
→
say at position i: −
γ = (∅, . . . , ∅, γ, ∅, . . . , ∅), where γ ⊢ n must be a hook, i.e.
→
−
γ
a i
γ = (1a , n − a) for some 0 ≤ a ≤ n − 1. Furthermore, χ→
− = (−1) ψj and so the
δj
character ψ i is a p-constant character of H.
Now, suppose 1 ≤ a ≤ n − 2. Then
→
−
γ
a−1 i i
ψ1 ψj + (−1)a ψ1i ψji = 0,
χ−
→ = (−1)
η−
j,1
→
−
→
−
γ
γ
n−2 i i
n−1 i
ψ1 ψj implies
ψj = χ−
a contradiction. If γ = (1n ), then χ→
→ = (−1)
− = (−1)
η−
j,1
δj
→
−
γ
i i
i
that ψ1i = −1 is the degree of ψ i , an absurd. So, γ = (n). We get χ→
− = ψj = ψ1 ψj
→
−
δj
γ
i i
i
and so ψ i is linear. Now, χ−
→ = ψk ψj implies ψk = 1 for all k 6= j. It follows that
µ−
j,k
either ψ i = 1H (and so χ = 1G ) or p = 2 and G = Cyc(2) ≀ Sym(n).
6
MARCO ANTONIO PELLEGRINI
→
Hence, we may now assume p ≤ n. Our next goal is to prove that −
γ has a
− defined
unique non-empty component. The choice of the conjugacy classes C→
δ
→
below is functional also for proving Proposition 3.5. Assume that −
γ has at least
two non-empty components and write n = pk + t with 0 ≤ t < p. If t = 0, take
→
−
γ
→
−
−
= 0, a contradiction. It t > 0,
α = ((pk), ∅, . . . , ∅). Then C→
−
α ⊆ Σp (G), but χ→
α
take the classes of p-singular elements of G corresponding to the following r-tuples:
−
→ = ((t, pk), ∅, . . . , ∅) , −
→ = ((t), (pk), ∅, . . . , ∅) ,
α
α
1
2
→ = ((t), ∅, . . . , ∅, (pk)) ;
−
→ = ((t), ∅, (pk), ∅, . . . , ∅) , . . . , −
α
α
r
3
−
→
−
→
→
−
β2 = (∅, (t, pk), ∅, . . . , ∅) , β3 = (∅, ∅, (t, pk), ∅, . . . , ∅) , . . . , βr = (∅, . . . , ∅, (t, pk)) .
→
−
γ
→
−
From χ−
→ 6= 0 it follows that γ has at most two non-empty components. Then
α
1
→
−
γ has exactly two non-empty components, say in position i and j, that must be
hooks of length, respectively, pk and t. For any s = 2, . . . , r we have
→
−
→
−
γ
γ
i j
i j
χ−
→ = ǫψs ψ1 6= 0,
→ = ǫψ1 ψ1 = χ−
α
α
s
1
for some ǫ = ±1, and so ψ i = 1H (i.e. i = 1). Now,
→
−
→
−
j
γ
γ
j
χα
→ = ǫψs
−
→ = ǫψ1 = χ−
1
βs
j
for all s = 2, . . . , r also implies ψ = 1H and j = 1, a contradiction.
→
We conclude that −
γ has a unique non-empty component, γ ⊢ n at the i-th
position. We prove now that actually i = 1. For all s = 2, . . . , r we have
→
−
→
−
γ
γ
i i
i i
χα
−
→ = νψs ψ1 6= 0,
−
→ = νψ1 ψ1 = χα
s
1
→
for some ν ∈ {±1, ±2}, which implies ψ i = 1H and so −
γ = (γ, ∅, . . . , ∅) for some
n
partition γ of n. As observed before, in this case H ≤ ker(χ) and the corresponding
character χ̄ must be a p-constant character of Sym(n).
We are left to consider the case when p = 2 divides |H| and χ̄ is a 2-constant
character of Sym(n): this happens only when n = 2, 3 and χ̄ is linear. However, G =
Cyc(2) ≀ Sym(2) is a 2-group and by Lemma 2.3, χ = 1G . Also G = Cyc(2) ≀ Sym(3)
is isomorphic to Cyc(2) × Sym(4) and, by Proposition 2.4, we get χ = 1G .
Corollary 3.4. Let G be a group of type Bn with n ≥ 2. Let χ be an irreducible
character of G of non-zero p-defect. Then, χ is p-constant for some prime p ≤ n if
and only if p > 2, Cyc(2)n ≤ ker(χ) and the corresponding character χ̄ of Sym(n)
is p-constant.
The groups Dn are subgroups of index 2 in Bn . Let α, β be partitions of some
0 ≤ t ≤ n such that |α| + |β| = n and let χ(α,β) be the corresponding irreducible
character of H = Cyc(2) ≀ Sym(n). If α 6= β, the characters χ(α,β) and χ(β,α)
restrict to the same irreducible character of G = Dn . The character χ(α,α) splits
(α,α)
(α,α)
on G as the sum of two irreducible characters χ+ , χ−
of G. Furthermore, a
class C(α,β) of H belongs to G only if β has even length and a class C(α,β) of H
splits into two classes of G if and only if β = ∅ and the parts of α are all even (see
[10]).
Proposition 3.5. Let G be a group of type Dn with n ≥ 4. Let χ be an irreducible
character of G of non-zero p-defect. Then, χ is p-constant for some prime p ≤ n
if and only if χ is the restriction to G of an irreducible p-constant character of Bn .
IRREDUCIBLE p-CONSTANT CHARACTERS OF FINITE REFLECTION GROUPS
7
Proof. Suppose that χ ∈ Υp (G). Our first goal is to show that χ is the restriction
to G of a character χ(α,∅) of H = Bn .
So, let χ be the restriction to G of some character χ(α,β) (or one of the two
(α,α)
characters χ± ). Let n = pk + t with 0 ≤ t < p. If pk and t are not both even,
we can prove that β = ∅, arguing as as done proving Theorem 3.3, since the classes
C((t,pk),∅) and C(∅,(t,pk)) belong to G.
Suppose now that β 6= ∅ and that both pk and t are even. If t ≥ 2, we consider
→
−
→
−
→
→
the classes C−
η1 , C−
η2 of G, where η1 = (∅, (t, pk)) and η2 = ((pk), (1, t − 1)). From
(α,β)
χ−
→
η1
6= 0 we get that α, β are both hooks of length, respectively, pk and t and
(α,β)
(α,β)
(α,β)
,a
= −χη−
6= 0 we get β = (t) or (1t ). In this case, we have χη−
from χ−
→
→
→
η2
2
1
contradiction.
Now, let t = 0 and n = pk ≥ 4 be even (whence k ≥ 2). Consider the classes
→
−
→
−
→ −
−
→
− , C→
− , where δ1 = (∅, (p, pk − p)) and δ2 = (∅, (n/2, n/2)). Observe that δ1 = δ2
C→
δ
δ
1
2
(α,β)
if, and only if, n = 2p. From χ→
−
δ1
6= 0 we get that α, β are both hooks of length,
(α,β)
= 0, unless n = 2p. So,
respectively, pk − p and p. However in this case, χ→
−
δ
→ 2
−
(α,β)
− , where δ3 = ((1, p − 1, p), ∅): χ→
suppose n = 2p and take the class C→
= 0,
−
δ
3
(α,β)
δ3
(α,β)
always
and χ→
unless β = (p) or (1p ). In this case, comparison between χ→
−
−
δ3
δ2
produces an absurd.
We conclude that χ is the irreducible restriction to G of χ(α,∅) for some α ⊢ n. As
we already remarked, the characters χ(α,∅) are obtained by inflation from Sym(n).
Hence, we are left to exclude that the character χ(α,∅) |G is p-constant, even if the
corresponding character χα of Sym(n) is not p-constant. In particular we have to
−−→ of p-singular elements of G, such that
consider suitable conjugacy classes C−
(α,∅)
α ⊢ n does not consist of even parts. However, this was already done in the
original proof of Proposition 3.2 in [9] and [5], where the conjugacy classes that
were analyzed never consist of even parts.
Finally, we consider the dihedral groups I2 (n) of order 2n. A direct check of their
character table, that can be found, for instance, in [4], gives the following result.
Proposition 3.6. Let G be a group of type I2 (n) with n ≥ 3 and let χ be a nonlinear irreducible character of G of non-zero p-defect. Then, χ is p-constant for
some prime p dividing 2n if and only if p = 3 = |G|3 and χ = ψn/3 , in the notation
of [4, page 182]. In particular, cχ = −1.
Proof of Theorem 1.2. Let G be a finite reflection group. Suppose that 1G 6= χ ∈
Υp (G), for some prime p dividing |G|. Since G can be written as direct product of
irreducible reflection groups G = H1 × . . . × Hr , by Proposition 2.4, p divides |Hi |
for a Q
unique value of i ∈ {1, . . . , r}. Furthermore, χ = ψ ⊗ 1K , where ψ ∈ Υp (Hi ),
K = j6=i Hj and cχ = cψ .
Now, by Propositions 3.1 and 3.2, cψ = ±1 if Hi is of exceptional type or
Hi = Sym(n) for n ≥ 3. If Hi is of type Bn , then ψ is obtained by inflation from a
p-constant character of Sym(n) (Corollary 3.4) and so, also in this case, cψ = ±1.
Clearly, cψ = ±1 even when Hi is of type Dn (Proposition 3.5). Finally, if Hi is a
dihedral group, by Proposition 3.6 we have cψ = −1.
8
MARCO ANTONIO PELLEGRINI
Proof of Theorem 1.3. As done proving Theorem 1.2, write G as direct product of
irreducible reflection groups G = H1 × . . . × Hr . Since p = 2 divides the order
of any Hi , by Proposition 2.4, r = 1. Now, by Propositions 2.4, 2.5 and 3.5 and
Corollary 3.4, G must be a dihedral group I2 (n) where n ≥ 1 (I2 (1) ∼
= Cyc(2)).
Now, by Proposition 3.6 we obtain that χ must be a linear character of G. Finally,
by Proposition 2.1, G has order 2n, where n is odd. The statement follows from
the fact that, when n is odd, I2 (n) has only two linear characters.
4. Open problems
Our final aim is to characterize the finite groups for which Υp (G) 6= {1G }. From
Propositions 2.1 and 2.4 and Lemma 2.2, we are reduced to consider non-linear
characters of indecomposable groups with trivial center.
The condition cχ = ±1 seems to be too weak for a possible classification. Namely,
by Theorem 1.2, the p-constant characters (of non-zero p-defect) of all finite reflection groups have this property. On the other hand, one may focus on p-constant
characters χ such that |cχ | > 1. Using [1, SmallGroups], the following conjecture
has been verified for all groups of size less than 1250:
Conjecture 4.1. Let G be a finite group which admits an irreducible p-constant
character χ such that |cχ | > 1. Then p 6= 2 and the Sylow p-subgroups of G are
homocyclic.
Notice that when cχ = ±1 we can observe a different behavior. For instance, the
irreducible character χ of Alt(6) of degree 9 is 2-constant with cχ = +1 and the
irreducible character of PSL3 (2) of degree 7 is 2-constant with cχ = −1. However,
in both cases the Sylow 2-subgroups are dihedral of order 8.
The smallest group that admits an irreducible p-constant character with |cχ | > 1
is obtained taking p = 3 in the following:
Lemma 4.2. For any odd prime p, there exists a finite group G admitting an
irreducible p-constant character χ with cχ = +2.
Proof. For any prime p, take a 2-dimensional space V defined over the field Fp . The
group Q8 = {±1, ±i, ±j, ±k} admits an irreducible representation Φ with space of
representation V , given by
a b
0 −1
Φ(i) =
, Φ(j) =
,
b −a
1 0
where a, b ∈ Fp are chosen such that a2 + b2 + 1 = 0. Define
1
0
G(p) =
: v ∈ V, g ∈ Q8 ∼
= V ⋊ Q8 .
v Φ(g)
The group Q8 acts on V as a group of fixed-point free automorphisms and so G(p)
is a Frobenius group. The elements of G(p) have order {1, 2, 4, p}. The irreducible
character χ of G(p) of degree 2, obtained by inflation from Q8 , is p-constant with
cχ = +2.
The smallest group that satisfies Conjecture 4.1 with a Sylow p-subgroup which
is not elementary abelian is G248,253 whose structure is (Cyc(9) × Cyc(9)) ⋊ Q8 ,
while the smallest one with a Sylow p-subgroup of rank > 2 is G248,730 whose
structure is (Cyc(3) × Cyc(3) × Cyc(3) × Cyc(3)) ⋊ Q8 (here Gn,k denotes the k-th
group of size n, according to [1]). In both cases, p = 3 and χ(1) = cχ = +2.
IRREDUCIBLE p-CONSTANT CHARACTERS OF FINITE REFLECTION GROUPS
9
The smallest group admitting an irreducible p-constant character with cχ < −1
is M10 (cχ = −2 and χ(1) = 16), while the smallest perfect group with |cχ | > 1 is
the following.
Example 4.3. Consider the group H = hh1 , h2 , h3 i ≤ SL2 (11), where
3 1
0 1
2 0
h1 =
, h2 =
, h3 =
.
1 −3
−1 0
0 6
The group H is isomorphic to SL2 (5) and acts regularly on V = F211 , hence the
group G = V ⋊ H is a Frobenius group with a single class of 11-singular elements.
We get that Υ11 (G) = Irr(G). Also, for χ(1) = 1, 2, 3, 4, 5, 6 we obtain cχ = χ(1)
and for χ(1) = 120 we get cχ = −1. Clearly the Sylow 11-subgroups of G are
elementary abelian of order 112 .
Looking at the perfect groups, we can state a stronger version of Conjecture 4.1.
Indeed, a direct check on the perfect groups of size < 15360, whose character tables
are contained in the package AtlasRep 1.5.0, seems to confirm the following.
Conjecture 4.4. Let G be a finite perfect group admitting an irreducible p-constant
character χ such that |cχ | > 1. Then p = 3, cχ = ±2 and the Sylow 3-subgroups of
G are isomorphic to Cyc(3) × Cyc(3).
Notice that the only simple sporadic group admitting irreducible p-constant characters with cχ = ±2 is the Mathieu group M22 , that satisfies Conjecture 4.4. The
other known simple groups satisfying Conjecture 4.4 are the groups PSL3 (q) when
gcd(q − 1, 9) = 3 and PSU3 (q) when gcd(q + 1, 9) = 3. Indeed, in these cases,
the Sylow 3-subgroups are isomorphic to Cyc(3) × Cyc(3) and using the character
tables described in [11] we may prove that following.
Proposition 4.5. Let G ∈ {PSL3 (q), PSU3 (q)} and let r be a prime dividing |G|
such that gcd(r, q) = 1. Let δ = +1 if G = PSL3 (q), δ = −1 if G = PSU3 (q). Set
d = gcd(q − δ, 3). Then 1G 6= χ ∈ Υr (G) if and only if one of the following holds:
(i) r divides (q 2 + δq + 1)/d and
(a) χ(1) = q 2 + δq: in this case cχ = −1;
(b) χ(1) = q 3 : in this case cχ = δ.
(ii) 2 6= r divides (q + δ) and
(a) χ(1) = q 3 : in this case cχ = −δ;
(b) d = 1, gcd(q + δ, 9) = r = 3 and χ(1) = q 3 − δ: in this case cχ = δ.
(iii) d = r = gcd(q − δ, 9) = 3 and
(a) χ(1) = q 2 + δq: in this case cχ = +2;
(b) χ(1) = q 3 : in this case cχ = δ;
2
(c) χ(1) = (q+δ)(q 3+δq+1) : in this case cχ = −δ.
References
[1] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.6.4 of 4 May
2013, (http://www.gap-system.org).
[2] J.E. Humphreys, Reflection groups and Coxeter groups, Cambridge Studies in Advanced
Mathematics, 29. Cambridge University Press, Cambridge, 1990.
[3] G. James and A. Kerber, The representation theory of the symmetric group, Encyclopedia
of Mathematics and its Applications, 16. Addison-Wesley Publishing Co., Reading, Mass.,
1981.
10
MARCO ANTONIO PELLEGRINI
[4] G. James and M. Liebeck, Representations and characters of groups, Cambridge Mathematical Textbooks. Cambridge University Press, Cambridge, 1993.
[5] C. Lassueur, G. Malle and E. Schulte, Simple endotrivial modules for quasi-simple groups, J.
reine angew. Math. 712 (2016), 141–174.
[6] F. Lübeck and G. Malle, A Murnaghan-Nakayama rule for values of unipotent characters in
classical groups, Represent. Theory 20 (2016), 139–161.
[7] A. Mendes, J. Remmel and J. Wagner, A λ-ring Frobenius characteristic for G ≀ Sn , Electron.
J. Combin. 11 (2004), no. 1, Research Paper 56, 33 pp.
[8] M.A. Pellegrini and A.E. Zalesski, On characters of Chevalley groups vanishing at the nonsemisimple elements, Internat. J. Algebra Comput. 26 (2016), 789–841.
[9] M.A. Pellegrini and A. Zalesski, Irreducible characters of finite simple groups constant at the
p-singular elements, Rend. Semin. Mat. Univ. Padova 136 (2016), 35–50.
[10] G. Pfeiffer, Character tables of Weyl groups in GAP, Bayreuth. Math. Schr. 47 (1994), 165–
222.
[11] W.A. Simpson, J.S. Frame, The character tables for SL(3, q), SU(3, q 2 ), PSL(3, q), PSU(3, q 2 ),
Canad. J. Math. 25 (1973), 486–494.
Dipartimento di Matematica e Fisica, Università Cattolica del Sacro Cuore, Via
Musei 41, 25121 Brescia, Italy
E-mail address: [email protected]
| 4math.GR
|
A SHM method for detecting damage with incomplete observations
based on VARX modelling and Granger causality
1
U. Ugalde1, J. Anduaga1, F. Martínez1 and A. Iturrospe2
1
2
Department of Sensors, IK4-Ikerlan, Mondragon, Spain ([email protected])
Department of Electronics and Computer Sciences, Mondragon Goi Eskola Politeknikoa, Mondragon, Spain
Abstract—A SHM method is proposed that minimises the required number of sensors for detecting
damage. The damage detection method consists of two steps. In an initial characterization step,
substructuring approach is applied to the healthy structure in order to isolate the substructures of
interest and later, each substructure is identified by a Vector Auto-Regressive with eXogenous inputs
(VARX) model measuring all DOFs. Then, pairwise conditional Granger causality analysis is carried
out with data measured from substructural DOFs to evaluate the information loss when measurements
from all DOFs are not available. This analysis allows selecting those accelerometers that can be
suppressed minimising the information loss. In the evaluation phase, vibration data from the reduced
set of sensors is compared to the estimated data obtained from the healthy substructure’s VARX model,
and as a result a damage indicator is computed. The proposed detection method is validated by finite
element simulations in a lattice structure model.
1 INTRODUCTION
Structural Health Monitoring (SHM) is a technological area which implements a damage detection and
characterization strategies for engineering structures [1]. SHM is regarded as a very important engineering field in order to
secure structural and operational safety; issuing early warnings on damage or deterioration, avoiding costly repairs or even
catastrophic collapses [2].
Most of the existing vibration based SHM methods can be classified into two different approaches: global approaches
and local approaches [3]. In the global approaches, the goal is to monitor the health of the entire structure. These global
methods have been tested and implemented in different types of structures during the last 30-40 years [4]. However, for
many large systems, global monitoring is not viable due to the lack of sensitivity of global features regarding local
damages, inaccuracies of developed models, high cost of the computational operations etc [5]. On the other hand, local
SHM methods are focused on evaluating the state of reduced parts within the entire structures by means of substructuring
methods. This approach aims to overcome global method’s limitations, dividing the whole structure into substructures and
analysing each one individually.
Several papers have been published proposing substructuring methods for large scale structures. Koh et al [6]
presented a “divide and conquer” strategy to monitor large structures based on the division of the whole structure into
isolated substructures. For each substructure, structural parameters are identified using the extended Kalman filter (EKF).
However, the EKF usually requires a prior analytical model, which is not always available [7]. Most recently, Xing et al
[8] proposed another damage detection method using Auto-Regressive Moving Average with eXogenous input (ARMAX)
1
models. The method is based on the extraction of natural frequencies from the ARMAX models and in their posterior
analysis. Nonetheless it is preferably applicable to small and simple structures.
The authors of this article proposed in [9] a damage localization method based on the substructuring approach and
Vector Auto-Regressive with eXogenous input (VARX) models. VARX models incorporate data from multiple DOFs and
the method allows locating and quantifying damages within larger substructures. It is necessary to measure all
substructural DOFs and consequently the amount of sensors that must be placed remain too high.
In general, if the number of sensors placed on a structure increases, the obtained information is more detailed.
However, the high cost of data acquisition systems and accessibility limitations constrain in many cases the use of large
number of sensors on a structure. Optimal sensor placement (OPS) methods aims to select a set of minimum number of
sensors from all possibilities, such that the data collected can provide adequate information for the identification of the
structural behaviour [10]. Different criteria can be used to evaluate the suitability of sensor locations, such as modal
assurance criterion (MAC), measured energy per mode, information entropy (IE) or mutual information (MI).
Krammer [11] proposed an optimal sensor placement method based on modal properties. The measured mode shape
vectors must be as linearly independent as possible in order to distinguish the measured or identified modes. MAC based
sensor optimization methods aim to select points from all candidates to keep the original properties of the structure as well
as it is possible.
Measured energy per mode criterion based OPS methods chooses those sensor positions with large amplitudes, which
is critical in harsh and noisy circumstances. Heo et al. [12] focused on a kinetic energy to determine the optimal sensor
placement on a long span bridge.
The two methods mentioned above are generally used when an analytical model (FEM) of the monitored structure is
available. Nevertheless, if only experimental measurements are available, methods like the ones that are based on mutual
information or information entropy can be used.
The mutual information (MI) of two random variables gives a measure of the mutual dependence between them and
quantifies the “amount of information” obtained about one variable through the other variable. The optimal sensor
locations can be obtained analysing which configuration gives the minimal mutual information values between sensors
[13].
In SHM, it is desirable to design a certain sensor placement configuration that assures the measurement of
representative data about the structural model parameters. The information entropy, as the measure of the uncertainty in
the system parameters, gives the amount of useful information in the measured data. Information entropy theory is used to
select the most informative sensors [14][15].
Granger causality (G-causality) is a statistical concept broadly applied in other scientific fields like neuroscience [16],
climatology [17] or econometrics [18] with the purpose of analyse the causality among different time series. As mutual
information and information entropy, does not require an analytical model and it is based on auto-regressive models.
In this paper, a SHM method is proposed for detecting damage with incomplete observations. Substructuring
approach and VARX modelling are used to isolate and identify the healthy substructure. Pairwise conditional Granger
causality analysis is carried out with data from all substructural DOFs to determinate which variables are the most proper
ones to be observed. Damage is detected using the healthy VARX model and the data measured from the selected nodes.
The rest of the paper is organised as follows. In section 2, the proposed method is presented. In section 3, the method
is evaluated by series of simulations and the obtained results are discussed. Finally the concluding remarks are exposed
in section 4.
2 THE PROPOSED METHOD
A SHM method is proposed for detecting damage with incomplete observations. The substructuring approach is
applied to isolate multi-DOF substructures from the healthy structure and each substructure is modelled by a
corresponding VARX model as in [9]. In the healthy condition all DOFs of the substructure are measured and pairwise
conditional Granger causality analysis is carried out with the data in order to determine which variables can be suppressed
assuring the minimum worsening as it is possible. During the evaluation phase, a reduced set of data is measured again and
by means of the healthy substructure’s VARX model the condition of the substructure is evaluated.
2
2.1. Substructure pairwise conditional Granger causality analysis
Granger causality analysis measures the causal influence between stochastic processes and determines also the
functional connectivity between them [19].
Traditionally, Granger causality analysis has been carried out between two stochastic processes. Assuming two jointly
distributed stochastic processes (X=X1,X2…Xn and Y=Y1,Y2...Yn ) it is said that Y does not cause X (according to Granger)
if X is only conditioned by its own past and therefore, X is independent of the past of Y. Otherwise, if the past of Y
contains information about the future of X, it is said that Y causes X according to Granger.
More recently Barnett and Seth [20] developed a Multivariate Granger Causality (MVGC) toolbox, where time and
frequency domain based multivariate G-causality approaches are available.
Within time domain multivariate G-causality approaches, pairwise conditional Granger causality is one of the most
popular ones. In pairwise conditional G-causality, it is supposed that the universe U is splits into three jointly distributed
processes (X, Y, Z):
X
U Y
Z
(1)
The analysis wishes to eliminate any joint effect of Z on the inference of Granger causality from Y to X. Full and
reduced VAR models are estimated from the multivariate processes and as is shown in equation 2 and 3, the difference
between them is that variable Y is not included in the reduced model:
p
p
p
k 1
k 1
k 1
(2)
X n Axx,k X nk Axy ,k Ynk Axz ,k Z nk x ,n
p
p
k 1
k 1
(3)
X n Axx ,k X nk Axz ,k Z n k x,n
A and A’ matrices correspond to the coefficients of the full and reduced VAR models, whereas
residuals of both models. Furthermore, p is the model order.
and are the
The causality from Y to X conditioned on Z (FY X|Z) is stated as:
| '
FY X |Z ln | xx
|
(4)
xx |
where xx and xx are the covariance matrices of the residuals of the full and reduced VAR models.
Figure 1 describes the proposed method for the selection of a certain substructural variable from a set of variables for
its posterior suppression. In this method a G-causality value from each substructural variable to the whole set of
substructural variables is analysed by a pairwise conditional G-causality approach. MVGC toolbox [20] is adapted to
carry out the analyses.
The first step is the measurement of the displacements from the substructure (U). Once U is measured, it is split into X,
Y and Z variables. X contains the displacement data that must be estimated, Y contains the variables that their causality
will be analysed and Z holds the rest of measurements. As it is shown in figure 1, different variables are assigned to X, Y
and Z in each iteration and the corresponding individual causality value (FY X|Z) is calculated. When all combinations
have been analysed, the individual causality values that are calculated setting the same variable in Y are added together in
order to obtain a general causality value (FG Y X|Z) for each variable.
3
The suppression of the variables with the lowest general causality values (FG Y X|Z) cause the minimum worsening in
the VAR model, so these variables are the most appropriate ones to be suppressed.
Figure 1. Selection of variables for their suppression –flow chart
2.2. Substructure damage detection with incomplete observations
A SHM method that aims to detect damages within substructures by means of incomplete measurements is described
in this section. The method evaluates the condition of the structure using the VARX model of the healthy substructure
and a set of variables that have been selected applying the approach described in the previous section.
We assume that the monitored substructure can be represented by the VARX model of equation 5. The VARX model
contains n endogenous variables (y) and m exogenous variables (x), whereas p and q are the endogenous and exogenous
model orders. Ai and Bi are endogenous and exogenous coefficient matrices.
We also assume that variable y1 has been selected to be suppressed, so it is not measured never again during the
evaluation phase. Thus, y2…yn and x1, x2… xm are the measured displacements from the substructure, whereas yˆ1 , yˆ2 ... yˆn
are the estimated displacement data.
yˆ1 (k )
yˆ1 (k 1)
yˆ1 (k 2)
yˆ1 ( k p)
yˆ (k )
y (k 1)
y (k 2)
2
A 2
A 2
... A y2 ( k p)
1
2
p
...
...
...
...
yˆ n (k )
yn (k 1)
yn (k 2)
yn ( k p )
x1 (k 1)
x1 (k 2)
x1 (k q )
x (k 1)
x (k 2)
2
B 2
... B x2 (k q )
B1
2
q
...
...
...
xm (k 1)
xm (k 2)
xm (k q )
4
(5)
As it is shown in equation 5, the displacement data corresponding to the endogenous variables at time instant k is
estimated combining the VARX model and the measured and estimated variables (if it corresponds to the suppressed
variable) at previous time instants. This process is carried out repeatedly from k equal to p until k equal to N+p in order
to estimate N samples of the endogenous variables yˆ1 , yˆ 2 ... yˆ n .
Once the total displacement data have been estimated, a damage indicator (DI) is computed as in equation 6. This
damage indicator is obtained as the total mean deviation between the measured endogenous variables and their
corresponding estimated values:
M
N
(6)
DI
( | yˆ
k 1
i 1
ki
y
ki
|)
M *N
N represents the amount of estimated samples for each endogenous variable and M represents the amount of measured
variables. Taking into account the assumptions that we have made during this section, the measured endogenous
variables are (y2…yn) and their corresponding estimated values are ( yˆ 2 ... yˆ n ), so M is equal to n-1.
3 NUMERICAL RESULTS
A linear and time invariant two-dimensional lattice structure is monitored in this section (see figure 2). The structure
consists of stainless steel bars that are connected together by rigid joints and we assume that the forces can only be
transmitted along the axial direction of the bars and the load can only be applied at the two ends of each bar. The structural
behaviour is described by a lumped parameter model, where we assume that all object are rigid bodies and all interactions
between the rigid bodies take place via springs.
As is shown in figure 2, a ten DOFs substructure is selected from the entire structure for its monitoring.
Figure 2. Isolated substructure in the structural model
The healthy structure is excited in the third mass (outside the substructure) by a Gaussian white noise and the
displacements are measured for each substructure DOF using a data sampling frequency of 1000 Hz during 2 seconds.
By means of the substructuring method explained in [9], the selected substructure is isolated from the general structure.
The isolated substructure is represented as a four exogenous and six endogenous variables VARX model. The exogenous
variables are the measured absolute displacements in z4x, z4y, z5x and z5y and the endogenous variables are the measured
absolute displacement in z6x, z6y, z7x, z7y, z8x and z8y. On the other hand, A1 and A2 are 6 x 6 endogenous coefficient matrices
and B1 is a 6 x 4 exogenous coefficient matrix. The healthy VARX model is estimated by the Multivariate Least-Square
estimator (MLS) [21].
5
z6 x
z
6y
z7 x
z7 y
z8 x
z8 y
(k )
z6 x
z
(k )
6y
z7 x
(k )
A1
(k )
z7 y
z8 x
(k )
(k )
z8 y
(k 1)
z6 x
z
(k 1)
6y
z7 x
(k 1)
A2
(k 1)
z7 y
z8 x
(k 1)
( k 1)
z8 y
( k 2)
(7)
( k 2)
z4 x (k 1)
z (k 1)
( k 2)
4y
B1
z5 x (k 1)
( k 2)
( k 2)
z5 y (k 1)
( k 2)
3.1. Substructure pairwise conditional Granger causality analysis
The pairwise conditional Granger causality analysis between all measured substructural DOFs (healthy state) is done in
this section in order to determine which are the most appropriate ones to be suppressed.
The bidirectional displacement data corresponding to each endogenous node is assigned to an individual data set. The
first data set contains displacement data from node 6 (z6x, z6y), the second data set contains the displacement data from
node 7 (z7x, z7y) and the last data set is formed by displacement data from node 8 (z8x, z8y).
Once the data sets are defined, an individual pairwise conditional Granger causality analysis is carried out between
them as is explained in section 2. As is shown in table 1, six combinations are analysed individually and their respective
individual G-causality value (FY X|Z) is calculated.
Table 1. Substructure’s individual pairwise conditional Granger causality analysis
FY X|Z
X
Y
Z
z7 x , z7 y
z6 x , z6 y
z8 x , z8 y
1.22e
z8 x , z8 y
z6 x , z6 y
z7 x , z7 y
1.26e
z6 x , z6 y
z7 x , z7 y
z8 x , z8 y
8.52e
z8 x , z8 y
z7 x , z7 y
z6 x , z6 y
1.06e
z6 x , z6 y
z8 x , z8 y
z7 x , z7 y
8.67e
z7 x , z7 y
z8 x , z8 y
z6 x , z6 y
1.09e
2
2
3
2
3
2
The individual G-causality values that are calculated assigning the same data set to Y are added together in order to
obtain a general causality value (FG Y X|Z) for each set of displacements. The result is shown in table 2.
Table 2. Substructure’s global pairwise conditional Granger causality analysis
FG Y X|Z
Y
z6 x , z6 y
2.480e
z7 x , z7 y
1.912e
z8 x , z8 y
1.957e
2
2
2
The set that contains displacement data of node 6 has the higher global G-causality value, whereas data sets
corresponding to nodes 7 and 8 have similar values between them, but significantly lower than the previous case.
Therefore, data from node 6 (z6x, z6y) is the least appropriate to be suppressed.
6
3.2. Substructure damage detection with incomplete observations
Different scenarios (healthy and damaged) are assessed by the developed SHM method. All considered damages are
stiffness losses of a specific spring (20%) within the structure. Furthermore, three different damage locations are
evaluated. In some of them, the damaged springs are within the substructure (k6,7 and k7,8) and in the others, they
correspond to external springs (k1,3).
As in the healthy state, the damaged structures are excited in the third mass (outside the substructure) by a Gaussian
white noise. The displacements are measured for all substructural DOFs using a data sampling frequency of 1000 Hz
during 2 seconds.
Granger causality analysis has determined that displacement data from nodes 7 or 8, even both, are appropriate to be
suppressed. Therefore, the condition of the simulated scenarios is evaluated by the developed damage detection method
suppressing firstly data from node 7, secondly data from node 8 and finally displacement data from nodes 7 and 8.
Although a particular case (suppression of data from nodes 7 and 8) is described below, the steps that must be followed
are equals in the other two cases. These steps are summarised in figure 1.
The displacement data of the endogenous variables ( zˆ6 x , zˆ6 y , zˆ7 x , zˆ7 y , zˆ8 x , zˆ8 y ) at time instant k is estimated combining the
healthy VARX model (A1, A2, B1) and the measured (z4x, z4y, z5x, z5y, z6x, z6y) and estimated ( zˆ7 x , zˆ7 y , zˆ8 x , zˆ8 y ) variables at
previous time instants. This process is carried out repeatedly from k equal to 2 until 2002 in order to estimate 2000
samples of the endogenous variables.
zˆ6 x
zˆ
6y
zˆ7 x
zˆ7 y
zˆ8 x
zˆ8 y
(k )
z6 x
z
(k )
6y
zˆ7 x
(k )
A1
(k )
zˆ7 y
zˆ8 x
(k )
(k )
zˆ8 y
(k 1)
z6 x
z
(k 1)
6y
zˆ7 x
(k 1)
A2
(k 1)
zˆ7 y
zˆ8 x
(k 1)
( k 1)
zˆ8 y
(k 2)
(8)
(k 2)
z4 x (k 1)
z (k 1)
(k 2)
4y
B1
z5 x (k 1)
(k 2)
(k 2)
z5 y (k 1)
(k 2)
Once the total displacement data ( zˆ6 x , zˆ6 y , zˆ7 x , zˆ7 y , zˆ8 x , zˆ8 y ) have been estimated, a damage indicator (DI) is computed as
in equation 6. This damage indicator is obtained as the total mean deviation between the measured endogenous variables
(z6x, z6y) and their corresponding estimated values ( zˆ6 x , zˆ6 y ).
Figure 3 shows the obtained results for three kinds of scenarios; healthy (a), damage out of the substructure (b) and
damage within substructure (c). Blue and green bars represent the DI values when data from node 7 or node 8 is
suppressed respectively, whereas the results obtained suppressing the data from both nodes is depicted by brown bars.
(a)
(b)
Figure 3. Calculated DI values for the simulated scenarios
7
(c)
Regarding to the results, when the whole structure is healthy or when the damage is introduced out of the substructure
(reducing k1,3), the calculated DI values are always almost zero.
On the other hand, if the damages are introduced within the substructure (k6,7 and k7,8) the DI values are much higher
than in the previous cases. Besides, the mentioned conclusions are valid even thought data from the two nodes is
suppressed.
Figure 3 shows that the DI values are practically the same if k7,8 is reduced and z7 and z8 are suppressed individually.
However, if k6,7 is modified, the corresponding DI values are different. This fact is related to the position of the damaged
element, because k7,8 correspond to a spring situated between node 7 and node 8, whereas k6,7 has more influence in node 7
than in node 8.
4 CONCLUSIONS
This paper proposes a SHM method to detect damages within substructures using incomplete measurements. In the
characterization step, which corresponds to the healthy state, the displacements from all substructural nodes are
measured and the corresponding VARX model is estimated. Furthermore, pairwise conditional Granger causality
analysis is carried out with the measured data in order to select which displacement data is the most appropriate to be
suppressed. In the evaluation phase, the healthy VARX model and the displacement data measured from the selected
nodes are used to determine the condition of the substructure.
A linear and time invariant model of a two-dimensional lattice structure is simulated to validate the proposed method.
The results show that the method allows detecting damages within substructures using a unique accelerometer.
The proposed method is also suited for three dimensional lattice structures, where the number of element’s
connections increases. Our research group is already applying this method in a real structure and the results will be
published soon.
5 ACKNOWLEDGEMENT
This study is partially funded by the Ministry of Economy and Competitiveness of the Spanish Government
(WINTURCON project, DPI2014-58427-C2-2-R). Any opinions, findings and conclusions expressed in this article are
those of the authors and do not necessarily reflect the views of the funding agency.
REFERENCES
[1]J. Ko and Y. Ni, "Technology developments in structural health monitoring of large-scale bridges," Engineering
structures, vol. 27, pp. 1715-1725, 2005.
[2]A. Mita, "Structural dynamics for health monitoring," Sankeisha Co., Ltd, Nagoya, vol. 114, 2003.
[3]L. Jankowski, Lukasz and I. P. P. o. Techniki, Dynamic load identification for structural health monitoring: Institute
of Fundamental Technological Research Polish Academy of Sciences, 2013.
[4]S. W. Doebling, C. R. Farrar, M. B. Prime, and others, "A summary review of vibration-based damage identification
methods," Shock and vibration digest, vol. 30, pp. 91-105, 1998.
[5]J. Hou, L. Jankowski, Lukasz, and J. Ou, "Structural damage identification by adding virtual masses," Structural and
Multidisciplinary Optimization, vol. 48, pp. 59-72, 2013.
[6]C. G. Koh, L. M. See, and T. Balendra, "Estimation of structural parameters in time domain: a substructure approach,"
Earthquake Engineering \& Structural Dynamics, vol. 20, pp. 787-801, 1991.
[7]A. W. Oreta and T.a. Tanabe, "Element identification of member properties of framed structures," Journal of
Structural Engineering, vol. 120, pp. 1961-1976, 1994.
[8]Z. Xing and A. Mita, "A substructure approach to local damage detection of shear structure," Structural Control and
Health Monitoring, vol. 19, pp. 309-318, 2012.
[9]U. Ugalde, J. Anduaga, Martínez, F, and A. Iturrospe, "Novel SHM method to locate damages in substructures based
on VARX models," in Journal of Physics: Conference Series, 2015, p. 012013.
8
[10]T.-H. Yi and H.-N. Li, "Methodology developments in sensor placement for health monitoring of civil
infrastructures," International Journal of Distributed Sensor Networks, vol. 2012, 2012.
[11]D. C. Kammer, "Sensor placement for on-orbit modal identification and correlation of large space structures,"
Journal of Guidance, Control, and Dynamics, vol. 14, pp. 251-259, 1991.
[12]G. Heo, M. Wang, and D. Satpathi, "Optimal transducer placement for health monitoring of long span bridge," Soil
Dynamics and Earthquake Engineering, vol. 16, pp. 495-502, 1997.
[13] R. J. Barthorpe and K. Worden, " Sensor Placement Optimization for Structural Health Monitoring, Encyclopedia of
Structural Health Monitoring, John Wiley & Sons, New York, NY, USA, 2009.
[14]C. Papadimitriou, J. L. Beck, and S.-K. Au, "Entropy-based optimal sensor location for structural model updating,"
Journal of Vibration and Control, vol. 6, pp. 781-800, 2000.
[15]K.-V. Yuen, L. S. Katafygiotis, C. Papadimitriou, and N. C. Mickleborough, "Optimal sensor placement
methodology for identification with unmeasured excitation," Journal of dynamic systems, measurement, and
control, vol. 123, pp. 677-686, 2001.
[16]A. K. Seth, A. B. Barrett, and L. Barnett, "Granger Causality Analysis in Neuroscience and Neuroimaging," The
Journal of Neuroscience, vol. 35, pp. 3293-3297, 2015.
[17]T. J. Mosedale, D. B. Stephenson, M. Collins, and T. C. Mills, "Granger causality of coupled climate processes:
Ocean feedback on the North Atlantic Oscillation," Journal of Climate, vol. 19, pp. 1182-1194, 2006.
[18]S. N. Sulku, "Causality Relation between Real Exchange Series and Emerging Stock Markets: Panel Common Factor
Analysis," Journal of Applied Economics \& Business Research, vol. 1, 2011.
[19]L. Barnett and A. K. Seth, "Granger causality for state-space models," Physical Review E, vol. 91, p. 040101, 2015.
[20]L. Barnett and A. K. Seth, "The MVGC multivariate Granger causality toolbox: a new approach to Granger-causal
inference," Journal of neuroscience methods, vol. 223, pp. 50-68, 2014.
[21] H. Lutkepohl, "New introduction to multiple time series analysis: Springer, 2007".
9
| 3cs.SY
|
Intelligent Information Systems
Proceedings of a Workshop held in
Augustów, Poland, 7-11 June, 1993
arXiv:1705.08440v1 [cs.AI] 23 May 2017
pages 210- 238
Knowledge Acquisition, Representation & Manipulation in
Decision Support Systems
M.Michalewicz, S.T.Wierzchoń, M.A. Klopotek
Institute of Computer Science
Polish Academy of Sciences
01-237 Warsaw, ul. Ordona 21, Poland
e-mail: [email protected]
Abstract. In this paper we present a methodology and discuss some implementation issues for a project on statistical/expert approach to data analysis and knowledge acquisition. We discuss some general assumptions underlying the project.
Further, the requirements for a user-friendly computer assistant are specified along
with the nature of tools aiding the researcher. Next we show some aspects of belief
network approach and Dempster-Shafer (DST) methodology introduced in practice to system SEAD. Specifically we present the application of DS methodology
to belief revision problem. Further a concept of an interface to probabilistic and
DS belief networks enabling a user to understand the communication with a belief
network based reasoning system is presented
1
Introduction
A statistical database is a collection of aggregated information at a given time. The
purpose of such database is to derive or infer facts; statistics is helpful in extracting
patterns from quantative (and qualitative) data [6]. Data are analysed to extract information and the result of such analysis must be validated to increase our knowledge about
modeled domain.
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
211
Several statistical packages (e.g. MINITAB SAS) are used for such statistical analysis
of data. However, they do not inform the user about underlying assumptions, seldom
warn against obvious misuse, do not provide guidance in the process of analysis, and
hardly assist in the interpretation of the results [8]. Clearly, there is a need for statistical
expert systems [11], [14], [15], [16], which would include the statistical expertise.
In [17] the authors provide a general characteristic of statistical expert systems. They
wrote:
“As statistical expert system is a computer program which can act in the role
of expert statistical consultant. That is, it can give expert advice on how to design
a study, what data should be collected to answer the research question, and how
to analyse the data. Moreover, since the fundamental tool of statistical analysis
is the computer, on which the expert system will be run, it is normally envisaged
that the expert system and data analysis routines will be an integrated whole.
thus, not only does the system advise on the analysis, but it also actually carries
it out, discussing the results and the direction of further analyses with the user.”
In this paper we deal with some aspects of creative data analysis. In particular we
are interested in applications of AI techniques for development of a software environment
which would allow to process both data and knowledge; such software should provide an
easy interface for formulating hypotheses and their statistical verification.
The paper is organized as follows. The next section discusses the architecture of a
creative data analysis system. Section 3 presents knowledge acquisition strategies. Section 4 introduces some aspects of new version of SEAD System, based on belief network
approach and DST. Section 5 is devoted to DS methodology: after describing the nature
of evidential reasoning we present mathematics of evidential resoning and next details of
the message propagation algorithm is introduced and its application to belief revision is
presented. In Section 6 a user interface to belief network is described.
2
Data Analysis System
Information systems are designed for storing and manipulating data (from some domain) and for drawing conclusions when a problem from that domain is formulated.
Clearly, this is of value insofar as the data are meaningful and the domain formalization
is sufficiently adequate.
There are at least two main categories of such systems, termed tentatively standard
decision aid tools (like statistical systems, pattern recognition programs, etc. and knowledge based systems. Although proven to be useful, systems of both categories suffer from
some drawbacks limiting their applicability.
212
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
The main disadvantage of systems from the first category is that a user is forced to (1)
formulate his (her) problem in a formal language, (2) check the input data validity, and
(3) translate solutions produced by the system to the language appropriate for a given
domain. In fact, to solve real-life problems the user must cooperate with professional task
analyzer.
When developing knowledge based systems an important obstacle concerning knowledge acquisition emerges. Knowledge bases generated by experts in different areas remain
incomplete, subjective and even contradictory. Additional difficulties arise in cases where
there is a need for processing data and knowledge simultanously.
In both categories of systems the process of hypotheses generation is left to the user.
The above approaches, considered separately, are especially inefficient when formalizing poorly recognized domains or domains with large amount of vagueness. In case of
medical diagnosis, for instance, although the characteristics of certain diseases are fixed,
the formal relationships among symptoms and diseases are hard to determine and serious
difficulties in formulating scientific hypotheses may occur. To formalize such domains one
must combine data, partial knowledge and tentative hypotheses.
In summary, an effective information system must provide the following features: data
storing, knowledge acquisition and manipulation, automated generation of hypotheses,
performing standard statistical analyses, including tests of research hypotheses, and easy
communication between the knowledge base and database.
Such experimental information system (called SEAD for Statistical and Expert system
for Analysis of Data) was recently implemented in the Institute of Computer Science,
Polish Academy of Sciences [26].
The system architecture reflects a typical understanding of the notion of science according which science is viewed as a progressive development of theoretical generalizations
of observations and experiments leading to new observations and new experiments. So,
scientific research means permanent moving within the cycle “observation – generalization – theory – application – observation”.
Various tools have been developed to support a researcher in different parts of his cycle, including tools supporting storing observations and experiments (data base systems)
and storing the acquired theoretical generalizations (knowledge base systems). Also the
tools supporting the knowledge acquisition are included, such as the ability to generalize
over a data base (statistical data analysis systems and numerous “learning from examples” systems), tools allowing verification against new incoming experiments results in
the data base (statistical analysis systems), or against new single cases (expert systems).
These tools are placed in specialized submodules, which are the main components of the
system: data base system, knowledge base system, expert system, knowledge acquisition
and verification tools.
The main feature of this system is implementation of all four components into one
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
213
organism, whose heart is DB pumping and absorbing data. Such a solution (based on
unified data formats and unified access methods) results in a research tool of quite new
and unique quality. The user can also employ each system’s component as an independent
unit or (s)he can add his (her) own modules.
3
Knowledge Acquisition Strategies
The completion of knowledge base by human experts is a bottleneck in the process
developing an expert systems. Usually, experts engaged in the process are excessively
cautious (especially when dealing with poorly recognized domains), they use different
semantics (often with vague and ambiguous notions) that results in a extremely long
completion period without any warrant that the set of rules is complete and that it
doesn’t contain contradictory assertions.
Another question concerns rules validity. According to methodological considerations,
to refine a rule we search for so-called idealizing law (being the heart of the rule) first,
and next we release systematically simplifying assumptions. Of course, such a process
must be stopped at a reasonable time, and since some assumptions still remain hidden,
the rule in its symbolic version is valid to some extent only. Hence we may address two
questions:
– were the most important assumptions embodied into the rule?, and
– how to estimate the degree of validity of such rule?
To answer these questions and to optimize the knowledge acquisition process a number
of tools have been developed in the SEAD system. These tools can be divided into two
categories described in sections 3.1 and 3.2.
3.1
Passive user
If certain aspect of the problem under the considerations is either ill-recognized or
the knowledge about it is too general, some data preprocessing might be necessary. the
aim of such preprocessing is to generate tentative rules directly from an existing data
base. At this stage the user plays a passive role contrary to the further stages realized
with his active collaboration. Here the user has a few possibilities:
– Discretization of a continuous variable under a criterion of diagnosis accuracy maximization.
– Some optimization but for a set of discrete variables.
– Identification of irrelevant variables.
– Generation of elementary hypotheses for specific values of variables.
– Generation of rules.
214
3.2
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
Active user
By an active knowledge acquisition we mean the simultaneous utilization of the DB
and the KB, and the statistical verification of the partial knowledge. The following tools
are helpful in these tasks.
– Tools for rules identification. These the search for rules with given condition or conclusion, the identification of rules contexually analogical to a given rule, etc. Any
hypothetical rule can be therefore compared by the user with those existing in the
knowledge base.
– Search for data in the DB relevant to a given hypothesis; particularly search for data
contradicting given rule. these operations are realized through queries to the DB. A
specialized interface allow simultaneous access to the KB and the DB.
– Rules formation from incomplete schemata (called suppositions). This is the active
strategy attempting to find a rule (or possibly a set of rules) satisfying constraints
imposed by the user. The constraints concern values of some attributes as well as
degrees of truth which should be attached to the rules.
– KB verification with respect to the contents of the DB. If the DB is sufficiently large
we are interested in how accurate the rules are, i.e. how strong is their expressive
power (measured in terms of misclassification error). To answer these questions we
employ the Statistical Module, which assists the user in his research (search for the
strongest relationships in the set of variables for a given group of objects, marginal
analysis for each variable, graphics for pairs of variables, et.).
3.3
Problems of Automatic Rule Generation
In very large statistical databases the generation of hypotheses (classification rules)
is done on the basis of small samples of data. However, all known methods for automatic generation of classification rules aim at the precise partition of the attributevalued space.They are create “long” conjunctions of the pairs (attribute, value), which
precisely classify all data sets [28]. Such approaches provide reasonable results for (large)
databases, where the classification problems are well defined. However, for databases
where classification is done on the basis of small samples of data (typical case for “soft
sciences”), this is not the case. In such databases there is a need for preprocessing data,
i.e. for generating a first iteration of hypotheses which are true with a high probability.
This first iteration of hypotheses should yield simple formulae expressed in a language
understandable for a user. The precise, “long” classification rules (hypotheses) would be
useless: they are precise only for a small sample of data and don’t provide a guidance
for the next iterations of hypotheses. Because of this, the development of new classification methods for a single attribute and small sets(i.e. 2 or 3) of attributes are of
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
215
particular interest. Summarizing, notice, that there are several problems connected with
automatic rule generation in small databases (these include also large databases, where
a data sampling was performed):
– the set of data is not representative for drawing conclusions strictly dividing the
attribute/value space,
– the generated rules are hardly interpretable, if at all possible; rules are detailed and
concerned with only several cases,
– a new sample (or a new set of data) can easily provide a contradiction,
– the set of rules usually is too large; this might be appropriate for an expert system
or diagnostic automata, but not for data analysis,
– in many methods the domains of attributes must be discrete; the process of discretization is usually left to the expert (user).
Taking into account the above drawbacks, we decided to construct a rule generating
algorithm which would start from a single attribute (“bottom up” approach). We present
the main idea of the method for creating the reasoning rules. This is based on (1) an
optimization process defined on the set of discrete values of a given attribute and (2) a
process of dividing a continous domain of an attribute into subsets; such optimal divisions
are converted into the set of classification rules [25]
4
SEAD.2 – the new version
Some experiences of exploitation of SEAD system caused several significant modifications, which were entered into the next version of SEAD:
4.1
Data Base
Data base, built for SEAD 1 by our group, compatible to dBase family, appeared to
be ineffective. Especially much difficulties was noticed when cooperation of data base
with expert and statistical parts (written in C language) was needed. We decided to
accommodate CODEBASE, which includes both libraries of procedures of access to data
bases like dBase IV, Fox Base or CLIPPER written in C and libraries for user interfaces
near to dBase IV or CLIPPER for MS DOS and MS WINDOWS.
4.2
Knowledge Base
Knowledge base now is the bayesian network with uncertainty propagation based on
Dempster-Schafer model, instead of rule-based knowledge base with uncertainty propagation near to MYCIN [37] used in former version of SEAD. This changes was caused by
following drawbacks of rule approaches:
216
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
– independence of variables is assumed, several appearances of the same variable in
chain causes some computation complications (each appearance of concrete variable
is treated as appearance of separate variable)
– independence of rules is assumed; data base appears as a rubbish where rules are
entered de facto without any control process.
– Shortliffe’s approach assumed the use certainty factor CF [37], nature of which was
defined not clearly enough. Even if rule’s CF was interpreted in terms of probability,
combination of rules lead very fast to loss of interpretability.
Meanwhile the belief networks has the following advantages:
– there exists well defined relationship between variables (conditional or not)
– these relationships are to be tested very simply, what can be used for new rules
generation for example
– one cannot add/remove rules without retribution, because it can break relationships
between variables
Approach based on belief networks and Dempster-Schafer methodology generates
many theoretical and practical problems. Some of them are showed below in section 5
4.3
Rule preceiving of bayesian knowledge base
It is easy to notice, that rule approach has several significant advantages. The most
important of them are following: simplicity, rule interpretability and naturalness and
comparatively easiness of introducing some knowledge acquisition mechanisms.
That was the reason of introducing to new version of SEAD systems some special
methods, which make possible for the user to preceive bayesian knowledge base as a set
of rules. These methods are described below in section 6.
5
5.1
DS Approach to Knowledge Representation and Manipulation
The Nature of Evidential Reasoning
The term evidential reasoning coined by Lowrance and Garvey in [23] covers a set of
techniques designed for processing and reasoning from row data (evidence) which are
currently accessible to an agent. It is based upon the Dempster-Shafer theory (DST)
of belief functions. This theory was founded by G. Shafer in his monograph [32] and
was highly influenced by the works of A. P. Dempster [9, 10]. DST is based on two
principal ideas: “the idea of obtaining degrees of belief for one question from subjective
probabilities for a related question and Dempster’s rule for combining such degrees of
belief when they are based on independent items of evidence” [34] p. 474.
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
217
To give a concise rationale for this theory (see [29]) for deeper discussion) we should
note that the information required to understand the current state of the world, or to
solve a real problem, comes from multiple sources, like real-time sensor data, accumulated domain knowledge and current contextual information. Information coming from
sensors, called evidence, provides a support for certain conclusions. The very nature of
evidence can be summarized as follows:(1) it is uncertain, i.e it allows for multiple possible explanations, (2) it is incomplete as the sensors rarely has a full view of situation,
(3) it may be completely or partial incorrect .
Because of its nature, evidence is not readily represented either by logical or by
classical probabilistic formalism. The ad hoc proposed methods for handling uncertain
information (like e.g. CF factors) cause difficulties in interpretation and in extending the
capabilities of such methodologies. DST, in contrast, provides a natural representation
for evidential information, a formal basis for drawing conclusions from evidence, and a
representation for belief.
In evidential reasoning a knowledge source is allowed to express probabilistic opinions
about the (partial) truth or falsity of statements composed of subsets of propositions
from a space of distinct, exhaustive possibilities (called the frame of discernment ). The
beliefs can be assigned to the individual (elementary) propositions in the space or to
disjunctions of these propositions (i.e. to compound propositions) or both. When belief
is assigned to a disjunction, a knowledge source is explicitly stating that it does not has
enough information to distribute this belief more precisely. Hence a knowledge source
can distribute its belief to statements whose granularity is appropriate to its state of
knowledge. Further, the statements to which belief is assigned are not necessarily distinct
from one another. The distribution of beliefs over a frame of discernment is called a body
of evidence.
The easiest way to formally represent a body of evidence is to define a mass function
m being a probability distribution over a set ofPsubsets of a frame of discernment Θ (as
such it must satisfy the obvious condition that {m(A) | A ⊆ Θ} = 1). Now, if the mass
function is positive on singletons of Θ only then our model reduces to standard probability
distribution. On the other extreme, a mass function can take the form m(A) = 1 for some
A ⊆ Θ. When A = Θ then such a function represents total ignorance about current state
of the world. When A ⊂ Θ, then the mass function describes our certainty that the truth
lies in A; when Θ is a Cartesian product of frames describing related questions then
m(A) = 1 defines a logical belief function. Thus, Boolean logic is another special case of
belief functions.
DST is therefore a generalization of both Boolean logic and probability calculus. This
generalization lies deep in the representation of belief functions and has effects on many
aspects of the theory. The heart of this theory is Dempster’s Rule of Combination, or
DRC for short, being a formal tool for fusing different bodies of evidence. The result is a
218
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
new body of evidence representing the consensus among the original bodies of evidence.
In special cases this rule behaves like Bayes’ rule and in other cases like set intersection.
Thus belief function provide a framework in which propositional logic and probabilistic
statements can be freely mixed and processed under common formalism.
Belief functions, however, provide more general models. Particularly, random sets are
a generalization of the fixed sets of logical models and upper (plausibility functions) and
lower (belief functions) bound intervals are a generalization of point probabilities.
Current automated reasoning systems are most effective when domain knowledge can
be modeled as a set of loosely interconnected concepts (propositions) what justifies an
incremental approach to updating beliefs. Applying DST formalism, independent opinions are expressed by independent bodies of evidence and dependent opinions can either
be expressed by a single body of evidence or by a network describing the interrelationships among several bodies of evidence (so called belief networks). Updating the belief
(by using Dempster’s Rule of combination) in one proposition affects the entire body of
evidence. This allows to involve whole the knowledge already accumulated in the system
(contrary to rule-based systems where only fragmentary knowledge is used during the
process).
In this paper we assume some familiarity with the DST, although the appropriate
definitions are included in the next section. In the sequel we will focus on the mathematics
of evidential reasoning, mainly on so-called belief revision.
5.2
Mathematics of Evidential Reasoning
In this section we briefly review a mathematical formalization of the evidential reasoning.
It consists of two main parts. In static part we simply represent our knowledge about
a problem in terms of belief functions and dynamic part allows to express knowledge
evolution.
Basic notions
X = {X1 , . . . , Xn } is a set of variables and Θi = Dom(Xi ) is a discrete set of possible
values of i-th variable. (The case of continuous variable is considered by [33] or [44].)
By A, B, C,. . . we shall denote subsets of the set X .Θ will denote the
Q configurations
space or frame of discernment, Θ = ×{Θ | i = 1 . . . n}, and Θ.A = {Θ | Xi ∈ A}
will denote the marginal frame. If x is a configuration from Θ then x.A stands for the
projection of x onto Θ.A; x.A is obtained by dropping from x the components being
members of Θi for all Xi ∈ X \ A.
Knowledge representation
Assume (Ω, B, P ) is a probability space, Ξ is a discrete space and X is a mapping which
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
219
assigns subsets of Ξ to points of Ω, i.e. X(ω) ⊆ Ξ for each ω ∈ Ω. Treating Ξ as the set
of elementary hypotheses we identify each subset of Ξ as a compound hypothesis. If Ω is
viewed as a set of observations then X(ω) = A can be interpreted as a rule “If observation
= ω then hypothesis = A”. According to classical approach the mapping X induces a
measure µ defined on POW - the power set of the set of all compound hypotheses:
µ(A) = P ({ω ∈ Ω | X(ω) ∈ A}),
A ∈ POW
(1)
Note that when X is a function, i.e. X(ω) ∈ Ξ for all ω ∈ Ω then POW can be reduced
to the power set of Ξ and (1) reduces to the definition of the random variable. Under
the general setting X constitutes so-called random set [47].
Of special importance are the next set functions derived from (1) and defined for all
A ⊆ Ξ, A 6= ∅:
m(A) = P ({ω ∈ Ω | X(ω) = A}),
(2)
Bel(A) = P ({ω ∈ Ω | X(ω) ⊆ A, X(ω) 6= ∅})
(3)
P l(A) = P ({ω ∈ Ω | (X(ω) ∩ A) 6= ∅}),
(4)
Assuming that all observations are relevant, i.e. X(ω) 6= ∅ for all ω ∈ Ω, we easily
recover the main properties of the basic probability assignment, m, i.e. m(∅) = 0, and
P
{m(A) | A ⊆ Ξ} = 1. In general case, when m(∅) > 0, the mass function should
be modified according to the assignment m(A) := m(A)/(1 − m(∅) and m(∅) = 0. The
m-function plays the role analogical to the probability distribution P (X = x), x ∈ ΘX ;
particularly when m(A) is positive on the singletons only then m is just a probability
distribution. The belief and plausibility functions can be computed from m by means of
the identities
X
X
Bel(A) =
{m(B) | B ⊆ A}, P l(A) =
{m(B) | B ∩ A 6= ∅}
(5)
Similarly plausibility function can be deduced from a belief function by means of the
equation P l(A) = 1 − Bel(Ξ − A), A ∈ Ξ.
Shafer [32] proposes further formulas enabling to express one set function in terms of
other set function and fast algorithms for doing this task are reported in [18].
In the sequel we shall use multivariate belief functions, i.e. belief function defined on
the family of subsets of the space Θ introduced at the beginning of this section.
Special class of multivariate belief functions form so called logical belief functions
satisfying the condition m(A) = 1 for some A ⊂ Θ. (When A = Θ then the mass
function represents vacuous belief function; it models total ignorance about the possible
location of X inΘ.) To be more illustrative we present four elementary situations:
220
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
(a) the logical expression “X1 = x1 and X2 = x2 . . . and Xn = xn ” is represented by
the element θ =< x1 , x2 , . . . , x >∈ Θ,
(b) the logical expression
Sn “X1 = x1 or X2 = x2 . . . or Xn = xn ”is represented by the
subset of the form i=1 ({xi } × Θ.(X − {Xi })),
(c) the logical expression “not (X1 = x1 )” is represented by the subset of the form
(Θ1 − {x1 }) × Θ.(X − {X1 }).
(d) the logical expression “if ((X1 = x1 ) or . . . or (Xn−1 = xn−1 ) then (Xn = xn )” is
represented by the subset of the form {x1 , x2 , . . . , xn } ∪ (Θ1 − {x1 }) × (Θ2 − {x2 }) ×
. . . × (Θn−1 − {xn−1 }) × Θn .
When the logical assertions cannot be stated with full certainty we convert them to so
called simple support function by setting m(A) = α and m(Θ) = 1 − α, where α ∈ (0, 1)
and A is a subset of Θ representing appropriate assertion.
Knowledge evolution
Up to this moment we have assumed that all the belief functions are defined on the
common space Θ. Suppose we have two belief functions defined on the frames Θ.A
and Θ.B, respectively. How to represent beliefs expressed by means of the attributes
(variables) from the set A in terms of the attributes from the set B? We answer these
questions now.
Minimal extension. Suppose that to describe a domain we used attributes from the
set A ⊂ X and as a result we obtained belief function BelA defined over the space Θ.A.
After a time we decided to use larger set of attributes B ⊃ A. To translate the function
BelA into the belief function BelB defined over the
Qspace Θ.B observe that using the
attributes from the set A we are able to create m = i∈A |Θi| (where |Θi| stands for the
cardinality of the set Θi ) elementary propositions q of the form “Xi1 = xi1 and Xi2 = xi2
and . . . and Xir = xir ” where Xij ∈ A, xij ∈ Θij and |A|= r. In other words the θ’s
are the only atomic propositions of the language LA . Since the values of the attributes
form the set B − A are unknown to a subject using the language LA then each atomic
proposition θ ∈ Θ.A translates into the compound proposition θ × Θ.(B − A) ⊂ Θ.B.
Similarly each A ⊆ Θ.A translates into the subset A × Θ.(B − A) of Θ.B. Thus the mass
function mA characterizing the belief function BelA translates into the mass function
mB of the form
mB (B) =
mA (a) if B = A × Θ.(B − A), A ⊆ θ.A
0
otherwise
(6)
In the sequel we shall denote the function mB derived from mA by means of the above
procedure as m↑AB .
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
221
Marginalization. Assume now that we reduce the set B of attributes to its subset
A. Obviously each atomic proposition θ ∈ Θ.B points to a proposition ξ ∈ Θ.A if its
projection onto Θ.A (defined as “dropping coordinates form B − A”) equals to ξ. Thus
the mass function mB translates to a mass mA via the equation
X
mA (A) =
{mB (B) | B ⊆ Θ.B, proj(B) = A}
(7)
In the sequel we shall denote the function mA derived from mB by means of the above
procedure as m↓A .
B
Fusing evidence. When m1 and m2 are two separate and independent bodies of evidence defined on a common frame Θ, to arrive to the final conclusions we combine them
by means of Dempster’s Rule of Combination (DRC for short). That is we crate new
mass function
m(A) = (m1P⊕ m2 )(A)
= κ · {m1 (A ∪ B) · m1 (A ∪ C) | A, B ⊆ (Θ − A), A ∩ B = ∅}
(8)
where κ is a normalizing constants defined as κ−1 = 1 − m(∅).
Note that when m1 and m2 are categorical mass functions, i.e. m1 (A) = 1 and
m1 (B) = 1 for some A, B ⊆ Θ then the rule produces categorical mass function with
m(A ∩ B) = 1. Hence DRC generalizes classical sets intersection.
To illustrate the notions introduced in this section assume that p and q are two
propositional variables, and their frames Ξ and Λ are such that Ξ = Λ = {t, f } where
t and f stand respectively for “true” and “false”. The logical implication can be described by the mass function mp→q on Θ = Ξ × Λ with one focal (i.e. positive) element
mp→q ({{f, f }, {f, t}, {t, t}}) = 1. Similarly the fact “p is true” is represented by the categorical mass function mp on the frame Ξ, that is mp ({t}) = 1. To combine these two
functions we must extend mp to the bpa on Θ: m↑Θ
p ({t, f }, {t, t}}) = 1. Now, combining
mp→q with m↑Θ
we
obtain
the
bpa
m
of
the
form
m({t, t}) = 1. Projecting this m onto
p
the space Λ we state that m↓Λ ({t}) = 1 what means that “q is true” what agrees with
the standard logical inference pattern Modus Ponens. Note that if we combine mp→q
with m-function representing statement “p is false” then the resulting function (after
projection on the frame Λ) has the form m({{f }, {t}}), i.e. nothing can be said about
the truth of the proposition q (what again agrees with classical logic). Similarly combining mp→q with the m function stating that “q is false” we obtain a counterpart of the
Modus Tollens pattern.
This example shows that the formalism of belief functions perfectly agrees with the
propositional calculus. In general, reasoning with categorical belief functions corresponds
to the idea of using set valued description of variables to estimation and testing hypotheses - cf [31].
222
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
Conditioning. When Bel is a belief function over Θ and Belo is a categorical belief
function focused on a subset A of Θ then the combination of Bel with Belo results in the
conditional belief function Bel(· | A). Its properties are described in [32], [46] and [38].
5.3
Reasoning under Uncertainty
With the DRC the problem of reasoning can be stated as follows: Given a collection of
unrelated pieces of evidence ε1 , ε2 , . . . , εm (each of which is translated to the appropriate
belief function) we turn them into a single body of the form ε1 ∧ε2 ∧· · ·∧εm characterized
by a belief function being the orthogonal sum the appropriate belief functions. Adding
to the resulting body some findings which are possibly certain, we obtain a conditional
belief function. This closely corresponds to the probabilistic reasoning where the bodies
of evidence are represented by the (conditional) probabilities from which the joint probability distribution is derived; having this distribution we are looking for a distribution
conditioned on a subset of Θ.
To formalize the problem observe that each body εi is represented as a belief function
over a subset Hi ⊆ X . Denoting H the family of subsets Hi , i = 1 . . . m we are interested
in a belief function
BelX (· | e) = (⊕{Bel|H ∈ H})↓{X}
(9)
where e is a subset of Θ representing current findings. This corresponds to deriving
a conditional probability
distribution P (X | e) from the joint probability distribution
Q
P (X1 , . . . , Xm ) = {P (Xi | Π(Xi ) | i = 1 . . . m} given the set of observations (instantiated variables) e; here Π(Xi ) is the set of immediate causes for Xi – see [27]. Note that
under probabilistic setting the subsets Hi are defined as the sum of {Xi } and Π(Xi );
further the operator ⊕ is replaced by the multiplication and the projection operator
↓{X}
corresponds to the summation over all variables form the set X − {X}. Hence the
equation (9) nicely summarizes both the probabilistic and strictly evidential problem of
reasoning.
Because immediate derivation of BelX (· | e) is numerically very expensive, a number
of procedures has been designed to this problem - cf. [27], [22], [36].
The simplest and more attractive procedure seems to be that of Shafer and Shenoy.
It relies upon the observation that deleting a single variable Y from the set X results in
a belief function (here X ′ = X − {Y })
′
Bel↓X = (⊕{BelH | H ∈ H, Y 6∈ H}) ⊕ (⊕BelH | H ∈ H, Y ∈ H)↓X
′
and follows from the fact that (1) the operator ⊕ is commutative and associative, and
(2) if BelA and BelB are two belief function defined over Θ.A and Θ.B, respectively,
↓(A∩B)
then (BelA ⊕ BelB )↓A = BelA ⊕ BelB
. Note that after performing the above
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
223
′
computations the original family H reduces
S to new family H consisting of all H’s such
′
that Y 6∈ H, and the new set H = ( {H | Y ∈ H}) − {Y }. In its very nature the
procedure is closely related to the approach used by [2] in solving nonserial dynamic
problems and it says that to find a margin over Θ.{X} it suffices to delete one by one
variables from the set X −{X}, each time combining belief functions over relatively small
frame being the sum of these H’s which contain the currently removed variable. This last
operation is allowed because the projection operator is such that (3) if I ⊂ J are two
subsets of X then Bel↓I = (Bel↓J )↓I . The three properties (1), (2) and (3) can be viewed
as the necessary and sufficient conditions for local computations described below- cf. [36].
Obviously, the numerical complexity of this new procedure hardly depends on the
order in which the variables from the set X ′ are removed - see [2] for details. Shafer and
Shenoy (1986) observed that an optimal order can be recovered if we arrange the original
family H into a join tree – consult [24] – (Markov tree in their nomenclature) which has
a nice property stating that when L ∈ H is a leaf node in the tree and P ∈ H is its
parent node then the variables from the set L − P occurs only in the set L. (Of course
when H is not an acyclic hypergraph it must be embedded in such a hypergraph first –
see [48] for an appropriate algorithm).
′
Now the procedure of finding Bel↓X can be summarized as follows. Assume that H is
organized into a join tree rooted on the node corresponding to the variable X. Each node
in the tree can be imagined as a separate processor equipped with local information in the
form of a component belief function (when a node was added to H in the process of the
tree formation its local information has the form of vacuous belief function). Each node
communicates with its neighbors by passing messages (belief functions) to the neighbors.
There are two kinds op operations: a propagation operation which describes how messages
are passed from node to node so that all of the local information is globally distributed,
and a fusion rule which describes how the messages incoming to a node are combined
to make marginal belief functions and outgoing messages. In other words propagation
takes place along the edges of the tree and fusion takes place within the nodes. Both the
operations can be formalized as follows. Let N (H; T ) denote the set of all neighbors of
the node H in the join tree T . The message which the node H sends to a neighbor H ′
equals to
′
M H→H = (BelH ⊕ (⊕{M G→H | G ∈ (N (H; T ) − {H ′ })}))↓(H∩H
′
)
(10)
′
The operation in the parentheses corresponds to fusion and (·)↓(H∩H ) represents propagation. Note that (10) allows for parallel belief updating. When a node received messages
from all its neighbors, a marginal belief is computed
Bel↓H = BelH ⊕ (⊕{M H
′
→H
| H ′ ∈ N (H; T )})
(11)
This procedure with some modifications reducing the number of summations ⊕, described
in [47]., has been successfully implemented in our system.
224
5.4
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
Belief Revision
The aim of belief revision is, according to [27] Ch. 5 to identify a composite set of
propositions (one from each variable) which “best” explains the evidence at hand. Under
probabilistic context this formalizes as follows. Given e, the set of instantiated variables
(evidence), let x stand for an assignment of values to the variables in X consistent with
e; such an x is said to be explanation, interpretation or extension of e. The most probable
explanation (MPE) of the evidence at hand is such an extension xe which maximizes the
conditional probability P (X = x | e). This task can be performed locally by letting each
variable X in X compute the function
β(x | e) = max{P (x, x’) | e}
x’
(12)
where x’ stands for an explanation projected on the space Θ(X − {X}). That 12 can be
computed locally follows from the fact that the pair (·, max) satisfies the requirements
(1)-(3) specified in section 5.3. Note that in 12 the operation of summation – requested
when computing conditional probabilities P (X | e) – was replaced by the maximum
operator.
When we use belief function the maximum operator appears both in the ⊕-summation
and in the projection operator. Hence to fit the rule (9) to the formula (12) we must
redefine both the operators. First, instead ⊕-summation we must define new ⊕ − max
operator
(m1 ⊕max m2 )(A) = max{m1 (A∪C)·m2 (A∪D) | C, D ⊆ (Θ−A), C∩D = ∅}(13)
Here, contrary to the ⊕-summation, normalization is unnecessary as the relative strengths
are more important than the absolute values of commitment to a given set of propositions.
Similarly, the projection operator takes now the form (we assume that A ⊂ B)
m↓maxB = max{mB (B) | B ⊆ Θ.B, proj(B) = A}
A
(14)
Using these operators in equation (10) we obtain a message passing algorithm for belief
updating. This algorithm can be used in two main modes described below.
Explanatory mode. Rooting the join tree on a variable X we are searching for the
instantiations of other variables which explain computed value ox X at best.
Hypothesizing. Again we root the tree on a variable X and we assume that X takes
a value xo . Our algorithm allow to find the best explanation for such an assignment (i.e.
we are searching an answer for the question “which settings of the remaining variables
explain the condition “X = xo ” at best?”)
Conditioning. We choose a group of variables, say ε, and we instantiate them to the
values e ∈ Θ.ε. Now our algorithm allows to answer the question “what will happen if
ε = e?”.
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
6
225
Viewing Belief Network as a Set of Rules
A natural phenomenon of human expression of knowledge is a widespread usage of
implications: most mathematical theorems are of the form ”if < premise > then <
conclusion >”, though obviously they are equivalent to for example a CNF or DNF
representation. Furthermore, humans use implication-like statements even if described
phenomena fail to be deterministic, e.g. in describing course of chemical reactions, laws
in social sciences etc. The term ”rule” for implication-like expression of knowledge found
a wide-spread use in artificial intelligence, e.g. in MYCIN-like expert systems.
Taking into account this phenomenon, also our group made an effort to find a way
for describing the contents of a belief-network based knowledge bases (which is very
different in nature from modular production rules based knowledge bases) in terms of
rule-like constructs.
This transcription should support the user by helping him:
–
–
–
–
to
to
to
to
understand the contents of the knowledge base
formulate queries to the knowledge base
enter and update the knowledge base and
understand the justifications of the system for results of its reasoning
The last item has just been subject of the preceding sections.
To understand better the the nature of the rule-like transcription let us first give the
general definition of a belief network:
6.1
The Concept of a Belief Network
We generalize here the definition of belief network from [12] (bayesian networks), while
using the denotation of [36].
Definition 1. We define a mapping ⊙ : V V × V V → V V called decombination such
that:
if BEL12 = BEL1 ⊙BEL2 then BEL1 = BEL2 ⊙ BEL12 .
In case of probabilities, decombination means memberwise division:
P r12 (A) = P r1 (A)/P r2 (A).
In case of DS
P belief functions, let us remaind that the commonality function Q is defined
as Q(A) = B;A⊆B m(B). Hence for Bel12 = Bel1 ⊕Bel2 then Q12 (A) = c·Q1 (A)·Q2 (A)
(c - normalizing factor). Let us remaind also that by a pseudo-belief function a function
over powerset is understood which differs from the proper belief function by the fact,
226
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
that also negative mass function (m) values are allowed, but only to such extent that
the commonality function Q remains non-negative. Then let us define the operator ⊖ as
yieldiing a DS pseudo-belief function such that: whenever Bel12 = Bel1 ⊖ Bel2 then
Q12 (A) = c · Q1 (A)/Q2 (A)
(c - normalizing constant) for non-zero Q2 (A). It is easy to check that such a function
always exists. Obviously, Bel1 = Bel1 ⊕ Bel12 . The operator ⊖ means then the decombination of two DS belief functions. Both for probabilities and for DS belief functions
decombination may be not uniquely determined. Moreover, for DS belief functions not
always a decombined DS belief function will exist. Hence we extend the domain to DS
pseudo-belief functions which is closed under this operator. We claim here without a
proof (which is simple) that DS pseudo-belief functions fit the axiomatic framework of
Shenoy/Shafer. Also, we claim that if an (ordinary) DS belief function is represented by
a factorization in DS pseudo-belief functions, then any propagation of uncertainty yields
the very same results as when it would have been factorized into ordinary DS belief
functions. Let us define now the concept of pseudo-conditioning.
Definition 2. By pseudo-conditioning | of a belief function BEL on a set of variables h
we understand the transformation: BEL|h = BEL⊙BEL↓h .
Notably, pseudo-conditioning means in case of probability functions proper conditioning.
In case of DS pseudo-belief functions the operator | has meaning entirely different from
traditionally used notion of conditionality - pseudo-conditioning is a technical term used
exclusively for valuation of nodes in belief networks. Notice: some other authors e.g. [3]
recognized also the necessity of introduction of two different notions in the context of the
Shenoy/Shafer axiomatic framework (compare a priori and a posteriori conditionals in
[3]). [3] introduces 3 additional axioms governing the ’a priori’ conditionality to enable
propagation with them. Our pseudo-conditionality is bound only to the assumption of
executability of the ⊙ operation and does not assume any further properties of it. We
will discuss the consequences of this difference elsewhere. Let us define now the general
notion of belief networks:
Definition 3. A belief network is a pair (D,BEL) where D is a dag (directed acyclic
graph) and BEL is a belief distribution called the underlying distribution. Each node i in
D corresponds to a variable Xi in BEL, a set of nodes I corresponds to a set of variables
XI and xi , xI denote values drawn from the domain of Xi and from the (cross product)
domain of XI respectively. Each node in the network is regarded as a storage cell for
any distribution BEL↓{Xi }∪Xπ(i) |Xπ(i) where Xπ(i) is a set of nodes corresponding to the
parent nodes π(i) of i. The underlying distribution represented by a belief network is
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
computed via:
BEL =
n
K
227
BEL↓{Xi }∪Xπ(i) |Xπ(i)
i=1
The notion of belief network just introduced possesses several important characteristics:
– The relationship between the global belief function BEL and its constituting factors BEL↓{Xi }∪Xπ(i) |Xπ(i) is local: this means that having an empirical model of
BEL and knowing the structure D of he distribution BEL we can estimate factors
BEL↓{Xi }∪Xπ(i) |Xπ(i) separately for each one by projection of the empirical model
onto the subset {Xi } ∪ Xπ(i) of variables.
– Removal of a leave node from D (and hence the respective factor from the above
”product”) is equivalent to projection of BEL onto the space spanned by variables
associated with the remaining odes: this enables to apply the technique of uncertainty
propagation by edge reversals.
– A belief network reflects causal dependencies among variables (by directions of arrows)
– We can reason about conditional independence of disjunctive sets of variables given
another set of variables using only graphical properties of dag-representation D and
without referring to numerical properties of the underlying distribution.
In case of probabilistic distributions the first to properties are obvious. The last
property has been studied very carefully for probabilistic belief networks by Geiger,
Verma and Pearl [12]. It is easily checked that the defining formula for the underlying
distribution in probabilistic case reduces to:
P (x1 , ..., xn ) =
n
Y
P (xi |xπ(i) )
i=1
In their paper they introduce the notion of d-separation (implying conditional independence) as follows:
Definition 4. [12] A trail in a dag is a sequence of links that form a path in the
underlying undirected graph. A node β is called a head-to-head node with respect to a
trail t if there are two consecutive links α → β and β ← γ on that t.
Definition 5. [12] A trail t connecting nodes α and β is said to be active given a set of
nodes L, if (1) every head-to-head-node wrt t either is or has a descendent in L and (2)
every other node on t is outside L. Otherwise t is said to be blocked (given L).
Definition 6. [12] If J,K and L are three disjoint sets of nodes in a dag D, then L is said
to d-separate J from K, denoted I(J, K|L)D iff no trail between a node in J and a node
in K is active given L.
228
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
It has been shown in [13] that
Theorem 7. [13] Let L be a set of nodes in a dag D, and let α, β ∈
/ L be two additional
nodes in D. Then α and β are connected via an active trail (given L) iff α and β are
connected via a simple (i.e. not possessing cycles in the underlying undirected graph)
active trail (given L).
Definition 8. [12] If XJ , XK , XL are three disjoint sets of variables of a distribution P,
then XJ , XK are said to be conditionally independent given XL (denoted I(XJ , XK |XL )P
iff P (xJ , xK |xL ) = P (xJ |xL ).P (xK |xL ) for all possible values of XJ , XK , XL for which
P (xL ) > 0. I(XJ , XK |XL )P is called a (conditional independence) statement
Theorem 9. [12] Let PD = {P |(D,P) is a Bayesian network}. Then:
I(J, K|L)D iff I(XJ , XK |XL )P for all P ∈ PD .
The ”only if” part (soundness) states that whenever I(J, K|L)D holds in D, it must
represent an independence that holds in every underlying distribution.
The ”if” part (completeness) asserts that any independence that is not detected by
d-separation cannot be shared by all distributions in PD and hence cannot be revealed
by non-numeric methods.
In case of DS belief functions, our defining equation for the underlying distribution
of a belief network has the form:
Bel =
n
M
Bel↓{Xi }∪Xπ(i) |Xπ(i)
i=1
It is easily seen that the first two properties hold for DS belief networks as defined in this
paper. The notion of d-separation and its relationship with conditional independence
are easily transferred onto DS belief networks.This definition differs significantly from
what is generally considered to be a DS belief network (e.g. in [36]). Usually, a DS belief
network is considered to have the defining equation of the form:
Bel =
n
M
Beli
i=1
where Beli is a DS belief function in some subset of variables of Bel without any assumption of the nature of Beli . The deviation from traditional approach seems to be
significant one and requires some explanation. First of all the traditional DS belief network has in general none of the above-mentioned three properties hold. The advantages of
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
229
our DS belief network definition are obvious. The question remains whether there are any
disadvantages. Ones connected with uncertainty propagation within the Shenoy/Shafer
axiomatic framework [36] is of particular interest for implementation of this expert system.
Shenoy and Shafer [36] consider it unimportant whether or not the factorization
should refer to conditional probabilities in case of probabilistic belief networks. We shall
make at this point the remark that for expert system inference engine it is of primary
importance how the contents of the knowledge base should be understood by the user as
beside computation an expert system is expected at least to justify its conclusions and
it can do so only referring to elements of the knowledge base. So if a belief network (or a
hypergraph) is to be used as the knowledge base, as much elements as possible have to
refer to experience of the user.
In our opinion, the major reason for this remark of Shenoy and Shafer is that in fact
the Dempster-Shafer belief function cannot be decomposed in terms of the traditional
conditional belief function as defined in the previous section 5. This can only be done if a
pseudo-conditioning like ours is introduced. But an intriguing question remains whether
the traditional DS belief networks extend essentially the class of DS belief functions
suitable for Shenoy/Shafer propagation of uncertainty. The sad result that really
Theorem 10. [20] Traditional DS belief network induced hypergraph (as considered by
Shenoy and Shafer [36]) may for a given joint belief distribution have simpler structure
than (be properly covered by) the closest hypergraph induced by a new DS belief network
(as defined above.)
This fact, however, is compensated completely by another one. Shenoy/Shafer propagation does not run in hypergraphs, but in hypertrees. And:
Theorem 11. [20] No traditional DS belief network induced hypertree (as considered by
Shenoy and Shafer [36]) may for a given joint belief distribution have simpler structure
than (be properly covered by) the closest hypertree induced by a new DS belief network
(as defined above.)
This fully justifies, in our opinion, the usage of the new belief network definition.
6.2
Understanding a Belief Network in Terms of Rules
A dag structure of a belief network was presented in Fig.1. Let us consider a fragment of
a belief network in Fig.2 consisting of a node (and the associated variable, or attribute, or
feature) Y and all of its parents: X and Z. Let the values of them range for Y: {y1 , y2 , y3 },
230
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
p3
✉
✲
p1
✉
p5
p4
✉
p2
✲
❅
❘
❅
✻ ❅
✲
✉
❅
❅
✉
p6
✉
❄
✉
✲
p7
❄
❅
❅
❅
❘
❅
✉
p8
Fig. 1. Belief Network - An Example
✉X
❅
❅
❅
❅
❅
❘ Y✠
❅
✉
✉Z
Fig. 2. A Typical Node of a Belief Network
for X:{x1 , x2 }, for Z:{z1 , z2 , z3 , z4 }, resp.
Then a table of valuations will be associated with the node Y. In probabilistic case
the valuation will take the form given in table 1.
In general, it will be an exhaustive table of conditional probabilities of each value of a
node given each value combination of its parents. In case of Dempster-Shafer valuation,
only some (that is focal point) conditional valuations will be given, e.g. as in table 2.
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
Table 1. Probabilistic Node Valuation
X
x1
x1
x1
x1
x1
x1
x1
x1
x1
x1
x1
x1
x2
x2
x2
x2
x2
x2
x2
x2
x2
x2
x2
x2
Z
z1
z1
z1
z2
z2
z2
z3
z3
z3
z4
z4
z4
z1
z1
z1
z2
z2
z2
z3
z3
z3
z4
z4
z4
Y Cond. Probability P (y|x, z)
y1
p111
y2
p112
y3
p113
y1
p121
y2
p122
y3
p123
y1
p131
y2
p132
y3
p133
y1
p141
y2
p142
y3
p143
y1
p211
y2
p212
y3
p213
y1
p221
y2
p222
y3
p223
y1
p231
y2
p232
y3
p233
y1
p241
y2
p242
y3
p243
Table 2. Dempster-Shafer Valuation of a Node
X
{(x1 ,
{(x1 ,
(x1 ,
{(x2 ,
{(x1 ,
(x1 ,
(x2 ,
(x2 ,
Z
z1 ,
z1 ,
z3 ,
z2 ,
z4 ,
z4 ,
z4 ,
z4 ,
Y bpa function m↓{Y,X,Z}|{X,Z}(y, x, z)
y1 )}
m1
y2 ),
y2 )}
m2
y3 )}
m3
y1 ),
y3 ),
y1 ),
y3 )}
m4
231
232
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
Each table of this form may be displayed for presentation purposes as a ”beam” of
rules:
in probabilistic case:
IF X = x1 AND Z = z1 THEN Y = y1 WITH p111
IF X = x1 AND Z = z1 THEN Y = y2 WITH p112
IF X = x1 AND Z = z1 THEN Y = y3 WITH p113
IF X = x2 AND Z = z1 THEN Y = y1 WITH p211
.........
IF X = x2 AND Z = z4 THEN Y = y3 WITH p243
and in Dempster-Shafer case for the valuation in table 2:
IF X=x1 AND Z=z1 THEN Y=y1 WITH Q1
IF X=x1 AND Z=z1 THEN Y=y2
AND IF X=x1 AND Z=z3 THEN Y=y2 WITH Q2
IF X=x2 AND Z=z2 THEN Y=y3 WITH Q3
IF X=x1 AND Z=z4 THEN Y=y1
AND IF X=x1 AND Z=z4 THEN Y=y3
AND IF X=x2 AND Z=z4 THEN Y=y1
AND IF X=x2 AND Z=z4 THEN Y=y3
WITH Q4
while measures Q (commonality) are calculated according to the principles of the
Dempster-Shafer Theory [32].
First of all notice that we are talking always about a ”beam of rules” and not of a
single rule - this was rarely considered in production rule systems.
In probabilistic case, it should be clear from the very beginning what is meant by a
single rule (of the beam): the certainty factor after the word ”with” expresses the proportion of cases fitting the ”rule” among those meeting the premise. At the same time,
the above examples exhaust the richness of the language needed for presenting contents
of the knowledge base to the user.
The Dempster-Shafer case is more subtle. Here, all the variables (X,Y,Z) are treated as
potentially set-valued, and the equality sign in the expressions should be rather treated as
”is containing” and not as set equality sign. The certainty factor after the word ”with” expresses again the proportion of cases fitting the ”rule” among those meeting the premise.
At the same time, the above examples exhaust the richness of the language needed for
presenting contents of the knowledge base to the user.
The representation language for knowledge base has not only syntactic restrictions,
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
233
but also semantica ones: within a rule, only direct predecessors of a given node may occur
in the premise part of a rule.
6.3
Rule-oriented Queries to the Knowledge Base
A knowledge base in belief network notation means something more than just the
sum of rules. The built-in system of uncertainty propagation can for the given network
give the level of probability (belief) for more general formulae without any restrictions
provided by network topology:
– for logical expressions (with operators AND, OR, NOT) in atomic formulae of the
form <variable> = <value>.
- the system may deduce what is the unconditional probability/belief of such a formulae to hold,
- the same question may be also answered conditionally: what is the probability/belief
of the expression given some of the variables of the network have restricted value
ranges,
- the same question may be also answered conditionally: what is the probability/belief
of the expression given some constraints expressed in terms of logical expressions are
imposed onto the network,
– for beams of rules of the type IF attr1=? AND attr2=? AND attr3=? ... THEN
attr=? - the complete beam can be inferred unconditionally, or conditionally or for
a given set of constraints as described for logical expressions.
For probabilistic case, the methods of implementation of the above query-answering
is described in [27]. Implementation for Dempster-Shafer Theory runs along the same
lines. Let us consider as an example the network in Fig.1. Let us assume that each of the
nodes p1 ,...,p8 takes only one of two values: ’y’ and ’n’.
We intend to ask the question: p5 =′ y ′ .or.p1 =′ n′ . The belief network is to be
amended temporarily by a node x1 , taking values ’y’ i ’n’, which we connect with p5 and
p1 (see Fig.3). With this node, we associate the rules:
IF p5 =′ t′ AND p1 =′ n′ THEN x1 =′ t′ WITH 1.
IF p5 =′ t′ AND p1 =′ t′ THEN x1 =′ t′ WITH 1.
IF p5 =′ n′ AND p1 =′ n′ THEN x1 =′ t′ WITH 1.
IF p5 =′ n′ AND p1 =′ t′ THEN x1 =′ n′ WITH 1.
and for the remaining rules the probability is equal 0.
234
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
p3
p5
p4
✉ ✲ ❡
✿
✘
✘✘
✘
❅
✘✘
✘
✘
❅
❘✘✘✘
❅
✻ ✘
p1
p✘
2 ✘✘
❅
❅ p6
✘✘
✉✘ ✲
✉
✉
❄
✉
✲
x1
✉
✲
❄
✉
p7
✲
❅
❅
❅
❘
❅
✉
p8
Fig. 3. A Belief Network with Query-Node: x1 : p5 =′ y ′ .or.p1 =′ n′
The answer is the final valuation of the node x1 after uncertainty propagation.
Now let’s turn to checking validity of the query of the form:
x3 : if (p5 =′ t′ .or. p1 =′ n′ ) .and. p3 =′ n′ then p6 =′ t′
with respect to contents of the knowledge base. Node x1 connected with p5 and p1 , node
x2 connected with p3 and p1 and node x3 connected with p6 and p1 (see Fig.4). Let nodes
x1 and x2 take values ’y’ or ’n’, and the node x3 - ’y’, ’n’, ’ ?’. Node x1 is associated with
a beam of rules as mentioned above, and x2 - with:
IF p3 =′ t′ AND x1 =′ t′ THEN x2 =′ t′ WITH 1.
IF p3 =′ t′ AND x1 =′ n′ THEN x2 =′ n′ WITH 1.
IF p3 =′ n′ AND x1 =′ t′ THEN x2 =′ n′ WITH 1.
IF p3 =′ n′ AND x1 =′ n′ THEN x2 =′ n′ WITH 1.
with the remaining rule probabilities equal 0.
Node x3 is associated with rule beam:
IF p6 =′ t′ AND x2 =′ t′ THEN x3 =′ t′ WITH 1.
IF p6 =′ t′ AND x2 =′ n′ THEN x3 =′ n′ WITH 1.
IF p6 =′ n′ AND x2 =′ t′ THEN x3 =′ ?′ WITH 1.
IF p6 =′ n′ AND x2 =′ n′ THEN x3 =′ ?′ WITH 1.
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
235
with the remaining rule probabilities equal 0.
p3
p4
p5
✉
✲
✲
✉ ✲
✉
❳❳
✿
✘
❳❳❳
✘✘
✘
❅❳
✘
❳
✘
✘
❳❳✘
❅
❳✘
❘✘✘
❅
✻ ✘
❳❳
❳❳❳
p1
p✘
2 ✘✘
❅
❅ p6
❳❳
✘
✘
③
✘
✉
✲
✉
✉
❄ ❳
❍
❅❍❍
❅ ❍❍
❄
❍❍
❅
❘
❅
❥
❍
✉
✲
✉
p7
p8
x1
❡
x2
❄
❡
❄
❡
x3
Fig. 4. x3 : if (p5 =′ t′ .or. p1 =′ n′ ) .and. p3 =′ n′ then p6 =′ t′
The final answer is the valuation of the node x3 after uncertainty propagation. The
probability of x3 =′ t′ indicates how frequently the rule fires and is correct,probability
of the event x3 =′ ?′ indicates how often the rule did not fire (and only for this reason is
considered by logicians to be true).
probability of x3 =′ n′ indicates, how frequently the rule fires, but is in error. (The
respective 3-valued logic is considered in [19].
6.4
Belief Network Construction
To construct a belief network correctly, one should:
– properly uncover the causal structure governing the attribute, and thereafter
– calculate the valuations of each of the node of the network.
For the probabilistic case, a number of respective causal-structure-from-data reconstruction algorithms have been elaborated [1], [4], [5], [7], [13], [30], [39], [40], [41], [42],
[43], [45].
236
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
Node valuation can be also calculated from data [5] as relative conditional frequency.
The Dempster-Shafer case is more difficult as clear relationship between data and the
joint belief distribution is still subject of disputes. Set this issue aside, proper Q-value
quotients may be exploited.
However, the knowledge base may be only partially reflected by data. Thenm there
emerges the necessity to manipulate manually the knowledge base. The user is then
requested to enter a consistent description of the network structure. Given this, the entrance of the valuations may be run precisely following the vary same pattern as the
contents of the knowledge base are presented to the user.
References
1. Acid S., deCampos L.M., Gonzales A., Molina R., Perez de la Blanca N.: Learning with
CASTLE, Symbolic and Quantitative Approaches to Uncertainty, Kruse R., Siegel P. (eds),
Lecture Notes In Computer Science 548, Springer-Verlag 1991, 99-106
2. Bertelè, U., and Brioschi, F., 1972. Nonserial Dynamic Programming. Academic Press, New
York.
3. Cano J., Delgado M., Moral S.: An axiomatic framework for propagating uncertainty in
directed acyclic networks, International Journal of Approximate Reasoning. 1993:8, 253280.
4. Chow C.K., Liu C.N.: Approximating discrete probability distributions with dependence trees,
IEEE Transactions on Information Theory , Vol. IT-14, No.3, (maj 1968), 462-467
5. Chow C.K., Wagner T.J.: Consistency of an estimate of tree-dependent probability distribution, IEEE Transactions on Information Theory , IT-19, (1973), 369-371
6. Chowdhury, S.I., Statistical Expert Systems — A Special Application Area for KnowledgeBased Computer Methodology, Technical Report No. 104, Department of Computer and
Information Science, Linköping University, Sweden, 1987.
7. Cooper G.F., Herskovits E.: A Bayesian method for the induction of probabilistic networks
from data, Machine Learning 9, 309-347 (1992).
8. Darius, P., Expert Systems and Statistics, Proceedings SEAS Spring Meeting 1984, Vol.II,
pp.529-538.
9. Dempster, A.P., 1966. New methods for reasoning towards posterior distributions based on
sample data. Annals of Mathematical Statistics, 37:355-374.
10. Dempster, A.P., 1967. Upper and lower probabilities induced by a multivalued mapping.
Annals of Mathematical Statistics, 38: 325-339.
11. French,J.C., Support for Scientific Database Management in Statistical and Scientific
Databases, Z. Michalewicz (Editor), Ellis Horwood, 1991, pp.51-82.
12. Geiger D., Verma T, Pearl J.: d-Separation: From theorems to algorithms, M.Henrion,
R.D.Shachter, L.N.Kamal, J.F.Lemmer (Eds): Uncertainty in Artificial Intelligence 5, Elsevier Science Publishers B.V. (North-Holland), 1990, pp. 139-148.
KNOWLEDGE ACQUISITION, REPRESENTATION & MANIPULATION
237
13. Geiger D., Paz A., Pearl J.: Learning causal trees from dependence information, Proc. of
AAAI (pp.770-776), Boston MA, MIT-Press.
14. Hand, D.J., Expert Systems in Statistics, The Knowledge Engineering Review, 1, pp.1-10,
1986
15. Hand, D.J., A Statistical Knowledge Enhancement System, Journal of the Royal Statistical
Society, Series A, 150, pp.334-345, 1987
16. Hand, D.J., Emergent Themes in Statistical Expert Systems, in Knowledge, Data, and
Computer Assisted Decisions, M. Schader and W. Gaul (Editors), Springer-Verlag, Berln,
pp.279-288, 1990.
17. Hand, D.J. and Cubitt, R., Statistical Expert Systems, in Statistical and Scientific Databases,
Z. Michalewicz (Editor), Ellis Horwood, 1991, pp.421-429.
18. Kennes, R., 1990. Computational aspects of the Mobius transform of a graph. Technical
Report TR/IRIDIA/90-13, IRIDIA, B-1050 Bruxelles, Belgium.
19. Klopotek M.A.: Beliefs and Probability in Bacchus’ l.p. Logic: A 3-Valued Logic Solution to
Apparent Counter-intuition, [in:] R. Trappl ed: Cybernetics and Systems Research, Proc.
11t h European Meeting on Cybernetics and System Research EMCSR’92, Wien, Österreich,
20. April 1992. World Scientific Singapore, New Jersey, London, HongKong vol. 1, 519-526
20. Klopotek M.A.: Beliefs in Markov Trees - From Local Computations to Local Valuation,
to appear in Proceedings of 12th European Meeting on Cybernetics and System Research,
Vienna, 5-8 April 1994
21. Kong, A., 1986. Multivariate belief functions and graphical models. Doctoral Dissertation,
dept. of Statistics, Harvard Univ.
22. Lauritzen, D.J., and Spiegelhalter, S.L., 1988. Fast manipulation of probabilities with local
representations - with applications to expert systems (with discussion). J. Roy. Statist. Soc.,
B-50, 205-247.
23. Lowrance, J.D. and Garvey, T.D., 1982. Evidential reasoning: An implementation for multisensor integration. Tech. Rep. 307, AI Center, SRI International, Menlo Park, CA.
24. Maier, D., 1983. The Theory of Relational Databases. Rockville Computer Sci. Press.
25. Michalewicz, M. and Michalewicz, Z., Automated Rules Generation, Proceedings of the Sixth
International Symposium on Methodologies for Intelligent Systems, Charlotte, NC, October
16-19, 1991, Springer-Verlag Lecture Notes in Artificial Intelligence, Vol.542, pp.389-398.
26. Michalewicz, M., Wierzchoń, S. T., Syropiatko, E., Pacan, A., Matuszewski, A., Klopotek,
M., Statistical-Expert System for Data Analysis and Knowledge Acquisition, in Statistical
and Scientific Databases, Z. Michalewicz (Editor), Ellis Horwood, 1991.
27. Pearl J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Influence,
Morgan and Kaufmann, 1988
28. Quinlann, J.R., Induction of Decision Trees, Machine Learning, Vol.1, No.1, 1986.
29. Ruspini, E.H. et al., 1992. Understanding evidential reasoning. Intn. J. Approximate Reasoning, 6, 401-424.
30. Rebane G., Pearl J.: The recovery of causal poly-trees from statistical data, w Uncertainty
in Artificial Intelligence 3, Kanal L.N., Levit T.S., Lemmer J.F. eds., Elsevier Science Publishers B.V., (North Holland) 1989, 175-182
31. Schweppe, F.C., 1973. Uncertain Dynamic Systems. Prentice-Hall, New Jersey,
238
M.MICHALEWICZ, S.T.WIERZCHOŃ, M.A. KlOPOTEK
32. Shafer, G., 1976. A Mathematical Theory of Evidence. Princeton University Press, Princeton.
33. Shafer, G., 1979. Allocations of probability. Ann. Probab., 7, 827-839.
34. Shafer, G., 1990. Belief functions. In: G. Shafer and J. Pearl (Eds.) ”Readings in Uncertain
Reasoning”, Morgan Kaufmann, Palo Alto, pp. 473-481.
35. Shafer, G., and Shenoy, P., 1990. Axioms for probability and belief function propagation. In:
R.D. Shachter (eds.), Uncertainty in Artificial Intelligence 4, North-Holland, Amsterdam,
pp. 169-198.
36. Shenoy P.P., Shafer G.: Axioms for probability and belief-function propagation, [w:] Shachter
R.D., Levitt T.S., Kanal L.N., Lemmer J.F. (eds): Uncertainty in Artificial Intelligence 4,
Elsevier Science Publishers B.V. (North Holland), 1990,
37. Shortliffe, E.H. and Buchanan, B.G., Model of Inexact Reasoning in Medicine, Mathematical
Biosciences, Vol.23, pp.251-379, 1975
38. Smets, P., (1988). Belief functions. In: P. Smets, A. Mamdani, D. Dubois and H. Prade
(Eds.), Non-Standard Logics for Automated Reasoning, New York: Academic Press, pp.
253-286.
39. Spirtes P., Glymour C., Scheines R.: Causality from probability, [w:] G.McKee (Ed.):Evolving
knowledge in natural and artificial intelligence, London: Pitman, 1990.
40. Spirtes P., Glymour C., Scheines R.: An algorithm for fast recovery of sparse causal graphs,
Social Science Computer Review 9, 62-72.
41. Srinivas S.,Russel S., Agogino A.: Automated construction of sparse Bayesian networks from
unstructured probabilistic models and domain information, Henrion M., Shachter R.D.,Kanal
L.N., Lemmer J.F.: Uncertainty in Artificial Intelligence 5, Elsevier Science Publishers B.V
(North-Holland) (1990), 295- 308
42. Valiveti R.S., OommenB.J.: On using the chi-squared statistics for determining statistic
dependence, Pattern Recognition Vol. 25 No. 11, pp. 1389-1400, 1992
43. Verna T.S., Pearl J.: Equivalence and synthesis of causal models, Proc. of the Conference
on Uncertainty in Artificial Intelligence (pp. 220-227), Cambridge MA, 1990
44. Wasserman, L.A., 1990. Prior envelopes based on belief functions. Ann. Stat., 18, 454-464.
45. Wermuth N., Lauritzen S.: Graphical and recursive models for contingency tables, Biometrica
72, 537-552
46. Wierzchon, S.T., 1987. An Inference rule based on Sugeno measure. In: J.C. Bezdek (ed.)
Analysis of Fuzzy Information, CRC Press, vol. 1, pp. 85-96.
47. Wierzchon, S.T., 1993a. Representation and processing of uncertain information in the
Dempster-Shafer framework (forthcoming)
48. Wierzchon, S.T., 1993b. Constraint propagation over restricted space of configurations, in:
M. Fedrizzi, J. Kacprzyk and R.R. Yager (eds.) Advances in the Dempster-Shafer Theory
of Evidence, J. Willey, N.Y.
| 2cs.AI
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.